This action might not be possible to undo. Are you sure you want to continue?
Ruin Probabilities
Seren Asmussen
World Scientific
Ruin Probabilities
ADVANCED SERIES ON STATISTICAL SCIENCE & APPLIED PROBABILITY
Editor: Ole E. BarndorffNielsen
Published Vol. 1: Random Walks of Infinitely Many Particles by P. Revesz Vol. 2: Ruin Probabilities by S. Asmussen Vol. 3: Essentials of Stochastic Finance : Facts, Models, Theory by Albert N. Shiryaev Vol. 4: Principles of Statistical Inference from a NeoFisherian Perspective by L. Pace and A. Salvan Vol. 5: Local Stereology by Eva B. Vedel Jensen Vol. 6: Elementary Stochastic Calculus  With Finance in View by T. Mikosch Vol. 7: Stochastic Methods in Hydrology: Rain, Landforms and Floods eds. O. E. Barndorff Nielsen et al. Vol. 8: Statistical Experiments and Decisions : Asymptotic Theory by A. N. Shiryaev and V. G. Spokoiny
Ruin P robabilities
Soren Asmussen
Mathematical Statistics Centre for Mathematical Sciences Lund University
Sweden
World Scientific
Singapore • NewJersey • London • Hong Kong
Published by World Scientific Publishing Co. Pte. Ltd. P O Box 128, Fatter Road , Singapore 912805 USA office: Suite 1B, 1060 Main Street, River Edge, NJ 07661 UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE
Library of Congress CataloginginPublication Data Asmussen, Soren
Ruin probabilities / Soren Asmussen. p. cm.  (Advanced series on statistical science and applied probability ; vol. 2) Includes bibliographical references and index. ISBN 9810222939 (alk. paper) 1. InsuranceMathematics. 2. Risk. I. Tide. II. Advanced series on statistical science & applied probability ; vol. 2. HG8781 .A83 2000 368'.01dc2l 00038176
British Library CataloguinginPublication Data A catalogue record for this book is available from the British Library.
First published 2000 Reprinted 2001
Copyright ® 2000 by World Scientific Publishing Co. Pte. Ltd. All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means, electronic or mechanical, including photocopying, recording or any information storage and retrieval system now known or to be invented, without written permission from the Publisher.
For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy is not required from the publisher.
Printed by Fulsland Offset Printing (S) Pte Ltd, Singapore
Contents
Preface I ix
Introduction 1 1 The risk process . . . . . . . . . . . . . .. . . . .. .. . . . . 1 2 Claim size distributions .. . . . . . . . .. . . . . . . . . . . . 5 3 The arrival process . . . . . . . . . . . . . . . . . . . . . . . . 11 4 A summary of main results and methods . . . . .. . . . . . . 13 5 Conventions . .. . .. .. . . . . . . . . . . . . . . . . . . . . 19
II Some general tools and results 23 1 Martingales . .. . .. .. . . . . . .. . . . . . . . . . . . . . 24 2 Likelihood ratios and change of measure . . .. . . . . . .. . 26 3 Duality with other applied probability models . . .. . . . . . 30 4 Random walks in discrete or continuous time . . . . . . . . . . 33 5 Markov additive processes . . . . . . . .. . . . . . . . . . . . 39 6 The ladder height distribution . . . .. . .. .. . . . . . . . . 47
III The compound Poisson model 57 1 Introduction . . . . . . . . .. .. .. . .. .. . . . . . . 58 . . . . . . . . . . . . . . . 61 3 Special cases of the PollaczeckKhinchine formula . . . . . . . 62 4 Change of measure via exponential families . . . .... . .. . 67 5 Lundberg conjugation . .. . . . . . . . . . . . . . . . . . . . . 69 6 Further topics related to the adjustment coefficient .. . . . . 75 7 Various approximations for the ruin probability . . . . . . . . 79 8 Comparing the risks of different claim size distributions . . . . 83 9 Sensitivity estimates . . . . . . . . . . . . . . . . . . . . . . . 10 Estimation of the adjustment coefficient . . . . . . . . . . . . 86 93 2 The PollaczeckKhinchine formula
v
vi
CONTENTS
IV The probability of ruin within finite time 97 1 Exponential claims . . . . . . . . . . . . . . . . . . . . . . . . 98 2 The ruin probability with no initial reserve . . . . . . . . . . . 103 3 Laplace transforms . . . . . . . . . . . . . . . . . . . . . . . . 108 4 When does ruin occur? . . . . . . . . . . . . . . . . . . . . . . 110 5 Diffusion approximations . . . . . . . . . . . . .. . . .. . . . 117 6 Corrected diffusion approximations . . . . . . . . . . .. . . . 121 7 How does ruin occur ? . . .. . . . . . . . . . . . . . . . . . . . 127 V Renewal arrivals 131 1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . 131 2 Exponential claims. The compound Poisson model with negative claims . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 3 Change of measure via exponential families . . . . . . . . . . . 137 4 The duality with queueing theory .. .. .. . . . .. . . . . . 141 VI Risk theory in a Markovian environment 145 1 Model and examples . . . . . . . . . . . .. . .. . . . . . . . 145 2 The ladder height distribution . . . . . . . . . .. . . . . . . . 152 3 Change of measure via exponential families ........... 160 4 Comparisons with the compound Poisson model ........ 168 5 The Markovian arrival process . . . . . . .. .. . . ... . . . 173 6 Risk theory in a periodic environment .. . . . .. . . . . . . . 176 7 Dual queueing models .... ... ................ 185 VII Premiums depending on the current reserve 189 1 Introduction . . . . . . . . . . . . . . . . . . . .. . . . . . . . 189 2 The model with interest . . . . . .. . . . . . . . . . .. . . . 196 3 The local adjustment coefficient. Logarithmic asymptotics . . 201 VIII Matrixanalytic methods 215 1 Definition and basic properties of phasetype distributions .. 215 2 Renewal theory . . . . . . . . . . . . . . . . . . . . . . . . . . 223 3 The compound Poisson model . . . . . . . . . .. . . . . . . . 227 4 The renewal model . . . . . . . . . . . . . . . .. . . . . . . . 229 5 Markovmodulated input . . .. . . . . . . . . . . . . . . . . . 234 6 Matrixexponential distributions . . . . . . . . . . . .. . . . 240 7 Reservedependent premiums . . . . .. . . . .. . . . . . . . 244
. .. . . . . . . . . . . . . . 259 3 The renewal model . 294 XI Miscellaneous topics 297 1 The ruin problem for Bernoulli random walk and Brownian motion. . . . . . . . . . . . . . . . . . . . . . . . . The twobarrier ruin problem . . . . . . . . ... .. . 331 A2 WienerHopf factorization . 287 4 Importance sampling for the finite horizon case . . . ... .CONTENTS vii IX Ruin probabilities in the presence of heavy tails 251 1 Subexponential distributions . . . 316 5 Principles for premium calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . 292 6 Sensitivity analysis . . . . . . . . . 326 Appendix 331 Al Renewal theory . . . .. .. . . . . . . .. . . . . . . 323 6 Reinsurance . . . . . . . . . . . . . . . . . . .. . . . . . . 281 2 Simulation via the PollaczeckKhinchine formula . . . . . . . . 279 X Simulation methodology 281 1 Generalities . . . . . . . . . . . . .. . . . . . . . . . . . . . . 290 5 Regenerative simulation . . . . . . 261 4 Models with dependent input .. . 297 2 Further applications of martingales . . . 344 AS Complements on phasetype distributions . . . . . . . . . . . .. . . . . . . . . . . . . .. . . . . . . . . . . .. . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . 336 A3 Matrixexponentials . . 306 4 The distribution of the aggregate claims . . . . . .. . . . . . . . . . 264 5 Finitehorizon ruin probabilities . . . . . . . . . .. . . . . . 340 A4 Some linear algebra . . . . . . . . . . . 304 3 Large deviations . . . . . . . . . . . . . . . . . . . . . 251 2 The compound Poisson model . .. . . . . . . . . . . . 285 3 Importance sampling via Lundberg conjugation . . . . . . 350 Bibliography Index 363 383 . . .. . . .. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . 271 6 Reservedependent premiums . . . . . . . . . . . .
This page is intentionally left blank .
this applies to longrange dependence which is intensely studied in the neighboring ix . and my belief was that this could be done rather quickly. that it can only say something about very simple models and questions. One reason for writing this book is a feeling that the area has in the recent years achieved a considerable mathematical maturity. Apart from these remarks. it would not be fair not to say that the practical relevance of the area has been questioned repeatedly.Preface The most important to say about the history of this book is: it took too long time to write it! In 1991. But the pace was much slower than expected. the book is basically mathematical in its flavour. and the series editor Ole BarndorffNielsen for their patience. and other projects absorbed my interest. if the formulations occasionally give a different impression. which has in particular removed one of the standard criticisms of the area. and the result is now that the book is much more related to my own research than the initial outline. However. 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!). It has obviously not been possible to cover all subareas. Thus. In particular. I have deliberately stayed away from discussing the practical relevance of the theory. The course was never realized. Let me take this opportunity to thank above all my publisher World Scientific Publishing Co. and has been an active area of research from the days of Lundberg all the way up to today. it is not by intention. I was invited to give a course on ruin probabilities at the Laboratory of Insurance Mathematics. Since I was to produce some handouts for the students anyway. the idea was close to expand these to a short book on the subject. 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. Risk theory in general and ruin probablities in particular is traditionally considered as part of insurance mathematics. University of Copenhagen. As an excuse: many of these projects were related to the book.
The book does not go into the broader aspects of the interface between insurance mathematics and mathematical finance. incorporate 11. it has not been possible to incorporate more numerical examples than the few there are. read Chapter I. VI.1.3. see e.g. Asmussen.se Lund February 2000 Soren Asmussen .maths . In addition. Hojgaard & Taksar [206].6 (to understand the PollaczeckKhinchine formula in 111. an area which is becoming increasingly important. Hojgaard & Taksar [35] and Paulsen & Gjessing [284]. I regret that due to time constraints. It is obvious that such a system involves a number of inconsistencies and omissions.lth. some basic discussion can be found in the books by Biihlmann [82] and Gerber [157]. for which I apologize to the reader and the authors of the many papers who ought to have been on the list.2. Resnick & Samorodnitsky [303] and references therein.15. http:// www. some papers not cited in the text but judged to be of interest are included in the Bibliography.x PREFACE field of queueing theory.4a. The present book is in between these two possibilities. IV. For a brief orientation. Finally. In the classical setting of CramerLundberg models. Another interesting area which is not covered is dynamic control. see in particular Michna [259]. IV. the standard stochastic control setting of diffusion models has been considered. I intend to keep a list of misprints and remarks posted on my web page. A book like this can be organized in many ways. another by method.5.lth.se/matstat / staff/asmus and I am therefore grateful to get relevant material sent by email to asmusfmaths . Good luck! I have tried to be fairly exhaustive in citing references close to the text. [381]). For a second reading. VII. IV. the first part of 11.89. One is by model.13 and IX.13. 111. VIII.45. Concerning ruin probabilities.2. VII. Here is a suggestion on how to get started with the book. e.g. IX. Chapters IIIVII introduce some of the main models and give a first derivation of some of their properties.13. 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). 111. X.14. The rest is up to your specific interests.2 more properly). see also Schmidli [325] and the references in Asmussen & Taksar [52]. The main motivation comes from statistical data for network traffic (e. Willinger et al. for the effects on tail probabilities.g.13 and XI. More recently.
5. A number of other figures were supplied by Christian Geisler Asmussen . 3 is reprinted from Asmussen & Nielsen [39] and parts of IX. Section VIII.2 by Rafal Kulik . were produced by Lone Juul Hansen .1 by Bjarne Hojgaard and the table in Example 111. of which there are not many at this stage .6 is reprinted from Asmussen & Schmidt [49] and parts of IX. More substantial remarks. 1 is almost identical to Section 2 of Asmussen [26] and reprinted with permission of Blackwell Publishers. 111 .8 . supported by Center for Mathematical Physics and Stochastics (MaPhySto). 5 from Asmussen & Kliippelberg [36] with the permission from Elsevier Science .PREFACE xi The second printing differs from the first only by minor corrections.1 and X. not least the more complicated ones. many of which were pointed out by Hanspeter Schmidli .4 from Asmussen.6 by my 1999 simulation class in Lund.6. Parts of II. . Schmidli & Schmidt [47] with the permission from Applied Probability Trust .5 from Asmussen [21] with permission from CRC Press. Section VII . Fig. IV. Aarhus. Parts of X. as well as some additional references continue to be at the web page.3 are reprinted from Asmussen & Rubinstein [46] and parts of VIII. Fig. Lund September 2001 Soren Asmussen Acknowledgements Many of the figures .
This page is intentionally left blank .
4) O<t<oo O<t<T 1 . For mathematical purposes.Chapter I Introduction 1 The risk process In this chapter . MT = sup St. M = (1.3) sup St. is a model for the time evolution of the reserves of an insurance company. and give a very brief summary of some of the models.1) We also refer to t/) ( u) and 0(u. T) as ruin probabilities with infinite horizon and finite horizon .i(u. (1. A risk reserve process { Rt}t>o.T) = P inf Rt < 0 I . (1. They are the main topics of study of the present book. results and topics to be studied in the rest of the book.Rt. We denote throughout the initial reserve by u = Ro. (1. as defined in broad terms .2) (O<t<T Ro=ul. we introduce some general notation and terminology. The probability O(u) of ultimate ruin is the probability that the reserve ever drops below zero. t/i(u) = P (infRt < 0) = P (infR t < 0 t>0 t>0 The probability of ruin before time T is t. it is frequently more convenient to work with the claim surplus process {St}t>0 defined by St = u . respectively. Letting T(u) = inf {t > 0 : Rt < 0} = inf It > 0 : St > u}.
say.2 CHAPTER I. However..pt. the ruin probabilities can then alternatively be written as .E Uk. t] is finite. respectively. (1.7) k=1 k=1 The sample paths of {Rt} and {St} and the connection between the two processes are illustrated in Fig.6) Sofar we have not imposed any assumptions on the risk reserve process. Putting things together. per unit time. and Nt = min {n > 0 : 0rn+1 > t} = max {n > 0: Un < t}• The size of the nth claim is denoted by Un.b(u) = P (r(u) < oo) = P(M > u). the number Nt of arrivals in [0. We denote the interarrival times of claims by T2. Thus. the time of arrival of the nth claim is an = T1 + • • • + Tn. INTRODUCTION be the time to ruin and the maxima with infinite and finite horizon.1 . we see that Nt Nt Rt = u + pt . (1. That is. the following setup will cover the vast majority of the book: • There are only finitely many claims in finite time intervals.i(u. St = E Uk . and T1 is the time of the first claim.5) i. Figure 1. .. T3. 1. (1.T) = F (MT > u) = P(r(u) < T). • Premiums flow in at rate p.1.
If 77 < 0.b(u) = 1 for all u. We shall not deal with this case either. VII. since any modeling involves some approximative assumptions. and hence O(u) < 1 for all sufficiently large u. (1. For the purpose of studying ruin probabilities this distinction is. t * oo. .20%. though many results are straightforward to generalize from the compound Poisson model. say 10% . then M < oo a. If 77 > 0.8) The interpretation of p is as the average amount of claim per unit time.. we shall. one may well argue that Brownian motion in itself could be a reasonable model.(. of course. on Fig. We study this case in Ch.s. for example. • Brownian motion or more general diffusions.8) holds. not discuss whether this actually corresponds to practice.1 the slope of {Rt} should depend also on the level). We shall discuss Brownian motion somewhat in Chapter IV. a basic references is Gerber [127]. and hence . • General Levy processes (defined as continuous time processes with stationary independent increments) where the jump component has infinite Levy measure. immaterial.) V 0. 1.1.. and in fact: Proposition 1. but as an approximation to the risk process rather than as a model of intrinsic merit. The models we consider will typically have the property that there exists a constant p such that Nt a E Uk k=1 p. allowing a countable infinity of jumps on Fig. It would appear obvious. Thus. Some main examples of models not incorporated in the above setup are: • Models with a premium depending on the reserve (i. However. and the basic ruin probabilities are derived in XI. 1. A further basic quantity is the safety loading (or the security loading) n defined as the relative amount by which the premium rate p exceeds p. however. however.1. rl= pP P It is sometimes stated in the theoretical literature that the typical values of the safety loading 77 are relatively small.1.e. that the insurance company should try to ensure 77 > 0.1 Assume that (1. one could well replace Rt by Rtnr(u) or RtA.s. THE RISK PROCESS 3 Note that it is a matter of taste (or mathematical convenience) whether one allows {Rt} and/or {St} to continue its evolution after the time T(u) of ruin. then M = oo a.
.. _ St __ k =1 Uk pt a4.oo t 0 J (provided the limit exists).s. (1. 0 We shall only encounter a few instances of a Cox process. t t p  p' t ^ oo. and that . and independent of {Nt}. it is not too difficult to show that p as defined by (1.11) . and here (1. The simplest concrete example (to be studied in Chapter III) is the compound Poisson model. this needs to be verified in each separate case.d.3 Assume p 54 1 and define Rt = Rt1p. corresponding to the Pdlya process. are i. are i.s.2 (Cox PROCESSES) Here {Nt} is a Poisson process with random rate /3(t) (say) at time t.. U2. However..v.b(u) < 1 for all u when rl > 0. 0(u. where {Nt} is a Poisson process with rate . namely that M = oo a. Then the connection between the ruin probabilities for the given risk process {Rt} and those ^(u).i. St In concrete models.d. tb(u) = 1 for all u holds also when rl = 0.i. Nt)}.8).s. (1.10) is a property which we will typically encounter.8) that F N. This case is referred to as the mixed Poisson process. we obtain typically a somewhat stronger conclusion. and independent of {(0(t).T) for {Rt} is given by V)(u) = t/i (u). . zP(u . in connection with risk processes in a Markovian or periodic environment (Chapter VI). Here it is easy to see that p = .6EU (on the average. However.4 CHAPTER I. If u oo. then this limit is > 0 which implies St a$ oo and hence M = oo a. U2.T) = i. The simplest example is 3(t) = V where V is a r . if {(3(t)} is nonergodic.8) is given by ^t p = EU • lim it (3(s) ds t.Tp)..10) Again. If U1. . . with the most notable special case being V having a Gamma distribution. (1. Thus p may well be random for such processes. rl > 0. M < oo a.10) hold with p constant. Proposition 1. INTRODUCTION Proof It follows from (1. not all models considered in the literature have this feature: Example 1.Q (say) and U1. then similarly limSt/t < 0. If 77 < 0. namely. k=1 (1.Q claims arrive per unit time and the mean of a single claim is EU) and that also Nt t aoo t lira EEUk = p.i(u.
Mitteilungen der Verein der Schweizerischen Versicherungsmathematiker and the Scandinavian Actuarial Journal. The Swedish school was pioneering not only in risk theory. Schmidli. Sundt [354]. Rolski. Taylor [364]. Daykin et al. De Vylder [110]. Insurance: Mathematics and Economics. 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]. in particular. Gerber [157]. Cox processes are treated extensively in Grandell [171]. [330]. Note that life insurance (e. Heilmann [191]. We roughly classify these into two groups . Schmidt & Teugels [307] and Seal [326]. in a number of models. was largely initiated in Sweden in the first half of the century. Some early surveys are given in Cramer [91].2. but in probability and applied probability as a whole. often referred to as collective risk theory or just risk theory. lighttailed distributions (sometimes the term . and we do not get near to the topic anywhere in this book. the assumption > 0 is equivalent to p < 1. some main texts (typically incorporating some ruin theory but emphasizing the topic to a varying degree) are Bowers et al. [101]. which is feasible since in most cases the process { Rt } has a similar structure as {Rt} (for example. Segerdahl [334] and Philipson [289]. In the even more general area of nonlife insurance mathematics. Buhlmann [82]... Daykin. while the first mathematically substantial results were given in Lundberg [251] and Cramer [91]. The term risk theory is often interpreted in a broader sense than as just to comprise the study of ruin probabilities. the recent survey by Grandell [173] and references therein. the research literature is often published in journals like Astin Bulletin . Note that when p = 1. we shall be able to identify p with the traffic intensity of an associated queue. see also Chapter XI.g. many results and methods in random walk theory originate from there and the area was ahead of related ones like queueing theory. Besides in standard journals in probability and applied probability. the role of the result is to justify to take p = 1. Hipp & Michel [198]. Gerber [159]) has a rather different flavour. [134]. Pentikainen & Pesonen [101]. Embrechts et al. Grandell [171].. Since { Rt } has premium rate 1. and in fact p < 1 is the fundamental assumption of queueing theory ensuring steadystate behaviour (existence of a limiting stationary distribution). For mixed Poisson processes and Polya processes.g. Straub [353]. Some of the main general ideas were laid down by Lundberg [250]. Notes and references The study of ruin probabilities. see e . Some main later textbooks are (in alphabetical order) Buhlmann [82].. An idea of the additional topics and problems one may incorporate under risk theory can be obtained from the survey paper [273] by Norberg. U2. CLAIM SIZE DISTRIBUTIONS 5 The proof is trivial. the claim arrivals are Poisson or renewal at the same time). [76].
P B[s]= (8Is ) . 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'. On the more heuristical side. INTRODUCTION 'Cramertype conditions' is used).2 (THE GAMMA DISTRIBUTION) The gamma distribution with parameters p. if 1 °O AB Jbos x B(dx) > 0.3) . Example 2 . 6 has density r(p)xPleax b(x) P and m.f. but different more restrictive definitions are often used: subexponential.x given X > x is again exponential with rate b (this is essentially equivalent to the failure rate being constant).g.2) = 0. for the compound Poisson model with exponential claim sizes the ruin probability . 2a Lighttailed distributions Example 2. (2. regularly varying (see below) or even regularly varying with infinite variance. a simple stopping time argument shows that this implies that the conditional distribution of the overshoot ST(u) . the m. For example in the compound Poisson model. The crucial feature is the lack of memory: if X is exponential with rate 6. Equivalently. i. s<8. where B(bo.f. In contrast. In particular. a fact which turns out to contain considerable information.g. then the conditional distribution of X . and can also be interpreted as the (constant) failure rate b(x)/B(x).e. Here lighttailed means that the tail B(x) = 1 . B is heavytailed if b[s] = oo for all s > 0. As in a number of other applied probability areas.B(x) satisfies B(x) = O(e8x) for some s > 0.1) The parameter 6 is referred to as the rate or the intensity. the exponential distribution is by far the simplest to deal with in risk theory as well.u at the time of ruin given r(u) is again exponential u with rate 8.8. B[s] is finite for some s > 0.O(u) can be found in closed form.6 CHAPTER I.2 and /LB is the mean of B. and heavytailed distributions.1 (THE EXPONENTIAL DISTRIBUTION) Here the density is b(x) = beax (2.
u Example 2 . are i. 0 < ai < 1. where X1. p) °° where r (x. if p is integer and X has the gamma distribution p. or just the Erlang(p) distribution. In particular. u .y i=1 where >i ai = 1.3 (THE HYPEREXPONENTIAL DISTRIBUTION) This is defined as a finite mixture of exponential distributions. and exponential with rate d. is > 1. i = 1.. This special case is referred to as the Erlang distribution with p stages. The exact form of the tail B(x) is given by the incomplete Gamma function r(x. B(x) = r(p) Asymptotically. P b(x) = r` aibiea.d. we develop computationally tractable results mainly for the Erlang case (p = 1. .ate (b2 ): L• i=o In the present text.1 Poisson events in [0.. the squared coefficient of variation (s. by Grandell & Segerdahl [175] and Thorin [369]..2) can be considered as the pth power of the exponential density (2. p).2.c. CLAIM SIZE DISTRIBUTIONS 7 The mean EX is p/b and the variance Var X is p/b2. JP 1 B(x) r(p ) XP ie ax In the sense of the theory of infinitely divisible distributions. X2. p) = J tPletdt. the Gamma density (2.. then X v Xl + • • • + X. In particular. Ruin probabilities for the general case has been studied. 2.. An important property of the hyperexponential distribution is that its s. .i..) VarX1 (EX )2 p is < 1 for p > 1.c.v. .). x] so that B(x) = r` e.v. one has r(bx. An appealing feature is its simple connection to the Poisson process: B(x) = P(Xi + • • • + XP > x) is the probability of at most p . among others. p. > 1 for p < 1 and = 1 for p = 1 (the exponential case).1) (or the 1/pth root if p < 1).. 0.
We give some theory for matrixu exponential distribution in VIII.7) q1 b(x) = cjxieWWx + djxi cos(ajx)ea'x + > ejxi sin(bjx)e`ix . q2 q3 (2.6. The parameters of a phasetype distribution is the set E of transient states. Example 2 .(2. This class of distributions is popular in older literature on both risk theory and queues. are b(x) = aeTxt. 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. (or. Important special cases are the exponential.f. The density and c. which is slightly smaller but more amenable to probabilistic reasoning.e.8) are realvalued. equivalently.8 CHAPTER I.7) are possibly complexvalued but the parameters in (2. it is notable from a practical point of view because of reinsurance: if excessofloss reinsurance has been arranged with retention level xo. a. a rational Laplace transform) if B[s] _ p(s)/q(s) with p(s) and q(s) being polynomials of finite degree.f. INTRODUCTION Example 2 .. Example 2 .d. resp. We give a more comprehensive treatment in VIII. However. The couple (a. then the claim size which is relevant from the point of view of the insurance company itself is U A xo rather than U u (the excess (U . T) or sometimes the triple (E.6 (DISTRIBUTIONS WITH BOUNDED SUPPORT) This example (i. of which one is absorbing and the rest transient. but the current trend in applied probability is to restrict attention to the class of phasetype distributions.g. the restriction T of the intensity matrix of the Markov process to E and the row vector a = (ai)iEE of initial probabilities.. . 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. See XI. 1)' is the column vector with 1 at all entries. the Erlang and the hyperexponential distributions.xo)+ is covered by the reinsurer). T) is called the representation. B(x) > 0 for x < xo) is of course a trivial instance of a lighttailed distribution.8) j=1 j=1 j=1 where the parameters in (2.4 (PHASETYPE DISTRIBUTIONS) A phasetype distribution is the distribution of the absorption time in a Markov process with finitely many states. B(x) = aeTxe where t = Te and e = (1 . there exists a xo < oo such that B(x) = 0 for x > xo.6.5 (DISTRIBUTIONS WITH RATIONAL TRANSFORMS) A distribution B has a rational m.
However.1.11) ex log logx 2r p 1 1 2 ( a ) f 1 (lox_P)2} (2. we obtain the Weibull distribution B(x) = eCx'.pl = 1 W (logx . p is defined as the distribution of ev where V .12) Sometimes also a location parameter a > 0 and a scale parameter A > 0 is allowed. the mean u is eµ+a /2 and the second moment is e2µ+2o2. b(x) = crx''le`xr. (2. or equivalently as the distribution of a°U+µ where U . (2.u l b(x) = d dx or J ax lor 1 exp Asymptotically. one being B(x) (1 + X)b(x) (1 + x)a+1' x > 0. a)/A)a+1' x > a. and then b(x) = 0. CLAIM SIZE DISTRIBUTIONS 9 2b Heavytailed distributions Example 2. Here failure rates b(x) = b(x)/B(x) play an important role.1). Writing c = d/r. Example 2 .8 (THE LOGNORMAL DISTRIBUTION) The lognormal distribution with parameters a2. the exponential distribution representing the simplest example since here b(x) is constant.p a 1 (2.2.7 (THE WEIBULL DISTRIBUTION) This distribution originates from reliability theory. the tail is B (x ) 2 x. a2).N(0.9 (THE PARETO DISTRIBUTION) Here the essence is that the tail B(x) decreases like a power of x.13) u . 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).9) which is heavytailed when 0 < r < I. In particular. All moments are finite. x < a. It follows that the density is 't (1ogX .N(p. (2. b(x) _ A(1 + (x a The pth moment is finite if and only if p < a .10) The loinormal distribution has moments of all orders. There are various variants of the definition around. u Example 2 .
{ s () 1s+3s29s3log(1+2s I p=3. The simplest examples correspond to p small and integervalued. 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 . u Example 2 . in particular.16) 11 Example 2. u . Choudhury & Whitt [1] as the class of distributions of r. 6 is defined as the distribution of et' where V has the gamma density (2.L( x ).13).14) The pth moment is finite if p < 5 and infinite if p > 5.x6+lr(p) (2.10 CHAPTER I.10 (THE LOGGAMMA DISTRIBUTION) The loggamma distribution with parameters p.17) where L (x) is slowly varying. the loggamma distribution is a Pareto distribution. where Y is Pareto distributed with a = (p .12 (DISTRIBUTIONS WITH REGULARLY VARYING TAILS) The tail B(x) of a distribution B is said to be regularly varying with exponent a if B(x) . A = 1 and X is standard exponential. satisfies L(xt)/L(x) 4 1. B(x) = O(xP). the loggamma distribution (with exponent 5) and a Pareto mixture of exponentials. x 4 oo (any L having a limit in (0.'s of the form YX. In general.2).1)/p. (2. Thus. another standard example is (log x)'). (2. in particular. INTRODUCTION Example 2. The density is 8p(log x)pi b(x) .11 (PARETO MIXTURES OF EXPONENTIALS) This class was introduced by Abate. i.e. x + 00. the density is { 3 (1 .(1 + 2x + 2x2)e2x) p = 2 (2.(1 + Zx + $ p = 3.15) x2 + 16x3 ) a3x/2) 3 (1 . For p = 1.12) (here L (x) * 1) and ( 2. examples of distributions with regularly varying tails are the Pareto distribution (2.v.
18) B(x) It can be proved (see IX. Also.3. but can never be sure whether this is also so for atypical levels for which far less detailed statistical information is available. We give some discussion on standard methods to distinguish between light and heavy tails in Section 4f. However. (2. 3 The arrival process For the purpose of modeling a risk process . U2. we may know that such a process (with a covariance function estimated from data) is a reasonable description of the behaviour of the system under study in typical conditions. and based upon such information it seems questionable to extrapolate to tail behaviour... the subexponential class of distributions provide a convenient framework for studying large classes of heavyu tailed distributions. but the model also admits a natural interpretation : a large portfolio of insurance holders . The reason is in part mathematical since this model is the easiest to analyze.. 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.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. which each have a ( timehomogeneous) small rate of experiencing a . THE ARRIVAL PROCESS 11 Example 2. this phenomenon represents one of the true controversies of the area.. Thus. it will be seen that we obtain completely different results depending on whether the claim size distribution is exponentially bounded or heavytailed. the knowledge of the claim size distribution will typically be based upon statistical data. Similar discussion applies to the distribution of the accumulated claims (XI.13 (THE SUBEXPONENTIAL CLASS OF DISTRIBUTIONS) We say that a distribution B is subexponential if xroo lim B `2^ = 2. one may argue that this difficulty is not resticted to ruin probability theory alone. When studying ruin probabilities. for example the lognormal distribution is subexponential (but not regularly varying).1) that any distribution with a regularly varying tail is subexponential.1. the claim size distribution represents of course only one aspect (though a major one). Namely. We return to a closer study in IX. and so is the Weibull distribution with 0 < r < 1. From a practical point of view. By far the most prominent case is the compound Poisson (CramerLundberg) model where {Nt} is Poisson and independent of the claim sizes U1. though the proof of this is nontrivial.
Some of them have concentrated on the marginal distribution of NT (say T = one year ). where {/3 (t)}too is an arbitrary stochastic process . 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 . are i. the first extension to be studied in detail was {Nt } to be renewal (the interarrival times T1 . Mathematically.. Nevertheless .. Another one is Cox processes. with the extension to premiums depending on the reserve. Historically. IV (and.8 (t) is a periodic function of t. gives rise to an arrival process which is very close to a Poisson process.3(t) fluctuating over time.. in particular to allow for certain inhomogeneities. In others.6. its basic feature is to allow more variation (bursty arrivals ) than inherent in the simple Poisson process. found the Poisson distribution to be inadequate and suggested various other univariate distributions as alternatives . the negative binomial distribution. epidemics in life insurance etc. when Jt = i.e. with a common term {Nt} is a Markovmodulated Poisson process . not many detailed studies of the goodnessoffit of the Poisson model in insurance are available . The one we focus on (Chapter VI) is a Markovian environment : the environmental conditions are described by a finite Markov process {Jt }too. and also that the ruin problem may be hard to analyze . which facilitate the analysis. such that 8(t) = . 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. in just the same way as the Poisson process arises in telephone traffic (a large number of subscribers each calling with a small rate).g. however. it is more questionable whether it provides a model with a similar intuitive content as the Poisson model. 5. The compound Poisson model is studied in detail in Chapters III. T2. This model . e. we study this case in VI . To the author 's knowledge . to be studied in Chapter V. The point of view we take here is Markov dependent random walks in continuous time (Markov additive processes ). Cox processes are. In order to prove reasonably substantial and interesting results . An obvious example is 3(t) depending on the time of the year (the season). getting away from the simple Poisson process seems a crucial step in making the model more realistic. A more appealing way to allow for inhomogeneity is by means of an intensity .i. see 11. the periodic and the Markov modulated models also have attractive features . radioactive decay (a huge number of atoms each splitting with a tiny rate ) and many other applications.12 CHAPTER I.d. too general and one neeed to specialize to more concrete assumptions . has some mathematically appealing random walk features . However . so that . This model can be intuitively understood in some simple cases like { Jt} describing weather conditions in car insurance . but with a general not necessarily exponential distribution ).. INTRODUCTION claim . This applies also to the case where the claim size distribution depends on the time of the year or . in Chapter VII).(3.
and here (4.4. 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. extreme value theory. More generally. point processes and so on. genetics models.1) holds as well provided the risk process has a premium rule depending on the reserve. others are quite different. stochastic differential equations. with Poisson arrivals and constant release rule p(x) = 1. A general release rule p(x) means that {Vt} decreases according to the differential equation V = p(V) in between jumps. The M/G/1 workload process { Vt } may also be seen as one of the simplest storage models. and the limit t 4 oo is the steadystate limit. The ones which appear most related to risk theory are queueing theory and dam/storage processes.1).0 (u. this amounts to Vo having the stationary distribution of {Vt}). Similarly. and a lot of information on steadystate r. In fact. It should be noted.1) where V is the limit in distribution of Vt as t + oo. A SUMMARY OF MAIN RESULTS AND METHODS 13 the environment (VI. 0(u) = P(V > u). dam/storage processes. interacting particle systems. Mathematically. reliability.v.'s like V is available. others being branching processes. 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 . it is a recurrent theme of this book to stress this connection which is often neglected in the specialized literature on risk theory. time series and Gaussian processes. The study of the steady state is by far the most dominant topic of queueing and storage theory. A stochastic process {Vt } is said to be in the steady state if it is strictly stationary (in the Markov case. ruin probabilities for risk processes with an input process which is renewal. R = p(R) in between jumps.T) = P(VT > u).1) permitting to translate freely between risk theory and the queueing/storage setting. queueing theory. Some of these have a certain resemblance in flavour and methodology. Markovmodulated or periodic can be related to queues with similar characteristics. it is desirable to have a set of formulas like (4. stochastic geometry. (4. and which seems well motivated from a practical point of view as well. Thus. methods or modeling ideas developed in one area often has relevance for the other one as well. In the setting of (4. this gives only f0 O°i (u)du which is of limited .6) . that quite often the emphasis is on computing expected values like EV. however.
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.1) in the setting of a general premium rule p(x): the events {VT > u} and {r (u) < T} coincide when the risk process and the storage process are coupled in a suitable way (via timereversion ). • The compound Poisson model with constant premium rate p = 1 and B being phasetype with a just few phases .3.v. e .. see Corollary VII. • The compound Poisson model with a claim size distribution degenerate at one point.T).2). have to some extent a different flavour.6.1. 3.14 CHAPTER I.1 . Similarly. The qualifier 'with just a few phases ' refers to the fact that the diagonalization has to be carried out numerically in higher dimensions. The infinite horizon (steady state ) case is covered by letting T oo. 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.1). the ideal is to be able to come up with closed form solutions for the ruin probabilities 0(u). the two areas. which gives a sample path version of (4. Example VIII. • The compound Poisson model with premium rate p(x) depending on the reserve and exponential claim size distribution B.p(y) y^ Jo p(x) can be written in closed form.1 is a sample path relation should be stressed : in this way the approach also applies to models having supplementary r. Thus . though overlapping.8. Vi(u. 4b Exact solutions Of course . B(x) = ebx. The fact that Theorem H. which can be expanded into a sum of exponential terms by diagonalization (see. the functions w x f d 1 exdx () .3.3. Here Vi(u) is given in terms of a matrixexponential function ( Corollary VIII. Here ?P(u) is explicit provided that . as is typically the case. . see Boxma & Cohen [74] and Abate & Whitt [3]. 3. • The compound Poisson model with some rather special heavytailed claim size distributions. A prototype of the duality results in this book is Theorem 11. Here O(u) = peryu where 3 is the arrival intensity. INTRODUCTION intrinsic interest .g. see Corollary III. p = 0/8 and y = 8 .'s like the environmental process {Jt} in a Markovmodulated setting.3.
1) are so complicated that they should rather be viewed as basis for numerical methods than as closedform solutions. However. T) themselves. Here are some of the main approaches: Laplace transform inversion Often. [s. the second best alternative is a numerical procedure which allows to calculate the exact values of the ruin probabilities.4. .S(u) 1S(oo) f °D exp {. Given this can be done. the formulas ( IV. f {eXp U LX 2. the only example of something like an explicit expression is the compound Poisson model with constant premium rate p = 1 and exponential claim size distribution .f f 2µ(y)/a2(y) dy} dx  (4.b(u)du .u(y)/a2(y) dy} 4c Numerical methods Next to a closedform solution.ff 2µ(y)/a2(y) dy} dx .2) is the natural scale. Also Brownian models or certain skip free random walks lead to explicit solutions (see XI .1) is the explicit form of the ruin probability when {Rt} is a diffusion with infinitesimal drift and variance µ(x). a2 (x): Ip (u) = where S(u) = f °O exp {. 191). but are somewhat out of the mainstream of the area . (u.Lef$er function. see VIII. 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. T) du dT 0 TO 00 in closed form than the ruin probabilities z/'(u). We don't discuss Laplace transform inversion much. 1). O(u. where Furrer [150] recently computed ii(u) as an infinite series involving the Mittag. Ab(u). • An astable Levy process with drift .7. Embrechts. For the finite horizon ruin probability 0(u. Grubel & Pitts [132] and Grubel & Hermesmeier [180] (see also the Bibliographical Notes in [307] p. Abate & Whitt [2]. relevant references are Grubel [179]. it is easier to find the Laplace transforms = f e8 . say the fast Fourier transform (FFT) as implemented in Grubel [179] for infinite horizon ruin probabilities for the renewal model. T). T) can then be calculated numerically by some method for transform inversion. A notable fact ( see again XI.
0(u) is then given in terms of a matrixexponential function euu (here U is some suitable matrix) which can be computed by diagonalization. One example where this is feasible is the renewal equation for tl'(u) (Corollary III. it states that i/i(u) . and in particular the naive idea of conditioning upon process behaviour in [0. T) as the solution to a differential. . In the compound Poisson model with p = 1. 4d Approximations The CramdrLundberg approximation This is one of the most celebrated result of risk theory (and probability theory as a whole). U is explicit in terms of the model parameters.or integral equation. whereas for the renewal arrival model and the Markovian environment model U has to be calculated numerically. u * oo. For the compound Poisson model with p = 1 and claim size distribution B with moment generating function (m. 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. 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.Ce"u.16 CHAPTER L INTRODUCTION Matrixanalytic methods This approach is relevant when the claim size distribution is of phasetype (or matrixexponential). see VIII.3) in the compound Poisson model which is an integral equation of Volterra type.4) 00['Y]1)'Y = 0.3) where C = (1 .1) and y > 0 is the solution of the Lundberg equation (4. Differential.7. dt] most often leads to equations involving both differential and integral terms.p)/(13B'[ry] .g. most often it is more difficult to come up with reasonably simple equations than one may believe at a first sight.and integral equations The idea is here to express 'O(u) or '(u.3.) B[s]. and in quite a few cases (Chapter VIII). as the solution of linear differential equations or by some series expansion (not necessarily the straightforward Eo U'u/n! one!). and carry out the solution by some standard numerical method. However.f. (4. which can equivalently be written as f3 [7] = 1 +13 .
often for all u > 0 and not just for large u. u > oo.6) 4e Bounds and inequalities The outstanding result in the area is Lundberg's inequality (u) < e"lu. and use the fact that first passage probabilities are more readily calculated for diffusions than for the risk process itself.6) are by far the best one can do in terms of finite horizon ruin probabilities '(u. 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.2. It has generalizations to the models with renewal arrivals. This list of approximations does by no means exhaust the topic. a Markovian environment or periodically varying parameters. For example. in such cases the evaluation of C is more cumbersome. the exact solution is as easy to compute as the CramerLundberg approximation at least in the first two of these three models. In the case of heavytailed distributions. in some cases the results are even more complete than for light tails. but typically not very precise in their first naive implementation.7 and IV. However. corrected diffusion approximations (see IV. the claim size distribution should have an exponentially decreasing tail B(x). J B dx. Diffusion approximations are easy to calculate. . In fact. The CramerLundberg approximation is renowned not only for its mathematical beauty but also for being very precise.4. See Chapter IX. other approaches are thus required. T). (4. In particular. incorporating correction terms may change the picture dramatically. Approximations for O(u) as well as for 1(u. some further possibilities are surveyed in 111 . Large claims approximations In order for the CramerLundberg approximation to be valid. when the claim size distribution is of phasetype. However. T) for large u are available in most of the models we discuss. for the compound Poisson model ^(u) p pu In fact . 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.
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. more importantly. . For example. . This procedure in itself is fairly straightforward.3). . to have smaller ruin probabilities than when B is nondegenerate with the same mean m. We return to various extensions and sharpenings of Lundberg's inequality (finite horizon versions. can we trust the confidence intervals for the large values of u which are of interest? In the present author's opinion. they have however to be estimated from data.. 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. . 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. UNT may be viewed as an interpolation in or smoothing of the histogram).g. In practice. However. However.18 CHAPTER I. 4f Statistical methods Any of the approaches and results above assume that the parameters of the model are completely known. say degenerate at m. and to plot the empirical mean residual life 1 N . UNT are i. it is a general principle that adding random variation to a model increases the risk.k (U(`) .U(k)) i =k+ i . obtained say by observing the risk process in [0.) at various places and in various settings. . This is proved for the compound Poisson model in 111. fitting a parametric model to U1. it has the advantage of not involving approximations and also. How do we produce a confidence interval? And. though not too many precise mathematical results have been obtained. given NT. The standard suggestion is to observe that the mean residual life E[U .. e.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. in the compound Poisson model. which is a standard statistical problem since the claim sizes Ui.i. T]. one expects a model with a deterministic claim size distribution B. empirical evidence shows that the general principle holds in a broad variety of settings. lower bounds etc. of being somewhat easier to generalize beyond the compound Poisson setting. . When comparing different risk models. the difficulty comes in when drawing inference about the ruin probabilities.8.d. For example.. INTRODUCTION Compared to the CramerLundberg approximation (4.
Thus Proposition 4.. Truncation to a finite horizon has been used. formula (5. 5 Conventions Numbering and reference system The basic principles are just as in the author's earlier book Applied Probability and Queues (Wiley 1987. The infinite horizon case presents a difficulty. in all other chapters than VI where we just write . (or a functional of the expectation of a set of r. good methods exist in a number of models and are based upon representing the ruin probability zb(u) as expected value of a r. to observe whether one or the other limiting behaviour is apparent in the tail. For example. in this book referred to as [APQ]). and look at them to see whether they exhibit the expected behaviour or some surprises come up. The problem is entirely analogous to estimating steadystate characteristics by simulation in queueing/storage theory. but is not very satisfying. UN.3 (or just VI.5. and also discuss how to develop methods which are efficient in terms of producing a small variance for a fixed simulation budget. having small probability) and that therefore naive simulation is expensive or even infeasible in terms of computer time. 4g Simulation The development of modern computers have made simulation a popular experimental tool in all branches of applied probability and statistics. claims U1. respectively.2. The chapter number is specified only when it is not the current one. Klnppelberg & Mikosch [134].i. and of course the method is relevant in risk theory as well. CONVENTIONS 19 as function of U(k). formula VI.(5.3) or Section 3 of Chapter VI are referred to as Proposition VI.. See further Embrechts.3) and Section VI. and in fact methods from that area can often be used in risk theory as well .. .3). the more typical situation is to perform a Monte Carlo experiment to estimate probabilities (or expectations or distributions) which are not analytically available. reference [14].. Simulation may be used just to get some vague insight in the process under study: simulate one or several sample paths.. We look at a variety of such methods in Chapter X. .e. Still.v.2. because it appears to require an infinitely long simulation. A main problem is that ruin is typically a rare event (i. However. where U(1) < . this is a straightforward way to estimate finite horizon ruin probabilities.d. < U(N) are the order statistics based upon N i.v's) which can be generated by simulation.4.
for a probability distribution IIGII = 1. infinitely often l. EX2/(EX)2. formula (5. n i oo. (A.g. IIGII the total mass (variation ) of a (signed ) measure G .Used in asymptotic relations to indicate that the ratio between two expressions is 1 in the limit.B(x) = P(X > x) of B.v.g.f.r. if B(x) . B(x) = P(X < x) = fx. The Laplace transform is b[s]. left hand side (of equation) m. oo).s. independent identically distributed i. with respect to w. as for typical claim size distributions.g. .h. moment generating function. b[s] is defined always if Rs < 0 and sometimes in a larger strip (for example. (moment generating function) fm e82B(dx) of the distribution B. If.3) or Section 3. squared coefficient of variation. A different type of asymptotics: less precise.f.2.f. log E[s] where b[s] is the m.s. In particular.t. say a heuristic approxi1 + h + h2/2. or a more precise one like eh . h + 0.e. then for 1s < 5).g.g. i. random variable s.20 CHAPTER L INTRODUCTION Proposition 4. right hand side (of equation) r.The same symbol B is used for a probability measure B(dx) = P(X E dx) and its c. and for a defective probability distribution IIGII < 1.29) refer to the Appendix.d. n! 27r nn+1/2en. cumulative distribution function P(X < x) c. E expectation. r. References like Proposition A. w.c.f.ceax. mation.d.h.d. B(x) the tail 1 .4. see under b[s] below.f.v. Abbreviations c.i. . i.f. B[s] the m. E. cumulant generating function. with probability Mathematical notation P probability. B is concentrated on [0.p. B(dy).o.
e. Usually. the value just before t. A. Unless otherwise stated. and column vectors have lowercase Roman letters like t.e. an example or a remark. Usually. only have finitely many jumps in each finite interval. (xi)diag denotes the diagonal matrix with the xi on the diagonal (xi)row denotes the row vector with the xi as components (xi). Notation like f3i and 3(t) in Chapter VI has a similar . i. i.e. CONVENTIONS {6B the mean EX = f xB(dx) of B ABA' the nth moment EXn = f x"B(dx) of B. . the ith unit row vector is e'i. often the term 'cadlag' (continues a droite avec limites a gauche) is used for the Dproperty. 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. R(s) the real part of a complex number s. oo) the space of Rvalued functions which are rightcontionuous and have left limits. the ith entry is 1 and all other 0. . all stochastic processes considered in this book are assumed to have sample paths in this space. a2) the normal distribution with mean p and variance oa2. (the dimension is usually clear from the context and left unspecified in the notation). Matrices and vectors are denoted by bold letters. 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.A] means E[XI(A)]. 0 marks the end of a proof.5. matrices have uppercase Roman or Greek letters like T. E[X. row vectors have lowercase Greek letters like a. though slightly more complicated. a. N(it. 7r. . xa. 21 D [0. intensity interpretation. F o r a given set x1.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).. I(A) the indicator function of the event A. Thus. the processes we consider are piecewise continuous.. Xt_ the left limit limstt X8f i. of numbers. In the Frenchinspired literature.
p the net amount /3pB of claims per unit time. e.1. cf.g. . EL the probability measure and its corresponding expectation corresponding to the exponential change of measure given by Lundberg conjugation. 'q the safety loading . I. or quantities with a similar time average interpretation. J the rate parameter of B for the exponential case B(x) = eby.5. 111.22 CHAPTER L INTRODUCTION B the claim size distribution. though slightly more complicated. VI. Notation like BE and B(t) in Chapter VI has a similar. cf.5. interpretation. FL. cf. ry The adjustment coefficient.1. I.
When encountered for the first time in connection with the compound Poisson model in Chapter III. 5 on random walks and Markov additive processes can be skipped until reading Chapter VI on the Markovian environment model. Sections 4. the level of the exposition is. a parallel selfcontained treatment is given of the facts needed there. strictly speaking. however. All results are proved elsewhere . The general theory is.Chapter II Some general tools and results The present chapter collects and surveys some topics which repeatedly show up in the study of ruin probabilities. The reader should therefore observe that it is possible to skip most of the chapter. used in Chapter VI on risk processes in a Markovian (or periodic) environment. 23 . however. The topic is. in most cases via likelihood ratio arguments. fundamental ( at least in the author' s opinion) and the probability involved is rather simple and intuitive. More precisely. somewhat more advanced than in the rest of the book. in particular at a first reading of the book. The duality results in Section 3 (and. the relevance for the mainstream of exposition is the following: The martingale approach in Section 1 is essentially only used here. Due to the generality of the theory. 5) are. The likelihood ratio approach in Section 2 is basic for most of the models under study. Sections 4. however. in part. not crucial for the rest of the book.
V) (u. StUit. The more general Theorem 6. T(u) < oo] + 0 = eryuE [e7Vu). however. {e'YS° }t>0 is a martingale. T(u) < T] + E [eryST . Proposition 1. 1 Martingales We consider the claim surplus process {St} of a general risk process. Example 1 .2) takes the form 1 = E [e'ys(). the time to ruin r(u) is inf It > 0 : St > u}.5 can be skipped. (b) St a$ oo on {T(u) = oo}.T(u) < cc] = e7uE {e7f(u) I T(u) < cc] z/.1 is basic for the study of the compound Poisson model in Chapter III.u denote the overshoot.)AT = E [e7ST(°). We get 1 = Ee7So = E e'Y S(. Our first result is a representation formula for O(u) obtained by using the martingale optional stopping theorem . As usual. Let e(u) = ST(u) . T(u) > T] .QµB < 1. (1.1 Assume that (a) for some ry > 0. and in the limit (1. Thus N. T) = P(T(u) < T). and the ruin probabilities are ip(u) = P (T(u) < oo). claim size distribution B and p = .2 Consider the compound Poisson model with Poisson arrival rate . the second term converges to 0 by (b) and dominated convergence (e7ST < eryu on {r(u) > T}).0.24 CHAPTER II.(u). f1 . SOME GENERAL TOOLS AND RESULTS The ladder height formula in Theorem 6. using r(u) A T invokes no problems because r(u) A T is bounded by T). Then e7u (u) = E[e74(u)j7(u) < oo] Proof We shall use optional stopping at time r(u)AT (we cannot use the stopping time T(u) directly because P(T(u) = oo) > 0 and also because the conditions of the optional stopping time theorem present a problem.2) As T > oo..
Proof Since c(a) = /3 (B[a] .1) . Then {St } is Brownian motion with variance constant o2 and drift p < 0. Now at the time r(u) of ruin {St} upcrosses level u by making a jump .2(c)). Under the conditions of Proposi Proof Just note that C(u) > 0..1 is satisfied. and thus Ee'rs° = 1. are i.Ft = a(S" : v < t).1. condition (a) of Proposition 1. and p =.= e"(') where K(a) = .ap.1 are satisfied. Since {St} has stationary independent increments. Thus.1) shows that Eels. Example 1 . of the normal distribution. the martingale property now follows just as in Example 1.a it is immediately seen that y = S . it follows that E [e7st+v I J] = e"rstE [e7(st+vSt) I Ft] = e7StEe"rs° = elst where .Q(B[a] . the ruin probability is O(u) = pe. Eeas° = e"(°) where n(a) = a2a2/2 .2 and drift p > 0. Thus. A simple calculation (see Proposition III.a = a .d.2.1) . 1.4 (LUNDBERG ' S INEQUALITY ) tion 1 . By standard formulas for the m. From this it is readily seen (see III.5 For the compound Poisson model with B exponential.Q and the U. From this it is immediately seen that the solution to the Lundberg equation ic(y) = 0 is y = 2p/a2.Q.i. u Corollary 1.a.1.3/6 < 1. MARTINGALES 25 where {Nt} is a Poisson process with rate .x. O(u ) < e7". and (b) follows from p < 1 and the law of large numbers (see Proposition III. u Corollary 1. The available information on this jump is that the distribution given r(u) = t and S.(„)_ = x is that of a claim size U given U > u .g.r" where y = S .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.f. the conditional distribution of the overshoot e(u) = U . and thus by the memoryless property of the exponential distribution . and thus Ee7s° = 1. with common distribution B (and independent of {Nt}). Thus 00 E [e'rt (") I T(u) < oo] = I e5e  dx = f 5edx . B(x) _ edx.u + x is again just exponential with rate S. the conditions of Proposition 1./3. Since {St} has stationary independent increments.
(2. and by the law of large numbers for the Poisson process . I. Grandell [171]. However. P correspond to the claim surplus process of two compound Poisson risk processes with Poisson rates /3. as shown by the following example this setup is too restrictive: typically'. F). F(S) = P(S) = 1. cf. . 0 and claim size distributions B. P are then singular (concentrated on two disjoint measurable sets). Example 2 .26 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Corollary 1. then S and S are disjoint . Theorem 2.v.1) 'though not always: it is not difficult to construct a counterexample say in terms of transient Markov processes. 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. Two such processes may be represented by probability measures F.3 below). Grandell & Schmidli [131].. which we equip with the natural filtration {. Embrechts. u Notes and references The first use of martingales in risk theory is due to Gerber [156]. and F.1 Let F. A somewhat similar u argument gives singularity when B $ B. A E Ft. But if a $ ^ . The number Nt F) of jumps > e before time t is a (measurable) r. hence so is Nt = limfyo N2`i. then z/'(u) = e7" where 'y = 21A/a2. Proof Just note that ^(u) = 0 by continuity of Brownian motion. and in analogy with the theory of measures on finite dimensional spaces one could study conditions for the RadonNikodym derivative dP/dP to exist. we look for a process {Lt} (the likelihood ratio process) such that P(A) = E[Lt.6 N S = { lim Nt I t +00 t gJ are both in F. oo). the parameters of the two processes can be reconstructed from a single infinite path. Thus the sets S = I tlim +oot t =. [172]. More recent references are Dassios & Embrechts [98]. and is further exploited in his book [157]. The interesting concept is therefore to look for absolute continuity only on finite time intervals (possibly random. A].Ft}too and the Borel afield F. B.F).e.6 If {Rt} is Brownian motion with variance constant a2 and drift p > 0. Delbaen & Haezendonck [103] and Schmidli [320]. P on (DE. on (DE.
F8] = L8 and the martingale property. This proves (i). . 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 . G l ] = E [_I(G)E[LTIFT] ] = E { _I(G)Lr ] = P(G). Lt < 0] can only be nonnegative if P(A) = 0.1) holds.Ft}. Lets < t. P be as in Proposition 2. the restriction of P to (DE.e. F the Borel o•field and P a given probability measure on (DE. then {Lt} is a nonnegative martingale w.2. A E F.r.2(i). A E Ft .3 Let {Lt}. that the restriction of P to (DE. then { 1 P(G) = EG . Finally.Tt) is absolutely continuous w. Proposition 2. ELt = 1 follows by taking A = DE in (2. we have E [ LTIFT]1 = LT on {T < T}. P) such that LLt = 1. The truth of this for all A E Y.r. ({. If r is a stopping time and G E PT. then there exists a unique probability measure Pon . A E F8. using the martingale property in the fourth step.r. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 27 (i. A]. F) such that ELt = 1. A] = E[Lt.F).2) Proof Assume first G C {T < T} for some fixed deterministic T < oo. _.t. Then Ft (A) = E[Lt. A].Y) such that P(A) = Pt(A). if for some probability measure P and some {. under the assumptions of (ii) we have for A E rg and s < t that A E Ft as well and hence E[L8. G C {T < oo}. 1 J (2. Hence E [_ .t. (i) If {Lt}t> o is a nonnegative martingale w. Proof Under the assumptions of (i). Then Lt > 0 and ELt = 1 ensure that Pt is a probability measure on (DE. Hence the family {Pt} is t>o consistent and hence extendable to a probability measure F on (DE. (ii) Conversely.1) and nonnegativity by letting A = {Lt < 0}. ({Ft} . Ft). . Conversely. implies that E[LtI.F such that (2. define P by Pt (A) = E[Lt.Pt)) The following result gives the connection to martingales.1) holds. Then P(A) = E[Lt.A] = EE[LtI(A)IF8] = EI(A)E[LtIFB] = EI(A)L8 = PS(A). G ] = E [LT .t.Pt}adapted process {Lt}t>o (2. By the martingale property.2 Let {Ft}t>o be the natural filtration on DE..
t.s. is Markov w.r. The crucial step is to obtain information on the process evolving according to F. A change of measure is performed by finding a process {Lt} which is a martingale w.28 CHAPTER II. For the definition.1: Corollary 2.Ft} . To this end. Rt) or Xt = (Jt. Consider a (timehomogeneous) Markov process {Xt} with state space E.1) is that it seems in general easier to deal with the (unconditional) expectation E[eryVu).1) in Proposition 1. Now just rewrite the r. where {Rt} is the risk reserve process. St).1. Lr(u) 11 The advantage of (2. In the context of ruin probabilities.3 we obtain a likelihood ratio representation of the ruin probability V) (u) parallel to the martingale representation (1. . 1 Since everything is non negative. Xt = St. r(u) < oo] occuring there than with the (conditional) expectation E[e'r{(u ) Jr(u) < oo] in (1..Gn {r <T} . SOME GENERAL TOOLS AND RESULTS In the general case . first in the Markov case and next (Sections 4. we have F(G) = V )(u). the natural filtration {.4) Proof Letting G = {r(u) < oo}. is nonnegative and has Ey Lt = 1 for all x. say. {St} = {u .t.1). First we ask when the Markov property is preserved. 5) for processes with some randomwalklike structure.4 Under condition (a) of Proposition 1. u From Theorem 2. and this problem will now be studied.2) by noting that 1 = ersr(„) = e1'ue7Ou). in continuous time (the discrete time case is parallel but slightly simpler). T(u) < oo]. applying (2.. and letting T * oo.3) to G of{r < T} we get 1111 F(Gn {r <T}) = E[ 1 . we need the concept of a multiplicative functional.O(u) = eryuE[e'YC(u). t. (2.4) compared to (1.2) follows by monotone convergence. of (2. one would typically have Xt = Rt. The problem is thus to investigate which characteristics of {Xt} and {Lt} ensure a given set of properties of the changed probability measure.Rt} the claim surplus process and {Jt} a process of supplementary variables possibly needed to make the process Markovian.h. (2. both sides are increasing in T.r. we assume for simplicity that {Xt} has Dpaths. Xt = (Jt. each F.
(Xtitl) with all t(i) < t + s.(A) = Ex [Lt. The converse follows since the class of r.Ft].7). t and let Px be the probability measure given by t.2. which in turn is the same as Ex[Lt+8Zt • (V8 o Bt)] = Ex[Lt • (L8 o 91)Z1 • (Y8 o et)] (2. since Ext [L8Y8] = E[(Y8 o et)(L8 o 8t)I.v.Ft] = Ex.5) is equivalent to Ex[Lt+8Vt+8] = E8[Lt • (L8 o 91)Vt+8] for any .Pt+8measurable r. let {Lt} be a nonnegative martingale with Ex Lt = 1 for all x.Ftmeasurable. Vt+e. o 9t = V. oo). ({Xt+u} 0<u<8) Theorem 2.T9measurable Y8. A]. o 9tI. The precise meaning of this is the following: being . t] * [0. (2. Lt has the form Lt = 'Pt ({x }0<u<t) for some mapping cot : DE[O.s. (2.5) Pxa. Zt.v. s.. where Ot is the shift operator.'s of the form Zt • (Y8 o 0t) comprises all r. Similarly. Y8. By definition of Px.v.t.5 Let {Xt} be Markov w. Indeed. t.'s of the form fl' f. since Zt • (Y8 o Ot ) is .Ft+8measurable. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 29 on DE and define {Lt} to be a multiplicative functional if {Lt} is adapted to {. and then L.5) are Tt+e measurable. which is the same as Ex[Zt(Y8 o Bt)] = E8[ZtEx. nonnegative and Lt+8 = Lt•(Lso9t) (2.Ft }. Then the family {Px}xEE defines a timehomogeneous Markov process if and only if {Lt} is a multiplicative functional.6) implies (2. or. the natural filtration {Ft} on DE. for all x. 0 .r.6) for any . the Markov property can be written E. Proof Since both sides of (2.v.F8measurable r.8) which is the same as (2.7).Y8f t < s.. Ex[Lt+8Zt(Y8 o et)] = Ex[LtZt(Y8 o 0t)(L8 o Bt)].7) for any Ftmeasurable Zt and any . (2. this in turn means Ex[Lt+8Zt(V8 oet)] = Ex[LtZtExt[L8Y8]].v.[Y.YB] for any Ftmeasurable r. (2.
.. . In between jumps. aN where or* = T UN_k+l.1) The initial condition is arbitrary. 3 Duality with other applied probability models In this section. We work on a finite time interval [0. } E[Lt+B I. . More precisely . Thus R = Ro + f p(R8) ds . u Notes and references The results of the present section are essentially known in a very general Markov process formulation.. reflection at zero and initiar condition Vo = 0. (3.e. R = p(R)). In between jumps. T] in the following setup: The risk process {Rt}o<t<T has arrivals at epochs or. {Vt} . The corresponding claim sizes are Ul.At where At = k: vk <t U.. . with a proof somewhat different from the present one. it xEE suffices to assume that {Lt} is a multiplicative functional with Ex Lt = 1 for all x. t. t] = LtExt L8 = Lt.6 For {u .. The result is a sample path relation. SOME GENERAL TOOLS AND RESULTS to define a timehomogeneous Markov process.. then Remark 2... and just after time or* {Vt} makes an upwards jump of size UU = UN _k+l. Ro = u (say). . 0 < vl < . A more elementary version along the lines of Theorem 2.30 CHAPTER H.. < aN < T. the arrival epochs are Qi. and thus for the moment no parametric assumptions (on say the structure of the arrival process) are needed.5 can be found in Kuchler & Sorensen [240]. . see Dynkin [128] and Kunita [239]. we shall establish a general connection between ruin probabilities and certain stochastic processes which occurs for example as models for queueing and storage. and the time to ruin is 7(u) = inf {t > 0: Rt < 0}. UN. the premium rate is p(r) when the reserve is r (i. (using the Markov property in the second step) so that the martingale property is automatic.. Indeed.Ft] = LtE[L8 o 9t I.. The storage process {Vt }o<t<T is essentially defined by time reversion. CN. The formulation has applications to virtually all the risk models studied in this book.
3.. . instead of (3. In particular.. The sample path relation between these two processes is illustrated in Fig.___ ._. Note that these definitions make {Rt} rightcontinuous (as standard) and {Vt} leftcontinuous.__.3) (u) Proof Let rt' denote the solution of R = p(R) subject to r0 = u.. DUALITY WITH OTHER APPLIED PROBABILITY MODELS 31 decreases at rate p(r) when Vt = r (i..1. 3.1 Define r(u) = inf It > 0: Rt < 0} (r(u) = oo if Rt > 0 for all t < T) and let ii(u.T) = P(VT > u).e. That is.T) = inf Rt < 0 P (O<t<T P(r(u) < T) be the ruin probability. {Vt} remains at 0 until the next arrival).AT_t._: 1} 0 011 =T01N ^N3 To 0 011 014 01N Figure 3. (3.__... V)(u....____•_.1 The events {T(u) < T} and {VT > u} coincide.x.11 4. V = p(V)).2) k: ok <t and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0.1) we have Vt = At  f P(Vs)ds where A= U= AT . Then rt°) > rt°) for all t when u > v. Theorem 3. (3. :.
2 from Harrison & Resnick [188]. Suppose next VT < u (this situation corresponds to the broken path of {Rt} in Fig. and then '0 (u) = P(V > u).2 Consider the compound Poisson risk model with a general premium rule p(r).U1 > roil . Resnick & Tweedie [79]. Then Vo. Nevertheless. 3.3). we have r(u) > T. the distinction between right. and since ruin can only occur at the times of claims. the connection between risk theory and other applied probability areas appears first to have been noted by Prabhu [293] in a queueing context. Hence if n satisfies VVN_n+1 = 0 (such a n exists. Then the time reversibility of the Poisson process ensures that {At } and {At } have the same distribution (for finitedimensional distributions. Theorem 3. If VaN > 0. We get: Corollary 3.U1 = Rol. say V. see Siegmund [344]. we have RQ„ < 0 so that indeed r(u) < T. this represents a model for storage.Ul < roil  Ul = RQ„ Va1V_1 < RQ2. if nothing else n = N). and a general premium rule p(r) when the reserve is r.1 and its proof is from Asmussen & Schock Petersen [50].32 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Suppose first VT > u (this situation corresponds to the solid path of {Rt} in Fig. one may feel that the interaction between the different areas has been surprisingly limited even up to today.T l . and so on. Then the storage process {Vt} has a proper limit in distribution. we can repeat the argument and get VoN_1 > Ra2 and so on. Thus we may think of {Vt} as having compound Poisson input and being defined for all t < oo. The arrival epochs correspond to rainfalls.1 with Ro = u = u2). with distribution B.3 and being i. say of water in a dam though other interpretations like the amount of goods stored are also possible. Historically. . and in between rainfalls water is released at rate p(r) when Vt (the content) is r. u Notes and references Some main reference on storage processes are Harrison & Resnick [187] and Brockwell. Proof Let T ^ oo in (3. Corollary 3.1 with Ro = u = ul). = r(VT) . Historically. u A basic example is when {Rt} is the risk reserve process corresponding to claims arriving at Poisson rate . Then similarly VVN = r0. The results can be viewed as special cases of Siegmund duality.d. if and only if O(u) < 1 for all u. Some further relevant more general references are Asmussen [21] and Asmussen & Sigman [51].and left continuity is immaterial because the probability of a Poisson arrival at any fixed time t is zero). Hence RQ„ > 0 for all n < N. 3.i.
= Xo + Y1 + • • • + Y.. 0). I. generated by Z1.min n=0.1)). ZN = . . (b) 1/i(u) = P(•r(u) < oo) > 0 as u * oo.2) (for a rigorous proof. as long as the random walk only takes nonnegative values..N From this the result immediately follows. WN = YN ... . there is an analogue of Theorem 3.1. with common distribution F (say)... and is reset to 0 once the r. n=0. W1.e.1) Thus {Wn}n=o. Z2.... Z2. (c) The Lindley process {WN} generated by Zl = Y1.. Theorem 4. WN be the Lindley process generated by Z1 = YN. Z2 . In particular... . W1. is defined by assigning Wo some arbitrary value > 0 and letting Wn+1 = (Wn + Zn+1)+• (4. of (4. Then the events {r(u) < N} and {WN > u} coincide..d.. evolves as a random walk with increments Z1i Z2.. Here F is a general probability distribution on R (the special case of F being concentrated on {1.. . R valued sequence Z1. has a proper limit W in distribution as n + oo.. . Xo = 0. the Lindley process Wo. if Wo = 0 then (Z1+•••+Zn) WN = Zl+•••+ZN.. 1} is often referred to as simple random walk or Bernoulli random walk)..d... just verify that the r . Let further N be fixed and let Wo.....YNn+1) n=0. 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....w... N min (Y1 + • • • + YNn) n=0..2) satisfies the same recursion as in (4..2 The following assertions are equivalent: (a) 0(u) = P(r(u) < oo) < 1 for all u > 0..1 in terms of Lindley processes .2). hits (oo. Z2 = Y2.Yl min (YN .1.. ..4. ... Proof By (4.1 Let r(u) = inf In: u + Y1 + • • • + Yn < 0}. W2.h... For discrete time random walks .. {Wn}n=0. N min (Y1 + + Yn). can be viewed as the reflected version of the random walk with increments Z1..1..s.1.. 0 Corollary 4....i. i.. .N (4. . where the Yi are i .1. Most often.. .. For a given i.. Z2 = YN_1 i .Y1 according to Wo = 0.1.
. doubly u infinite (n'= 0.. the Y1.. .l. . .) sup n=0.=o.d.. In general.1 is equivalent to WN D MN = (Z1 + .s .. ..2.2 and Theorem 4. .. Y1) has the same distribution as (Y1.. a sufficient condition for (e) is that EY is welldefined and > 0..34 CHAPTER II... .. ±1. a Markovian change of measure as in Theorem 2. (d) #.1. N.5 does not necessarily lead to a random walk: if. YN).. Next consider change of measure via likelihood ratios.. . Clearly.. .N so that WN _P4 M = supra=0. 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.. either M = oo a.the result is a sample path relation as is Theorem 3.l. .. 176) but appears to be rather intractable.s. W v m and P(W > u) = P (m > u) = 0(u). Similarly.1 have the same distribution for n = 0. The following result gives the necessary and sufficient condition for {Ln} to define a new random walk: . 0 By the law of large numbers. the Lindley processes in Corollary 4.o... (e).s. e.1 is actually not necessary .. w. (e)Yi+•••+Yn 74 . Combining these facts gives easily the equivalence of (a)(d).i.) and defines Zn = Yn. equivalently. or M < oo a. .g. . For a random walk.1. (Yi + • • • + Yn) > oo a. In that case . YN in Theorem 4. (Z1 + • • • + Zn) = m and P(W > u) = P(M > u) = i (u ).ooa.... SOME GENERAL TOOLS AND RESULTS (d) m = inf. Thus the assertion of Theorem 4..1. assumption on the Z1. The converse follows from general random walk theory since it is standard that lim sup (Y1 + • • + Yn) = oo when Y1 + • • • + Yn 74 oo. there is a more general version of Corollary 4.. then the restrictions of Fx. + Z. One then assumes Yn to be a stationary sequence.3 The i. Proof Since (YN. Px to Fn are equivalent (have the same null sets) so that the likelihood ratio Ln exists. the condition 00 F(YI+•••+ Yn<0)<00 n=1 is known to be necessary and sufficient ((APQ] p.1.g.s. ±2. Remark 4 . ZN or..
h(Yn) (4.3) holds for n = 1.s.4 Let {Ln} be a multiplicative functional of a random walk with E_L.4. the changed increment distribution is F(x ) = E[h(Y). Then the change of measure in Theorem 2. y )..f.4) ({Ln} is the familiar Wald martingale . where h (y) = g(0. The corresponding likelihood ratio is Ln = exp {a (Y1 + • • • + Yn) .Y2) = h(1'i)h(I'a). In particular. 100 ). and define Ln by (4. Y < x].5 Consider a random walk and an a such that c(a) = log F[a] = log Ee° ' is finite.5 corresponds to a new random walk with changed increment distribution F(x) = e'(a) Jr e"'F(dy) ..nrc(a )} (4. this means E(g(x. Y1). we get L2 = L1 (L1 o91 ) = h(Y1)g(X1. (4. of F). ..5 corresponds to a new random walk if and only if Ln = h(Y1) . the random walk property implies Ex f (Y1) = Eo f (Y1 ). Since L1 has the form g (Xo. for some function h with Eh(Y) = 1. Y) f (Y)] for all f and x.. (a) = log F(a] is the c.. We get: Corollary 4. Breiman [78] p. For n = 2. = 1 for all n and x. cf. RANDOM WALKS INDISCRETE OR CONTINUOUS TIME 35 Proposition 4. 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.g.4. Proof If (4. implying g (x.s. Conversely.3) holds... Y) = h(Y ) a. Then the change of measure in Theorem 2. In that case. e.4). Y ) f (Y)] = E[g(O.3) 1Pxa. and so onforn =3. u A particular important example is exponential change of measure (h(y) = e°y'(") where r.g.
the tradition in the area is to use continuous time models. but we omit the details ). v2 = 0 and v = 3B). 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. or imbedded into continuous time processes . the pure jump process is given by its Levy measure v(dx).5) Note that the structure of such a process admits a complete description. {Xt} can be written as the independent sum of a pure drift {pt}. with premium rate p. Xt (n)t(n1) being independent whenever t(O) < t(1 ) < . e f x:IxJ>e} v(dx) < oo (4. The simplest case is 3 = JhvMM < oo.. The traditional formal definition is that {Xt} is Rvalued with the increments Xt(1)_t(o). given by a the increment distribution F(x) = P(Xn+l . In continuous time.1). (4. Roughly. Xt =Xo+pt+oBt+Mt. SOME GENERAL TOOLS AND RESULTS Discrete time random walks have classical applications in queueing theory via the Lindley process representation of the waiting time . the claim surplus process for the compound Poisson risk model . In risk theory. a positive measure on R with the properties e J x2v(dx) < oo.e.Xt)I.).Ft] = Eof (X. this description needs some amendments.Xn < x). {Xt} is a random walk. i. see Chapter V. .. < t(n) and with Xt( i)_t(i_l) having distribution depending only on t(i) . (4. which corresponds to the compound Poisson case: here jumps of {Mt} occur at rate 0 and have distribution B = v/0 (in particular . However.36 CHAPTER II. An equivalent characterisation is {Xt} being Markov with state space R and E [f (Xt+e . say by recording the reserve or claim surplus just before or just after claims (see Chapter V for some fundamental examples). they arise as models for the reserve or claim surplus at a discrete sequence of instants. Xt(2)_t(l).6) More precisely.. the interpretation is that the rate of a jump of size x is v(dx) (if f of Ixlv (dx) = oo. a Brownian component {Bt} (scaled by a variance constant) and a pure jump process {Mt}. we are .7) for all e > 0. corresponds to a process with stationary independent increments and u = p. In discrete time. say the beginning of each month or year .. however.t(i . The appropriate generalization of random walks to continuous time is processes with stationary independent increments (Levy processes). .
8) O<t<T (assuming Wo = Xo = 0 for simplicity).1)v(dx) (4. defined as a system with a single server working at a unit rate. having Poisson arrivals with rate . cf. WT = XT . (ex .6).7 If {Xt} has stationary independent increments as in (4. T) = P(VT > u).min Xt (4.v.6 In the compound Poisson risk model with constant premium rate p(r) .3 and decreases linearly at rate 1 in between jumps.e. 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. O(u.4. VT + V for some r. oo]. Corollary 4.1. where c(a) = ap + a2a2/2 + f 00 provided the Levy measure of the jump part {Mt} satisfies f". and b(u) = P(V > u).10) . i. jxJ v(dx) < oo. Then the storage process {Vt} has constant release rate 1. and the reflected version is then defined by means of the abstract reflection operator as in (4. 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. [The condition for V < oo a.s. V E [0.2). Proposition 4. Here workload refers to the fact that we can interpret Vt as the amount of time the server will have to work until the system is empty provided no new customers arrive. 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. is easily seen to be f3pB < 1.] Processes with a more complicated path structure like Brownian motion or jump processes with unbounded Levy measure are not covered by Section 3. 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). Chapter III. Furthermore. 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. A different interpretation is as the workload or virtual waiting time process in an M/G/1 queue. then Ee'(xtxo) = Eoeaxt = etx(a).Q and distribution B of the service times of the arriving customers.
Xt)g(s. This is of course no coincidence since the distribution of Xl . we use the characterization (4. Q2 = v2.f. 8 Assume that {Xt} has stationary independent increments and that {Lt} is a nonnegative multiplicative functional of the form Lt = g(t.Xt)L8 o 0tIFt] = E [f (Xt+s .5) and get E [f(Xt+B . we show in the compound Poisson case ( IlvIl < oo) in Proposition III.11) (eax . v(dx) = e9xv (dx).Xt)I'Ftl = E [f(Xt+B . Theorem 4 .Xo is necessarily infinitely divisible when {Xt} has stationary independent increments.g.1. . Xt Xo) with E2Lt = 1 for all x.1 that E eaMt = exp fmoo In the general case . Xt +B . By explicit calculation . use the representation as limit of compound Poisson processes. Proof For the first statement .Xt)I Ftl = Eof (X8)g(s.xo)tk ( e). Eea(µt + QBt) = et{aµ +a2OZ/2}. Chung [86]).1 ) v(dx) . X8) = Eof (X8)L8 = Eof (X8)• For the second.4 o) aµ + ((a + 0 ) 2  0 2 )o 2 /2+ r w J 00 (e (a + 9)x  a 9x )v(dx) 00 a(µ + O 2) + a2a2 / 2 + J (eax . then the changed parameters in the representation (4. 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) . e.10) is the LevyKhinchine representation of the c. In particular. let e" (a ) = Eoeaxl. 5 has stationary independent increments as well.g.38 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Proof By standard formulas for the normal distribution.6) are µ = µ + Oo2 .1)eexv(dx). u Note that (4. Then the Markov process given by Theorem 2. (4. t.. if Lt = e9(xt . of an infinitely divisible distribution (see.
Example 4 . .. Thus (since µ = p = 1.St)g(Jt+s)I. U2. and let the Px refer to the claim surplus process of another compound Poisson risk process with Poisson rate. Example 4 . b with b(x) > 0 for all x such that b(x) > 0). 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.(3B(dx). 0.l3 and claim u size distribution B.4).3 and claim size distribution B # B.11 For an example of a likelihood ratio not covered by Theorem 4.1) For shorthand . Recalling that U1. one reason is that in parts of the applied probability literature. Ei instead of P2. <t whenever the RadonNikodym derivative dB/dB exists (e. St)} where {Jt} is a Markov process with state space E (say) and the increments of {St} are governed by {Jt} in the sense that E [f (St+8 . then the martingale {eex(t)tk(e)} is the continuous u time analogue of the Wald martingale (4.5. where . is defined as a bivariate Markov process {Xt} = {(Jt. .8. dB/dB = b/b when B. B have densities b. we write Pi.. b = a = 0) the changed process is the claim surplus process of another compound Poisson risk process with Poisson rate . the corresponding claim sizes . v(dx) _ . MAP stands for the Markovian arrival process discussed below.3 and claim size distribution B. (5. the structure of MAP's is completely understood when E is finite: 2and only there . abbreviated as MAP in this section2. u 5 Markov additive processes A Markov additive processes. a = 0.0.. corresponding to p = 1.o[f (S8)g(J8)]. Then we can write v(dx) _ /3eOxB(dx) = / (dx). it is then easily seen that Lt = H dB(Ui) i:o.10 Let Xt be the claim surplus process of a compound Poisson risk process with Poisson rate .g.2. As for processes with stationary independent increments .9 If X0 = 0. MARKOV ADDITIVE PROCESSES 39 Remark 4.3 =. Ei.(3 = . B(dx) = B[9] B(dx).0 in the following.Ft] = Ejt...3B[B]. are the arrival times and U1.
. {Jt} is specified by its intensity matrix A = (Aij)i. oo). {St} evolves like a process with stationary independent increments and the parameters pi.Sr_1.o(Ji = j. In continuous time (assuming Dpaths).jEE• On an interval [t. let {Jt} be standard Brownian motion on the line. .i.1 For a MAP in discrete time and with E finite. which we omit and refer to Neveu [272] or cinlar [87].) If E is infinite a MAP may be much more complicated. SOME GENERAL TOOLS AND RESULTS In discrete time.f.[a) = (Ei[easl. In addition. with the Y„ being interpreted as interarrival times..it = A. the distribution of which has some distribution Bij.. Proposition 5.. As a generalization of the m. Y2. If all Fij are concentrated on (0.. v. consider the matrix Ft [a] with ijth element least Ei . a MAP is specified by the measurevalued matrix (kernel) F(dx) whose ijth element is the defective probability distribution Fij(dx) = Pi. t+s) where Jt . a jump of {Jt} from i to j # i has probability qij of giving rise to a jump of {St} at the same time. Y1 E dx) where Y„ = S„ .J1=j)= Fij (dx) Pij In simulation language.jEE (here pij = Pi(J1 = j)) and the probability measures Hij(dx)=P(Y1 EdxlJo=i.40 CHAPTER H. vi(dx) in (4. a MAP is the same as a semiMarkov or Markov renewal process.9 EE = (iii&ij[a])i j EE . Jn = j. An alternative description is in terms of the transition matrix P = (piA.6) depending on i. 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. by generating Yn according to Hij when J„_1 = i. 1 J1 ='^])iJEE = (Fij[a])i . Fn[a] = F[a]n where P[a] = P . (That a process with this description is a MAP is obvious.g. As an example. the converse requires a proof. this means that the MAP can be simulated by first simulating the Markov chain {J„} and next the Y1.
kEE Jn = k]Ek[e"Y" .1 )v(dx).qkj + k?^j qkj Bkj [a] } = Ei [east. By PerronFrobenius theory (see A. qij.1) } (recall that qjj = 0). pi. J1 = A which in matrix formulation is the same as Fn+1 [a] = Fn[a]F[a]. Bij (i. which in conjunction with Fo[a] = I implies Ft[a] = etK[a) according to the standard solution formula for systems of linear differential equations.4c). Then. In matrix formulation .(')(a)) diag + (). up to o( h) terms. j E E) and So = 0. we infer that in the discrete time case the . Proof Let {Stt) } be a process with stationary independent increments and pa rameters pi . u Proposition 5. Jt = k] { 1 . vi(dx). Then the matrix Pt[a] with ijth element Ei [east. Jt = k] { xk kEE j la] . 013 . a= . \ diag Ft[a] = Ft[a]K. where K[a] = A+ (r. u In the following.5.ijgij(Bij[a] . Jn+1 = A] = 41 Ei[ e 5„. Jt = j] Ejesh'^ + E Ak j hEi [ease . Jt = j] is given by etK[a].2 Let E be finite and consider a continuous time Markov additive process with parameters A. assume that the Markov chain/process {Jt} is ergodic. vi(dx) (i E E). Sn ) yields Ei[easn+ '. Jt = j] (1 + htc (j) (a)) j + Ak j qk j (Bk +h E Ei [east . 00 r(i) (a) = api + a2ot /2 + f (e° . MARKOV ADDITIVE PROCESSES Proof Conditioning upon (Jn.1)) . this means that F't+h [a] = Ft[a] II+h(rc(i)(a)) +hA+h(Aijgij(Bij[a]1)) I. aSt h = (1 + Ajjh) Ei [east .
The corresponding left and right eigenvectors v("). Jt = j] . u Let k(a) denote the derivative of h() w.t. a.Jt+v = easttK( a)E [ee (st+vst)vK(a)h(a) jt+v I ^tJ = easttt(a)EJt (easesvK(a )h^a)1 = easttK(a)h^a). and appropriate generalizations of the Wald martingale (and the associated change of measure) can be defined in terms of .e=e°tk. and we shall take V(a)h(a) = 1.r. Proof For the first assertion. . Proposition 5. Proof By PerronFrobenius theory (see A.tK(a)h(a) J jj it L o is a martingale. 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.42 CHAPTER II.4. Yrh(a ) = 1. Corollary 5. In particular. The function ic(a) plays in many respects the same role as the cumulant g. Eie"sth^a) = e'Pt[a]h( a) = e.h(a)vva)etw(a). Jeast. h(") are only given up to a constants. just note that [a]h(a) = eietK (a)h(a) = etK(a)h(a). Corollary 5.f.4 Eie"sth(a) = h=a)et?(").etx It then follows that E feast+^(t+v)K(a)h(a) I ^tl l . as will be seen from the following results.5 EiSt = tK'(0) + ki . (5.4c).3 Ei [east. cf. of a random walk.2) where 7r = v(°) is the stationary distribution. We also get an analogue of the Wald martingale for random walks: Proposition 5. cf. Then h(°) = e. Corollary 5.Eikjt = ttc'(0) + ki . and write k = k(°). Furthermore. h(") may be chosen with strictly positive components. its derivatives are 'asymptotic cumulants'. Since v(").7. we are free to impose two normalizations.c(a) (and h(")).
E=ST = tc'(0)E7.. there is typically a function h = h(") on E and a ^c(a) such that Ey a"st t" (") * h(x). Corollary 5.. For the second .a) + ttc (a)2hia ) Multiplying by v=. 43 Ei [Steast h(a) + east k^a)1 = et"(a) (kia) + tic (a)hia)) .4) . .5.4. one obtains a generalization of Wald's identity EST = Er • ES.St]2 = t2/c'(0 ) 2 + 2ttc (0)vk . the distribution of Jo). ] = t2tc (0)2 + 2tK'(0)vk + ttc"(0) + O(1) .g.") }) . Since it is easily seen by an asymptotic independence argument that E„ [Stkjt] u = trc'(0) E„kjt + O(1). 8 Also for E being infinite (possibly uncountable ). t a oo. for a random walk: Corollary 5. u The argument is slightly heuristic (e. (5.Eikjr . (5.3) 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. summing and letting a = 0 yields E„ [St + 2Stkj.2ttc (0)Evkjt + 0(i).3) Let a = 0 and recall that h(°) = e so that 0=°) = h(o) = 1. subtraction yields Vary St = tic"(0) + O(1). [E. t im v^"St = '(0) Proof The first assertion is immediate by dividing by tin Corollary 5. tam E tSt a (0). In the same way. MARKOV ADDITIVE PROCESSES Proof By differentiation in Proposition 5. Remark 5 .5 yields + W (a)k.5. More precisely. Ee"st typically grows asymptotically exponential with a rate ic(a) independent of the initial condition (i. we differentiate (5.e.7 No matter the initial distribution v of Jo. Squaring in Corollary 5.6 For any stopping time T with finite mean.+ k. the existence of exponential moments is assumed ) but can be made rigorous by passing to characteristic functions.
4 that { h(Jt) easttK(a) L o (5. Remark 5. h(Jo) Lo is a Px martingale for each x E E. Then {Lt } is a multiplicative functional. where {Jt} is deterministic period motion on E = [0. some extra conditions are imposed.3b and Remark VI. inconvenient due to the unboundedness of ea8 so we shall not aim for complete rigour but interpret C in a broader sense. In view of this discussion . V. let ha(i.e.. and the family {f LEE given by Theorem 2.1) one then ( at least heuristically) obtains lim Ex eaSv v a) K( v+oo nEx easttK(a)EJt east(vt)K(a) u[J = Ex easttk(a)h(Jt) It then follows as in the proof of Proposition 5.6) We shall not exploit this approach systematically.6. St)} be a MAP and let 0 be such that h(Jt) OStt. (5.10 Let {(Jt. St) } as follows. however.s. see.. 1) (i. G is defined as Gf (x) = lim Exf (Xt) .for the present purposes. xEE . u forsEE). Usually. this is. 0) = h(i )( 1 + ttc(a)). From (5. 0 Proposition 5. Given a function h on E. s) = ea8h(i). For t small . An example beyond the finite case occurs for periodic risk processes in VI.44 CHAPTER IL SOME GENERAL TOOLS AND RESULTS for all x E E.5) is a martingale . We then want to determine h and x(a) such that Ejeasth (Jt) = etK(a)h(i).5. we take the martingale property as our basic condition below (though this is automatic in the finite case). gha(i.9 The condition that (5. First.6. this leads to h(i) + tcha( i. 0) = n(a) h(i).5) is a martingale can be expressed via the infinitesimal generator G of {Xt } = { (Jt.5 defines a new MAP. however. Jt = (s+t) mod 1 P8a. in particular that f is bounded.f (x) tyo t provided the limit exists.(9) {Lt}t>o = .
Qi < oo and Bi a probability measure.5. We omit the details. In the infinite case .1) eft ea' f ij (dx) = Hij (dx) Hij [0] . in the discrete time case.. . 1 + q(b . ^i = of qij Bij [0] 1 + qij ( Bij [0] .1) . (5. 0<q<1.12 The expression for A means h(e) Aij = hie) Aij [1 + gij(Bij[0] i 0 j. In particular. 0<b<oo.10 is given by P = eK(e) Oh e) F[e]Oh('). Bi. then also vi (dx) is compound Poisson with e Ox ^i = /3iBi[0]. one can directly verify that (5.c(0)e = tc(0)e . vi (dx) = f3 Bi(dx) with . if vi(dx) is compound Poisson. That 0 < qij < 1 follows from the inequality qb <1. u Theorem 5. this gives a direct verification that A is an intensity matrix: the offdiagonal elements are nonnegative because Aij > 0. qij = r.tc(0)e = 0 . 0 < qij < 1 and Bij [0] > 0.11 Consider the irreducible case with E finite.1) holds for the P.7) In particular. MARKOV ADDITIVE PROCESSES Proof That { Lt} is a multiplicative functional follows from L8 ogt = h(Jt+s) es(St+ . Here Oh(e) is the diagonal matrix with the h=e) on the diagonal..ic(0)e = ic(0)Oh e) h(e) .7(dx) Bij [0] Bij(dx) in the continuous time case .St)sl(e) h(Jt) 45 The proof that we have a MAP is contained in the proof of Theorem 5. Then the MAP in Proposition 5. Bi [0] Remark 5. Bi(dx) = Bi(dx). That the rows sum to 1 follows from Ae = Oh(e) K[O]h(B) . Ai = µi + 0Q.11 below in the finite case. and by A = Oh(°) K [0]Oh(e ) vi(dx) = e"xvi (dx).(0)j.
.8) h(. v. In matrix notation . v= .t. Letting a = 0 yields the stated expression for A.Bay [0]) That k(') (a + 0) . Jt = A. Further Fib (dx ) = P=(YI E dx.8).tc(0)' )Ah() = Oh(o) K[a + 0]Oh() . in continuous time ( 5. Jl = j] :(Yi E dx. are probability measures .r. This shows that F.tc (') (0) corresponds to the stated parameters µ. Ji = j) h(e) eeyK(B)p h(8) h(e) eexK ( h=e) e)Fi. Yi E dx..tc(') (8)/ d)ag h 7 Aiiii (Bii[a + 0] .13) for matrixexponentials . Here the stated formula for P follows immediately by letting t = 1. H1. Hence the same is true for H=j and H.. First note that the ijth element of Ft[a] is etK(e)Ej [e(a+B)st E:[east Jt = j] = Ej[Lteas' . this implies k[a] = A 1 ) (K[a + 0] . SOME GENERAL TOOLS AND RESULTS Proof of Theorem 5.tc(0)I. Jt = j] = hie) .11. is absolutely continuous w. since Hij.8.46 CHAPTER II. (dx).. a = 0 in (5. (dx) of a process with stationary independent increments follows from Theorem 4. F:j with a density proportional to eei . . this means that Ft[a] = etw ( e)Ohc) Ft[a + 9]oh (e) (5. Jl = j) = Ei[Lt. Now we can write K[a] =A+A ) ( K[a + 0] . Similarly.K [O])Oh(e) (0) l + ( A + (tc(') (a + 0) .e) Consider first the discrete time case .8) yields et'[a] = Ohie )et (K[a +e]K(e)I)Oh(°) By a general formula (A. it follows that indeed the normalizing constant is H1 [0].
1) = Aij4ij(Bij[a] . 6 The ladder height distribution We consider the claim surplus process {St } of a general risk process and the time 7. Though the literature on MAP's is extensive. [261].. 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]. the literature on the continuous time case tends more to deal with special cases. [262] in discrete time. there is. hardly a single comprehensive treatment. . Notes and references The earliest paper on treatment of MAP's in the present spirit we know of is Nagaev [265].7). 0]. 7+ < oo). and is typically defective. Write r+ = T(0) and define the associated ladder height ST+ and ladder height distribution by G+(x) = 11 (S. 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). oo).+ < x.6. however. which. is slightly less general than the present setting. For the Wald identity in Corollary 5. has no mass on (oo. h. < x) = 11 (S. however. Note that G+ is concentrated on (0. see also Fuh & Lai [149] and Moustakides [264]. [226] and Miller [260]. i.)Ajjgij(Bij[a+0] .6.1).Bij[0]) = hjel)ijgijBij[0](Bij[a] . [225]. Conditions for analogues of Corollary 5.(u) = inf {t > 0 : St > u} to ruin in the particular case u = 0 . Much of the pioneering was done in the sixties in papers like Keilson & Wishart [224].e.. an extensive bibliography on aspects of the theory can be found in Asmussen [16]. THE LADDER HEIGHT DISTRIBUTION 47 Finally note that by (5.3 for an infinite E are given by Ney & Nummelin [266].
1 For the compound Poisson model with p = 01LB < 1.ST+(1) and so on. Here bo(x) _ B(x)/µB.B(x) denotes the tail of B.1 The term ladder height is motivated from the shape of the process {Mt} of relative maxima. The main result of this section is Theorem 6. and the maximum M is the total height of the ladder. i.5 below. it follows that for g > 0 measurable. R+ is concentrated on (oo. SOME GENERAL TOOLS AND RESULTS M ST+(2) Sr. the second ladder height (step) is ST+(2) .. we shall first consider the compound Poisson model in the notation of Example 1. Recall that B(x) = 1 . a fact which turns out to be extremely useful.00 ). In simple cases like the compound Poisson model. The first ladder step is precisely ST+.2. see Fig. = ST+(1) Figure 6.1) The interpretation of R+(A ) is as the expected time {St} spends in the set A before T+.(3B(x ) = pbo(x) on (0.i.e. at present we concentrate on the first ladder height. the sum of all the ladder steps (if rl > 0. In other cases like the Markovian environment model.1. In any case.. i. they have a semiMarkov structure (but in complete generality. where basically only stationarity is assumed.1.e. the ladder heights are i. Theorem 6 . 6. Also. by approximation with step functions . 0]. o 00 (6. the second ladder point is ST+(2) where r+(2) is the time of the next relative maximum after r+(1) = r+. For the proof of Theorem 6. G+ is given by the defective density g + (x) =. oo).48 CHAPTER K. has no mass on ( 0. define the prer+occupation measure R+ by R+(A) = E f o "o I(St E A. On Fig. which gives an explicit expression for G+ in a very general setting.2) .T+ > t)dt = E f 0T+I(St E A) dt.d. the dependence structure seems too complicated to be useful). To illustrate the ideas. there are only finitely many).1. 0 f T+ (6. Thus. g(y)R+(dy) = E f g(St)dt. 6.
see Fig. 49 Proof Let T be fixed and define St = ST . ST < ST_t. .2(a): T+ > t Figure 6.O<t<T) = P(STEA. 0 < t < T) P(STEA.T+>T) = P(STEA.ST<St. 6. P(STEA.ST_t.6. THE LADDER HEIGHT DISTRIBUTION Lemma 6 . since the distribution of the Poisson process is invariant under time reversion. That is. has the same distribution as {St}o<t<T. 0]. 0 < t < T.2 R+ is the restriction of Lebesgue measure to (00. {St }o<t<T is constructed from {St}o<t<T by timereversion and hence. St S* t a Figure 6.O<t<T).0<t<T) = F(ST E A.2.2(b): r+ < t Thus.ST<St.St<0.
50 CHAPTER II. G+(A) = Q f 0 B(A .y)R+(dy) 00 Proof A jump of {St} at time t and of size U contributes to the event IS. s. we get G+ (A) = f 00 /3 dt E[B(A . and since the jump rate is /3. That is.T+ > t] dt 0 T+ _ /3E f g( St) dt = 0 f g(y) R+(dy) 0 00 where g(y) = B(A .3 Lemma 6 .3 where the bold lines correspond to minimal values.y) (here we used the fact that the probability of a jump at u t is zero in the second step.St _). this is just the Lebesgue measure of A. The probability of this given { Su}u<t is B(A .St _)I(r+ > t). Figure 6. T+ > t] 0 _ /3 f E[B( A . and (6. 6. cf. SOME GENERAL TOOLS AND RESULTS Integrating w. for A C (0.t dT. oo).3 G+ is the restriction of /3R+*B to (0. U + St_ E A. Fig. E A} precisely when r+ > t. it follows that R+ (A) is the expected time when ST is in A and at a minimum at the same time .. But since St 4 oo a.r. . oo).St)..2) in the last).
1 With r+(y) = I(y < 0) the density of R+. Nt St =>Uk k=1 t where Nt = max{k = O. THE LADDER HEIGHT DISTRIBUTION 51 Proof of Theorem 6. . Uk) (k = 1.. oo) x (0. 6 . Uk) for those k for which ak .e. We call M * stationary if M* o B8 has the same distribution as M* for all s > 0..M o 08 shifted by s is defined the obvious way. assuming basically stationarity in time and space. obviously. 2. 6 . i. . as a point process on [0. The traditional representation of the input sequence {(TT. Fig.1.s. 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. U k)} k=1 a is as a marked point process M *. oo).* ) and the second the mark (the claim size Uk ). {St+8 . h]} /h (by stationarity. we define the arrival rate as E# { k : ak E [0 . this is equivalent to the risk process {St*} being stationary in the sense of (6.z)B(dz) _ f I(x < z)B(dz) _ f (x). The points in the plane (marked by x on Fig. The marked point process . 0 Generalizing the setup. 4 (the points in the plane are (ak .4).3 yields g+ (x) = . cf.Q f r+(x .T+ < oo). Lemma 6. i. we consider the claim surplus process {St }t>o of a risk reserve process in a very general setup.s > 0). this does not depend on h). 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..:T1 +•••+Tk <t}.6.) where ak = Ti + • • • + Tk .4) are (ak...e. the first component representing time (the arrival time o.S8 )t> o = {St }t>o for all s > 0. In the stationary case.
Section 5) which has pure jump structure corresponding to pi = a = 0. vi(dx) = .QiBi(dx)... Uk) k=1. h] Eco(M*) = 1 E f co(M o Bt)dt. The two fundamental formulas connecting M* and M are Eco(M) = aE E.4 Given a stationary marked point process M*. most often one takes h = 1). SOME GENERAL TOOLS AND RESULTS M* U.4 Consider a finite Markov additive process (cf.e. h. k: vk E [0. Oh becomes the approximate probability F(ri < h) of an arrival in [0. h] and the sum approximately ^o(M*)I(ul < h).s. = 0 . o. 0. letting h J. See. Assume {Jt} irreducible so that a stationary distribution 7r = (1i)iGE exists.5) does not depend on h.. where TI = 0. V(M* o eak ). This more or less gives a proof that indeed (6. of (6. ..g. Sigman [348] for these and further aspects of Palm theory. i. we define its Palm version M as a marked point process having the conditional distribution of M* given an arrival at time 0 . the r. As above .2. Example 6 . i 1 U2 Us 1_ 0 or Q2 $ U3 *1 L 0 7 X I 11 1 Figure 6. Note also that (again by stationarity) the Palm distribution also represents the conditional distribution of M* o Ot given an arrival at time t. and let T = T2 denote the first proper interarrival time . We represent M by the sequence (Tk.52 CHAPTER II.5) represents the conditional distribution of M* given vi = 0. e. where T is the first arrival time > 0 of M and h > 0 an arbitrary constant (in the literature.
p.= i. let the arrivals and their marks be generated by {Jt} starting from Jo = j. Then the ladder height distribution G+ is given by the (defective) density g+(x) = . Assume that St * . j) and let the initial mark Ul have distribution Bi when i = j and Bij otherwise.s. qij when {Jt} jumps from i to j and have mark distribution Bij. First choose (Jo_. the probability aij of Jt .. an arrival for M* occurs before time t + dt w.e.6. .O for i # j. It follows that we can describe the Palm version M as follows . Jt = j is iri(3i /.oo a. If Jt_ = i.p. THE LADDER HEIGHT DISTRIBUTION 53 Interpreting jump times as arrival times and jump sizes as marks.O for i = j and iriAijgij/. After that.OF(x). 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 .6iBi + Aijgij Bij j#i iEE iEE 0 Theorem 6 . v.O fo "o F(x)dx = . 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.*(0) with initial reserve u = 0 is p = /3EU0. having the Palm distribution of the claim size and F (x) = F(Uo < x) its distribution ./. aij for (i. Note in particular that the Palm distribution of the mark size (i.6 Under the assumptions of Theorem 6. Before giving the proof.p. let U0 be a r. Jo) w. and by some additional arrivals which occur w.5. This follows by noting that iP*(0) = IIG+JI = J0 "o g+(x)dx = . and that p = 0EU0 < 1. we note: Corollary 6.OEU0. the ruin probability . A stationary marked point process M* is obtained by assigning Jo distribution Tr. we get a marked point process generated by Poisson arrivals at rate /3i and mark distribution Bi when Jt = i.
(left limit) when 0 < it < t and is illustrated on Fig .St*_ u.Q_k and has size U_ k.. the arrival times 0 < 0'1 < Q2 < . { Su}0<u<t is distributed as a process {Su} . Let p(t) be the conditional probability that ST+ E A.0<u<t) = P(StEA.1] here the r .St<Su. Then clearly * G+ (A) = P(ST+ E A) = Consider a process { f p(t)f3dt.Su<0.0<u<tIAt) = P(St EA. oo) x (0 .5). SOME GENERAL TOOLS AND RESULTS V` (0) = E E Uk k: ak E [0. .. that M* and M have doubly infinite time (i. 6.o.$St_ u.o. A standard argument for stationary processes ([78] p. h. CHAPTER H. 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 EU0.5. 0<u<t) = P(St EA.54 By (6. are point processes on (oo . and the kth preceding claim arrives at time t .s.. in (oo. The last property is referred to as insensitivity in the applied probability literature.A.. . which makes an upwards jump at time .0<u<tIAt) = P(St EA.Mt).. oo) p(t) = P(St EA. The sample path relation between { Su } and { Su } amounts to S„ = St . We then represent M by the mark (claim size ) Uo of the arrival at time 0.Su< 0.5... in (0. . oo)). (k = St}t>o 1.o<u<t where a claim arrives at time t and has size Uo. moves down linearly at a unit rate in between jumps and starts from S0 = U. Proof of Theorem 6.0<u<t) = P(St EA. has a very simple interpretation as the average amount of claims received per unit time .g. Now conditionally upon At . T+ = t given the event At that an arrival at t occurs .e.l. the mark at time Qk is denoted by Uk. 105) shows that one can assume w.). 2. It follows that for A C (0. oo ) and the arrival times 0 > 0_1 > a_2 > ..St <.Su_ <0. 0).
2 therefore immediately shows that L(dy) is Lebesgue measure on (oo. G' (A) = 3 f P(St E A. cf. NIt)dt . Thus. and since by assumption St * oo a. Mt)dt = i3EL(A) o"o . t a oo. Uo].5.5 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.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. In Fig. Since So = U0.5 where it = { St < Su. 6. 6. and we let L(dy) be the random measure L(A) = fo°° I(St E A. the support of L has right endpoint U0. the left endpoint of the support is oo. 0 < u < t } is the event that { Su } has a relative minimum at t .. time instants corresponding to such minimal values have been marked with bold lines in the path of { St}.s. Fig.
SOME GENERAL TOOLS AND RESULTS = OE f 0 I(Uo>y)I (yEA)dy = Q f IP (Uo>y)dy A 0o a fA P(y) dy• 0 Notes and references Theorem 6. [147]. .2.56 CHAPTER II. [263] (a special case of the result appear in Proposition VI. A further relevant reference related to Corollary 6.6 is Bjork & Grandell [67].5 is due to Schmidt & coworkers [48].1).
Thus . A common view of the literature is to consider such processes as perturbed compound Poisson risk processes .4 below . {Rt} and the associated claims surplus process {St} are given by Nt Nt Rt = u+t EUi.6) and simulation methods ( Chapter X). exact matrixexponential solutions under the assumption that B is phasetype (see further VIII. with common distribution B. • the premium rate is p = 1. say. . and assume that • { Nt}t>o is a Poisson process with rate j3. i. 3). Some possibilities are numerical Laplace transform inversion via Corollary 3. • the claim sizes U1. St = uRt = EUi t. see Chapter IV. and independent of {Nt}.. being of the form Rt = Rt+Bt + Jt where {Rt } is a compound 57 .e. U2. 4. 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. Panjer's recursion ( Corollary XI. For finite horizon ruin probabilities ..d.Chapter III The compound Poisson model We consider throughout this chapter a risk reserve process {Rt } t>o in the terminology and notation of Chapter I. are i. i=1 i=1 An important omission of the discussion in this chapter is the numerical evaluation of the ruin probability.
For (c). for (d) just note that the kth cumulant of St is tic(k) (0). of the claim surplus St . m. (d) The kth cumulant of St is tf3p(k) for k > 2.t = t(p .+Uk)P(Nt = k) k=O e8t k=O B[s]k .'s etc. 0 . P = PAB = 1/(1 + rl) Proposition 1.t = E E [ U k k=1 k=1 Nt . where K(k) (0) is the kth derivative of is at 0.1). and this immediately yields (a).f. Schmidli [319].g. Write pB^) = EUn' YB = Pali = EU.)3t (fit' k t} = etk(8) exp {st '3t + B[s]f Finally. (c) Ee8St = et" (8) where c(s) = f3(B[s] . Proof It was noted in Chapter I that p . 1 Introduction For later reference. 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). and Schlegel [316]. Furrer [150]. we get Ee8st = 00 e8t c` Ee8 (U1+. We do not spell out in detail such generalizations. and that B(k)[0] = Pak).1 (a) ESt = t(13µ$ .s.Rt. [324]. THE COMPOUND POISSON MODEL Poisson risk process. Dufresne & Gerber [126]. See e.t = fltpB .1). we shall start by giving the basic formulas for moments.g. e . say stable Levy motion.58 CHAPTER III.. (b) Var St = t. {Bt} a Brownian motion and {Rt} a pure jump process.1) .t = E[Ntµs] ..6pBa).1 is the expected claim surplus per unit time.1) = t(p .u . A more formal proof goes as follows: Nt r Nt ESt = E > U k . cumulants .
Sn+0 .h < St < S(n+1)h + h.3EU01 = 1µs where rt is the safety loading. which is often used in the literature for obtaining information about {St} and the ruin probabilities. u + v]. II. St = oo.2 (DRIFT AND OSCILLATION) St/ta3'p1 ast >oo. the Uk .1. meaning that the increments are stationary and independent. . if t = nh + v with 0 < v < h.1 = . we need the following lemma: Lemma 1. For the proof. lim supt. v > 0. then Snh .Tk.3 If nh < t < (n + 1)h. St = oo. Here is one immediate application: Proposition 1. We return to this approach in Chapter V. The right hand inequality in (1.. S„+V > S„ .3) is proved similarly. Indeed.d. then St 4 co. where Tk is the time between the kth and the (k ...1 is the same as if {St} was a random walk indexed by t = 0. however. we have Sok . INTRODUCTION 59 The linear way the index t enters in the formulas in Proposition 1. then lien inft. obviously 0(u) = F(maxk Sok > u). Obviously.Sok_l = Uk .4.i. cf.Tk are i.. then St 00. In this way. (d) If 17 = 0. The connections to random walks are in fact fundamental. (a) No matter the value of 77. The point of view in the present chapter is. and the value is then precisely v. so that {Sok } is a random walk with mean EUET = EU. we get a discrete time random walk imbedded in the claim surplus process {St}. (c) If 77 > 0. rather to view {St} directly as a random walk in continuous time.V. In particular. Proof We first note that for u. 1. 2.1)th claim. then St> SnhV>Snhh. and there are at least two ways to exploit this: Recalling that ok is the time of the kth claim. For example. (b) If 77 < 0.S„ attains its minimal value when there are no arrivals in (u.
. or by a general result on discrete skeletons ([APQ] p. 169) stating that lim infra. 2h. Notes and references All material of the present section is standard. Proof The case of 17 < 0 is immediate since then M = oo by Proposition 1.1(b)) that the assertion holds as t 4 oo through values of the form t = 0. For any fixed h. Corollary 1. However. p.2: Proposition 1.2.. and < 1 for all u when 77 > 0.t . if P(M > 0) = 1. hence by induction i. The general case now follows either by another easy application of Lemma 1.5 The limiting distribution of St .p. and (b). Assuming that each risk generates claims at Poisson intensity /3 and pays premium 1 per unit time. it is seen that upcrossing occurs at least twice. lim supn_. h A similar argument for lim sup proves (a).1. and hence by the strong law of large numbers. There is also a central limit version of Proposition 1.6 Often it is of interest to consider size fluctuations. then {St} upcrosses level 0 a..1. This contradicts u St400.. If rl > 0. h.. THE COMPOUND POISSON MODEL Proof of Proposition 1. {Snh}n=o. Thus using Lemma 1.4 The ruin probability 0(u) is 1 for all u when 77 < 0.1. . and hence it folz lows from standard central limit theory and the expression Var(St) = tf3pB (Proposition 1.60 CHAPTER III.. (c) are immediate consequences of (a). it suffices to prove 4'(0) = F(M > 0) < 1.o. at least once.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).3. Remark 1 . is a discretetime random walk for any h > 0. where the size of the portfolio at time t is M(t). Part (d) follows by a (slightly more intricate) general random walk result ([APQ].s. 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 a4' ESh = h(p .2. Considering the next downcrossing (which occurs w. is a discrete time random walk. 0 Snh = 00.... Snh u = 00 (the lemma is not needed for (d)). u 307). {Snh}n=o.3. 1 since St 4 oo) and repeating the argument.1). this case can be reduced to the compound Poisson model by an easy operational time transformation u T1(t) where T(s) = )3 fo M(t)dt.
6. 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 . Summing over n.just before ruin is again B0. This follows simply by noting that the process repeats itself after reaching a relative maximum. Fig.34 or A.e.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. IV. Note that this . cf. The decomposition of M as a sum of ladder heights now yields: 00 Theorem 2 . the formula for the distribution of M follows . Combined with i/i(u) = P ( M > u). and we further get information about the joint conditional distribution of the surplus and the deficit. Here bo(x) _ Proof The probability that M is attained in precisely n ladder steps and does not exceed x is G+ (x)(1 .1. which we henceforth refer to as the PollaczeckKhinchine formula.1) representing the distribution of M as a geometric compound. we may view the ladder heights as a terminating renewal process and M becomes then the lifetime.1 provides a representation formula for 0(u).IIG +II)EG+ . 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}. that r(0) < oo) is Bo: taking y = 0 shows that the conditional distribution of (minus) the surplus ST(o). d. p < 1. oo ). n=0 (2.P) E PnBon(u) . equivalently. i. [APQ] Ch. B(x)/aB. The expression for g+ was proved in Theorem 11. The following results generalizes the fact that the conditional distribution of the deficit ST(o) just after ruin given that ruin occurs (i. Theorem 2. Note that the distribution B0 with density bo is familiar from renewal theory as the limiting stationary distribution of the overshoot (forwards recurrence time ). 1e. the ladder heights are i. We assume throughout rl > 0 or. 1 The distribution of M is (1. 0 Alternatively. As a vehicle for computing tIi(u).6. (2. It is crucial to note that for the compound Poisson model. but we shall be able to extract substantial information from the formula.. and we shall here exploit the decomposition of the maximum M as sum of ladder heights. Thus .1.2. nevertheless.. 11.
see for example [APQ]. V is uniform on (0. The proof of Theorem 11.62 CHAPTER III.1 is traditionally carried out for the imbedded discrete time random walk. In the risk theory literature.6. the form of G+ is surprisingly insensitive to the form of {St} and holds in a certain general marked point process setup. We assume rt > 0 throughout. Asmussen & Schmidt [49]. (1 .d. Beekman [61]. in this setting there is no decomposition of M as a sum of i. there is a general marked point process version. see Schmidli [323] and references there. ST(o )) given r (0) < oo is the same as the distribution of (VW. (a) 11 (ST(o)_ > x. cf. Theorem 2 . f +b (b) the joint distribution of (ST( o). (d) the marginal distribution of ST(o)_ is B0. 1) and W has distribution Fw given by dFyy/ dB(x) = x/µB. ST(o) > y. the PollaczeckKhinchine formula is often referred to as Beekman 's convolution formula. Theorem A1.V)W) where V.6. For the study of the joint distribution of the surplus ST(u)_ just before ruin and the deficit ST(„).5. 2 The joint distribution of (ST(o )_. [62]. However. and the conditional distribution of ST(o)_ given ST(o)_ = z is Bo z) The proof is given in IV. where it requires slightly more calculation. 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.5. ladder heights so that the results do not appear not too useful for estimating 0(u) for u>0. (c) the marginal distribution of ST(o)_ is Bo . Theorem 2. cf.2(a) is from Dufresne & Gerber [125]. Again.i. ST(o)) is given by the following four equivalent statements: B(z) dz.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. As shown in Theorem 11. 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).just after ruin. . W are independent. Feller [143] or Wolff [384]. 7r(0 ) < oo) = Q 3 Special cases of the PollaczeckKhinchine formula The model and notation is the same as in the preceding sections. cf.
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.p. For a failure at x.1 e ax = n1 (n .e.. Integrating from u to oo. hence without memory. a further relevant reference is Bjork & Grandell [67].1 0(0) = p = Nl2B = 1 1 +71 Proof Just note that (recall that T+ = r(0)) 00 z/^(0) = I' (r+ < oo) = IIG+II = )3 f(x)dx =l3LB• Notes and references The fact that tp(u) only depends on B through µB is often referred to as an insensitivity property. then. I.3. the current ladder step must terminate which occurs at rate S and there must be no further ones which occurs w. Thus .p)pSe a ( l v)x = p( S . however .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. B0 is exponential with rate S and the result can now be proved from the Pollaczeck Khinchine formula by elementary calculations .2 If B is exponential with rate S.6. the formula for P(O) holds in a more general setting. the result follows . Alternatively. Let r ( x) be the failure rate of M at x > 0. 0 . use Laplace transforms. also be seen probabilistically without summing infinite series . The result can. 3b Exponential claims Corollary 3.O)e(b0)x. But claims are exponential . As shown in 11.p.1)1 00 ( 1 . and hence this overshoot has the same distribution as the claims themselves .p) = S . Bon is the Erlang distribution with n phases and thus the density of M at x > 0 is (1 .p) E pn S n x n. 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. 1 . Thus r(x) = S(1 .
and weights 1/2 for each.1) For a heavytailed B. u . E. Then the first term on the r.y)G+(dy) For the last identity in (3.4) can be derived by elementary algebra from (3.+ <u. 2 is one of the main classical early results in the area.3) below.g. (3.S.s.3.y)/3B (y) dy. II. (3. THE COMPOUND POISSON MODEL In VIII. the survival probability Z(u) = 1 .3)). we show that expression for /'(u) which are explicit (up to matrix exponentials) come out in a similar way also when B is phasetype.+ <U. (u) 35eu + 35e6u. Corollary 3.p + G+ * Z(u) = 1 . cf. and conditioning upon S. We mention in particular the following: (a) check that ip (u) = pe (60)u is solution of the renewal equation (3.2) Notes and references Corollary 3.4) zu P(M > u .i(u) satisfies the defective renewal equation Z(u) = 1 .h.4) is similar (equivalently.+ = y yields P(M>u. then 24 1 V.S.64 CHAPTER III. 0 Proof Write o (u) as P(M>u) = P(S.3 The ruin probability Vi(u) satisfies the defective renewal equation ik (u) = 6+ (u) + G+ * 0(u) = Q f B(y) dy + u 0 f u 0(u .3) Equivalently.p + f u Z(u .+ >u.T+ <oo). (b) use stopped martingales .1 p pBo(u). just insert the explicit form of G+. 3c Some classical analytical results Recall the notation G+(u) = f^°° G+(dx). (Example VIII. The case of (3.y)G+(dy ) = f U V(u . u + oo.3.2). A variety of proofs are available . we use the PollaczeckKhinchine formula in Chapter IX to show that b(u) . is ?7+ ( u). (3. (3. if 3 = 3 and B is a mixture of two exponential distributions with rates 3 and 7.3).y)f3 (y) dy.1.T+ <oo)+P(M> u. T+ <00) (3.
206207).p)s .4) can be derived by elementary but tedious manipulations.g. either of these sets of formulas are what many authors call the PollaczeckKhinchine formula. In view of (3. 191)./3B[s] which is the same as (3. eau B(u) du = f PB 3PB SPB 0 o (3. Some relevant references are Abate & Whitt [2]. [APQ] pp.7) s +. e.1 Bo[s] = f oc. In fact.7) and Corollary 3.P)PB 2(1 .s . [APQ] pp.3 .p)2 3(1 .3 is standard .p = (1 .6) 00 = (I .g. (3. Bo of B0 as m e8u B(du) = B[s] .. for example.p) E p"Bo[s]" = 1 . by analytical manipulations (L'Hospital's rule) from (3.P)pB' (3.3. Of course.pBo[s] no (1 .5 can be found in virtually any queueing book.4 The Laplace transform of the ruin probability is 65 fo Hence Ee8M 00 e8uiP(u)du . numerical inversion of the Laplace transform is one of the classical approaches for computing ruin probabilities. 0 Notes and references Corollary 3.s .(3B[s] 1 .. We omit the details (see. which yields the survival probability as 00 f u }t Z(u) = f f3eRtdt 0 from which (3. Griibel & Pitts [132].8) Proof This can be shown. Also (3.Ps s(.7).Ee8M) f ao e8' ( u)du = a8uP (M > u)du = 0 o 1 ( 1+ (1 . The approach there is to condition upon the first claim occuring at time t and having size x . SPECIAL CASES OF POLLACZECKKHINCHINE Corollary 3.3 .PPB2) EM2 = PPB) + QZPBl 2(1 .5 The first two moments of M are 2 EM . Corollary 3./3B[s] .f. see e. it is not surprising that such arguments are more cumbersome since the ladder height representation is not used.)3B[s]) (3. Griibel [179] and Thorin & Wikstad [370] (see also the Bibliographical Notes in [307] p. Embrechts.5).p)s s /3 .5). . g.5) Proof We first find the m.(3 . 111112 or Feller [143].
u)]k d 1 u) _ a) n ( du ( k! (1  . we may assume p = 1 so that the stated formula in terms of the survival probability Z(u) = 1 .1).u/p)]k ko k! Proof By replacing {St} by {Stu/p} if necessary.6 If B is degenerate at p.1). of (3.z/'(u) takes the form Z(u) L^J L.Q (k 1 k= n  [O(k . Z^ =eR(k.1)! k=1 u1 .u)]k1 ku+1) [/3( k .9) shown for n . differentiation yields Z(u) _ /3Z(u) . .u) a)Qea" + (1 .9).3+ 18+ J0 Z(uy).Q) 3e. eO('u) [)3(k .66 CHAPTER III.u) [N(k ./3Z(u .h.y<1)dy 0<u<1 1 < u < oo uu ulhu 1a+/3 J0 uZ(y)dy U Z(y) dy 113+0 For 0 < u < 1.s. then p) 1: ep(k u/.u) [p(k .3I( 0<y<1)dy Z(y)/3I(0<u.u)]k k! k0 The renewal equation (3.)3(1 .1 < u < n and let Z(u ) denote the r.4) for Z( u) means f lhu Z(u) = 1.Q) k=0 k! E e0( = /32(u) .u + 1 )]k = QZ(u) . For n < u < n + 1. Assume (3. 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. THE COMPOUND POISSON MODEL 3d Deterministic claims Corollary 3.u)]k k! (1 L3) 1: e_O(ku) NIN (k (k .9) follows for 0 < u < 1./32(u .
The question then naturally arises whether ie is the c. 4 Change of measure via exponential families If X is a random variable with c. 0 Notes and references Corollary 3. 00 the standard definition of the exponential family {F9} generated by F is FB(dx) = e°xK(e)F(dx). and define rce by (4. co(a) = rc(a + 9) .rc(9) = .1) or equivalently. it follows that Z(u) = 2(u) for n<u<n+1.1) .3B[9]. We could first tentatively consider the claim surplus X = St for a single t. (4. say t = 1: recall from Proposition 1.a. or equivalently BB[a] = B[^+ Repeating for t 54 1.(9) is welldefined.4) works as well.f. and thus (4. (4. K(a) = logEe'X = 109f 00 eaxF(dx) = logF[a]. The answer is yes: inserting in (4. See also Iversen & Staalhagen [208] for a discussion of computational aspects and further references.3B = .4.r. we just have to multiply (4.a. in terms of the c.(9).1 that c(a) = /3(B[a] .2) shows that the solution is Ox [O]0].4.f.2).g. F and c. but will now be repeated for the sake of selfcontainedness. Formalizing this for the purpose of studying the whole process {St}.2) (Here 9 is any such number such that r.3) by t.g. of F9.f. we set up .6 is identical to the formula for the M/D/1 waiting time distribution derived by Erlang [139].Qe(Bo[a] .f. 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) . CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 67 Since Z(n) = 2(n) by the induction hypothesis.1) .d.g. (4.) The adaptation of this construction to stochastic processes with stationary independent increment as {St} has been carried out in 11. B9(dx) = B[9] B(dx).4) .
THE COMPOUND POISSON MODEL Definition 4.i. and thus (4.7) Proof We must prove that if Z is FTmeasurable.10) . and define 09.. . n) for a given n. (4.r. But let Xk = SkT/n .68 CHAPTER III. t < T.i.2. (4. (4. Xn).f. .FTn) = Q(SkTIn : k = 0.(9)} (4. = exp {BST .5) for the density. G].9) Proof We first note that for any fixed t. with T taking the role of n) is the analogue of the expression exp{8(x1 + • • • + xn) .r.0e and claim size distribution Be. the PBT) are mutually equivalent on. G C {T < oo}.1) and multiply from 1 to n). The identity (4.nr. . then EBZ = E [Ze9ST _T"(9)I .4). and PBT) the restriction of PB to FT. replications from Fe (replace x by xi in (4. (4.8) By standard measure theory.5) for the density of n i. with common c. .6) F(G) = Po (G) = EB [exp {BST + Ttc(0)} .Tic (0)} .8) follows by discrete exponential family theory. Z is measurable w. . in particular the expression (4.FT. Let FT = o(St : t < T) denote the o•algebra spanned by the St.. oo) governing a given compound Poisson risk process with arrival intensity. Then P(G) = Fo(G) = EB [exp {BST + TK(O)} . v(Xi. .1 Let P be the probability measure on D[0..d. (4. The following result (Proposition 4.3 and claim size distribution B.2 For any fixed T. BB by (4. Ti(a)/n. Then the Xk are i.d. it suffices to consider the case where Z is measurable w. Then FB denotes the probability measure governing the compound Poisson risk process with arrival intensity.3 Let T be any stopping time and let G E FT. and dP(T) dP^T) That is.1.S(k_1)Tln.t.t. for G E FT. EeeBSt + tk(B) = 1.7) now follows by taking Z = eBST+TK(e)I(G) u Theorem 4 . Proposition 4. G].g. the corresponding expectation operator is E9.
Then G E FT.(Y) = 13(B['Y] . 77 Thus. GT C_ Jr < T}.r)rc(9)}I .1) . subject to the basic assumption ij > 0 of a positive safety loading.g. (a) rc (a) (b) KL(a) 'Y 'Y Figure 5. Now consider a general G.1) _ 1 + a. Ee [exp { BST +Trc(9)} I(G) FT)] = 1. the typical shape of rc is as in Fig.7) holds. according to what has just been proved. Then GT = G n Jr < T} satisfies GT E FT. 5. Given FT.9) holds for G as well.5. Thus by (4. t = T . Thus.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. (0) + rc'(0)a = 0 + ES1 a = a (p .1(a). so that PG = EeE0 [exp { 9ST+Trc(9)}I(G)I FT)] = Ee [exp { BST + rrrc(O)} I(G)EB [ exp {9 (ST . .ST) + (T . The behaviour at zero is given by the first order Taylor expansion c(a) r.9) holds with G replaced by GT. c(a) is a convex function of a.10). LUNDBERG CONJUGATION 69 Now assume first that G C Jr < T} for some deterministic T.r is deterministic.f. (4.FT]] = EB [exp { BST + Trc(9)} I(G)] . 5 Lundberg conjugation Being a c. and hence (4.7 1 Some discussion further supporting this statement is given in the next section.
the Lundberg exponent. 5. and (4.3.g.4) ELS1 = #L(0) cf. (5. Taking T = r(u). we further note that ( 5. An established terminology is to call y the adjustment coefficient but there are various alternatives around.a = i(a + 7). THE COMPOUND POISSON MODEL exists . e.2) 7 Figure 5. we write FL instead of F7.2 s As support for memory.1) .s). Note that KL (a) = /L (BL [a] . Fig.1(b). Thus B[7] = 6/. Thus.1(b).2)) is 7 = 5/3.2 is B(7) = 1 + ^.1) is precisely what is needed for one of the terms in the exponent .70 CHAPTER III. the claim surplus process has positive drift > 0. b[s] = 5/(b . Example 5 .3. 5. 5. (5.QL instead of /37 and so on in the following . an equivalent version illustrated in Fig. Lundberg conjugation corresponds to interchanging the rates of the interarrival times and the claim sizes. G = {T(u) < oo} in Theorem 4. (5.1 Consider the case of exponential claims.4) yields /3L = b and that BL is again exponential with rate bL =. It is then readily seen that the nonzero solution of (5. Fig. Equation (5.1) (or (5.1) is known as the Lundberg equation and plays a fundamental role in risk theory .3. u It is a crucial fact that when governed by FL. .3) cf.
T = T+. V)(u) < e7u. LUNDBERG CONJUGATION 71 to vanish so that Theorem 4.7) 0 and all that is needed to check is that ( 5.P Y j o' xeryxOB (x) dx /3k [Y] .5).G+ L) (x) G+L) (x) IL(+) µ+L) L) where G+L) is the FL.4).(oo) (in the sense of weak convergence w.1.(u)} .6 ). G = {S.Ce7u as u 4 oo. To this end. we therefore have ELe7t(u) + C where C ELe7 (00) = µ+) f e7(1 . e(u) has a limit i. take first 0 = ry.e7x)G+(dx). Letting e(u) = ST(u) . we can rewrite this as 0(u) = e"ELe7^(u).t. (5.8) .G+L)(x)) dx ry^+L) J 00 f 0 (1 .2 (LUNDBERG'S INEQUALITY) For all u > 0.u be the overshoot and noting that PL(T(u) < oo) = 1 by (5.3. see A .1e. which shows that G(L) (dx) = e7xG +(dx) = e7x /3 (x) dx. where C .3 (THE CRAMERLUNDBERG APPROXIMATION) i'(u) .5.1p .ascending ladder height distribution and µ+ its mean.3 takes a particular simple form.7) is the same as (5. 0 Theorem 5 .r. Proof Just note that e(u) > 0 in (5. Since a7' is continuous and bounded.+ E A} in Theorem 4. (5. PL ) with density 1 .6) Proof By renewal theory. ST+ E A] . T(u) < oo] . V) (u) = P(T(u) < oo) = EL [exp {ryS.5) Theorem 5 .1 (5. Then P(ST+ E A) = EL [exp { 7S?+} . (5.
1) (5.1)) and 7µ+L) = 'y/3 [7] 7 1/0 = /3B ['y] .7). of course. . this solves the problem of evaluating (5. (5. we get L where 00 (1 .12) Example 5 . Using (5. u .3 (this was found already in Example 5.11) so that I 7B ['Y](B[7]1) BI [7]Q VP (7) 72 7 (using (5.4 Consider first the exponential case b(x) = Seax. From this it follows.1 .72 CHAPTER III.1 = ^7 The accuracy of Lundberg's inequality in the exponential case thus depends on how close p is to one.10) VW = JI c* e° (x) dx = a (B[a] .1 above) and that C = p. THE COMPOUND POISSON MODEL In principle.8) yields +L) J0 xel'B ( x) dx (5.")G + (dx ) = 1  J0 00 3B(x) dx = 1p. Then 0(u) = pe(a_Q)u where p = /3/S. that 7 = S . Noting that SIG(L)II = 1 because of (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.e. but some tedious (though elementary) calculations remain to bring the expressions on a final form.4). or equivalently of how close the safety loading 77 is to zero.
5. LUNDBERG CONJUGATION Remark 5.5 Noting that PL  1 = ,3LIBL  1 = #ci (0 ) = k (ry) _ ,QB' ['Y]  1 ,
73
we can rewrite the CramerLundberg constant C in the nice symmetrical form G, _'(0)1  1  p K'(7) PL1
(5.13)
In Chapter IV, we shall need the following result which follows by a variant of the calculations in the proof of Theorem 5.3: 1  aB[ry  a]  1 Lemma 5 . 6 For a # ry, ELea^ (°°) = 7 aK'(7) 7  a Proof Replacing 7 by a in (5.7) and using ( 5.8), we obtain 1 (I 1  ^ e('ra) x,3 (x)dx) (L ) ELea^*) = a \\\ f
using integration by parts as in (3.6) in the last step . Inserting (5.12), the result follows. u
Notes and references The results of this section are classical, with Lundberg's inequality being given first in Lundberg [251] and the CramerLundberg approximation in Cramer [91]. Therefore, extensions and generalizations are main topics in the area of ruin probabilities, and in particular numerous such results can be found later in this book; in particular, see Sections IV.4, V.3, VI.3, VI.6.
The mathematical approach we have taken is less standard in risk theory (some of the classical ones can be found in the next subsection). The techniques are basically standard ones from sequential analysis, see for example Wald [376] and Siegmund [346].
5a Alternative proofs
For the sake of completeness, we shall here give some classical proofs, first one of Lundberg's inequality which is slightly longer but maybe also slightly more elementary:
74 CHAPTER III. THE COMPOUND POISSON MODEL
Alternative proof of Lundberg 's inequality Let X the value of {St} just after the first claim , F(x) = P(X < x). Then , since X is the independent difference U  T between an interarrival time T and a claim U, ,3+ry F'[7} = Ee7 ( UT) = Ee7U • Ee7T = B['Y] a = 1' where the last equality follows from c(ry) = 1. Let 0( n) (u) denote the probability of ruin after at most n claims. Conditioning upon the value x of X and considering the cases x > u and x < u separately yields
,0(n +1) (u) = F(u) +
Ju
0 (n) (u  x) F(dx).
We claim that this implies /,(n) (u) < e 7u, which completes the proof since Vi(u) = limniw 1/J(n) (u). Indeed , this is obvious for n = 0 since 00)(u) = 0. Assuming it proved for n, we get
„/, (n+1)(u) <
F(u) + e7u
00
Ju
e7(u=) F(dx)
00
<
f
e7x F(dx)
+ fu
e  7(u z) F(dx)
u
o0
= e 7uE[ 'Y] = e 7u.
Of further proofs of Lundberg's inequality, we mention in particular the martingale approach, see II.1. Next consider the CramerLundberg approximation. Here the most standard proof is via the renewal equation in Corollary 3.3 (however, as will be seen, the calculations needed to identify the constant C are precisely the same as above): Alternative proof of the CramerLundberg's approximation Recall from Corollary
3.3 that
(u) = )3
J OO B(x) dx + J U Vi(u  x)/3 (x) dx.
u 0
Multiplying by e7u and letting Z(u) = e7" O(u), we can rewrite this as
u Z(u) =
z(u) = e7u/
J
B(x)dx, F(dx) = e7x,QB(x)dx,
u
z(u)
f +
J
e7(ux ),Y' 1 • l•(u  x) • e7'/B(x) dx,
0
= z(u) +
J0 u Z(u  x)F(dx),
6. MORE ON THE ADJUSTMENT COEFFICIENT 75
i.e. Z = z+F*Z. Note that by (5.11) and the Lundberg equation, ry is precisely the correct exponent which will ensure that F is a proper distribution (IIFII = 1). It is then a matter of routine to verify the conditions of the key renewal theorem (Proposition A1.1) to conclude that Z (u) has the limit C = f z(x)dx/µF, so that it only remains to check that C reduces to the expression given above. However, µF is immediately seen to be the same as a+ calculated in (5.10), whereas
L
00
z(u) du =
f
J
/3e7udu "o
J "o B(x) dx = J "o B(x)dx J y,0eludu
u 0 0
B(x)^ (e7x  1) dx = ^' (B[7]  1)  As] [0 µs] = l y P^
using the Lundberg equation and the calculations in (5.11). Easy calculus now gives (5.6). u
6 Further topics related to the adjustment coefficient
6a On the existence of y
In order that the adjustment coefficient y exists, it is of course necessary that B is lighttailed in the sense of I.2a, i.e. that b[a] < oo for some a > 0. This excludes heavytailed distributions like the lognormal or Pareto, but may in many other cases not appear all that restrictive, and the following possibilities then occur: 1. B[a] < oo for all a < oo. 2. There exists a* < oo such that b[a] < oo for all a < a* and b[a] = 00 for all a > a*. 3. There exists a* < oo such that fl[a] < oo for all a < a* and b[a] = 00 for all a > a*. In particular , monotone convergence yields b[a] T oo as a T oo in case 1, and B[a] T oo as a f a* in case 2 (in exponential family theory , this is often referred to as the steep case). Thus the existence of y is automatic in cases 1 , 2; standard examples are distributions with finite support or tail satisfying B(x) = o(eax)
76 CHAPTER III. THE COMPOUND POISSON MODEL
for all a in case 1, and phasetype or Gamma distributions in case 2. Case 3 may be felt to be rather atypical, but some nonpathological examples exist, for example the inverse Gaussian distribution (see Example 9.7 below for details). In case 3, y exists provided B[a*] > 1+a*/,3 and not otherwise, that is, dependent on whether 0 is larger or smaller than the threshold value a*/(B[a*]  1). Notes and references Ruin probabilities in case 3 with y nonexistent are studied, e.g., by Borovkov [73] p. 132 and Embrechts & Veraverbeeke [136]. To the present authors mind, this is a somewhat special situation and therefore not treated in this book.
6b Bounds and approximations for 'y
Proposition 6.1 ry <
2(1  aps) 2µs
OMB PB)
Proof From U > 0 it follows that B[a] = Eea' > 1 + µsa + pB2)a2/2. Hence 1 = a(B[7]  1) > Q (YPB +72µs)/2) = 3µs + OYµa2) 2 (6.1) 7 'Y from which the results immediately follows. u
The upper bound in Proposition 6.1 is also an approximation for small safety loadings (heavy traffic, cf. Section 7c): Proposition 6.2 Let B be fixed but assume that 0 = ,3(77) varies with the safety loading such that 0 = 1 Then as 77 .0, µB(1 +rl) 2) Y = Y(77) 277 PB Further, the CramerLundberg constant satisfies C = C(r1)  1. Proof Since O(u) + 1 as r7 , 0, it follows from Lundberg's inequality that y * 0. Hence by Taylor expansion, the inequality in (6.1) is also an approximation so that OAY]  1) N Q (711s + 72µB2) /2) = p + 3,,,(2) B 'y 7 2 2(1  p) _ 271µB
QPB PB)
6. MORE ON THE ADJUSTMENT COEFFICIENT 77
That C 4 1 easily follows from y 4 0 and C = ELe7V°O) (in the limit, b(oo) is distributed as the overshoot corresponding to q = 0 ). For an alternative analytic proof, note that C  1P = rlµB 73B' [7]  1 B' [ry)  1/0 711µB µB +7µB2 )  µB(1 +77 ) 'l = 1. 277q
77
7PBIPB
 77
13 Obviously, the approximation (6.2) is easier to calculate than y itself. However, it needs to be used with caution say in Lundberg's inequality or the CramerLundberg approximation, in particular when u is large.
6c A refinement of Lundberg 's inequality
The following result gives a sharpening of Lundberg 's inequality (because obviously C+ < 1) as well as a supplementary lower bound:
Theorem 6 .3 C_eryu < ,)(u) < C+ eryu where
= B(x) = C_ x>o f °° e7( Yx)B(dy )' C+
B(x) xuo f 0 e'r( vx)B(dy)
Proof Let H(dt, dx ) be the PLdistribution of the time r(u) of ruin and the reserve u  S7(„)_ just before ruin . Given r(u) = t, u  ST (u) = x, a claim occurs at time t and has distribution BL(dy)/BL(x), y > x. Hence ELe7£(u) 0
J
°o
H(dt, dx)
fX
eY(Y x) 00 f°° B(dy) x
BL dy
BL(x)
o
f
f H(dt, dx)
L ^ H(dt, dx) f e7B( x)B(dy) Jo oc, < C+
J0 0 o" H(dt, dx) = C. o" J
The upper bound then follows from ik(u) = e7uELeVu), and the proof of the u lower bound is similar.
78 CHAPTER III. THE COMPOUND POISSON MODEL
Example 6.4 If B(x) = eax, then an explicit calculation shows easily that B(x) _ e6X fz ° e7(Yx)B(dy) f x' e(6,6)(Yx)8esydy = 5 = P. Hence C_ = C+ = p so that the bounds in Theorem 6.3 collapse and yield the exact expression pey" for O(u). u The following concluding example illustrates a variety of the topics discussed above (though from a general point of view the calculations are deceivingly simple: typically, 7 and other quantities will have to be calculated numerically). Example 6.5 Assume as for (3.1) that /3 = 3 and b(x) = 2 .3e3x + 2 .7e7x, and recall that the ruin probability is 24 5su 5eu + 3e *(u) = 3 Since the dominant term is 24/35 • e", it follows immediately that 7 = 1 and C = 24/35 = 0.686 (also, bounding aS" by a" confirms Lundberg's inequality). For a direct verification, note that the Lundberg equation is
7 = /3(B['Y]1)
= 3\
2.337
+2.7771
which after some elementary algebra leads to the cubic equation 273  1472 + 127 = 0 with roots 0, 1, 6. Thus indeed 7 = 1 (6 is not in the domain of convergence of B[7] and therefore excluded). Further, 1P = B [7] 181B = 13 2.3+2.71 = 1 3 1 7 I 7'
_ 17
2 (3 a )2 + 2 (7  a)2 «=7=1 2 1p _ 7 _ 24
36 '
3.171 35* 36 For Theorem 6.3, note that the function QB[Y]1 f°°{L 3e_3x+
u
• 7e7x 1 dx
J
3 + 3e4u
f 0c, ex .
I 2 . 3e3x + 2 . 7e7x l dx
l J
9/2 + 7/2e4u
7. VARIOUS APPROXIMATIONS FOR THE RUIN PROBABILITY 79
attains its minimum C_ = 2/3 = 0.667 for u = oo and its maximum C+ = 3/4 = 0.750 for u = 0, so that 0.667 < C < 0.750 in accordance with C = 0.686.
Notes and references Theorem 6.3 is from Taylor [360]. Closely related results are given in a queueing setting in Kingman [231], Ross [308] and Rossberg & Siegel [309]. Some further references on variants and extensions of Lundberg's inequality are Kaas & Govaaerts [217], Willmot [382], Dickson [114] and Kalashnikov [218], [220], all of which also go into aspects of the heavytailed case.
7 Various approximations for the ruin probabil
ity
7a The BeekmanBowers approximation
The idea is to write i (u) as F(M > u), fit a gamma distribution with parameters A, 6 to the distribution of M by matching the two first moments and use the approximation
0(u)
f
u
Sa
r(A)
xa  leax dx.
According to Corollary 3.5, this means that A, 8 are given by A/S = a1, 2A/52 = a2 (2) PIB3) ^ZP(B)2 __ PPB a2 al 2(1  P)PB 3(1  P)µ8 + 2(1  p)2' i.e. S = 2a1 /a2, A = 2a2 1/a2.
Notes and references The approximation was introduced by Beekman [60], with the present version suggested by Bowers in the discussion of [60].
7b De Vylder's approximation
Given a risk process with parameters ,(3, B, p = 1, the idea is to approximate the ruin probability with the one for a different process with exponential claims, say with rate parameter S, arrival intensity a and premium rate p. In order to make the processes look so much as possible alike, we make the first three cumulants match, which according to Proposition 1.1 means p=AUB1=P1,
2N
(2) 6^= =OP
,
/3,4)
.
p = (/3max 0)µB.PBo [s (/3max .1 As /3 f Nmax.3 )1 } _ 1p 1 . heavy traffic conditions mean that the safety loading q is positive but small. Mathematically.(3)2 P PB 2µB 2µB Letting /3* = /3/P. [174]) shows that it may produce surprisingly good results.g.(bA*)".p .s(/3max .b(u) = p*e. the premiums exceed only slightly the expected claims.Ps(/3max . p* _ . 7c The heavy traffic approximation The term heavy traffic comes from queueing theory. Proposition 7.8µBo  Ss' u where 6 = µB/µBo = 2µa/µB 2) . we have according to the PollaczeckKhinchine formula in the form (3./3)PBo PB . 1924.3* /S. That is. Proposition 1.)3 )PBo µB  .P ./3)] 1 . but has an obvious interpretation also in risk theory: on the average./3)M converges in distribution to the 2a exponential distribution with rate S = B' Proof Note first that 1 . (/3max . the approximating risk process has ruin probability z.1. numerical evidence (e.2) was suggested by De Vylder [109]. and hence the ruin probability approximation is b(u) e(bAln)u.80 CHAPTER III. Notes and references The approximation (7. we shall represent this situation with a limit where /3 T fl but B is fixed.3 and Corollary 3. or equivalently that /3 is only slightly smaller than /3max = 1/µ8. Grandell [171] pp.1. THE COMPOUND POISSON MODEL These three equations have solutions 9/3µB2)3 30µa2)2 3µa2) (3) P+ (3) ' 0 .7) that Ee$(Amex /j)M _ 1p _ 1p Eo [s (0max 1 .p + p { 1 1p ti 1 . Letting Bo be the stationary excess life distribution. Though of course it is based upon purely empirical grounds. cf.
1 1 . VIII). The present situation of Poisson arrivals is somewhat more elementary to deal with than the renewal case (see e . or equivalently that 0 is small compared to µB . Numerical evidence shows that the fit of (7./3)u * v. the first results of heavy traffic type seem to be due to Hadwiger [184].ze a2unµB laB (7. VARIOUS APPROXIMATIONS FOR THE RUIN PROBABILITY 81 Corollary 7. We return to heavy traffic from a different point of view (diffusion approximations) in Chapter IV and give further references there .2 If . light traffic conditions mean that the safety loading rl is positive and large . It is worth noting that this is essentially the same as the approximation (2) z/i(u) Ce. light traffic is of some interest as a complement to heavy traffic . Of course.7. These results suggest the approximation Vi(u) e6(0_. Mathematically. [APQ] Ch. but has an obvious interpretation also in risk theory: on the average .2. the term light traffic comes from queueing theory. obviously Corollary 7.B AB ) 6()3max _'3) = However . Notes and references Heavy traffic limit theory for queues goes back to Kingman [230]. in risk theory heavy traffic is most often argued to be the typical case rather than light traffic . That is ./3)u).g. 2 provides the better mathematical foundation. as well as it is needed for the interpolation approximation to be studied in the next subsection. However . then P(u) 4 e6„ Proof Write z'(u) as P((/3max . u * oo in such a way that (3max .l3)M > (/3max .ryu .Q T /3max. .3) is reasonable for g being say 1020% and u being small or moderate.p _ 2rl11B PB p. the premiums are much larger than the expected claims . In the setting of risk theory. This follows since rl = 1/p . and hence 2µ2B 1 . we shall represent this situation with a limit where 3 10 but B is fixed.4) suggested by the Cramer Lundberg approximation and Proposition 6.0)u. while the approximation may be far off for large u. 7d The light traffic approximation As for heavy traffic .p.
z/' (u) convergence P(U . Light traffic does not appear to have been studied in risk theory. i. 0(u) /3 J B(x)dx = /3E[U .T > u) = J o" B(x + u)/3eax dx .= 1 aJ 1 a 0+ 1 = = p.u. Indeed. [97]. cf. U > u] = /3iE(U .e. Omax max m. (7. see Daley & Rolski [96]. THE COMPOUND POISSON MODEL Proposition 7. ( 3 J O B dx. 7e Interpolating between light and heavy traffic We shall now outline an idea of how the heavy and light traffic approximations can be combined. En'=2 • • • = O(/32) so that only the first terms matters. 0 u Notes and references Light traffic limit theory for queues was initiated by Bloomfield & Cox [69].5) u Proof According to the PollaczeckKhinchine formula. Again. The crude idea of interpolating between light and heavy traffic leads to 0 (u) C1 .(3 10. .u)+. Another way to understand that the present analysis is much simpler than in these references is the fact that in the queueing setting light traffic theory is much easier for virtual waiting times (the probability of the conditioning event {M > 0} is explicit) than for actual waiting times . ao n=1 00 n=1 (u) P) anllBBon(U) onPaBon(u) • Asymptotically. by monotone time T of the first claim . the Poisson case is much easier than the renewal case.Q limIP ( u) + Q lim z/'(u) Amax &0 amax ATAm.82 CHAPTER III.3 As .5) follow by integration by parts. Asmussen [19] and references there. Sigman [347]. u Note that heuristically the light traffic approximation in Proposition 7. 10 ( u The alternative expressions in (7.3 is the same which comes out by saying that basically ruin can only occur at the F(U .T > u). and hence 00 (U) /3pBBo (u) = 0 / B(x)dx. For a more comprehensive treatment.
3n. available. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS 83 which is clearly useless . ^IE) exist: 1 (B) HT QmsxQ hm J e e6" 2µE/µE2)'" = e(1 6)" =  Q1Qm.8. Instead. however. ^ LT Q maxQ m"^ Qlo V LT) ( CHT(v) (say).6) is . we may ask which one carries the larger risk in the sense of larger values of the ruin probability V(') (u) for a fixed value of 0. (7. z/i(E) (u) = pe(QmaxQ)u. that is.O0 M. no empirical study of the fit of (7. (U).3).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.3 and use similar notation for %(B) (u) = (u).x . Substituting v = u(. Another main queueing paper is Whitt [380]. The adaptation to risk theory is new. Al . . [84]. B(2).O(E)(u) 1 (1 . ) M. even if the safety loading is not very small. the idea of interpolating between light and heavy traffic is due to Burman & Smith [83 ]. 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. with rate 1/µB = /3max. "/Qmex Cu) CLT(u ( /3max 0) + O16 CHT( U(Qmaz . _(E) (u). Let OLT) (u) denote the light traffic approximation given by Proposition 7. 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). to get nondegenerate limits . . one may hope that some correction of the heavy traffic approximation has been obtained. Notes and references In the queueing setting . Thus ./3)) . where further references can be found . we see that the following limits HT) (u').Wmax f(x ) dx + pee6mQ. 8 Comparing the risks of different claim size distributions Given two claim size distributions B(1).VHT) ( ax QmQ ) h (B) ( .
then Bill = B(2). this ordering measures difference in variability. U(2) such that U(l) has distribution B('). Here convex ordering is useful: Proposition 8. Recall that B(') is said to be stochastically smaller than B(2) (in symbols. B(2)) if f fdB(1) < f fdB(2) for any convex function f. Proposition 8. cf. we can assume that 1) < St 2l for all t. the proof is complete. Finally.1 If B(') <d B(2).6.s. B(' <.' 1)(u) < V)(2) (U) for all u. Proof According to the above characterization of stochastical ordering. then . Rather than measuring difference in size.84 CHAPTER III. Proposition 8. or the existence of random variables U(l). B(2) and PB(1) = µB(2). then i.2 If B(') <j. whereas (consider x2) B(2) has the larger variance. u Of course. B(') <d B(2)) if B(1)(x) < B(2)(x) for all x.ill(u) < V)(2) (U) for all u. U(2) distribution B(2) and U(1) < U(2) a. we shall need various ordering properties of distributions. one can interpret f x°° B(y) dy as the net stoploss premium in a stoploss or excessofloss reinsurance arrangement with retention limit x. Taking probabilities. . In the literature on risk theory. In terms of the time to ruin. an equivalent characterization is f f dB(') < f f dB (2) for any nondecreasing convex function f. 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). most often the term stoploss ordering is used instead of increasing convex ordering because for a given distribution B. equivalent characterizations are f f dB(') < f f dB (2) for any nondecreasing function f.1 is quite weak. In particular (consider the convex functions x and x) the definition implies that B(1) and B(2) must have the same mean. this implies St T(l)(u) > r(2)(u) for all u so that 17(I) (U) < oo} C_ {T(2)(u) < oo}. XI. THE COMPOUND POISSON MODEL To this end. Bill is said to be convexly smaller than B(2) (in symbols. B(2)) in the increasing convex order if f BM (y) dy < f 00 Bi2i (y) dy x x for all x. A weaker concept is increasing convex ordering: B(1) is said to be smaller than B(2) (in symbols. B(') <i. we have the convex ordering. for more detail and background on which we refer to Stoyan [352] or Shaked & Shantikumar [337].
it is seen that asymptotically in heavy traffic larger claim size variance leads to larger ruin probabilities.1 and µB at 1 so that the safety loading 11 is 10%. A general picture that emerges from these results and numerical studies like in Example 8. we have Bol) (x) f ' B(1) (y) dy < ' f' B(2) (y) dy = Bo2) (x)• µ 85 I. B(2).5 If '0(1)(u) < p(2) (U) for all u and a. Proof Consider the light traffic approximation in Proposition 7. and consider the following claim size distributions: B1: the standard exponential distribution with density ay. The problem is to specify what 'variation' means. Then V. This u implies that D <. Proof If f is convex. B(2). Hence by the PollaczeckKhinchine formula . . from which the result immediately follows. with fixed mean. A first attempt would of course be to identify 'variation' with variance. then /'(')(u) < 0(2)(u) for all u.e. say to p. 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.u) = (1 _ P) E /3npnBo( 1):n(u) n=1 00 < (1.p ) E /3"µ"Bo2)* n(u) _ V(2) (u) n=1 = Corollary 8.1. larger variance is paramount to larger second moment. B.2 is the following: Proposition 8. then B(1) <. The heavy traffic approximation (7. Example 8.. Corollary 8. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS Proof Since the means are equal.4) certainly supports this view: noting that. A partial converse to Proposition 8.8.3 provides another instance of this.6 Fix /3 at 1/1. (D) (u) < O(B) (U ) for all u. we have by Jensen 's inequality that E f (U) > f ( EU). and here is one more result of the same flavor: Corollary 8.6 below is that (in a rough formulation) increased variation in B increases the risk (assuming that we fix the mean).(1) (..4 Let D refer to the distribution degenerate at 'LB . Bo1) <_d Bo2) which implies the same order relation for all convolution powers.3 If B(1) <.
4142. We return to ordering of ruin probabilities in a special problem in VI.01%. i. B3 the comparison is as expected from the intutition concerning the variability of these distributions. In terms of variances o2. 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). with the hyperexponential distribution being more variable than the exponential distribution and the Erlang distribution less.0' U0. we have 0r3 = 2 < or2 = 1 < 02 = 10 < 04 = 00 so that in this sense B4 is the most variable. = 0. B4: the Pareto distribution with density 3/(1 + 2x)5/2.e. 9 Sensitivity estimates In a broad setting.9A2e'2r where A. 1%. B.001 u0.e'\1x + 0. and consider a = 5%.) = a. the behaviour of which is governed by a parameter 9. One then obtains the following table: U005 U0. Let ua denote the a fractile of the ruin function. 11 Notes and references Further relevant references are Goovaerts et al.4.86 CHAPTER III. THE COMPOUND POISSON MODEL B2: the hyperexponential distribution with density 0.01%. B3: the Erlang distribution with density 4xe2x. van Heerwarden [189].1%. A2 = 3.000. all distributions have mean 1..lA. Note to make the figures comparable. 1/)(u. Kluppelberg [234]. and this is presumably a consequence of a heavier tail rather than larger variance. 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. For B1i B2.1358. However. 0. Pellerey [287] and (for the convex ordering) Makowski [ 252]. 0. [166]. which appears to be smaller than the range of interest in insurance risk (certainly not in queueing applications!). A standard example from queueing theory is . in comparison to B2 the effect on the ua does not show before a = 0.
and hence the effect of changing p from 1 to 1 + Ap corresponds to changing /3 to /3/(1 + Op) /3(1 . Example 9. a0 as ao 80 19P .2 Consider a risk process { Rt} with a general premium rate p.Ap). Thus at p = 1. Then if t is large . obtained say in the natural way as the empirical arrival rate Nt/t in [0. and hence a _ e(60)u + u e(60)u = ( i + which is of the order of magnitude uV. increasing in u. a2/t). Similar conclusions will be found below. while /3 = j3 is an estimate.9.3.1. For example. SENSITIVITY ESTIMATES 87 a queueing network. if = a e(6A)u. or we may be interested in aV)/0/3 as a measure of the uncertainty on '0 if 0 is only approximatively known.19P a/ . we may be interested in a'/ap for assesing the effects of a small change in the premium. it follows that ' is approximatively normal N(0.01/2u.. Thus. with 0 the vector of service rates at different nodes and routing probabilities. i. Then a p ao = 00 Qa/. say estimated from data.1 Consider the case of claims which are exponential with rate 8 (the premium rate is one). 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. In the present setting. the premium rate p and the claim size distribution B. the distribution of %3 0 is approximatively normal N(0„ Q/t). where the partial derivatives are evaluated at p = 1. the standard deviation on the normalized estimate ^/1' (the relative error ) is approximatively .(u) for large u. a/3 0 . Proof This is an easy time transformation argument in a similar way as in Proposition 1. where Q2 = fl ( l2 1113 / _ Ou2v)2. In particular . Then ib = Pe(613)u. u Proposition 9. and s(9) the expected sojourn time of a customer in the network.e. Then the arrival rate /3(P) for { R(P) } is )31p. t]. Assume for example that 8 is known. Let R(P) = Rtli.
we cannot expect in general to find explicit expressions like in Example 9.3. but we shall concentrate on a special structure covering a number of important cases.()^ 1 . and the proofs of (9. () Proof According to (9. x > 0 (9.^)] 1(/3+y)we(9+'y.4). e.Owe (9. Consider first the adjustment coefficient y as function of 3. THE COMPOUND POISSON MODEL As a consequence.w(6. Similar notation for partial derivatived are used below. but must look for approximations for the sensitivities 0.g. 5) (Q+'Y)[we(0+7. we can rewrite the Lundberg equation as w(9+ y.1 or Proposition 9. In the case of the claim size distribution B. namely that of a twoparameter exponential family of the form Bo.3 or/and B. (9.r. (9. Consider first the case of 8/8/3: . Differentiating w.10) below. mathematically a proof is needed basically to show that two limits (u * oo and the differentiation as limit of finite differences) are interchangeable. /3 yields w e(e + Y. The most intuitive approach is to rely on the accuracy of the CramerLundberg approximation . this intuition is indeed correct. 4) (9 . Proposition 9.6 below for some discussion of this assumption). ^) . various parametric families of claim size distributions could be considered. (.3.3 70 = 'Ye = = 7 /3(1 we(e +'y. 9.6) As will be seen below.5) are similar.()YC = 1 +y/ /3 \ Q2 From this (9. so that heuristically we obtain '00 50ryu = Coe"u .2) (see Remark 9.3) follows by straightforward algebra.uypCe7u urypO.()wC(e. 3) ( 9 .((dx ) = exp {Ox + (t(x) . ()} p(dx) . Of course.0 = t/'(u) and the CramerLundberg constant C.w(O.88 CHAPTER III. However . it suffices to fix the premium at p = 1 and consider only the effects of changing . Viei '0(. and write yp = 8y/8/3 and so on .()(0 +'0) ' (9 .t. u Now consider the ruin probability 0 = 0 (u) itself.(/3 + y)we(9 + 7. () = log(1 + y//3). (3+'y)PC (0+7. for the ruin probabilities .
w(O. () exp {w(9 + a.St (U) Ee.10) (9. But from the proof of Theorem 5.11) Ee. BarndorffNielsen [58]).g.QB(x) dx.2 of the Appendix ). we multiply by e7" and let Z(u) = elt" cp(u). u 0 Then Z = z + F * Z and F is a proper probability distribution .x)B(x)dx. Z= zl + z2 where zl (u) = e7u J m B(x)dx.(e"U = = wS(O.x).w(9.9. Hence by a variant of the key renewal theorem (Proposition A1. 11 For the following. we note the formulas Ee. ()] exp {w (O + y. Combining these estimates .4 As u oo.8) (Section 5). w((9 + a. O} (9. (9.x)B(x) dx + J U W(u .([a] = exp {w(9 + a. ()} . SENSITIVITY ESTIMATES Proposition 9. and alsoo zl(u) + 0 because of B['y ] < oo.w(9.3(x) dx.p)/C'y. () .8).x).8) Letting cp = e0/e/3 and differentiating (9. By dominated convergence. Z(u)/u a C//3PF where PF is the mean of F. the proof is complete.3) for z/'(u). () . PF = (1 . 0(u) = /3 Ju"O B(x) dx + f 0 0(u . () . u 0 Proceeding in a similar way as in the proof of the CramerLundberg approximation based upon (9. it holds that 89 a ue ryu a/3 Q(1 P) 7C2 Proof We shall use the renewal equation (3.12)).9) (9.C). we get p(u) = J "O B(x) dx + J U O(u .x) F(dx ) f u J C F(dx) = C as u 4 oo. z2(U) = e7" J u b(u .4t (U)e°`U = which are wellknown and easy to show (see e. ()} . Be.3 (see in particular (5. z2(u) _ 1 ^ e'ri`i7i( u .we(9 . Further write de = [we (9 +'y. () . F(dx) = e'yy/3B(x)dx.
z = zl + z2.8) that cp(u) . 01 (i+) do = +'Y. ()]B(dy) dx x 0 0C T ON O .w (9.wc(O. 8^ ue7u.x). this implies Z = z + F * Z. )}B(dy)• Letting cp it thus follows from (9.w(e. 0 x Multiplying by e7" and letting Z(u) = e"uV(u). ()](e7v . ^)} [wc (0 + 7. ()} 1z(dy) = f [t(y) . By dominated convergence and (9.6e7u f "o f[t(y) .x)f3 f ^[t(y) .12) f exp {O y + (t(y) .e7x/3 f 00 [t(y) . 2 z 07P N ue7u (3C de . ()]B(dy) dx.w(0. C)] (1 + 7 ) Proposition 9.wc (O.lB(x) dx = e7uzl(u) + e7°zz(u) + V(u T where zl (u) = .2) holds.w( (0.w( (0. THE COMPOUND POISSON MODEL [we(e+7. Then as u > oo. C) . oo z2 (u) f C . u Z2(U) = e7° f u ^/i(u .11). F(dx) = e7x. ^) . ())B(dy) dx.5 Assume that (9.QB(x)dx.wc(9.1) B(dy) 'f '[t(y) .6C do 89 1p 8( 1p Proof By straightforward differentiation.9)(9. ()]e7vB(dy) 'fCd 7 c .90 CHAPTER III. 8 8() 8( (9.we (0.
It follows after some elementary calculus that p = a)3/5 and.15) (9.Y)a+1 ' (9. that C = a.3ary tog('Finally.1 . Z(u) /3C 91 o c'o e11(t (y) . ./35' a/i'y + aryl 625ry. < = a.a log S = log r(c) .) log(9) = %F(a) logs where %1 = F'/]F is the Digamma function.QS 1 . Example 9. We get w( (0.pa+1 .16) (9. a /(S .w((9. () ='I'(t.ry) 5a1 cry (5 .6 Consider the gamma density b (x ) = Sa xa.17) (9. ())B(dy) < oo.9. () = C/9 = a/S. by inserting in the above formulas.1edz = 1 exp {Sx + a log x .rye) S 5rya. Here (9.Sry a/32 + a/37 + /37 ..12) follows.14) de = d( 7!3 76 = 7e = log ( \ ( \5a_ / \SSry ) 72 . t(x) = logx. 9 = S. we (9.. ( 9. U 7µF from which the second assertion of (9.yu/3C2do u86 89 1p' az/) = 8z/. () = log r(a) .12) takes the form y) alp a . SENSITIVITY ESTIMATES as u 3 oo.13) (9. ue_Yu 'C2d( 8a 8( 1 p . and also zj (u) 4 0 because of f Hence.a/35a&y' ' (9.(log r(a) a log S)} • r(a) 1..C log(9).18) (05 + 57 _'3_y . and the proof of the first one u is similar. w(e.2) holds with p(dx) = xldx.
w(e.log c = 2 In particular.w(9. Straightforward but tedious calculations .CZ try)} 1 C C2 try ."62 .9) (() . () = Cc . C = . further yield . THE COMPOUND POISSON MODEL Example 9.92 CHAPTER III.([Y] .21og 2. ()} = exp {c (C . which we omit in part . C) . t(x) _ .1 = eXP {c(C .2) with µ(dx) = 2x3zrdx.22.2 log (0.2 .1 16 +ry c C22ry 2( = + 70 We (e.3. Be.7 Consider the inverse Gaussian density ( b(x) Zx37 exp This has the form (9.2a) } Thus the condition B[a*] > 1 + a* /.l3 of Section 6a needed for the existence of ry becomes e^Q > 1+62 / 2. C) = B = Yc = de = do = . 9 = .S[a] = exp {w (9 + a. for a < a* = z (.3Ee.
1) .. the exponent of the density in an exponential family has the form 01 tl (x) + • • • + 9ktk (x). Thus. then ryT < 0. and we estimate y by means of the empirical solution ryT to the Lundberg equation.a. or Ct(x). Notes and references The general area of sensitivity analysis (gradient estimation) is currently receiving considerable interest in queueing theory. thus.cue_7u)3C P Remark 9. by the LLN both F (NT = 0) and F (PT > 1) converge to 0 as T . (9.7 and references there.. B are assumed to be completely unknown. BT [a]= NT ^` e"U. let NT 16T = ^T . In general. kT (a) = /T (BT [a] .10. To this end.2) is motivated as follows.3C2de 1p' z a = c . the exponent is either Ox. and hence explicit or asymptotic estimates are in general not possible.8 The specific form of (9. That it is no restriction to assume k < 2 follows since if k > 2. [379] consider a special problem related to reinsurance. in u which case the extension just described applies. to our knowledge. Note that if NT = 0. queueing networks) are typically much more complicated than the one considered here. if 1 PT = /3TNT(U1+. Comparatively less work seems to have been done in risk theory. then BT and hence ryT is undefined. Thus. in which case we can just let t(x) = 0. . Finally if k = 1. ESTIMATION OF THE ADJUSTMENT COEFFICIENT Finally. the main tool is simulation. we can just fix k . However.2 of the parameters. the results presented here are new. Also. 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).12) takes the form a = a 93 ar. ae t 1lEY u S _ . sj=1 and let YT be defined by IKT('ryT) = 0. for which we refer to X. the models there (e. 10 Estimation of the adjustment coefficient We consider a nonparametric setup where /3.oo. we have assumed k = 2 and ti (x) = x.g. Van Wouve et al. However ..+UNT) > 1.
For the proof. vfoVFB[2y].2) rT(7) N N (0./^ B[27] .v. 16T where V1.3T .3/ T).1)2 + E[27] .B[7]) + B [7] .2) follows from NT/T a4' .B[7]2 V2 . Lemma 10 . then (10.B[7]) 0+ Iv/o(b[y].T y .B[7]2 }) ( T 0 .. B[2'Y]  /3T ) .1 As T 4 oo.1) .'s.3) Proof Since Var(eryU) = we have B[7].7 + (. .2 As T * oo. (10.If .(27)/K'(7)2.: N 0. Hence KT(7) = (F' + (OT a(B[7l 0))((BT [7] .: N ()3.b[Yp'V21 T { (E[7] . N ( n[7].i3)(B[7] 1) + (3(BT[7]  . it is easy to see that we can write \ V1 1 l _ .a BT[7] I B[7] I + .)vl+ N CO. 1) r. since NT /T . we need a lemma. V2 are independent N (0.94 CHAPTER III.1) 'YT .Q and Anscombe 's theorem. 7T a4' 7. a2 where a2 = /3r. If furthermore B[27] < oo.B[7]2 n Hence ( 10.'Y . More generally. B[27] . THE COMPOUND POISSON MODEL Theorem 10 . B [7]2 (10.1) .
To this end .e.. lcT(a ) 4 /c(a).c'(7) N (0' T (2(7) / N (0. Combining ( 10.4) and Lemma 10. NT i =1 n'(a) for all a so that for all sufficiently large T K7 . ESTIMATION OF THE ADJUSTMENT COEFFICIENT which is the same as (10. I. °7IT) .2.e. y + E) eventually.'T(a) = 1 E Uie°U' a$' EUe "u = B'[a]. OT a 95 u 4 /3.'(y).e) < 0 < r. it follows that 7T7 KT(7T) .E) < 4T(7T) < 4T(7 + E).1 By the law of large numbers. 6"Y (10. and the truth of this for all e > 0 implies ryT at 'y. Proof of Theorem 10. Let 0 < E < ry. NT BT [a] Hence r.(ry + e) and hence KT(7 . 0 are estimated from data .(ry . 7T E (y .10. where ryT is some point between ryT and ry. If ryT E (7  we have KT(7 .KT(7) kT(7) K'(7) . By the law of large numbers.E ) < 4T(7T) < (7 +0' which implies 'T(ry4) a$' r. BT[a] 3 B[a]. Then r. Now write KT(7T)  kT(7) = 4T(7T)( 7T 7).4) + E). first note that e7TU N (e7U u2e27Uo'2/T) 7 .E) < 0 < kT(7 + E) for all sufficiently large T .1 can be used to obtain error bounds on the ruin probabilities when the parameters . Theorem 10.Q.3).
ft.3 or equivalently p > 1/2 or 11 < 100%. Letting Wo = 0. it means 2 (8 . Wn). Embrechts & Mikosch [133]. This approach in fact applies also for many models more general than the compound Poisson one. Frees [146]..Q.5%).info< „< t S.96 CHAPTER III. i .) = a (e.Wn) are i . satisfies b(.C1e"a ( see e. A major restriction of the approach is the condition B[2ry] < oo which may be quite restrictive.e. various alternatives have been developed.e. One (see Schmidli [321]) is to let {Vt} be the workload process of an M /G/1 queue with the same arrival epochs as the risk process and service times U1.f.T VIT where r7ry. t]}. 6 < 2. U2. Vt = St . > 0 for some t E [Wn_ 1.T = 3TKT ( 21T)IKT (^T)2 is the empirical estimate of vy and fc. the nth busy cycle is then [Wn1. 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. Mammitzsch [253] and Pitts. Csorgo & Teugels [95]. = 1. Further work on estimation of y with different methods can be found in Csorgo & Steinebach [94]. wn = inf{t > W. For example . if B is exponential with rate 8 so that ry = 8 . Asmussen [23]) can then be used to produce an estimate of ry.0) < 5... and the known fact that the Y„ = max Vt tE[W„1.i. Herkenrath [192]. [197]. . For this reason . V..ueryuU ". Notes and references Theorem 10. Deheuvels & Steinebach [102].1 is from Grandell [170]..g.. Griibel & Embrechts [292]. i.g. Hipp [196].96 if a = 2.d. with a tail of the form P(Y > y) ...1 : Vt = 0.
See Fig. the premium rate is 1.g.Chapter IV The probability of ruin within finite time This chapter is concerned with the finite time ruin probabilities 0(u.1) . 'y) where c(a) attains it minimum value. generalizations to other models are either discussed in the Notes and References or in relevant chapters.s.f.1 where p = 13µB. exists. 97 . B[•] and mean AB. 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. Unless otherwise stated. defined as solution of c(ry) = 0 where ic(s) _ /3(B[s] . In particular. it is assumed that i > 0 and that the adjustment coefficient (Lundberg exponent) y. The notation is essentially as in Chapter III. Further let 'Yo be the unique point in (0.1 (the role of ryy will be explained in Section 4b). the Poisson intensity is 0 and the claim size distribution is B with m. The safety loading is q = 1/p . 0.
7. PL = 6/0 = 1/p > 1).9). E[T(u) I T(u) < 00 ] = ELT (U). PROBABILITY OF RUIN IN FINITE TIME Figure 0.. 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 Exponential claims Proposition 1.) = 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. Var[T(u) I T(u) < 00] = VarL T( U) . EL refer to the exponentially tilted process 3 with arrival intensity S and exponential claims with rate / (thus .1) (1.(4.t.r. 1 FL. By the likelihood identity III.1 The claims surplus is {St}. the time of ruin is T(u) and ^(u) = ST(t&) .u is the overshoot. we have for k = 1. 2 that E [T(u)k. In particular.(. using that the overshoot l.5 .1 In the compound Poisson model with exponential claims with rate S and safety loading 77 > 0. (u) is exponential with rate 0 w.(U) < 00] = ELT(u)ke'YS. . FL and independent of T(u).98 CHAPTER IV.
2) is aLELT( u) . u + ELe(u) _ PL . which leads to the quadratic 02 + (/3 .1)T(u))2 = UL where = s.V/ is as asserted.(yo) = 2V . the 1.1)T(u)) = VarLe(u) + (PL .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 ./3) .h.1./3 . Let 0 > yo be determined by ^c(0 ) = a.1)) ELST(u) ELT(u) (PL . the Laplace transform of the time to ruin is given by Eea7( u) = E [eaT (u).1 (6)3)2 which is the same as the r. EXPONENTIAL CLAIMS For (1 .h.s.s.12 Thus the l."(ry) = 26//32. is V1rLSr( u) +VarL ((PL . T(u) < oo] fora > r.I (1.6 + a)0 .s. where = eBu I 1 .B = a. Since Sr (u) and (PL . Wald's second moment identity yields 2 EL (Sr(u) .(PL .1 /3u + 1 u + 1 //3 = 6/3 6/01 For (1. 0 Proposition 1. 1). This means that /3(6/(6 .h.2 In the compound Poisson model with exponential claims with rate 6 and safety loading rl > 0.2).1)2VarLT(u) + 2 Ca 1I VarLT(u). .6a = 0 with solution 0 (the .1) .1)ELT(u).0) . of (1.6.1)T(u) are independent with QL the same mean .1//32 (6/)3 1)2 26(/3u + 1)/(6 . we have by Wald's identity that (note that ELSt = t(pL .
. PROBABILITY OF RUIN IN FINITE TIME sign of the square root is + because 0 > 0).. (1. T2 . . Using 5 = 6 . Fig.3) we have E [e«T(u ). Note that it follows from Proposition 1. T(u) < oo] = EB [exp {aT(u ) . Cf.Y1 Y2 Figure 1. But by the fundamental likelihood ratio identity ( Theorem 111..1 where Y1.1 . 1. are the ladder heights which form a terminating sequence of exponential r.3. St Ti F.v.. the result follows.OuEee 04(u) = ee u be BB+B where we used that PB(T(u) < oo ) = 1 because 0 > ryo and hence E9S1 = K'(0) u > 0. Ti.'s with rate 5.4.. and M(u)+1 is the index of the ladder segment corresponding to T(u).v. More precisely. Y(u) belonging to a convolution semigroup .4) The interpretation of this that T(u) can be written as the independent sum of T(0) plus a r. T(u) < oo] = e. M(u) T(u) = T + E Tk k=1 where T = T(0) is the length of the first ladder segment .100 CHAPTER IV.3 that we can write EeaT( u) = eeuEe 017(o)..9ST(u) +T(u)!c(0)} .0.. are the lengths of of the ladder segments 2...T+ Ti a t U T I 1 a i F. Y2.
4.1.d.T.0.. including the customer being currently served).6(u) = Vfl/j l(Su.i. T. .k + 1). U2.cos (u/. 2.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. .1(u.I ex cos B cos j O dB fo " .. If QT = N > 0. let (cf.T..T) 1 I fl(O)h(0) fdO where (1.T the service times of the customers awaiting service . i. where U1. Note that the case 6 # 1 is easily reduced to the case S = 1 via the formula V.. the following formula is convenient by allowing t.6. EXPONENTIAL CLAIMS 101 For numerical purposes .1. ...v. UN. then VT = U1.3 sin0 + 29) f3(0) = 1+/32/cos9.T + • • • + UN. [4]) 00 (x/2)2n+3 Ij (x) OnI(n+j)! . and exponential with rate S = 1.6) fl(9) f2(0) = = fexp {2iTcos9(1+/3)T+u(/cos91) cos (uisin9) .e. For j = 0. Then V(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 . Let {Qt} be the queue length process of the queue (number in system. density xN lex/(N . the conditional distribution of VT given QT = N is that of EN where the r. Proof We use the formula . UN. Corollary 11.1 )!. Since U1 . EN has an Erlang distribution with parameters (N..i (u. 1). .3 Assume that claims are exponential with rate b = 1.T are conditionally i.ST).T is the residual service time of the customer being currently served and U2 .T. cf. T) to be evaluated by numerical integration: Proposition 1. .
and define tj = e(1+R)Taj/2Ij(2vT T). f3(0) . (1.3(k +1)/ 2ei(k + l)6 (.112 l 1( k +1)/2 [ 31/ 2 cos(kO) .102 CHAPTER IV.)3k +1 tj g'(QT >.13(k +l)/2ei(k +1)9 R E .1 R [.i(k +1)e R [/3( klal/2e:0 (01 /2 e .1 00 ok+lR 00 j=k1 +1)/2e . let I _ j (x) = Ij (x).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. 8789) 00 E aj j= 00 = 1. 00 E '3j/2 cos(je) j=k+1 00 _ j=k+1 ^j/ zeij = .cos((k + 2)9)] d9.8 ) yields F(QT > k + 1) .1)] L _112 /(k+1)/2 [. k k2 + $k+1 E bj 00 t j . PROBABILITY OF RUIN IN FINITE TIME denote the modified Bessel function of order j.ie .cos (( k + 1)0)] f3(9) Hence the integral expression in (1.44).(31/2eie .(31/2 cos (( k + 2)9) .31 /2eie L 1)] 1 I/31/2eie .)3k+1 = e(1+0)T e201/2Tcos 7r 0 e )3(k +l)/2 [31/ 2 cos ( kO) .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 .38). 912. in particular equations (1. similar formulas are in [APQ] pp.
The rest of the proof is easy algebra. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Since P(QOO > k + 1) = flk+1.7) that _ [^ au ak+l (30 k L. 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). Seal [327] gives a different numerical integration fomula for 1 . however.T) which.3 was given in Asmussen [12] (as pointed out by BarndorffNielsen & Schmidli [59]. there are several misprints in the formula there. . however.2. from the accumulated claim distribution N. F(x. the numerical examples in [12] are correct). equivalently. 2 The ruin probability with no initial reserve In this section . it follows as in (1. we are concerned with describing the distribution of the ruin time T(0) in the case where the initial reserve is u = 0. We first prove two classical formulas which are remarkable by showing that the ruin probabilities can be reconstructed from the distributions of the St. E Fk. The first formula. T)..e = e' COS a cos(uf31/2 sin 0). oo (u)31/2e^e)k = )3k z cos(k9) = R k. T) in terms of F(. We allow a general claim size distribution B and recall that we have the explicit formula z/i(0) _ P(7(0) Goo) = p.0(u. u Notes and references Proposition 1. Ui < x I / (note that P(St < x ) = F(x + t. t) = P . 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. k! k=O k0 i/z Co Uk ate" o'/z e . is numerically unstable for large T. going back to Cramer. Related formulas are in Takacs [359]. or. t )).
2. resp.T)) does not {Stv)} depend on v. Stv^ _ Define M(v. 1 1 . [v. PROBABILITY OF RUIN IN FINITE TIME Theorem 2 .(0. ") } is at a minimum at time t. v].t)= {Stv) < SM. 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.1 In formulas.T))dv. . co ).(. f T lStv)} 0<t<T by a 'cyclic translation'. we define a new claim surplus process St StM NJ Figure 2. T]. T) = P(Tr(0) > T) = P(M(0.b (0.T)dx.T].3) with A = (0.1. meaning that we interchange the two segments of the arrival process of {St}o<t<_T corresponding to the intervals [0.T))dv E^T I(M(v.(6. Then 1 .i.0<w<t} St+v .T)) 1 fT P(M(v. Proof For any v E [0.S„ 0 <t<Tv STS„+St_T+v Tv<t<T as the event that IS. T T o where the second equality follows from II.T) T F(x. See Fig.104 CHAPTER IV.
v). T) = M(0. we can take v E (w E. Proof The event {ST < u} = { Ei T Ui < u + T j can occur in two ways: either ruin does not occur in [0. 0<t<v} = {ST < St .2.T) = F(u+T.Sv. For example.Tt))f(u+t. cf. T]. t). T. T)) dv. then M(v.2. T) occurs. in which case there is a last time o where St downcrosses level u. T) as {ST<St+ vS. T) occurs. we can write M(v. ST > 0. It is then clear from the cyclical nature of the problem that this holds irrespective of whether M(0.T) occurs.T)) dv f T I(M(0.T)f(I z /)(0. w) for some small E. this integral is 0 if STv) . v < t < T} n M(0.. v<t<T}n{ST<STSv+St. 0<t <Tv}n{ST<ST Sv+St T+v. 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 . where the last equality follows from ST < St on M(0.v<t<T} = {ST<StSv. Hence T TE f I( M(v. then i fT I(M(v.T) occurs or not as long as ST < 0. T)) dv = TEST = T fP(ST < x) dx T T NT 1 f P(ST < x) dx = 1 f P Ui T . v) = M(0. Indeed. We claim that if M(0. THE RUIN PROBABILITY WITH NO INITIAL RESERVE 105 Now consider the evaluation of fo I(M(v. v).xdx. T Theorem 2 . letting w = inf It > 0 : St_ = mino<w<T Sw}.ST on M (0.t)dt. or it occurs. T T o i =1 Let f (•. Fig 2. Obviously.. t) denote the density of F(•. v). If ST < 0.2 10(u. . there exist v such that M(v. T)).T) and Sv < 0 on M(0. It follows that if M(0 .
T) = C(z. which occurs w. ST_ _ z}. which is independent of St and has the stationary excess distribution B0.b(u. u which is the same as the assertion of the theorem. t + dt] occurs if and only if St E [u. E [t.T)+ J0 T (1V.3 Define r_ (z) = inf It > 0 : St = z}.T) = {St < 0.v. Proposition 2.2. 2. O(T .t). Let Z be a r. {S t > z.Tt))P(StE[u.z.2 .(0. The proof is combined with the proof of Theorem 111. Hence P(ST<u) = 1 .ST_ t_ and let A(z.2. 0 < t < T. For a fixed T > 0.T) = . u + dt] and there is no upcrossing of level u after time t. z > 0. {St > . PROBABILITY OF RUIN IN FINITE TIME u Q II T Figure 2. The following representation of T(0) will be used in the next section. 0 < t <T.106 CHAPTER IV. C*(z. ST_ _ z}. define St = ST . Then P(T(0) E • I T(0) < oo) = P(T_ (Z) E •).2 Here o. ST_ _ z} .p. 0 < t <T .u+dt]). Proof of Theorem 111.
Proof of Proposition 2.3.T(0)<oc) = f x F(U > y + z U > z) P(Sr(o)_ E [z.T + dT]. z + dz]) = P(A(z. {St }o<t<T have the same distribution . r(0) < oo) = 3R(z) dz JP(C(z. z + dz]. (2. 2. Hence integrating (2. z + dz]. u which is the assertion of Theorem 111.2. z + dz]. T(0) < oo) = OR(z) dz in (2.T))f3B(z) dz dT. and since {St}o<t<T.T)) = P(Cx. Thus P(Sr(o)_>x. z + dz].3 But by sample path inspection (cf.T)). ST(o)_ E [z. .3).T))dT = Off(z) dz P(T_ (z ) < oo) = 3B(z) dz. we therefore have P(A(z. It follows by division by P(ST(o)_ E [z.T) = C*(z. Fig. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Then 107 P(r(0) E [T.ST(o) >y.1) yields P(ST(o)_ E [z. 7( 0) < oo) = P (C(z)) dT. Figure 2.T).2. T + dT] I S7(o)_ E [z. A(z.1) z T .1) that P(T(0) E [T.2. T(0) < oo) B(y B(z) + z) f3B(z) dz = 3 f °^ B(y + z) dz = f3 + x v f B(z) dz.
Theorem 2.1 Eear( y) = eyr(a). a martingale proof is in Delbaen & Haezendonck [103]. Notes and references For Theorems 2. [329]. Note that T_ (y) < oo a.r(a). 3 Laplace transforms Throughout in this section.108 Hence CHAPTER IV.1 and the present proof is in the spirit of Ballot theorems. see in addition to Prabhu [293] also Seal [326].c(r(a)) l = l er( a)se+at } u yields 1 = eyr(a)Eear(y).3 was noted by Asmussen & Kl(ippelberg [36]. one based upon a result of Asmussen & Schmidt [49] generalizing Theorem 11. cf. z + dz].2 ga(x) = Qexr(a) f "o eyr(a)B(dy) x .s. PROBABILITY OF RUIN IN FINITE TIME ]P(7(0) E [T. 2.5 and one upon excursion theory for Markov processes (see IX.5a). Lemma 3.1. Lemma 3 .T+dT]). r(0) < oo. who instead of the present direct proof gave two arguments. Tak'ecs [359]. I L Let ga(x) be the density of the measure E[ear(°). Proof Optional stopping of the martingale I er (a) 9 t.6. ^(0) E dx] (recall that ^(0) = Sr(o)) and write ga[b] = f OD ebxga(x) dx. because of77>0. z + dz]. Proposition 2.2.3. Let T_ (y) be defined as Proposition 2. T(0) < oo) = dTP(T_(Z) E [T. T(0) < oo) 0 = dT f 0 P(C(z))P(Z E [z. r(a) denotes the solution < 'Yo of the equation a = ic(r (a)) = .T + dT] T(0) < oo) dT f ' P(C(z))P(Sr( o)_ E [z. (3.(3(B[r( a)] . some relevant references are Shtatland [338] and Gusak & Korolyuk [181].(yo).1) .1) where a > r. In the setting of general Levy processes.
ic(b)/b x(b) + a eb"E[eaT(" ). time T(u): u u Here is a classical result : the double m. (Laplace transform) of the ruin Corollary 3.r(a) = a [B[b] B[r(a)]] . Z = y] = EeaT.°° ga(x)dx. T(u) < oo] du = Proof Define Za(u) = E [eaT(" ). It is then easily seen that Za(u) is the solution of the renewal equation Za (u) = za (u) + fo Z.x)ga (x) dx where za(u) = f.r(a) The result now follows by inserting /3B[s] = ic( s) +/3+ s and ic(r(a)) =a.r(a) b . r(u) < oo). LAPLACE TRANSFORMS 109 Proof Let Z be the surplus . y + dy]. E[ear (o) I T(0) < oo .ga [b] 0 TO Using Lemma 3.3.g.T(0) < oo] = 20[b] = za[b] (9a[b] 9a[0])/b 1 .2 P(Z E [y. b .r(a) oo Q f ex(br(a))dx f00 eyr(a)B(dy) x 0 Q f evraB(dy) e(a))dx 0 Q cc ev(br (a)) .x)(a) B(dy)• Lemma 3 .f.ga [b] 1 . the result follows after simple algebra. Then by Proposition 2. Hence eb"du E[eaT("). Further by Theorem 111.(v) = ev''(a).5 f 00 o a/r(a) .3.3 ga[b] = c(b) Proof + b + a .1] evr(a)B(dy)[ b . Corollary 3. (u .2. rr(0) < oo) = 1_ r(a) Proof Let b = 0.4 E[eaT (o). u .ST(o)_ just before ruin . £(0) E dx) = /3B(x + dy) dx and hence ga(x) = f e r)/3B(x + dy) _ /3 f x e(v.3.
for any m T(u) u .1 Assume 77 > 0..w ) v/.mu D 2 4 N(0. i. For the proof. u 1 ET(u) 1 p1 u where Pw2 = 311B)m3• 7(u) . t T(u) T(u) T(u) t m = lim = lim = lim Utioo u + Sr(u) u+oo S. Later results then deal with more precise and refined versions of this statement. (4.. The first main result of the present section is that the value umL. the known results are even less explicit than for the exponential claims case. That is.mL > E T(u) < 00 ) 40.e. = (p . and take basically the form of approximations and inequalities. Then as u * oo.3LELU 1 1p' is in some appropriate sense critical as the most 'likely' time of ruin (here C is the CramerLundberg constant). For the second . and hence a. Proposition A1. (u) t.110 CHAPTER IV. T(u) a.(u ) = o(u) a.2 Assume ri < 0. This proves the first assertion of (4. for any c > 0 P( Further.h(u.3). mu ) ( 0 m < ML '(u) 1 m > rL. T(u)/u mL as u + oo.s. Theorem 4 . we need the following auxiliary result: Proposition 4. where _ 1 _ 1 1 C ML w(ry) 6B'[7J 1 . uoo u using e. By Proposition 111. St/t 1 1/m.UProof The assumption 11 < 0 ensures that P(T(u) < oo) = 1 and r(u) a4' oo.1.2.s.r(u) = Er(u) • ES. P = /3µB > 1.00 St = lim . Then given r(u) < 00. PROBABILITY OF RUIN IN FINITE TIME 4 When does ruin occur? For the general compound Poisson model.1) i.1)Er(u) . note that by Wald's identity u + EC(u) = ES.6. cf.
Thus.^ N (o. T) for T which are close to the critical value umL). though it is not easy to attribute priority to any particular author.1 is standard.6µB2) Z v m (3µB2) Z.s.3).g. WHEN DOES RUIN OCCUR? and that Ee(u)/u a 0.1).1 The l.4.2) follows immediately from u (4. 4a Segerdahl's normal approximation We shall now prove a classical result due to Segerdahl.5) St . the same conclusion holds with t replaced by r(u). cf.mL >E By Proposition 4.h. this can be rewritten as u + 1(u) . Theorem 7. 4). 1'r(U) .3.1. PL (•)+ 0. T (u) < 00 J 0(u) e7'PL U \ I T u) . Tu) T( u) . the result comes out not only by the present direct proof but also from any of the results in the following subsections. apB ) .N(0.t/m D (2) 111 . which may be viewed both as a refinement of Theorem 4.1). again Proposition A1.mL U > E.mu (2) '• m3/2 µB 7 . For (4 .r(u)/m T(u) ti µB2) Z.7 6  11 Proof of Theorem 4.2 of [86]) and (4. note first that ( Proposition 111.2.1 (by considering 0(u. of (4. and (4. and as a timedependent version of the CramerLundberg approximation. According to Anscombe' s theorem (e. .1) is T (u)  U mL P( T (u) < I > E. T(u) < oo f / 00) e7uE L [e_7 (t1). Notes and references Theorem 4. If Z . implying T(u) . proving (4.1).6.mu m .
v.L+YWLV'U) . Proof Define u' = u .r. we can replace T(u) by r(u'). O . oo). and thus in (4.u1/4)I(S(u') > u1 /4) h(oo) + 0.6). using that ul/4 . Then the distribution of T(u) . resp .T ( u')] = E[ T ( ul /4 .ST( u') = u1/4 . Let h(u) = E f (^(u)). we need the following auxiliary result: Proposition 4.4).T(u') given F.))I h(ul /4  ^(u)) I(6 (u') C ) f < ul /4 + f(e(u') . E9(Z) (4. Using ( 4.6) whenever f. Hence Ef (Vu )) 9 (T(u.5) For the proof. PROBABILITY OF RUIN IN FINITE TIME Corollary 4. Then h(u) 4 h(oo) = E f (6(oo)). one has 9 (r(u)_rnu) Ef (^(u)) * E. then e(u) and r(u) are asymptotically independent in the sense that. letting Z be a N(0.e(u') oo w . S( u ) < ul/4] < ET(ul / 4) = O(ul/4). oo ). with w2 as in (4.4 (SIAM'S LEMMA) If 71 < 0.um. we get E[ T (u) .) is readily seen to be degenerate at zero if ST(u•) > u and otherwise that of T(v) with v = u .VU T.t.l:(oo) (recall that rt < 0).ul/4. P because of ^(u') .mul h(oo)Eg(Z).(u. (oo. and similarly as above we get E[f(^(u)) I Fr(u.3).^(T(u')).112 CHAPTER IV. g are continuous and bounded on [0.3 (SEGERDAHL [333]) Let C be the CramerLundberg constant and define wL = f3LELU2mL = f3B"[ry]mL where ML = 1/(pL1) = 1/($B'[ry]1).a C4'(y )• ( 4. e'°'/b (u. Then for any y.w2) r.f ( (oo)) .)mu \ h(oo)Eg (r(ul) .
yK(ay)• (4. see Asmussen [12] and Malinovskii [254]. 0. 4b Gerber's time.dependent version of Lundberg's inequality For y > 0. Cf. y u) < . see also von Bahr [55 ] and Gut [182]. yy by 1 K. oo ) as u * oo. e7v" y < ^'(7) (4 .4) in the last. The precise condition for (4. u needs to be very large).3 in terms of Edgeworth expansions . 3 is due to Segerdahl [333]. y > k'(7) . just substitute T = umL + ywL in (4.4.z/)(u . Notes and references Corollary 4 .8) Note that ay > 7o and that 7y > •y (unless for the critical value y = 1/ML).7) to be good. PL(T(u ) < umL + ywL) 113 4 C4(y). y u) < e 7v" . WHEN DOES RUIN OCCUR? Proof of Corollary 4.3 ery"z/i(u . ELe7E (") . Segerdahl 's result suggests the approximation b(u. For refinements of Corollary 4. that for the fit of (4.T) Ce7"4 (T . also Hoglund [204].9) ( 4 . umL + ywL f) = e"P(T (u) < umL + ywL) = EL [e7V "). For practical purposes . CL Fig.oo. T(u) < umL + ywL f. The present proof is basically that of Siegmund [342]. Thus . 10) '5(u) .1. however . where we used Stain's lemma in the third step and (4.7) to be valid is that T varies with u in such a way that y(T) has a limit in (.7) whenever u is large and ly(T)l moderate or small (numerical evidence presented in [12 ] indicates . define ay.5 '(u . .5) and solve for y = y(T). Theorem 4.umL wI V"U u (4.7) To arrive at this . in practice one would trust (4.(ay) = 17 7y = ay .
which may be understood from Theorem 4.Y' (u.8 below . In view of Theorem 4.r.3 yields easily the following sharpening of (4.6 It may appear that the proof uses considerably less information on ay than is inherent in the definition (4. f Some urther discussion is given in XI.7 i. we have rc(ay) < 0 and get (u) . and hence t. An easy combination with the proof of Theorem 111. Numerical comparisons are in Grandell [172 ]. 5 is due to Gerber [156 ].yu ) = eayuEav [e . Hoglund [203] treats the renewal case. . 0.b (u. we arrive at the expression in (4. yu) is e 'Yyu/ . yu ) = < eayuEay [eay^ ( u)+T(U)K ( ay). the bound a7y° turns out to be rather crude .5.9): Proposition 4.t. From the proof it is seen that this amounts to that a should maximize ayic(a).114 CHAPTER IV. T(u) < yu] < eayu + yUr(ay) Y < eayuEav [ eT(u)K(av )L T(u) < yu} Similarly.v"U. yu) < C+(ay)e7a„ where l C+(ay) = sup f 00 eayR(xy)B( .8). yy is sometimes called the timedependent Lundberg exponent.2. and generalizations to more general models are given in Chapter VI. dy) Notes and references Theorem 4 . if y > 1/ic'(y). PROBABILITY OF RUIN IN FINITE TIME Proof Consider first the case y < 1/K'(y). a.8).1). see MartinLM [257] . Then ic(ay) > 0 (see Fig . However. who used a martingale argument. the point is that we want to select an a which produces the largest possible exponent in the inequalities.h(u. which shows that the correct rate of decay of tp(u. For a different proof. yu 11 < T(u) < oo j < eayu +Y UK(ay) Remark 4. yu < T (u) < oo 1 l e ayuEav [eT ( u)K(ay).ay4(u)+ T(u)K(ay ).6. u Differentiating w.
ayuEay f eay^ ( u)+T(u)K(ay). not inequalities. we have ryas = ay . (4. u 4 oo.12) < yu] Here the first expectation can be estimated similarly as in the proof of the CramerLundberg ' s approximation in Chapter III. The traditional application of the saddlepoint method is to derive approximations..e. and ii(u) .z. and in case of ruin probabilities the approach leads to the following result: Theorem 4 . As a motivation. (0) r1 (a) ' I.4.'(y ).: T. (4.5. We thereby obtain that T is 'in the center' of the Padistribution of T(u).yu) c ay . the choice of ay. Proposition 4. Ea . yu) = e.ayC() . (4. T(u) < yu] .13) . [eT(u )K( ay). then the solution &y < ay of . i.e. This idea is precisely what characterizes the saddlepoint method.6 with P replaced by Pay and FL by Pay.c(&) = ic(ay) is < 0. For any a > yo.^3 ]1/ Bay [lay .8 If y < 1/ic'(ry).ay y 'Yay  ay . then the relevant choice is precisely a = ay where y = T/u. WHEN DOES RUIN OCCUR? 115 4c Arfwedson's saddlepoint approximation Our next objective is to strengthen the timedependent Lundberg inequalities to approximations.. Using Lemma 111.ay a.i(u. and b(u. it is instructive to reinspect the choice of the change of measure in the proof. if we want EaT(u) . then ay > 0.yyu y l ay I 21ry/3B" [ay] V fU_ u + 00.(u.. yu ) ayay e ryyu ayay 27ry/3B"[ay] u Proof In view of Stam 's lemma. T(u) suggests heuristically that l t/.ay and get Ea e ayf (00) y _ 'Ya( ayKal lay C 1 .11) ' If y > 1/ r .2 yields EaT(u) u u r. the formula 0(u. yu ) eaauEaye .
c(ay)ul/2W p 2ir = eyu(ay) dz 1 rc(ay ) 2. The proof of (4.7ruw2 Inserting these estimates in (4.1.3(5/(S .l'B)y /(Pay .a) .1) .ay + ayl /BLay] .1) under Pay mation (4.1)3 = y3/3B"[ay]. we get heuristically that Eay Ler (u)r(ay).9 Assume that B(x) = eay. 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)K(ay) ay ayI&YI For the second term in (4. V < 01 Ir 00 er(ay)"1'2"'x eyur.a)2 .116 CHAPTER IV.13). .(ay) _ y(ay .1)3 = (jB"[ay]l (Pay .13) rigorously. Example 4. i B[7ay .ay ) r. (4.13).(j (1 . PROBABILITY OF RUIN IN FINITE TIME ry I i .B[ay] /ay &y y(ay .ay) ay +. (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 . The difficulties in making the proof precise is in part to show (4.I ay &y a ^c'(ay) a (1 +. it seems tempting to apply the normal approxiyu + ul/2wV.4).c'(a) _ /3a/(8 .11) follows.12) is 0 entirely similar. where V is normal(0. Then ic(a) = . Writing r(u) and W2 = I3ay{.1B[ay]1 ) y(ay . a nr=.a. T(U) < yu] = eyuk (ay)E''ay (ek(ay )"1/2WV. and the equation ic'(a) = 1/y is easily seen to have .
then { __ . DIFFUSION APPROXIMATIONS solution ay=5 117 V 1 (the sign of the square root is .= (s. in discrete time: if p = ES.p.g.because the c.3+52 1+/351/y' sy 7 B ii[ay] 25 _ 251/2(1 + y)3/2 (5 . (5. 2 = Var(Si ) the variance.f. . 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. c a 00.5./4 ^y for 1/i (u.. is the drift and o.1) . is undefined for a > 5). yu) when y < 1/ic'('y) = p/1 .. The mathematical result behind is Donsker's theorem for a simple random walk {Sn}n=o.11) gives the expression '31/4 ( .ay)3 0 3/2 and (4. and next to note that such an approximation in particular implies that the first passage probabilities are close.. A related result appears in BarndorffNielsen & Schmidli [59].tcp) Lo {Wo ( t)}t>0 . y) a''y" L '3 _ fl ) 51 /4(1 +1IY)3/4 \.8 is from Arfwedson [9].1. 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 =.i )( v s vc ('3 + s _2 / . 0 Notes and references Theorem 4.
t} _ {W_1(t)} . However. St = EN` U= . oo)). for the purpose of approximating ruin probabilities the centering around the mean (the tcp term in (5. of which a particular case is the claim surplus process (see the proof of Theorem 5.1.2) t>o where p = pp = p .1 As p J. We want an approximation of the claim surplus process itself.. p.3) whenever c = cp f oo as p 1 p.3. and this can be obtained under the assumption that the safety loading rt is small and positive.. cf.z } {W_1(t )}t>o (5.1)) is inconvenient. Letting c = a2/pp. (5. and consider the limit p j p. It is fairly straightforward to translate Donsker's theorem into a parallel statement for continuous time random walks (Levy processes). 0 . PROBABILITY OF RUIN IN FINITE TIME where {W( (t)} is Brownian motion with drift S and variance (diffusion constant) 1 (here 2 refers to weak convergence in D = D[0.1) with S. Mathematically.a = Snp) and the inequalities Sn )C . we shall represent this assumption on 77 by a family {StP) L of claim surplus processes indexed by the premium rate p. a2 =/3µB2) Proof The first step is to note that { WC (St P) . this is an easy consequence of (5. where p is the critical premium rate APBTheorem 5 .1 below). + {Wo(t ) . Lemma 111. Indeed .p.tcpp) y = { WC (Sct) pct) } {Wo( t)}t>o (5. This is the regime of the diffusion approximation (note that this is just the same as for the heavy traffic approximation for infinite horizon ruin probabilities studied in III. such that the claim size distribution B and the Poisson rate a are the same for all p (i. n/c < t < (n + 1)/c.tp).p/c < St(p) < S((n+l)/ c + Pp/c.e. we have o {i!t s: .3) takes the form LI S(P) { a2 to2/µ2 + t LI S (P) { a2 ta2/µ2 {W0(t)}.7c).118 CHAPTER IV.
h. any probability measure concentrated on the continuous functions. C. the continuous mapping theorem yields sup W Sz2 to lP 4 sup Wi(t)• O<t<T O<t<T a2 Since the r. and in fact some additional arguments are needed to justify (5. this implies P sup 0<t<T a 12 Stu2 /µ2 > u 4 P ( sup W_1( t) > u O<t<T But the l.5.e.4) Note that IG(.8 or [APQ] p. (5.h. ulpI/a2). However.1.r. u). ('. since ti(u) has infinite horizon .6) This is the same as the heavy traffic approximation derived in III.7c. 1. (. . we omit the details .t. Corollary 5.1. 119 It is wellknown (Corollary XI. is IG(T. 263) that the distribution IG(•. is 1/ip (ua2 /IpI. we obtain formally the approximation V.(u) ti IG(oo.u).2 suggests the approximation u 0(u. has a continuous distribution. and the r.( ^ I + e2( \ I . ulpl /a2) = e2"1µl / or2. [169] or [APQ] pp. TS(u)=inf{t>0: WW(t)>u}. 199.s. For practical purposes . see Grandell [ 168].h.1 . (ua2 To2 op \ IPI > IG ( T .f I \\\ J \ (5.2 As p j p.. DIFFUSION APPROXIMATIONS Now let Tp(u) = inf{t>0: S?)>u}.T) IG(Tp2/ a2).5). 196. u) of r( (u) (often referred to as the inverse Gaussian distribution) is given by IG(x. w. Corollary 5 . u) is defective when < 0. u) =PIT( (u) < x) = 1 ..1 I 7= .5) Note that letting T * oo in ( 5.s. ^ p2 Proof Since f 4 SUP0<t<T f (t) is continuous on D a.6) from Theorem 5. (5. Because of the direct argument in Chapter III.. the continuity argument above does not generalize immediately.Ta2 /p2).s.
The picture which emerges is that the approximations are not terribly precise. Then as 0 _+ 90. and two further standard references in the area are Grandell [168]. Michna & Weron [152] suggested an approximation by a stable Levy process rather than a Brownian motion. For claims with infinite variance. the B9. See for example Billingsley [64]. We conclude this section by giving a more general triangular array version of Theorem 5.1 and Section VIII.Po = 09µB6 . on the premium rule involving interest. the simplicity of (5.5) for the compound Poisson model which does not require much more computation. However. The proof is a straightforward combination of the proof of Theorem 5. Theorem 5. in particular for large u. a2 = ae = 00µa6 Notes and references Diffusion approximations of random walks via Donsker's theorem is a classical topic of probability theory. Furrer. that 00 4090.t.3 Consider a family {Ste) } oc claim surplus processes indexed by a parameter 9.00µB6 + 0. (5. and which is much more precise. in Asmussen [12]. such that the Poisson rate Oe. as 0 * 00 and that the U2 are uniformly integrable w. PROBABILITY OF RUIN IN FINITE TIME Checks of the numerical fits of (5. Further relevant references in this direction are Furrer [151] and Boxma & Cohen [75]. 0) { 2 StQ2 /µ2 D { W_ i(t)}t>o t>o D 2 where p = pe = pe . the claim size distribution B9 and the premium rate p9 depends on 0.Pe.r. B0 * Boo. for more general models it may be easier to generalize the diffusion approximation than the CramerLundberg approximation. However. we have ^A. as an example of such a generalization we mention the paper [129] by Emanuel et al.6) are presented.. The first application in risk theory is Iglehart [207]. pt? 4 peo.g.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. . in the next subsection we shall derive a refinement of (5. e.6 of [APQ]. In contrast.1. pe . [169].5) and (5.120 CHAPTER IV. All material of this section can be found in these references.6) therefore does not appear to of much practical relevance for the compound Poisson model. Assume further that 039µB6 < pe. In view of the excellent fit of the CramerLundberg approximation.
PB('r(u ) < oo) < 1 for 9 < 0.1 > 0. In this setup. Since Brownian motion is skip free. Determine yo > 0 by r. P9(r (u) < oo) = 1 for 9 > 0. . this is because in the regime of the diffusion approximation . whereas there we let the given 3B. 2. 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.1) . For each 9. The setup is the exponential family of compound risk processes with parameters ( B9 constructed in III.90) and the given risk process corresponds to Poo where 90 = 'yo. However . In terms of the given risk process with Poisson intensity .s and p = /3µB < 1. 0(0) = 0. let P9 refer to the risk process with parameters Q9 = QoB0[9] = QB[9 9o].Q (B[s] .ao (0) _ /c(s + 9 .c(s) = . B9(dx) =Bale] Bo(dx) e9z keo)z = B[9 .9(s) = Ico ( s + 9) .6. which we have seen to play an important role for example for the CramerLundberg approximation . this means the following: 1. Let PO refer to the risk process with parameters e9oz Qo = QB[90]. . Then r./c(9 . and we are studying b(u.4.T) = Peo(r(u) < T) for 90 < 0.90) .'(yo) = 0 and let 90 = 'Yo. Then EOU' = Boki[0] = Biki[eo]/E[9o] and "(s) = k(sBo)k(9o). this idea ignores (among other things) the presence of the overshoot e(u). Bo(dx) = B[eo]B(dx).6. claim size distribution B . it is more convenient here to use some value 9o < 0 and let 9 = 0 correspond to n = 0 (zero drift). 77 = 1/p . 3. risk process with safety loading 77 > 0 correspond to 9 = 0 . The objective of the corrected diffusion approximation is to take this and other deficits into consideration. and we want to consider the limit 77 10 corresponding to Oo f 0. 77 is close to zero. 9o T 0.90] B(dx).
.1) . PROBABILITY OF RUIN IN FINITE TIME Recall that IG(x.e.122 CHAPTER IV.. () where h (A. The corrected diffusion approximation to be derived is (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. the solution of r. IGu+u2.. write r = T(u). C) = 2A + (2 . (6. (01. (.Varo S1 = f30Eo U2 = S1. 1) • Since L eatIG (dt. The first step in the derivation is to note that µ = k (0) = r0 (00) .3 applies and yields 1061 U61 Stdlu2/CZdi {W_1(t)}t>0 t>0 which easily leads to 1 StU2 {W( J(t)1t>0 { u S1 t>o Y'(u.. bl IG(t81.() The idea of the proof is to improve upon this by an O (u1) term (in the following. 9otc0" (0) = 0061 = ul. _ ^(u) = ST . C. tu2 ) i IG (t. 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. . C .C. and Si = QoEoU2 = Q B"'['Yo Eo U3 ].2) . (U.1) IG(x.3) this implies (take u = 1) Ego exp { . Theorem 5. i. (. One has (6.S. for brevity. 0o to.(y) = 0. u) = IG(x/u2.2' where as ususal ry > 0 is the adjustment coefficient for the given risk process.T) 1+u2 (6. Vargo S. means up to o(u1) terms): . u) = euh(a . .u.7(u)/u2} eh(A.
9o T 0 in such as way that C = Sou is fixed. u is Eeazead2/++ Eeaz[1 + ab2/u] where the last expression coincides with the r. the formal Laplace transform inversion is heuristic: an additional argument would be required to infer that the remainder term in (6.v.exp { h(A. In ( 1) and (2).6.1 As u + oo.'yu /2)(1 + b2/u)} + Aug 1I J .3.s.2).v. To arrive at (6. 6 .4. . (6. The initial reserve u has been selected such that the infinite horizon ruin probability b(u) is 10% in (1) and (3).z .1 + 629. of a (defective) r.7. p = 0. it holds for any fixed A > 0 that Ego exp { Ab1rr(u)/u2} . . A numerical illustration is given in Fig.2 ). 1% in (2) and (4). 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. we get by formal Laplace transform inversion that C 2 u. .2) is indeed o(u1). But the Laplace transform of such a r. is the c. . that the saddlepoint approximation of BarndorffNielsen & Schmidli [59] is a serious competitor and is in fact preferable if 77 is large] . in (3) and (4). CORRECTED DIFFUSION APPROXIMATIONS 123 Proposition 6.s.52/u where Z has distribution IG (•. distributed as Z . The solid line represents the exact value . which is based upon exponential claims with mean µB = 1. just replace t by Tb1/u2.f. Note. we have p =. calculated using numerical integration and Proposition 1. 1. however .3 = 0. of (6.ry2 . that whereas the proof of Proposition 6. bl I IG I t +2 . the r.5) according to (6.d.h. however.h.5) Once this is established .1 below is exact.1 + u2 I Indeed. and the dotted line the corrected diffusion approximation (6.3).
the fit at p = 0.1 W IU.2 e.u2 2u3 (e . (Inc 0s 0. The proof of Proposition 6.0 0.08 a.07 0.05{ 0.01 0.4 may not be outstanding but nevertheless.OOIi O.1 It is seen that the numerical fit is extraordinary for p = 0.1 proceeds in several steps.19)2 11 20 20 i0 T 1n0 Figure 6.7 or at values of Vi(u) like 1% is unsatisfying.011 L1 60 T IM 11.111 W(U.(061 0.() Lemma 6. .199 0. see Asmussen [12]. A51 7(SAT 3 3 h(X..T) 111 0. it gives the right order of magnitude and the ordinary diffusion approximation hopelessly fails for this value of p.7.114 0. OM 0. BarndorffNielsen & Schmidli [59] and Asmussen & Hojgaard [34]. For further numerical illustrations.T1 00.08 0.aa1 .124 0. and all of the numerical studies the author knows of indicate that its fit at p = 0.00 0. Similarly.W21 0. PROBABILITY OF RUIN IN FINITE TIME 0.EB 0 p ex p ( 7 S h ^)u ..T) 0. Note that the ordinary diffusion approximation requires p to be close to 1 and '0 (u) to be not too small.02 I 90 120 160 2W A0 Z WT 40 80 120 160 100 240 280 T 111 WI.TI CHAPTER IV.^) .
+ h (A. the formulas Po(C(0) > x) Po(C(co) > x) imply 1 °° Po(ST(o) > x) = EIU fIP0 (U>y)dy .co (e) .(3)Eea LauT exp i 3J .T (co (8/u) . (6.7) 2 2 . in Lemma 6. C) 1 1 + u2/ 111 + 2u CZ Z  (2A + ()1/2 J 1 Proof It follows by a suitable variant of Stam's lemma (Proposition 4.6.3 EoU2 + 103OoEoU3 + " 2 6 Using d2 ..1) h(A. 1 = PB(T < oo) = Eo0 exp 125 {(B .4) that the r.C)C/u . () 62 Eeo exp u u2 J . 3 lim Eof (u) = EoC(oo) = a2 Ep = 3EoU2 uroo Proof By partial integration .00)(u +C)  'r (.2u (B3 .4 Ea. + a1b2 + . 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 .2 behaves like C l Eeo eXp r _ ^81T 1 Sl u2 1 u 2u3 [1+h(AC) S .C2 = 2). the result follows.6) u U3 Lemma 6 .61a2T (B3 .s. CORRECTED DIFFUSION APPROXIMATIONS Proof For a>0.h.(3) J t _ aa1T l + eh(A.. (6.co ((/u)) } Let 8 = (2a + (2)1/2 = h(). () + C and note that 2 KO (0) = 102. exp ue } al 1J 3 exP I [2).r0 (00)) } Replacing B by 8/u and Bo by C/u yields e(B() = E eo exp { (e ...
we get h(A.e h(aS)h (^^ 262 exp {_h(. [2+ (2 . Thus by Taylor expansion around ( = 90u.() I 1 + u2 ) y . l Lemma 6 . and inserting this and 9o 2 = S/u on the r. yu/2). Thus a2 y = 290 + O (u2).(2A + ()1/21 exp S h(A..7) and using eh(a. 2 and (6. () .\ + () 1 2 / .x. 5 exp { _h(A) (1 + / y u J)) exp 1.h.2. 0 The last step is to replace h(A.1 (y/2 + Oo)u . yields +90 62 0 + O(u 3) 2u2 +O(u 3).S) d e 62 .h (A. we get the correct asymptotic exponential decay parameter ^/ in the approximation ( 6.\+ (2 (3 e 2u [ (2.126 CHAPTER IV.4.s.2 (^/2 + 3y9o + 390) + O(u3). letting formally T * oo yields 7/)(u) C'e7u where C' = e7a2). yu/2) h(A. 2 + 00 = .6) and 7co (Oo) = ico('y + Bo) to get 0 = 21 (^/2 + 2y90) + 1112 (_Y3 + 3_Y200 + 3y9o) + O(u4). C) ( 1+ u2 The result follows by combining Lemma 6 . PROBABILITY OF RUIN IN FINITE TIME The last term is approximately (e 3 (3) 27.2) for O(u) (indeed . There are two reasons for this : in this way. yu/2) 11+ 62 I} S 1 \\\ u/11 l 62 (3 2u 2A Proof Use first (6. and the correction terms which need to be added cancels conveniently with some of the more complicated expressions in Lemma 6. () by h(\.() .2u [2A+ (2 3 .6  d h(A.
4. that is. the analogous analysis of finite horizon ruin probabilities O(u. Hogan [200] considered a variant of the corrected diffusion approximation which does not require exponential moments.5 in Lemma 6.T) has not been carried out and seems nontrivial. The corrected diffusion approximation was extended to the renewal model in Asmussen & Hojgaard [34].(i+ 62 exP{ h(A. () (i+a ) 2A + (2 . the 'typical' value (say in sense of the conditional mean) was umL.7. with the translation to risk processes being carried out by the author [12]. The answer is similar: the process behaved as if it changed its whole distribution to FL. 7 How does ruin occur? We saw in Section 4 that given that ruin occurs. 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. this case is in part simpler than the general random walk case because the ladder height distribution G+ can be found explicitly (as pBo) which avoids the numerical integration involving characteristic functions which was used in [345] to determine the constants. the approach to the finite horizon case is in part different and uses local central limit theorems. ()} . 'yu/2) 127 ( i+ M pz^ exP { h (A. Fuh [148] considers the closely related case of discrete time Markov additive processes. The adaptation to risk theory has not been carried out. ()} 3 h (A. HOW DOES RUIN OCCUR? exp { h (x. () I 1 + u2 )I 2u L 2A+C2_(2 exp { _h.e. u Notes and references Corrected diffusion approximations were introduced by Siegmund [345] in a discrete random walk setting. the same as for the unconditional Lundberg process. () I 1 + u 2 ) } S 1 . .1: Just insert Lemma 6. In Siegmund's book [346].1 (y/2 + Oo)u )} 1 (i + U ) [2+ C2 2u 62 S Pt^ exP { J 62(2 exp { h(A. i. We shall now generalize this question by asking what a sample path of the risk process looks like given it leads to ruin. and to the Markovmodulated model of Chapter VI in Asmussen [16].
^(u) is exponential with rate 8 w.. In fact. {ST(u)+t .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.r.128 CHAPTER IV. F(u)c] ti e' ru]PL (F(u)`) > 0. . Proof P(u) (F(u)c) = F(flu)c.vi(u) Ce'Yu Corollary 7.FT(u) = o' (T(u ). In the exponential case. the numerator becomes e'ruELe7^ (u)PL(F( u)t) = e7uCFL (F(u)°) when F(u) E .t. Then also P(u)(F(u)) + 1.EL[e7S.t.T. FL As example. + TMOO ).r.1 Let {F(u)}u>0 be any family of events with F(u) E F.3L and the claim size distribution from B to BL. Recall that . . u * oo.(u)_ and ^(u) are independent .ST(u)}t> o is just an independent copy of {St}t>o). and let M(u) be the index of the claim leading to ruin (thus T(u) = Ti + T2 + . 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.F. F(u)c] P(r(u) < oo) ?P(U) < EL[e7u. r(u) < oo) . Note that basically the difference between FT(u) and . (u) and satifying PL(F(u)) * 1.2 If B is exponential. then P(u) and FL coincide on .F.1. so in the in the proof.TT(u) _measurable. P(u) and rate = aL w.(u)_ and similarly the denominator is exactly equal to Ce7u..3 to . . Recall that 13L = (3B[ry] and BL(dx) = e'rxB(dx)/B[7]. we give a typical application of Theorem 7.(u) is not . Theorem 7 . {St}0< t<T(u)) Proof Write e'rsr(u ) = e'rue'r£(u). stating roughly that under F(u).(u)_ is that i. the Poisson rate changes from . PROBABILITY OF RUIN IN FINITE TIME changed its arrival rate from 0 to /3L and its claim size distribution from B to BL.(u).
129 M(u) >2 I(Tk < x) M(tu) p(u) M(u) >2 I(Uk < x) BL(x).3. HOW DOES RUIN OCCUR? Corollary 7. A somewhat similar study was carried out in the queueing setting by Anantharam [6]. the subject treated in this section leads into the area of large deviations theory. the queueing results are of a somewhat different type because of the presence of reflection at 0. however. take I(Tk < x) . see further XI.7. This is currently a very active area of research.eALx) M(u) k=1 u The proof of the second is similar. Proof For the first assertion.3 M(u) pcu) 1 . . From a mathematical point of view.(1 . Notes and references The results of the present section are part of a more general study carried out by the author [11].eaLx. who also treated the heavytailed case.
This page is intentionally left blank .
see A... {St} is the claim surplus process given by I. with common distribution B.i.. r(u) the time to ruin. Var(St) = 11Ba2A + I4AaB lim t goo t PA 131 .1 Define p = !µ.Then no matter the distribution Al of T1i B..1). the one corresponding to the stationary case by 00)(u). Thus the premium rate is 1. and M is the maximum of {St}. The ruin probability corresponding to the zerodelayed case is denoted by 1/'(u). the claim sizes U1.1. Then the arrival process is stationary which could be a reasonable assumption in many cases (for these and further basic facts from renewal theory. We use much of the same notation as in Chapter I.Q.7). . A different important possibility is Al to be the stationary delay distribution A° with density A(x)/µA. U2. Proposition 1. the Tn are independent. .. D'2. of the risk process form a renewal process: letting Tn = Qn .1 (T1 = a1). with the same distribution A (say) for T2.. .. In the socalled zerodelayed case.d. and the one corresponding to T1 = s by 1/i8 (u).Chapter V Renewal arrivals 1 Introduction The basic assumption of this chapter states that the arrival epochs O'1.(1. AA t*oo lim St = lim ESt t tioo t = p . are i. with Nt = # {n: Un <t} the number of arrivals before t. . T3. the distribution Al of T1 is A as well.
2 (DETERMINISTIC ARRIVALS) If A is degenerate. one could imagine that the claims are recorded only at discrete epochs (say each week or month) and thus each U. and (1 . after E.St] = a(p . we get similarly by using known facts about ENt and Var Nt that Nt Var(St) = Var E Nt U. s + t µA PA 0 Of course. Here are two special cases of the renewal model with a similar direct interpretation: Example 1. such that no arrivals occur in the off state.1). From this ( 1. the . Nt = Var(PBNt) + E(4Nt) Q2 2 0`2 A tpB B + o(t). but the arrival rate in the ON state is .1) ENt/t + 1/µA. stating that E[Nt+a .Nt] * a/PA. t 4oo Proof Obviously. CHAPTER V.1 gives the desired interpretation of the constant p as the expected claims per unit time. by the elementary renewal theorem (cf. 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.3 (SWITCHED POISSON ARRIVALS) Assume that the process has a random environment with two states ON. the definition 77 = 1/p .a is really the accumulated claims over a period u of length a. The simplest case is of course the Poisson case where A and Al are both exponential with rate 0. However . This has a direct physical interpretation (a large portfolio with claims arising with small rates and independently). say at a. Thus. The renewal model is often referedd to as the Sparre Andersen process. For (1 . 3) follows similarly by Blackwell 's renewal theorem.132 Furthermore for any a > 0. RENEWAL ARRIVALS lim E [St+a . Proposition 1. Nt + EVar U. A.0 > 0. OFF.t. Example 1 . 2). Nt ESt = E E UI Nt t = ENt•pB . If the environment is Markovian with transition rate A from on to off and u from OFF to ON.1) follows .
. S o<t<oo n=0.y)B(dy). Ao..4) w. and for historical reasons.r.4) with phase space {oN. (an arrival occurs necessarily in the ON state. A is phasetype (Example 1. the first term represents the probability F(U1 .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. and the present author agrees to a large extent to this criticism.1..4) fo Indeed.i. The values of the claim surplus process just after claims has the same distri bution as {Snd^ }• Since the claim surplus process {St} decreases in between max St = max ^d^. Therefore.d. u For later use. integrate (1.oFF}. 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 . (1.} with {S(d)} a discrete time random walk with increments distributed as the independent difference U .1. and then the whole process repeats itself).s < u). arrival times... we have From this the result immediately follows. we feel it reasonable to present at least some basic features of the model.. the fundamental connections to the theory of queues and random walks. U1 .s > u) of ruin at the time s of the first claim whereas the second is P(r(u) < oo. initial vector (1 0) and phase generator 11 However. The following representation of the ruin probability will be a basic vehicle for studying the ruin probabilities: Proposition 1. the relevance of the model has been questioned repeatedly.s. However.t. More precisely. if for nothing else then for the mathematical elegance of the subject. .2.1.T between a claim U and an interarrival time T. Proof The essence of the argument is that ruin can only occur at claim times. 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 . For the stationary case. in general the mechanism generating a renewal arrival process appears much harder to understand. INTRODUCTION 133 interarrival times become i.
That is . St = t .3* (say ) and the U.Ut.1 If. b=1 !=1 where {Nt } is a Poisson process with rate . i. A typical sample path of {Rt } is illustrated in Fig. with common distribution B* (say) concentrated on (0. 2. If . < 1. the claim surplus process are given by Nt Nt Rt = u+^U.a*PB• > 1. t. At the time of death . are independent of {Nt} and i. then 0 * (u) = 1 for all u > 0. The compound Poisson model with negative claims We first consider a variant of the compound Poisson model obtained essentially by signreversion . then 0*(u) = e 'r" where ry > 0 is the unique solution of 0 = k*(ry) = *(B*[ry] . Theorem 2 .d. 00). The initial reserve is obtained by prepayments from the policy holders. A simple sample path comparison will then provide us with the ruin probabilities for the renewal model with exponential claim size distribution.0* (u) = P (rr* (u) < oo) where rr* (u) = inf It > 0: Rt < 0} ) .1. the claims and the premium rate are negative so that the risk reserve process . each of which receive a payment at constant rate during the lifetime .1 r* (u) One situation where this model is argued to be relevant is life annuities. (2. we shall be able to compute the ruin probabilities i(i* (u) for this model very quickly (.1) . U Figure 2. the remaining part of the prepayment (if any ) is made available to the company.1) +ry.134 CHAPTER V. RENEWAL ARRIVALS 2 Exponential claims.3* pB. resp . Using Lundberg conjugation .
s.* (a) = log Ee'st I.2 sup St = inf St = 00 t>o t>o and hence 0* (u) = 1 follows. Since ic'(0) < 0. cf.2(b).(a) 7 Figure 2.. T_ ( u) < 001 e7"P(T_ (u) < oo) = e"V)* (u). B.g.1.UB. St=Rtu=St. Then the c. B*. the safety loading of { St} is > 0. EXPONENTIAL OR NEGATIVE CLAIMS [Note that r. B* [7] and let {St} be a compound Poisson risk process with parameters .2(a).(u ) < oo) = E {e7sr_ (u). If I3*pB* < 1. and thus 1 = P(T. Proof Define 135 St =u . 2. of {St} is c(a) = is*(a7). 0 Now return to the renewal model. Then { St } is the claim surplus process of a standard compound Poisson risk process with parameters 0 *. then by Proposition 111. Fig. Hence T_(u) < oo a.2. > 1 .0.3*. and the Lundberg conjugate of {St} is { St } and vice versa. 0) and has typically the shape on Fig. Define T_ (u) = inf It > 0 : St = u} . T_ (u) = inf { t > 0 : St = u 'r* (u). Let B(dx) = ^e7x B*(dx). Hence y exists and is unique. Then the function k* is defined on the whole of (oo. .Rt.2 Assume now .f. 2. (a) is*(a) (b) .
Ui ..136 CHAPTER V. the distribution of M(d) is a mixture of an atom at zero and an exponential distribution with rate parameter ry with weights 1 .. However. Then B* = A.a) = 1 . and (2 .1 it is seen that ruin is equivalent to one of these values being > u. and 5PA > 1...1.4. 3* = 6. A variant of the last part of the proof. with rate S (say). with the probability that a particular jump time is not followed by any later maximum values being 1 . Taking m.1 means that M* is exponentially distributed with rate ry.•... respectively.1) + ry = 0 which is easily seen to be the same as (2.2). which has the advantage of avoiding transforms and leading up to the basic ideas of the study of the phasetype case in VIII.Y a I.Ti = U1. T2 = U2.• • • .4 goes as follows: define 7r+ = P(M(d) > 0) and consider {St*} only when the process is at a maximum value. Now the value of {St*} just before the nth claim is To +T1* +. Hence M* max {To + Ti + • • • + Tn . RENEWAL ARRIVALS Theorem 2 .1..g. 1) means that 8(A[ry] .Tr+. alternatively termination occurs at a jump time (having rate 8). According to Theorem 2. To + max {Ul+•••+UnTI.u+ and lr+. and hence the failure rate . 2..7r+ 7r EeTo b/(Sa) + +.2 If B is exponential..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 .1..)(u) _ 1r+e7" where ry > 0 is the unique solution of 1 = Ee'Y(uT ) = S 8 A[.. u Hence P(M(d) > u) _ 1r+e'r".e.Tn} n=0.+Tn U1 Un..'s and noting that V)*(u) = P(M* > u) so that Theorem 2. the failure rate of this process is y.f. we get Ee'M(d) = Ee°M* _ Y/(. and from Fig .Un } = max St = t>0 n=0.Y] (2. then ... To + M(d) in the notation of Proposition 1.
B(dx). Hence the failure rate of M(') is 6(1 .3 A[a] B[a] F( d) [a +)3] F(d) [a] = Fad) [^] Letting M(u) = inf in = 1. a ladder step is the overshoot of a claim size. which states that for a given a.y/b.4.... the relevant exponential change of measure corresponds to changing the distribution F(d) of Y = U . 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^. Putting this equal to y. Furthermore. we have ] A[a )3] E«d'efl' = Bad> [a] A ad> [Q] = B[a +. This follow since. Bads (dx) = .7r+) = ry and hence P(M(d) > u) = P(M(d) > 0)e7u = 7r+e'r". consider instead the failure rate of M(d) and decompose M(d) into ladder steps as in II.2. we see that ry = 6(1.6.7r+). 3a The imbedded random walk The key steps have already been carried out in Corollary 11. The probability that the first ladder step is finite is 7r+. B^d) where Aad> (dt) = ^[ a] A(dt). : S(d) > u} . 111.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] . However. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 137 is b(1. 0 3 Change of measure via exponential families We shall discuss two points of view. letting P(d) refer to the renewal risk model with these changed parameters . resp.2.7r+) and hence r+ = 1. Thus a ladder step terminates at rate b and is followed by one more with probability 7r+..5. hence exponential with rate b.3. the imbedded discrete time random walk and Markov additive processes.
r. 187) and thereby for ^(u) to be nonlattice w. in the easiest nonexponential case where B is phasetype.t. In fact. This is known to be sufficient for ^(O) ]p (d) ([APQ] Proposition 3. (b) V)(u) . we get: Proposition 3. VIII. and claim (a) follows immediately from this and e (u) > 0. to converge in distribution since p(yd) (r(0) < oo) = 1 because of r (d)' (y) > 0. just note that F7(d) is nonlattice when F is so .1) is explicit given 7..1 For any a such that k(d)' (a) > 0. Corollary 3. ik. i .C(°)eryu where C(O) = C0[7] .1 implies Cu) = e«uE ( 7d)e«^(u) ...2 p.1).p)/($B'[7] . let 7 > 0 be the solution of r. For claim (b). (d) (7) _ 0. 00)(u) . cf. Proof Proposition 3. (a) '(u) < eryu. the evaluation of C is at the same level of difficulty as the evaluation of i/i(u) in matrixexponential form. O(u) = eauE (d)ea{ (u)+M(u)K (d)(a) .Ce"u where C = limu.C8e7u where Cs = Ce78B[7].(u) . E(d)e 1' (u).138 CHAPTER V. provided the distribution F of 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 . For the stationary case. Consider now the Lundberg case.e.u the overshoot . RENEWAL ARRIVALS be the number of claims leading to ruin and ^(u) u = SM(u) . We have the following versions of Lundberg' s inequality and the CramerLundberg approximation: Theorem 3 .4.2 In the zerodelayed case. 7µA .T is nonlattice.3 For the delayed case Tl = s.
According to Remark 11. we invoke the behavior at the 1 = h«(0. 3b Markov additive processes We take the Markov additive point of view of II. The expressions are slightly more complicated and we omit the details.(s. 0 0 .dt ) eadt = h ( s) .Sdt] = Ee'uh(T) means 1 = f ' e°^B(dy) f ' h( s)A(ds). (u + s .1) = C(O).4).1) (normalizing by h(0) = 1). E8h0 (Jdt. To determine boundary 0.dt(ah ( s) + h'(s)) so that Gha ( s. h(s) = e(a +x( a))8 (3..9. Here K.5./c. we get r u +8 e"8(u) 139 e7uB(u + s) + 4 0 + L 00 J e7(v8)e7(u+8v). IPA 0 Of course.h'(s). Let P8f E8 refer to the case Jo = s. 0) = ah (s) . For s > 0. 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] . Equating this to tch(s) and dividing by h(s) yields h'(s)/h(s) _ .0 ) = Eo[ha ( Jdt.5. 0) = tc(a)h(s). (s.(°) ( Ce8B[7] Ao(ds) similar manner. The underlying Markov process {Jt} for the Markov additive process {Xt} = {(Jt. delayed version of Lundberg's inequality can be obtained in a e7u.a .y) B(dy) 0 For the stationary case. where G is the infinitesimal generator of {Xt} = {(Jt.St)} can be defined by taking Jt as the residual time until the next arrival. Sdt) = h(s .3. St)} and h. we look for a function h(s) and a k (both depending on a) such that Gh. B(x) = o(e7x) and dominated convergence.. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES Proof Using (1. y) = e°yh(s).
. 5 For the compound Poisson case where A is exponential with rate .2) As in 11. [a + /3] A[b . resp. Further.rc(a)] = 1. ] = E. (3. An important exception is.140 CHAPTER V. T2.5. J n+1 u are independent with distribution Aa for the Tk and Ba for the Uk.2).S„):0<v< )be Lt = eaSt tK(a) h(Jt) = east .c(a)] B[a] Proof Pa.8.s(Jo = s) = 1 follows trivially from Lo = 1.1)a in agreement u with Chapter III. An easy extension of the argument shows that U1. T2 are independent with distributions Ba. .s and P(d). .13]A[b . .rc(a)] B[a] A[a . [e1U1 + 6T2ea ( U1s)stc ( a)e(a+K(a ))( T2s)I B[a +.tK( a)e.a .c(a)] which shows that U1. U. = J8 = T2.rc(a)] = B = Ba[13]Aa[5]...4 The probability measure Pa. however. .e. resp.s is the probability measure governing a renewal risk process with Jo = s and the interarrival distribution A and the service time distribution B changed to Aa.e(«+k(a))t esy A(dt). . A[a .. Ease AU1+6T2 [ AU1+6T2 = Ea a LT.. i. (3. RENEWAL ARRIVALS B[a]A[a . Aa as asserted . rc(a) = 0 (B[a] . since JT..s governing {(Jt. . Remark 3 . Ba where Aa (dt) .( a+r' (a))(Jt s) h(s) where c(a) is the solution of (3.2) means 1 = B[a]/3/(/3+a+rc (a)). we can now for each a define a new probability measure Pa. the determination of the adjustment coefficient ry where the defining equations rc(d) (ry) = 0 and rc(ry) = 0 are the same. Ba(dx) = B(dx). Proposition 3.a . Note that the changed distributions of A and B are in general not the same for Pa. St)}too by letting the likelihood ratio Lt restricted to Yt = a((J.
that is. Label the customers 1. Then J(rr(u)) = TM(u)+1 and hence Ws(u. The claim for the zerodelayed case follows by integration w. Then "^ e(ay+w(aY))8 Ys(u.4. and define yy = ay . the amount of .4. that is. the time from he arrives to the queue till he starts service. not after time T. defined as the single server queue with first in first out (FIFO. Let M(u) be the number of claims leading to ruin .6 Let y < let ay > 0 be the solution of ic'(ay) = 1/y. u The approach via the embedded random walk is standard. XII.yx(ay).r. and assume that T„ is the time between the arrivals of customers n .rc( ay)] = e(aa+(r ))sb[a ]e7yu L y1 In particular.t. .5.ay+ray))TM(.g. and U„ the service time of customer n. Using the Markov additive approach yields for example the following analogue of Theorem IV. The actual waiting time Wn of customer n is defined as his time spent in queue (excluding the service time).. . Notes and references 4 The duality with queueing theory We first review some basic facts about the GI/G/1 queue. [APQ] Ch. which is the same as the asserted inequality for 0.4. . A(ds). For the approach via Markov additive processes. or FCFS = first come first served) queueing discipline and renewal interarrival times.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. 2. yu) = F'ay. it is easily seen that ic(ay ) > 0.(u. T(u) < yu h(JT(u)) < eayu+yuk(ay ) ( Eia y Le(a(+k(ay))s v. yu).s eaysr(")+r(u ) K(ay) h (s) . yu ) < e7yu..1 and n. 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.5 Proposition 3. The virtual waiting time Vt at time t is the residual amount of work at time t. see in particular Dassios & Embrechts [98] and Asmussen & Rubinstein [45]. Proof As in the proof of Theorem IV. see e.)+1 e J j e(ay+w(ay))8 e .yu ) e7vu A[ay . for the zerodelayed case zp8(u.
and obviously z/'(u) = limN.2) (b) as t * oo. and we have P(V > u) = ?/iiol(u). equivalently.Tn)+ Proof The amount of residual work just before customer n arrives is VQ„ ."^ Vi(N) (u).3. (4. . Thus Vos}1 _ = (Wn + Un . Then P(r(u) < T) is the probability z/iiNi (u) of ruin after at most N claims. Thus.Tn)+. but we shall present a slightly different proof via the duality result given in Theorem II.. 0 Applying Theorem 11. we get: Corollary 4.1 and Corollary 11. 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 11.1.2 Let Mnd) = maxk=o.• • • Tk ). Vt converges in distribution to a random variable V.3) Proof Part (a) is contained in Theorem 11. 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. It then jumps to VQ„ . p < 1. Let the T there be the random time UN. Then: (a) as n + oo. the proposition follows.1. we have Wn = Van(left limit). (4. an+1) = [on.3 Assume rl > 0 or.+ Un. whereas in [On . (u). Also {Zt}o<t<T evolves like the leftcontinuous version of the virtual waiting time process up to just before the Nth arrival.1 Wn+1 = (Wn + U. (4. and we have P(W > u) = V. equivalently.n1 (U1 +• • •+Uk Tl .4): Proposition 4. but interchanging the set ..4..4. and combining with (4. the waiting time a customer would have if he arrived at time t.2. The traffic intensity of the queue is p = EU/ET. Wn converges i n distribution to a random variable W. in which case {V} remains at zero until time on+1.4: Proposition 4. then Wn v M. since customer n arrives at time on.4..142 CHAPTER V.1). on + Tn) the residual work decreases linearly until possibly zero is hit. If W1 = 0.1) The following result shows that {Wn} is a Lindley process in the sense of II.
{ 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). For part (b).2).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. It follows that P(WN > u) =.T*)+. Then the arrivals of {Rt} in [0. However. resp . as WN.T* < x) fK(x_y)F(dy) (x > 0 is crucial for the second equality!). we let T be deterministic .Ao in (b).4.. convergence in distribution hold for arbitrary initial conditions . and we get: ... and hence in particular ZT is distributed as the virtual waiting time just before the Nth arrival. Then K(x) = J x00K(x .T* = y yields K(x) = P ((W + U* . THE DUALITY WITH QUEUEING THEORY 143 (T1. u Now return to the Poisson case . i. (4. A. Letting n oo in Corollary 4. T1.e.... (4. Hence for x > 0. hence (since the residual lifetime at 0 and the age at T have the same distribution . Then the corresponding queue is M/G/1. Corollary 4. x > 0. we obtain: Corollary 4..le) the same is true for the timereversed point process which is the interarrival process for { Zt}o<t < T• Thus as before .5 (LINDLEY'S INTEGRAL EQUATION) Let F(x) = P(U1T1 < x). conditioning upon U* . In fact .(N)(u) has the limit tp(u) for all u. where U*.4 The steadystate actual waiting time W has the same distribution as M(d).oo in Proposition 4.y)F(dy). T] form a stationary renewal process with interarrival distribution A.. TN) with (TN. T1 .2. we get W = (W + U* . but this requires some additional arguments (involving regeneration at 0 but not difficult) that we omit.. namely W1 = 0 in (a) and Vo = 0.5) Proof Letting n . Ti) and similarly for the U.. by an obvious reversibility argument this does not affect the distribution . K(x) = P(W < x). cf.1.T* are independent and distributed as U1. which implies the convergence in distribution and (4.T*)+ < x) = P(W + U* .
the zerodelayed and the stationary renewal processes are identical. the actual and the virtual waiting time have the same distribution in the steady state. VIII).6 For the M/G/1 queue with p < 1. . RENEWAL ARRIVALS Corollary 4.5) to hold for all x E R and not just x > 0). 0 Notes and references The GI/G/1 queue is a favourite of almost any queueing book (see e . Proof For the Poisson case.g.5) is in fact a homogeneous WienerHopf equation. Hence '(u) = Ali(°)(u).g. Some early classical papers are Smith [350] and Lindley [246]. Asmussen [24] and references there.5) looks like the convolution equation K = F * K but is not the same (one would need (4. The equation (4.144 CHAPTER V. Note that (4. That is. W v V. see e. Cohen [88] or [APQ] Ch. 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. implying P(W > u) = P(V > u) for all u.
t St = E Ui . and can be computed as the positive solution of WA = 0. As in Chapter I. here it exists whenever A is irreducible which is assumed throughout. N. • The premium rate when Jt = i is pi. dv. M = supt>o St. The intensity matrix governing {Jt} is denoted by A = (A. • Claims arriving when Jt = i have distribution Bi.)iJEE and its stationary limiting distribution by lr. .f pi. i=1 0 and r(u) = inf It > 0: St > u}. Ire = 1. {Jt} describes the environmental conditions for the risk process. {St} denotes the claim surplus process.. Thus.(3i when Jt = i.T) = Pi (T(u) < T). The ruin probabilities with initial environment i are '+ki(u) = pi(T(u ) < oo) = Pi (M > u). 145 Oj( u.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 .
this is no restriction when studying infinite horizon ruin probabilities. and p is the overall average amount of claims per unit time.2. T(=)). and assume that weather conditions play a major role for the occurence of accidents.T(n)). we shall assume that pi = 1.5 below.a('). MARKOVIAN ENVIRONMENT where as usual Pi refers to the case Jo = i. we could distinguish between normal and icy road conditions. one expects that 3i > on and presumably also that Bn # Bi. which is clearly unrealistic. respectively. Example 1 .4) with representation (E(i). a(i). t(i) = T(')e are the exit rates. and we have f3. u .1) iEE Then pi is the average amount of claims received per unit time when the environment is in state i.. Bi. with rates Ani and Ain.1 implicitly assumes that the sojourn times of the environment in the normal and the icy states are exponential. r^ = P (1. For example. An example of how such a mechanism could be relevant in risk theory follows. i and corresponding arrival intensities Qn.2 (ALTERNATING RENEWAL ENVIRONMENT) The model of Example 1. cf. leading to E having two states n. Proposition 1.1 Consider car insurance. Example 1. say. assume that the sojourn time in the icy state has a more general distribution A(i). say.14. cf. the intensity matrix is A OW) T(i) T(n) t(n)a(i) where t(n) = T(n)e. Thus. P = E 7riPi. meaning that accidents occuring during icy road conditions lead to claim amounts which are different from the normal ones. Unless otherwise stated. We let p Pi = /ji/AB. the operational time argument given in Example 1.11 below. /3 = Nn when j E E(n). we can approximate A(i) with a phasetype distribution (cf.146 CHAPTER VI. Then the state space for the environment is the disjoint union of E(n) and E(i). in blockpartitioned form. = iii when j E E(i). Assume similarly that the sojourn time in the normal state has distribution A(n) which we approximate with a phasetype distribution with representation (E('). f3i and claim size distributions Bn. 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. According to Theorem A5.
3. and . where W = (w.1. .. the parameters are ^ij = aid/pi. u From now on. One way to model this would be to take A(') to be Coxian (cf. t(n) = T("i)e... Approximating each A('?) by a phasetype distribution with representation (E('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) . This amounts to a family (A(")) ?CH Of sojourn time distributions. i E E(n) }. i8f n1. resp.T(n)). say it is larger initially. i ) : n E H.3 Consider again the alternating renewal model for car insurance in Example 1. ..j = .5 (MARKOVMODULATED PREMIUMS) Returning for a short while to the case of general premium rates pi depending on the environment i.1. n8}. the state space E for the environment is { ('q.. and similarly for the normal period.a(n).tEH is a transition matrix.2.>.n. and 1/ii(u) = t/ii(u). we assume again pi = 1 so that the claim surplus is Nt St = ?Ui_t. (9) where q = CHI.. but assume now that the arrival intensity changes during the icy period. it = Jel(t)....3i/pi. Indeed. A('^). one could for example have H = {i1. is the probability that a long icy period is followed by a short normal one. say. St = SB=(t). MODEL AND EXAMPLES 147 Example 1 .Q. u Example 1 . Then for example wi. iq (visited in that order) and letfOil >. In the car insurance example.p.. The simplest model for the arrival intensity amounts to .4) with states i1.. 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. depending only on 77.. T(9) +wggt(9)0. Example VIII. 1 . u Example 1.3i.J017. let T 9(T) = f pi. 0 Then (by standard operational time arguments) {St } is a risk process in a Markovian environment with unit premium rate. such that the icy period is of two types (long and short) each with their sojourn time distribution A('L). w. such that a sojourn time of type rt is followed by one of type c w. Qi = . dt.
MARKOVIAN ENVIRONMENT We now turn to some more mathematically oriented basic discussion. 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.e(A(Oi)d'sg)xe.(3iBi(dx).8 As t oo. . )3*. Next we note a semiMarkov structure of the arrival process: 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).(Qi)diag)• More precisely. In particular. . vi(dx) = . the empirical distribution of the claims is B*. More precisely. o = 0. JT1 = j) = Qj • e. Nt Nt a . Pi (Ti E dx. The key property for much of the analysis presented below is the following immediate observation: Proposition 1.5. t l=1 Note that the last statement of the proposition just means that in the limit.148 CHAPTER VI. dx. N > 1(Ul < x) a4 B*(x). B* = 1 /^* Bi.6 The claim surplus process {St} of a risk process in a Markovian environment is a Markov additive process corresponding to the parameters µi = pi.7 The Pidistribution of T1 is phasetype with representation (ei. one can associate in a natural way a standard Poisson one by averaging over the environment. iEE iEE )3 These parameters are the ones which the statistician would estimate if he ignored the presence of Markovmodulation: Proposition 1.A . the Markov additive structure will be used for exponential change of measure and thereby versions of Lundberg's inequality and the CramerLundberg approximation. qij = 0 in the notation of Chapter 11. we put )3* = E 7fi/3i. Proof The result immediately follows by noting that T1 is obtained as the lifelength of {Jt} killed at the time of the first arrival and that the exit rate obviu ously is f3j in state j.
In particular. and furthermore in the limit JT. e. B*. we may view Nt`i as the number of events in a Poisson process where the accumulated intensity at time t is Niti..7. i. By Proposition A5. A.1. Bi.x) = Nt E > I (Uk) X) Nt Bi(x) 1=1 iEE k=1 iEE 1: t5 Bi( x) = B*(x). Proof According to Proposition 1. 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. aA. In particular.. However . Nt a' t t iEE Also. zli( (u) . Then it is standard that ti lt '4' iri as t > oo.. Conditioning . and let {St °i} refer to the one with parameters Pi. given {Jt}0<t<0. y Ni) i Nti) t a. Proposition 1. MODEL AND EXAMPLES 149 Proof Let ti = f1 I(JJ = i) ds be the time spent in state i up to time t and Nti) the number of claim arrivals in state i . Bi(x). the limiting distribution of the first claim size U1 is B*. cf. the Fidistribution of T1 in {St(a ) } is phase type with representation (E. {St} to the compound Poisson model with parameters 0 *. Hence Nt'> a .2... Then {St)} + {St*} in D[0.4. this converges to the exponential distribution with rate 0* as a * oo.aA . Bi. Example 11. ^j 7riNi. oo) as a 4 oo. N + oo Hence 1 Nt 1 N`+) Nits Nt E I ( Ut <. iEE 13 A different interpretation of B* is as the Palm distribution of the claim size. has distribution (7ri)3i //3*)iEE and is independent of Ti.* (u) for all u.6.9 Consider a Markovmodulated risk process {St} with param eters Ni.(/3i)aiag). denoting the sizes of the claims arriving in state i by U(') 1 standard law of large numbers yields U(') the N 1: I(Ukik < x) k=1 N a$.
2}. e. U.. T) + ?P* (u. B1=3E3+2E7... thick. is as in {St }. > 1..1 with periods with positive drift alternating with periods with negative drift.=1. with T2 being exponential with rate . Claims of type E3 arrive with intensity 2 . MARKOVIAN ENVIRONMENT upon FT. Continuing in this manner shows that the limiting distribution of (T. Thus. those of type E7 with intensity z s = 5 in state 1 and with intensity z .g. which also yield O(a) (u.l3* and U2 having distribution B*. marked by thin. resp.. shows similarly that in the limit (T2. and (at least when a is small such that state changes of the environment are infrequent). Computing the parameters of the averaged compound Poisson model. A= (  a a ) \ a a 5 5 J 9 3 2 a1=2. the overall drift is negative since it = (2 2) so that p = 71P1 + 112P2 = 7. 1. lines in the path of {St}.).. there are p = 2 background states of {Jt}.2 +2 2 = 3.. and in fact P1 = 31AB1 = 9 3 1 2 (5 3 3 1 1 2 1 5 7 1 81 70 ' _ 19 4 5 P2 = . The fact that indeed 0(a) (U) 3 0* (u) follows.10 Let E_{1.1. we first get that 3 (3* = 2. since E3 is a more dangerous claim size distribution than E7 (the mean is larger and the tail is heavier). 5 5 where E5 denotes the exponential distribution with intensity parameter 5 and a > 0 is arbitrary. That is. 9 .1 of [145].. from Theorem 3.s = 1o in state 2.31µB 2 = 2 5 3 7 70 Thus in state 1 where p. s 5 in state 2. 0 Example 1. 132=2. we may imagine that we have two types of claims such that the claim size distributions are E3 and E7.2.. the paths of the surplus process will exhibit the type of behaviour in Fig.150 CHAPTER VI. oo). T) for all u and T. 1.s = o in state 1 and with intensity 1 .2. From this the convergence in distribution follows by general facts on weak convergence in D[0.FT. state 1 appears as more dangerous than state 2. U2) are independent of . the company even suffers an average loss. B2=1E3+4E7. On Fig.
1 a. Hence B* = 415E3+5E7/ +4 ( 51E3 +5 E7) = 1E 3 +2E7. 0 The definition (1. Hence (i) Nti) 1 U(i) k' N(i)k=1 E t 4 St + t = iEE Nt t 1: 7ri Qi µs.1. the averaged compound Poisson model is the same as in III. 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. Proof In the notation of Proposition 1.(3. MODEL AND EXAMPLES 151 Figure 1.1. 01 /.3* = 3/4 of the claims occur in state 1 and the remaining fraction 1/4 in state 2. iEE . a fraction r.1 Thus.1) of the safety loading is (as for the renewal model in Chapter V) based upon an asymptotic consideration given by the following result: Proposition 1. = P. For (b).11 (a ) ESt/t * p . t + oo..1).s. (b) St/t * p . note first that EN Uk')/N a4' µgi. t * oo. That is.
The proof of Proposition 1. and so on. X2 =SW2 So. then M = 00 a.. Then by standard Markov process formulas (e..s. Since the X„ are independent .0i(u) < 1 for all i and u.. There seems still to be more to be done in this area. and hence oscillates between 0o and oo so that also here M = oo. The mainstream of the present chapter follows [16]. then M < oo a.s. having the Pidistribution of X.1(b) is essentially the same as the proof of the strong law of large numbers for cumulative processes. limit p . n n Thus {SWn l is a discrete time random walk with mean zero. + Xn SWn ](1 a . and hence 1/ii(u ) = 1 for all i and u. Theorem II. [315]. and involves a version of the .12 If 77 < 0.1)Eiw = 0.1 and the Corollary are standard.g. see the Notes to Section 7. also + . See Meier [258] and Ryden [314]. some early studies are in Janssen & Reinhard [211]. and a more comprehensive treatment in Asmussen [16]. and .4. EiX = 0. Proof The case 77 < 0 is trivial since then the a. If 77 > 0. see [APQ] p. with X2.Eiw o'o Eiw • E ^ifjµs.1 jEE = (p . 136 or A..ld. In risk theory. [302]..Jt=i}. X 1 =Sty.. .a form a renewal process .152 CHAPTER VI.2(a) p. .. [APQ]. MARKOVIAN ENVIRONMENT Corollary 1. Now let r) = 0. Statistical aspects are not treated here. s. dt .\ i and EiX1 Ei f 13 J. PB. The case 77 > 0 is similarly easy. 0 Notes and references The Markovmodulated Poisson process has become very popular in queueing theory during the last decade.. [212]. with some important improvements being obtained in Asmussen [17] in the queueing setting and being implemented numerically in Asmussen & Rolski [43]. Now obviously the w. let some state i be fixed and define w=wl=inf{t >0:Jt_#i. and hence M = 00.. 38) Eiw1 = 1/ir. and hence wn /n a4.1 of St / t is > 0.. 2 The ladder height distribution Our mathematical treatment of the ruin problem follows the model of Chapter III for the simple compound Poisson model. Eiw.Jt=i}. w2=inf {t>w1:Jt_#i. Proposition 1. X3.
00 (2.A) =ZI(St E. j. 6.g. oo)) = f R(i. which represents a nice simplified form of the ladder height distribution G+ when taking certain averages : starting {Jt} stationary.x).6. G+ is the matrix whose ijth element is E G +(i. Let further R denote the preT+ occupation kernel. we get the same ladder height distribution as for the averaged compound Poisson model. B* in Section 1. the definition of . oo) = J ao 0 G+(i. cf. For measurevalued matrices.j.5. j. k. (y.i. T+ < oo) and let G+ be the measurevalued matrix with ijth element G+(i.a/i. Proposition 2. j.. IIG+ II denotes the matrix with ijth element IIG+(i. n=0 (2.x. . T R(i.IIG +II)e.2. e.EA.1) 0 (b) G+ (y. Proposition 2. but is substantially more involved than for the compound Poisson case .6*. •). oo)).(u) = Pi(M < u) = e' E G+ (u)(I . only with the product of real numbers replaced by convolution of measures. j. The form of G+ turns out to be explicit (or at least computable). However . Define the ladder epoch T+ by T+ = inf It : St > 0} = r(0). for i.2(a) below ) where the ladder height distribution is evaluated by a time reversion argument. dx)/jBj(y .A) = Pt(ST+ E A. by specializing results for general stationary risk processes (Theorem II . let G+(i.dx). see also Example II. •)• kEE Also. •) * G +(k.Jr+ =j. and S (dx) the measure valued diagonal matrix with /3 Bj(dx) as ith diagonal element. That is.j. •) II = JG+(i.3*B *(y)dy.j E E.1 irG+(dy)e =.2) R(dx)S((y . Thus.4) we obtain the following result . THE LADDER HEIGHT DISTRIBUTION 153 PollaczeckKhinchine formula (see Proposition 2. we define the convolution operation by the same rule as for multiplication of realvalued matrices.Jt=j)dt. j.2 (a) The distribution of M is given by 00 1 .
and let further {my} be the Evalued process obtained by observing {Jt } only when {St*} is at a minimum value. St < S* for u < t.2 useful . JJ = j.1 The following observation is immediate: Proposition 2. {mx} is a non terminating Markov process on E. we need as in Chapters II. 0  x Figure 2. From this (2. MARKOVIAN ENVIRONMENT Proof The probability that there are n proper ladder steps not exceeding x and (x)ej. mx = j when for some (necessarily unique) t we have St = x. G+ the probability that there are no further ladder steps starting from environment j is e^ ( I . That is.2) is just the same as the proof of Lemma 11. To this end . the intensity matrix A* has ijth element * 7r ^i3 7ri and we have Pi(JT = j) = 7rj P2(JT = i)7ri (2. marked by thin. The u proof of (2. .6.3.3 When q > 0. resp.3) We let {St*} be defined as {St}. hence uniquely specified by its intensity matrix Q (say).IIG+II)e. lines in the path of {St}. thick.1) follows by summing over n and j.154 CHAPTER VI. and that the environment is j at the nth when we start from i is e . III to bring R and G+ on a more explicit form . To make Proposition 2. only with {Jt} replaced by {Jt } (the /3i and Bi are the same ). see Figure 2.1 for an illustration in the case of p = 2 environmental states of {Jt}. we need to invoke the timereversed version {Jt } of {Jt} .
4 Q satisfies the nonlinear matrix equation Q = W(Q) where 0 co(Q) = n* .(/3i)diag + T S(dx) eQx. ( Q( n)) converges monotonically to Q. The definitions are illustrated on Fig.*. 0 mms1   ^O \ T. } is a minimum value at v = t. 2.and a jump (claim arrival) occurs at time t. If there are no jumps in (t. we recursively define the depth of an excursion as 1 plus the maximal depth of a subexcursion. Q( n+l) _ ^. s]. and the excursion is said to have depth 1 if each of these subexcursions have depth 0. corresponding to two subexcursions of depth 0. = x}. Proof The argument relies on an interpretation in terms of excursions.(/3i) diag. Note that the integral in the definition of W(Q) is the matrix whose ith row is the ith row of _ 3 f e2Bi(dx). and the excursion ends at time s = inf {v > t : S.0. Furthermore. Figure 2. THE LADDER HEIGHT DISTRIBUTION 155 Proposition 2.2 .2. and S(dx) is the diagonal matrix with the f3iBi(dx) on the diagonal. An excursion of {St*} above level x starts at time t if St = x.2 where there are three excursions of depth 1. the sequence {Q(n)} A* defined by Q(O) = . In general. Otherwise each jump at a minimum level during the excursion starts a subexcursion.. For example the excursion of depth 2 has one subexcursion which is of depth 1. {S. we say that the excursion has depth 0.2.
(2. By considering minimum values within the excursion. the subintensity matrix of {min+i ) } is cp (Q(n)) = Q(n +l) which implies that qgj +1) = \!. 7rE Proof We shall show that Fi(Jt=j. or through an arrival starting an excursion terminating with J.5) A (note that we use A = {x : x E Al on the r.St <S*.4) To show Q = cp(Q).s. 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.u< t). Writing out in matrix notation . A) = L' U(j. mx+dx = j) occurs in two ways . The proof of Q = W(Q) then immediately carries over to show that the subintensity matrix of {mil) } is cp (Q(o)) = Q(l).6) . j. we first compute qij for i $ j. it becomes clear that pij = r [eQh] 0 ij Bi (dy) • (2. Theorem 2 . either due to a jump of {Jt } which occurs with intensity A= j. 0)). StEA . A). Fi(mh =i ) = 1 + =hflh+Qihpii+o(h) implies qii = 'iii /i +)3ipii.(01)diag = Q.T+>t) _ ^iF 7ri (JJ =i. Q = W(Q) follows. Now let {m ( n) } be {mx } killed at the first time i7n (say) a subexcursion of depth at least n occurs . Similarly. of the definition to make U be concentrated on (co. Suppose mx = i. h.5 R(i. p1^) Define a further kernel U by f U(i. A) = f Pi(mx = j) dx eie4xej dx A u (2. It follows that qij = A..4).156 CHAPTER VI.j +/3ipij. e. i.Qi + )%pij) Now just note that t pij and insert (2. Similarly by induction . It is clear that { mini } is a terminating Markov process and that { mio) } has subintensity matrix A* . Then a jump to j (i. = j.j.St EA..
Jo=i. To this end.7 It is instructive to see how Proposition 2.St EA. and to obtain a simple solution in the . oo)) = f o' eIXS((x + z. Jt = i. G+((z.6 (a) R(dx) = eKxdx. `` {K(n)} [the W(•) here is of course not the same as in Proposition 2.r.g. and we let k be the corresponding right eigenvector normalized by Irk = 1.0<u<t. K( n (d) the sequence converges monotonically to K. 0<u<t).z+>t) = P. and get irPi(Jt =j.=StSt. e.4].t.StEA. where A is the diagonal matrix with 7r on the diagonal: Corollary 2. 0 < u < t) = 7rjPj(Jt =i. St < St U. (Jo = j. {Jt }. 0 +1) = cp (K( n)) defined by K(o) = A . oo))dx. From Qe = 0. Remark 2. S. (b) for z > 0. the CramerLundberg approximation (Section 3).Qi)diag.(Jt=j. we shall see that nevertheless we have enough information to derive. THE LADDER HEIGHT DISTRIBUTION 157 from which the result immediately follows by integrating from 0 to oo w. u It is convenient at this stage to rewrite the above results in terms of the matrix K = 0'Q'A. St E A..0<u<t) = P. dt. consider stationary versions of {Jt}..(..6 is hardly all that explicit in general.St <Su.6(b): from 7rK = 0 we get 7rG+(dy)e = J W 7reKx(fiiBi(dy + x))diag dx • e 0 w(fiiB1(dy + x))col dx f 0 EirifiiBi(y)dy = fi*B*(y)dy• iEE 0 Though maybe Corollary 2.S„<0. x < 0... St EA. (c) the matrix K satisfies the nonlinear matrix equation K = W(K) where W( K) = A ( i) diag + fi J "O eKx S(dx). it is readily checked that 7r is a left eigenvector of K corresponding to the eigenvalue 0 (when p < 1). We may then assume Ju=Jtu.6). and this immediately yields (2.2.1 can be rederived using the more detailed form of G+ in Corollary 2.
158
CHAPTER VI. MARKOVIAN ENVIRONMENT
special case of phasetype claims (Chapter VIII). As preparation, we shall give at this place some simple consequences of Corollary 2.6. Lemma 2 .8 (I  IIG+II)e = (1  p)k. Proof Using Corollary 2.6(b) with z = 0, we get
IIG+II = feIxsux, oo dx.
In particular, multiplying by K and integrating by parts yields
0
(2.7)
I)S(dx) KIIG+II =  (eKx
T
= K  A + (,13i)diag 
Z
S(dx) = K A.
2.8)
0 OO
Let L = (kir  K)'. Then (k7r  K) k = k implies Lk = k. Now using (2.7), (2.8) and ireKx = ir, we get
kirIIG +IIe =
ao k f
7rS((x , oo))e = k (lri(3ips, ) rowe = pk,
0 KIIG+IIe = Ke,
(kirK)(I  IIG+II)e = kKepk+Ke = ( 1p)k.
Multiplying by L to the left, the proof is complete. u
Here is an alternative algorithm to the iteration scheme in Corollary 2.6 for computing K. Let IAI denote the determinant of the matrix A and d the number of states in E. Proposition 2.9 The following assertions are equivalent: (a) all d eigenvalues of K are distinct; (b) there exist d distinct solutions 8 1 , .. , sd E {s E C : its < 0} of (A + (131(Bi[s]  1))diag  sIl = 0. (2.9) I n that case , then Si, ... , sd are precisely the eigenvalues of K, and the corresponding left row eigenvectors al, ... , ad can be computed by
ai (A 
(fi(Bi[Si]

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

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

1
E(+(IIG+(e,j)IIG+[t,j;
In matrix formulation, this means that
C =
E L;i
e7f()
hj,r(_) L
 L
ryC M e
L
c+
(IIG+II  G +[ 7]) 0le
1
L
YC+M+e
'y(PL  1)
(ir KL) (I  G+[ y]) 0le,
164
CHAPTER VI. MARKOVIAN ENVIRONMENT
using Lemma 2.10 for the two last equalities. Inserting first Lemma 3.9 and next Lemma 2.8, this becomes 1 7r LA 1(YI  K)(I  IIG+II)e 'Y(PL  1) = 1 P 7r LA 1(yI  K) k = 1P 7rLO1k. Y(PL  1) (PL  1 ) Thus, to complete the proof it only remains to check that irL = vL A. The normalization vLhL = 1 ensures vLOe = 1. Finally, VLOAL = vLAA'K['Y]A = 0
since by definition vLK[y] = k(y)vL = 0.
u
3b Ramifications of Lundberg 's inequality
We consider first the timedependent version of Lundberg 's inequality, cf. IV.4. The idea is as there to substitute T = yu in 'Pi (u, T) and to replace the Lundberg exponent y by yy = ay  yk(ay ), where ay is the unique solution of rc(ay)= 1 Y Graphically, the situation is just as in Fig. 0.1 of Chapter IV. Thus, one has always yy > y, whereas ay > y, k( ay) > 0 when y < 1/k'(y), and ay < y, k(ay) < 0 when y > 1/k'(y). Theorem 3 .10 Let C+°) (y) _ 1
miniEE hiav)
Then 1 y< (y)
y>
Vi(u,yu)
Pi(u) 
C+°)(y)hiav)
e7vu,
(3.6)
V,i(u,yu)
< C+)(y)hiar )e 'Yvu,
(y) (3.7)
Proof Consider first the case y <
Then, since k (ay) > 0, (3 .1) yields
'12(u,yu)
hiav)]E'iav,i
h(ay ) J*(u)
exp {ayST(,L ) +r(u)k( ay)}; T(u) < yu
We further write G(u) for the vector with ith component Gi(u) = EiEE G+(i.3. However. Chie ryu < Vi(u ) < C+hie 7u. hj P .5) will produce the maximal ryy for which the argument works. dy)• o iEE jEE .5). for a vector <p(u) = (cpi (u))iEE of functions .y)G+(z.8 ) Then for all i E E and all u > 0.V)i(u. we let G+ * W(u) be the vector with ith component E(G+(i. yu < r(u) < 00] < hiav)C+o)( y)eavu+yuw(av) 0 Note that the proof appears to use less information than is inherent in the definition (3. (3. oo)) and. as in the classical case (3.9) For the proof. Our next objective is to improve upon the constant in front of a7u in Lundberg's inequality as well as to supplement with a lower bound: Theorem 3.00 su e7( ( 3. (u. av 'i [h.j) * coj)(u) _ f u ^Pj(u . yu) f h(av)e v avuE«v. r(u) yu o)(y)eavuEav. we shall need the matrices G+ and R of Section 2.7. r(u) < yu] hiay)C+ h=av)C+ o) (y)eayu+yuw(av).i I (a) exp {aye(u) + r(u)r. exp {e() + r(u))} . CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 165 hiav)e _avuE.i [e*(u)K(av). we have ic(ay) < 0 and get 'i(u) .j.i [eT(u)K(av ). if y > 1lk'(ry)..(ay)}.11 Let Bj (x) C_ = min 1 • inf jEE hj x>o f2° e'r( vx)Bj(dy) ' C+ _ mE 1 Bj(x) J Y x)Bj (dy). 1 Similarly. yu < r(u) < 00 h 4(u) < h(av)C+o)(y)eavuEav .
dx).u IMP:°) (u) I < oo.7. dx ) Bj (u . 0 G+(i.u Iv 2°)(u)I Pi(rr+(N + 1) < oo) + 0. j.x) x) jEE 0 E Qj f jEE R(i. Lemma 3 .& (u). dx) f e7( vu)Bj (dy . and define W(n+1) (u) = G(u) + (G+ * tp(n))(u).ery(&u+x)Bj (dy) Bj(u Bj (u . if r+ (n) is the nth ladder epoch. j.3jhj // f 00 R(i. °O .(0) ] (u) < sup Jt t. Hence lim cp(n) exists and equals U * G. Then cpin)(u) sit (u) as n + oo. dy) 00 C+ ijhj f R(i. j. = Eo G+ G. MARKOVIAN ENVIRONMENT Lemma 3 .x ) = Gi(u).13 For all i and u.12 Assume sup1.x ) R(i. 00 f C_ hj f e(Y)G+(i. 00 Thus C+ > hj f"o e7(Yu)G +(i. dy) = aj f Bj(dy . just note that the recursion <p(n+1) = G + G+ * (p(n) holds for the particular case where cpin)(u) is the probability of ruin after at most n ladder steps and that then obviously u cp2n) (u) + t. Proof Write UN = EN G+ .dy). dy) : 1(u) < C+ > hj u e(1tL)G+(i. Then iterating the defining equation ip(n+1) = G + G+ * V(n) we get W(N+1) = UN * G + G+N+1) * ^(o) However.j. n > oo. U = U". jEE u 0 j.166 CHAPTER VI. _ To see that the ith component of U * G(u) equals ?Pi (u). j. dx) 100 C .x) jEE 00 u 0 //^^ C+E. we have G *(N +1) * ^.j.
u)G+(i.ST).13) Hence.3.M > u) = Pi(ST<u. letting MT = maxo<t<T St. 9 for the last equality in (3. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES proving the upper inequality.Pi(MT > u) = Pi(MT < u.MT<u. it follows that Vi(u) < C_(yo) h=70)e7ou. dy) jEE o (3.11.tpi(u. from which the lower inequality follows by letting n * oo. T) < C+(')' o)hi7u)e7ou8T .M>u) = Ei [VGJT (u . and the proof of the lower one is similar. Then 0< Vi (u )  0i(u. (3. dy) jEE u U +C_ hje7( yu)G jEE"" +(i. j. Indeed.11) C_e7u 57 O+[i. this is obvious if n = 0. and let 8 = e'(70). Here is an estimate of the rate of convergence of the finite horizon ruin probabilities 'i (u. y]hj = C_ e7uhi. j.10: Theorem 3 . and assuming it shown for n.13. 13 and the second by the induction hypothesis . we have Vii (u) . (3. dy) (3. T) = Pi (7.8) with y replaced by yo and hi by h=7o ).y)G+(i. and using Lemma 3 . MT < u.13 Let first cp=°)(u) = C_ hie"u in Lemma 3. we get Wo n +1) (u) = ? 7 i ( U ) + E J u gyp. We claim by induction that then cpin) (u) > C_ hie7u for all n.10) C_ 1 f hje7(y. j.(u) < T ) to 0i (u) which is different from Theorem 3.10 ) by Lemma 3 . jEE estimating the first term in (3. j. u The proof of the upper inequality is similar .n) ( u . taking cps°) (u) = 0. ST < u] < C+(yo)e7ouEi [h^7o)e70ST1 l T J = C h(7o)e7ou8T .T) = Pi(M > u) . 14 Let yo > 0 be the solution of 'c'(yo ) = 0. 167 u Proof of Theorem 3.12) Proof We first note that just as in the proof of Theorem 3. +i .11). let C+(yo) be as in (3.
33 or Bi 0 Bj. The results to be presented show that quite often this is so.2) alone just amounts to an ordering of the states. (4..o. 4 Comparisons with the compound Poisson model 4a Ordering of the ruin functions For two risk functions 0'.31:5)32 . in part from the folklore principle that any added stochastic variation increases the risk. B2 <_s. this is not the case for (4. (4.. Further related discussion is given in Grigelionis [176]. we refer to ... 0"(u) = P(M" > u)) Now consider the risk process in a Markovian environment and define i' (u) _ >iEE irioi(u)..o.. MARKOVIAN ENVIRONMENT Notes and references The results and proofs are from Asmussen and Rolski [44]. M" of the corresponding two claim surplus proceses (note that 0'(u) _ P(M' > u). Bp. ". and finally in part from queueing theory. where it has been observed repeatedly that Markovmodulation increases waiting times and in fact some partial results had been obtained. Occasionally we strengthen (4.4) To avoid trivialities. this correponds to the usual stochastic ordering of the maxima M'..3) Bl <_s. The Markov process {Jt} is stochastically monotone (4. The motivation that such a result should be true came in part from numerical studies..0.o. <s. V)" if z/i'(u) <'c"" (u). (4. we also assume that there exist i # j such that either /3i <. [177].5) Note that whereas (4. but that in general the picture is more diverse.168 CHAPTER VI. It was long conjectured that 0* Vi.3p. For the notion of monotone Markov processes. is the one for the Markovmodulated one in the stationary case (the distribution of J0 is 7r). we define the stochastic ordering by 0' < s.1) Obviously.3). The conditions which play a role in the following are: .3) to B = Bi does not depend on i. < .2) (4.o. where o*(u) is the ruin probability for the averaged compound Poisson model defined in Section 1 and . u > 0.
9) (4. ^i 7ri = 1.x)dx _ /3*B*(u) + f u / ^ t=1 > 3 * B* ( ) + f (4..x) dx u o i =1 i=1 (4. Lemma 4 .2.* For the proof. Proof of Theorem 4. 7(0) < oo) = pirf+). Proschan & Walkup [140].. we need two lemmas. b1 < . Section 4.6) 7r= fl*B*(u) + p> s=1 +) fu 0 b (u  x)Bt (x) /pB. < bp and 7ri > 0 (i = 1.8) ^j Tri/iBd(x) . note that (4.6. 0 Here (4. COMPARISONS WITH THE COMPOUND POISSON MODEL 169 Stoyan [352]. 2 If al < ..7) 7ri. Theorem 4 .4) is automatic in some simple examples like birthdeath processes or p = 2 . Then V.6). . we obtain (cf.r (JT(o) = i. 3 (a) P. then P P P 7rjbj.2)(4. dx (4. = b.9) follows by considering the increasing functions 3iBi (x) and Oi (u . = aP or b1 = .. Conditions (4. 1:7riaibi > E 7riai i=1 i=1 j=1 The equality holds if and only if a1 = .1) which with basically the same proof can be found in Asmussen & Schmidt [49]..r (Sr(o) E dx Jr(o) = i.7) and Lemma 4.4) say basically that if i < j .3* f uB(x) z/^.10) Q*B*(u)+. the second follows from an extension of Theorem I1.1 Assume that conditions (4.x)B*(x) dx.1 for the first term in (4.. and it is in fact easy to show that Vii(u ) < t/j(u) (this is used in the derivation of (4.x) of i and using Lemma 4. .r(u x)dx. T(0) < oo) = Bi(x) dx/tcai . also Proposition 2..10) and (4.3 for the second) *(u) _ /3 *B* (u) +.2)(4.4. < a.. E 7r i Wi(u . then j is the more risky state .5 (cf. Comparing (4. Proposition 2..2..3iBi(x)YPi(u .. p)... it follows by a standard . where 7r2+) = QiµBilri/p.4) hold. (b) P.1. The first is a standard result going back to Chebycheff and appearing in a more general form in Esary.13* J0 u 0*(u . Lemma 4 . (4.. Conditioning upon the first ladder epoch.9 ) below).
h. it will hold for all sufficiently large u.0*• i=1 But it is intuitively clear (see Theorem 3..4 is the understanding of whether the stochastic monotonicity condition (4. (4. and from this the claim follows.s. i=1 i=1 7'r(0) _ EFioiwi(0) . 4b Ordering of adjustment coefficients Despite the fact that V)* (u) < *. Then the l.1 and Proposition 4.s. Using (4. Recall that . of (4. Rolski & Schmidt [32].. this ruin probability is /3iPBi. As is seen. u To see that Proposition 4.6). Then i/i*(u) < .8) we get P P '*' (0) = 3* + /3*1* (0) _ > lri'3qqi • E 7i/ipBi .0.* (0).. except possibly for a very special situation ..2.11) i=1` and that A has the form eAo for some fixed intensity matrix A0.4 is not vacuous.1 of [145] for a formal proof) that z/ii(u) converges to the ruin probability for the compound Poisson model with parameters . it is sufficient to show that 0'.170 CHAPTER VI. 0.4) is essential (the present author conjectures it is). What is missing in relation to Theorem 4.4 Assume that . that P P /^ 1r1NiµBi /^2 /^ ^i/ji pBi < 1il3i i=1 i=1 (4. Bi as e J.(0) < b *'(0) for e small enough. µB2 = 104. µB.6). Proof Since 0.h.11) is of order 104 and the r.. MARKOVIAN ENVIRONMENT argument from renewal theory that tk. let = ( 1/2 1/2 ) . Frey.3µi < 1 for all i.3i. 01 = 103. dominates the solution 0* to the renewal equation (4. of order 101. = 102.r (u ) fails for all sufficiently small e > 0. u Here is a counterexample showing that the inequality tp* (u) < V).(0) = V. Notes and references The results are from Asmussen. Q2 = 1. (u) is not in general true: Proposition 4. For u = 0. they are at present not quite complete. (u) may fail for some u./3*.
g. Proof Define X= f &ids.14) A„(O) iioo varXt t t By convexity.4(b) that the limit in (4. (4. This implies that A is strictly convex. in particular .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) > 0.4. Asmussen [20]) as discussed in 11. e. Hence if 5i 54 0 for some i E E.5 y < ry*. it follows by Proposition A1. Jt = i])' EE = vA+n(6.Jt=kI A (the return time of k) where k E E is some arbitrary but fixed state.g.13) implies A(a) > 0 for all a. (4. The adjustment coefficient y* for the averaged compound Poisson model is the solution > 0 of rc*(ry*) = 0 where rc*(a) _ 13*(B*[a] . with strict inequality unless rci (y*) does not depend on iEE.5.a = E irirci(a).1) . is nondegenerate unless bi does not depend on i E E... 0 .12) iEE Theorem 4. Xt)} is a Markov additive process (a socalled Markovian fluid model. 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] . It is clear that the distribution of X. cf. Now we can view {Xt} as a cumulative process (see A.14) is nonzero so that A"(0) > 0. and by Proposition II.13) (4.5.ld) with generic cycle w = inf{t>0: Jt_54 k.)a.5.a.2 we have (Ei[e"X'. (4.7) )i is convex with A'(0) = lim EXt tioo t = iEE 70i = 0. Further (see Corollary 11.(a) > 0 for all a 0 0. which in view of EiEE 1ibi = 0 is only possible if Si = 0 for all i E E.1) . Lemma 4. with strict inequality unless a = 0 or bi = 0 for all i E E. Then {(Jt.
e7r)1 (Ici(Y*))diage.15) Normalizing h by 7rh = 0.15) yields 0 = (Ii(y*)) diage + Aoh'(0) = (rci('Y*)) diage + (Ao . h'(0) = (Ao .12) and rc*(y*) = 0. Frey. we get rc (y*) > 0 which in a similar manner implies that u y < y*.16) Differentiating (4. note that y(a) + mins=1. Then > risi = 0 because of (4. Rolski & Schmidt [32]. a = 1 in Lemma 4. and our aim is to compute the sensitivity ay e a E=O A dual result deals with the limit a 4 oo.. we have 7rh' = 0. this implies that the solution y > 0 of K(y) = 0 must satisfy y < y*.p yi and compute 8y 8a a=0 In both cases. where A. 4c Sensitivity estimates for the adjustment coefficient Now assume that the intensity matrix for the environment is Ae = Ao/ e. 0 = ((ri(Y))diag + ery (4{('Y))diag)h + (A0 + e(?i'Y))diag)h'. (4.172 CHAPTER VI. Thus y(e) * y* as e 10.eir)h'(0).5 is from Asmussen & O'Cinneide [40].. In the case of e. Notes and references Theorem 4. Since ic is convex with rc'(0) < 0 . multiply the basic equation by a to obtain 0 = (A0 + e(r£i(y))diag)h. If rci(y* ) is not a constant function of i E E. Here we put a = 1/e.15) once more and letting e = 0 we get . Hence rc (y*) > 0. Further a(1) = rc(y*) by definition of A(.. the basic equation is (A + (rci(y))diag)h = 0. (4. improving upon more incomplete results from Asmussen. MARKOVIAN ENVIRONMENT Proof of Theorem 4.. Hence letting e = 0 in (4.6.) and rc (•).Qi and Bi are fixed . whereas the . h depend on the parameter (e or a).5. h(0) = e. The corresponding adjustment coefficient is denoted by ry(e). Let bi = rci(y*). y.
Rolski & Schmidt [32]. THE MARKOVIAN ARRIVAL PROCESS 173 0 = 27'(0)(ri(`Y *)) diage + 2(ci('Y* )) diag h' (0) + Aoh" (0) .8 If (4.5. (4. The analogue of Proposition 4.. and may have some relevance in risk theory as well (though this still remains to be implemented).7 8ry aE = 1 7r(ci ('Y*))diag ( Ao e7r)1(Xi(Y*))diage *=0 P Now turn to the case of a.17) (4. 0 = (Ao + ry'(ii(Y)) diag )h + (aAo + (Ki(7'))diag)h'. p. We get 0 = (aAo + ( lc&Y))diag)h.18) 0 = 27'(0)p+27r(rs. Frey. We assume that 0 < y < 7i. the intensity for such a transition (referred to as marked in the following) is denoted by Aii l and the remaining intensity . 5 The Markovian arrival process We shall here briefly survey an extension of the model. i = 2. (4.19) holds.19) Then 'y ^ ryl as a ^ 0 and we may take h(0) = el (the first unit vector).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). 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. (4.18).17) by 7r to the left to get (4. multiplying (4. . and we have proved: Proposition 4.20) Letting a = 0 in (4..16) yields Proposition 4.8 when ryi < 0 for some i is open.i(7' *))diagh'(0). which has recently received much attention in the queueing literature. . then 8a a=o All rci (0) Notes and references The results are from Asmussen. Inserting (4.
A ( 2) = A (2`1 ) ® A. Bii = Bi . Note that the case that 0 < qij < 1. u Example 5 .6i ) diag.1 (PHASETYPE RENEWAL ARRIVALS) Consider a risk process where the claim sizes are i. then {Nt} is a Markov additive process if and only if it corresponds to an arrival mechanism of the type just considered.2 (SUPERPOSITIONS) A nice feature of the setup is that it is closed under superposition of independent arrival streams . the claim surplus is a Markov additive process (cf. where qij is the probability that a transition i * j is accompanied by a claim with distribution. refer to notation) { Jt k) }. Jt2)) (2. For i = j. is neither 0 or 1 is covered by letting Bij have an atom of size qij at 0. let { Jt 1) }. and thus 1i = 0. We then let (see the Appendix for the Kronecker E = E(1) x E(2). the Markovmodulated compound Poisson model considered sofar corresponds to A(l) = (.2) A(1) = A(' 1) ® A(1. Jt = (Jtl).i. with common distribution B. A(l) = tv.2). but the point process of arrivals is not Poisson but renewal with interclaim times having common distribution A of phasetype with representation (v. Indeed. Here are some main examples: Example 5 . A(l) = T. A(1'k) A(2 k1). we may let {Jt} represent the phase processes of the individual interarrival times glued together (see further VIII. the definition of Bij is redundant for i i4 j.2 for details).4).^) etc. that Bii = Bi . and that are determined by A = A(l ) +A(2) where A is the intensity matrix the governing {Jt}. This is the only way in which arrivals can occur. MARKOVIAN ENVIRONMENT f o r a transition i + j by A .174 CHAPTER VI. The extension of the model can also be motivated via Markov additive processes: if {Nt} is the counting process of a point process.d. Bij = B. T).(13i )diag. In the above setting. A(1) = A . . B. the definition of Bi is redundant because of f3i = 0. Thus . Again . II. j(2) } be two independent environmental processes and let E(k). we use the convention that a1i = f3i where 3i is the Poisson rate in state i. and the marked transitions are then the ones corresponding to arrivals.
The versatility of the setup is even greater than for the Markovmodulated model. • upon a claim.5. after which it starts afresh. all Al i2. E = { WORKING. where ik = 0 means that the kth policy has not yet expired and ik = 1 that it has expired.1i2. 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 . Easy modifications apply to allow for • the time until expiration of the kth policy is general phasetype rather than exponential.. RETIRED. Thus. Example 5 .. THE MARKOVIAN ARRIVAL PROCESS Bij... 11. the kth policy enters a recovering state. INVALIDIZED.g... or..iN = C27 All other offdiagonal elements of A are zero so that all other Bii are redundant. Similarly. WIDOWED. assume that there is a finite number N of policies.kj = Bik) B13 4k = Bak) 175  (the definition of the remaining Bij.iN. DEAD etc...iN C17 AilO. However . claims occur only at state transitions for the environment so that AN2... u Example 5 . as the Markovian arrival process ( MAP). 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. i2i . DIVORCED.. In this way we can model. Bilo.. possibly having a general phasetype sojourn time. In fact . iN. iN = all BOi2.iN.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).4 (A SINGLE LIFE INSURANCE POLICY ) Consider the life insurance of a single policy holder which can be in one of several states. more recently.1i2 .}. and that the policy then expires.iil. MARRIED. This means that the environmental states are of the form i1i2 • • • iN with il....kl is redundant). E 10.. the idea of arrivals at transition epochs can be found in Hermann [193] and Rudemo [313]. say. e. superpositions of renewal processes. Assume further that the ith policy leads to a claim having distribution Ci after a time which is exponential.. with rate ai.iN are zero and all Bi are redundant.. iN.. Hermann [193 ] and Asmussen & Koole [37] showed that in some appropriate ..iil.iN = a2.
Without loss of generality. We denote throughout the initial season by s and by P(8) the corresponding governing probability measure for the risk process. we talk of s as the 'time of the year'.1) Then the average arrival rate is /3* and the safety loading rt is 77 = (p* . 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. Obviously.3*µs • p = f /3(v) dv 0 0 (6. . 1). Neuts [271] and Asmussen & Perry [42]. The basic assumptions are as follows: • The arrival intensity at time t of the year is 3(t) for a certain function /3(t). one needs to assume also (as a minimum) that they are measurable in t. • The premium rate at time t of the year is p(t). • Claims arriving at time t of the year have distribution B(t). 0 < t < 1. a claim arrives with rate /3(s + t) and is distributed according to B(8+0 . 1). from an application point of view.176 CHAPTER VI. 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 . Let 1 1 /3* _ f /3(t) dt. where i f00 xB(°) (dx) _ . Thus at time t the premium rate is p(s + t). Lucantoni et at. For the Markovmodulated model. but now exhibiting (deterministic) periodic fluctuations rather than (random ) Markovian ones. )3 t 1 J (6.p)/p. Lucantoni [248]. p * = 0 p(t) dt. p(t) and B(t) are defined also for t t [0. [248]. let the period be 1.2) Note that p is the average net claim amount per unit time and µ* = p//3* the average mean claim size. Sengupta [336]. we may assume that the functions /3(t). By periodic extension. for s E E = [0. B* = J f B(t) ((*) dt. one limitation for approximation purposes is the inequality Var Nt > ENt which needs not hold for all arrival streams. continuity would hold in presumably all reasonable examples. Some main queueing references using the MAP are Ramaswami [298].
6.3(t) = 3A(1 + sin 27rt).1 As an example to be used for numerical illustration throughout this section.t. It is easily seen that . B*. let . equivalently. Thus .w(t)) dt). the discussion in 111. respectively. The arrival process {Nt}t>0 is a timeinhomogeneous Poisson process with intensity function {/3(s + t)}t>0 . 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 (1.3) Note that A enters just as a scaling factor of the time axis.8. of the periodic model as arising from the compound Poisson model by adding some extra variability. or. The behaviour of the periodic model needs not to be seen as a violation of this principle. for Markovmodulated model typically the adjustment coefficient is larger than for the averaged model (cf.10.1. Many of the results given below indicate that the averaged and the periodic model share a number of main features. We u assume in the rest of this section that p(t) . and thus the averaged standard compound Poisson models have the same risk for all A. in agreement with the general principle of added variation increasing the risk (cf. not random. (6.2 Define T 6(T) = p(t ) dt.w(t).9).(3. The claim surplus process {St } two is defined in the obvious way as St = ^N° Ui . 0 Then (by standard operational time arguments ) {St} is a periodic risk process with unit premium rate and the same infinite horizon ruin probabilities. and we recall from there that the ruin probability is 24 1 *(u) _ 3 5eu + 35e6u. u Remark 6 .1) and Example 1. RISK THEORY IN A PERIODIC ENVIRONMENT 177 In a similar manner as in Proposition 1. since the added variation is deterministic. Section 4b). p* as an averaged version of the periodic model. In contrast. the average compound Poisson model is the same as in III. St = SeI(t). we shall see that for the periodic model increasing A increases the effect of the periodic fluctuations. it turns out that they have the same adjustment coefficient. one may think of the standard compound Poisson model with parameters 3*. In particular. In contrast.3* = 3A. p(t) = A and let B(t) be a mixture of two exponential distributions with intensities 3 and 7 and weights w(t) _ (1 +cos27rt)/2 and 1 . the conditional distribution . Thus. Example 6 .
.g.5.a) Proof Conditioning upon whether a claim occurs in [t.3(v)(B(vl [a] . Dassios & Embrechts [98] and Asmussen & Rolski [43].5 (see in particular Remark 11.. 6b Lundberg conjugation Motivated by the discussion in Chapter II. t + dt] or not. The exposition of the present chapter is basically an extract from [44].g.a. of the averaged compound Poisson model (the last expression is independent of s by periodicity).tc* (a)] dv then h (. we start by deriving formulas giving the m.1) a = J8 . . i. we obtain E. with the underlying Markov process {Jt} being deterministic period motion on E = [0.T) = P(8)(r(u) <T).3(s + t)dt[B(8+t)[a] .1) dv . J Theorem 6 .(8) [eaSt+dt I7t] = = (1 . Daykin et.178 CHAPTER VL MARKOVIAN ENVIRONMENT of U. but it turns out to have obvious benefits in terms of guidelining the analysis of the model as a parallel of the analysis for the Markovian environment risk process. a) is periodic on R.Q(v) (B(„) [a] .a be the c.e. e.s . with some variants in the proofs.adt +.1) . and the ruin probabilities are 0(8) (U) = P(s )(r(u) < 00). r(u) _ inf It > 0 : St > u} is the time to ruin . (6. As usual. Jt = (s + t) mod 1 P(8) .8). The claim surplus process {St} may be seen as a Markov additive process.al. 3 E(8)eaSt = h(s.(3(s + t)dt)e«St adt + /3(s + t)dt .^8 [. let f 8+1 tc *(a) _ (B* [a] . To this end.a) = exp { .g. 0 (5)(u. see the Notes to Section 7).f. of the claim surplus process. a) etw*(a) h(s+t.(1 .4) At a first sight this point of view may appear quite artificial. Notes and references The model has been studied in risk theory by.east B(8+t) [a] east .. [44] (the literature in the mathematical equivalent setting of queueing theory is somewhat more extensive.f.1]) . and define h(s. [101] .a . 1). given that the ith claim occurs at time t is B(8+t).
it then suffices to note that E(8)Le.s.6.t}t>o = h(s.(e) Let = h( h(Jo.3.* (a) h(s. dt log E(8)east a + f3(s + t) [B(8+t) [a] . With g the infinitesimal generator of {Xt} = {(Jt. Proof In the Markov additive sense of (6. a) Corollary 6. St)} . 0) exist and are finite.3(s + t)[D(8 +t)[a] .1)dv l og E(8) et where atetk•(a) h(t.1]) . St)} and .5. 9) is a P ( 8)martingale with mean one.adt +. a) Thus E(8)east = h(s + t. RISK THEORY IN A PERIODIC ENVIRONMENT E(8)east+ dt d Et. we can write Lo Jt. a) et..t.1)dv  o h(t. a) = h(s. E (8)east (a +. + v)(B([a] . a) h(s + t.c* (e) {Le.9 as follows.0(s + t)dt[B(8+t)[a] .4 For each 0 such that the integrals in the definition of h(t .6 . h(s + t.1]) . so that obviously {Lo. a). a) as well as the fact that rc = k` (a) is the correct exponential growth rate of Eeast can be derived via Remark 11.2. at + f log h(s + t. a) = exp I f t3(v)(kv)[a) . According to Remark 11. u Remark 6.t = 1 by Theorem 6.(8)east 179 = = = = = E(8)east (1 .5 The formula for h(s) = h(s. 0) P(8)a. a) .log h(s.9) eastt.t} is a multiplicative functional for the Markov process { (Jt.4). B) eoSt t.1].
(3(s)dt) +.180 CHAPTER VI. cf.0) = Kh(s). we put for short h(s) = h(s.'y). yo is determined by 0 = k* (70) = QB*. Proposition 6. [70] . it follows by Theorem II.4. Bet)(dx) = ^ B(t ) (dx). of St is as for the asserted periodic risk model. That is. as above E (s) ha(Jdt. Lemma 6 . MARKOVIAN ENVIRONMENT ha(s.3(v)( Bi"i [a] . (iv) finally. the restrictions of Plsi and Pest to Ft are equivalent with likelihood ratio Le. 0 < s < 1.f. That rc = is*(a) then follows by noting that h(1) _ u h(0) by periodicity. When a = y. say. y solves n* (y) = 0.3(s)ks)[a]h(s)} ah(s) 13(s)h(s) + h'(s) +.1) . .1. However. ( iii) use approximations with piecewiese constant /3(s). such that for any s and T < oo. B(s).a . A further important constant is the value yo (located in (0.T.3(s)dt • B(s)[a]h(s) = gha(s. correspond to a new periodic risk model with parameters ex .y) = eayh(s). Sdt) = h(s + dt) eadt (1 .5 that we can define a new Markov process {(Jt. see [44] for 11 a formal proof.2.3. For each 0 satisfying the conditions of Corollary 6.g. Proof (i) Check that m. St)} with governing probability measures Fes). Proposition 6. the requirement is cha(i. ry)) at which n* (a) attains its minimum. (ii) use Markovmodulated approximations (Section 6c). That is. Equating this to rch (s) and dividing by h(s) yields h(s ) = h(s) = a + . J s [.3(s)B(s) [a]h(s).3(s)h(s) + h'(s) +.6 The P(s). Now define 'y as the positive solution of the Lundberg equation for the averaged model. P(s) (T(u) < oo) = 1 for all u > 0.6 ( s ) exp { 0( s )&s) [a] + tc .60(t) = a(t)B(t)[0].7 When a > yo. 0) = h(s) + dt {ah(s) .tc] dv} (normalizing by h(0) = 1).
The proof involves machinery from the ergodic theory of Markov chains on a general state space.9) 0(')(u) = h(s.1) the distribution of (l: (oo). s E I. the mean number of claims per unit time is p« 181 = Jo 1. a) a > ry0 (6. Here and in the following. the Markov process {(^(u). have components with densities b(8)(x) satisfying inf sEI. xEJ 0 (s)b(8)(x) > 0. T) = h(s. Wu).g. considered with governing probability measures { E(8) }E[ . a) e«uE(8 ) e «^ .2.9) and noting that weak convergence entails convergence of E f (^(u). The relevant likelihood ratio representation of the ruin probabilities now follows immediately from Corollary 11.7) h(B(u). a) TI h(9(u). Corollary 6. J C R+ such that the B(8). e(cc)) Letting u > oo in (6.10) Then for each a. 9(u)) for any bounded continuous function (e. Lemma 6 . T(u) < (6. we need the following auxiliary result .9 Assume that there exist open intervals I C [0.9(u))} u>0. 0(u)) * (b(oo). RISK THEORY IN A PERIODIC ENVIRONMENT Proof According to (6. we get: . which is not used elsewhere in the book. u which is > 1 by convexity. say s0.8 The ruin probabilities can be computed as (u)+T(u)k'(a) ^/i(8) (u.1.8) (6. q) = eryx/h(q)). ^(u) = ST(u) .2). f (x. and we refer to [44]. 1). B(oo)). a)e«uE (a iP(s) (u) = h( s)e7uE(` ) h(O(u)) To obtain the CramerLundberg approximation from Corollary 3. and no matter what is the initial season s.6(v) dv Jo ' xe«xB (°) (dx) r^ xe«xB'(dx) = Q'B' [ a] = ^' J 0 = ^c"'(a) + 1. has a unique stationary distribution.u is the overshoot and 9(u) = (T(u) + s) mod 1 the season at the time of ruin.4. (6.6.
1.) C = E1 h(B(oo)) u + oo. elementary calculus yields h(s) = exp { A C 2^ cos 2irs  4^ sin 21rs + 11 cos 41rs .10) of Lemma 3. this provides an algorithm for computing C as a limit.Ch(s)ery".16. illustrating that the effect of seasonality increases with A. At this stage .182 CHAPTER VI. MARKOVIAN ENVIRONMENT Theorem 6. 10 shows that certainly ry is the correct Lundberg exponent. where e. Among other things.1.W. Vi(8) (u) . we obtain immediately the following version of Lundberg ' s inequality which is a direct parallel of the result given in Corollary 3.11) gives an interpretation of h(s ) as a measure of how the risks of different initial seasons s vary.10 Under the condition (6. Noting that ^(u) > 0 in ( 6. 11 7/'O (u) < C+°)h(s) ery".ir) } Plots of h for different values of A are given in Fig.9). A=1/4 A=1 A=4 0 Figure 6. 6. which may provide one among many motivations for the Markovmodulated approximation procedure to be considered in Section 6c.6 for the Markovmodulated model: Theorem 6 . For our basic Example 6 . (6. 1. where C(o) = 1 + info < t<i h(t) . it does not seem within the range of our methods to compute C explicitly.1 In contrast to h. Theorem 6 .11) Note that ( 6.
yu) 000 (u) .w)e4u .12 Let 00)(y) 1 Then info < t<i h(t. Consider first the timedependent version of Lundberg's inequality. the proofs are basically the same as in Section 3 and we refer to [44] for details. Lundberg's inequality can be con siderably sharpened and extended. #c( ay) < 0 when y > 1/tc'('y).11 as well as it supplements with a lower bound.w) . T) and replace the Lundberg exponent ry by ryy = ay . 1 ) and all u > 0..g.w)e4u dx 9w + 7(1 . (6. where ay is the unique solution of W(ay) =y• (6. r.13) Elementary convexity arguments show that we always have ryy > Y and ay > ry. ay) • (6. C_h(s)e7u < V. We state the results below. .14) < C+)(y)h(s) e7yu.12) As for the Markovian environment model. (ay).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. e7 ( yx)B(t)(dy) > Then for all $ E [0. we substitute T = yu in 0(u. we first note that the function fu° exu {w • 3e .13 to our basic example. 1 (6. whereas ay < y.42 so that 183 tp(8) (u) < 1.4. Just as in IV.16 In order to apply Theorem 6. we obtain Co) = 1.0(8) (u+ yu) (6. Theorem 6 .47r sin 27rs + 167r cos 47rs . e.7e .(8) (u.(ay) > 0 when y < 1/ic' (7).(s)(u) < C+h(s)e7".17) (6.3x + (1 . RISK THEORY IN A PERIODIC ENVIRONMENT Thus. Theorem 6.6.15) The next result improves upon the constant C+) in front of eryu in Theorem 6.42 • exp {J_ cos 27rs .167r I Cu.yr.w ) • 7e u{w • 3e3x + ( 1 .7x j dx _7x } _ 6w + 6(1 . in our basic example with A = 1.
but thereby also slightly longer. completing a cycle . C+ = 1. with s the initial season. Then 7oudT . Finally.1 sin 27rs + 1 cos 47rs . MARKOVIAN ENVIRONMENT attains its minimum 2 /3 for u = oo and its maximum 6 /(7 + 2w) for u = 0.16. This observation motivates to look for a more formal connection between the periodic model and the one evolving in a finite Markovian environment. 14 Let C+('yo) be as in (6.013. where the environment at time t is (s + t) mod 1 E [0.20 •exp { 2n cos 27rs .20). Of course.T) < C+('Yo)h( s. and in fact. we have the following result: Theorem 6 . Thus. 0 <'p(8)(u ) .I eu.16) with 'y replaced by yo and h(t) by h(t. 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.19 I eu.'Yo)e (6.19 } 0 <8<1 8 + cos 21rs Thus e.4^ sin 2irs + 16^ cos 41rs .0.184 CHAPTER VI..g.66.\ = 0 . 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'. .013. 1) for the environment). Some of the present proofs are more elementary by avoiding the general point process machinery of [44]. .\ 3 C+ = sup 6 exp { A (.1 sin 2irs + 16_ cos 47rs .. Thus C_ = 2 inf ex cos 2irs .66. exp 2^ cos 21rs . yo). and let 8 = er' (Y0).cos 27rs .9 3 0<8<1 p 27r 47r 167r 161r 2 _ _e.18) Notes and references The material is from Asmussen & Rolski [44]. n}.181 s(u) < 1. 1/i18 1 s (u) > 0. 6c Markovmodulated approximations A periodic risk model may be seen as a varying environment model. the nth Markovian environmental process {Jt} moves cyclically on {1.(8)(u.L sin 27rs + 1 I cos 47rs . 1). such a deterministic periodic environment may be seen as a special case of a Markovian one (allowing a continuous state space E = [0. .
since the settings are equivalent from a mathematical point of view.1 ((i 1)/n) ) and Bni = B . . and the ruin probability corresponding to the initial state i of the environment is then Y'yn)(t) = F (M(n) > t). We want to choose the /3ni and Bni in order to achieve good convergence to the periodic model. DUAL QUEUEING MODELS 185 within one unit of time on the average .7. Thus. To this end. AE= Aii'r?/7ri• The arrival intensity is /3i when Jt = i. one simple choice is Oni = 0( i .jEE of the risk process. 7 Dual queueing models The essence of the results of the present section is that the ruin probabilities i/ (u). Bi.20) be the claim surplus process of t>o the nth approximating Markovmodulated model.21) which serves as an approximation to 0(1)(u) whenever n is large and i/n s.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. Notes and references See Rolski [306]. it is desirable to have formulas permitting freely to translate from one setting into the other. so that the intensity matrix is A(n) given by n n 0 ••• 0 0 n n ••• 0 A(n) _ (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. This queue is commonly denoted as the Markovmodulated M/G/1 queue and has received considerable attention in the last decade. M(n) = Supt>o Stn). (6. z/'i (u. Let 0j. We let {Stn)} (6. 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.
Proposition 7.1 Assume V0 = 0.2 The relation between the steadystate distributions of the actual and the virtual waiting time distribution is given by F(W > u. Taking probabilities and using the stationarity yields 7riPi(T(u) < T. {Jt }o<t<T• Then we may assume that Jt = JTt. 2 . Proof Consider stationary versions of {Jt}o<t<T.1) over j. Jt ).oo in u (7. . JT = j) = LjPj (VT > u.. I* )3i P(V > u.. JT = j} and {VT > u. J*) is the steadystate limit of (Vt. The first conclusion of that result then states that the events {T(u) < T. 0 < t < T and that the risk process {Rt}o<t<T is coupled to the virtual waiting process {Vt}o<t<T as in the basic dualitylemma (Theorem 11. and the virtual waiting time (workload) process {Vt}too are defined exactly as for the renewal model in Chapter V. (VT > u I JT = 2). J* = i) for all j. let T . JJ = i). MARKOVIAN ENVIRONMENT • Customers arriving when Jt = i have service time distribution Bi.=1 . J* = i).2).n(VT > u.3).1) follows.0i (u . (7. JT = j) = 7rjPj(VT > u.P(V > u.T(V > u I J* = i). Jo = j.3) 7ri where (V.3. • The queueing discipline is FIFO. J* = i) = P. Proposition 7. JT = Z). . JT = i} coincide.1). (7. (7. For (7. Jo = i. just sum (7. In particular. JT = i) = 'P.186 CHAPTER VI. ii (u) = it /3 P(W > u. and for (7. Now let In denote the environment when customer n arrives and I* the steadystate limit. (7.2) and use that limF (VT > u. The actual waiting time process 1W1. Then Pi(T(u) < T.1) 7ri In particular.2) Oi(u) = 1. JT = i) = P(V > u.4) where 0* = >jEE 7rj/3j. I* = i).. T) = 7ri 1 P. and (7.
if T is large. 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. N * oo. and (7.=i) a4. and further references (to which we add Prabhu & Zhu [296]) can be found there. with (7. P(. a paper relying heavily on classical complex plane methods.1 is from Asmussen [16]. With {Vt} denoting the workload process of the periodic queue.6) (7.5) follows from (7. and of these.3). In the setting of the periodic model of Section 6.. P(W >u. that /3(t).4). Lemoine [242].T)(T(u) <T) = P(8)(VT > u).7) of that paper. DUAL QUEUEING MODELS 187 Proof Identifying the distribution of (W.I.7. u Notes and references One of the earliest papers drawing attention to the Markovmodulated M/G/1 queue is Burman & Smith [84]. we have 1: I(W. I*) with the timeaverage . >u. see in particular Harrison & Lemoine [186]. P(1')(r(u) < oo) = P(')(00) > u). B(t) have been periodically extended to negative t).g. A more probabilistic treatment was given by Asmussen [17]. J* = i) see W > u.7) (7. . on average /32TP(V > u.8) For treatments of periodic M/G/1 queue. a general formalism allowing this type of conclusion is 'conditional PASTA'. T].3) improving somewhat upon (2. I* = i.o. n=1 N However. see Regterschot & van Doorn [123]. and one has PI'>(rr(u) < T) = P('_T)(VT > u).4) can be found in Regterschot & de Smit [301]. The first comprehensive solution of the waiting time problem is Regterschot & de Smit [301]. (7. Proposition 7.I *=i).l.4) and (7. p < 1 then ensures that V(*) = limNloo VN+9 exists in distribution. The relation (7. and Rolski [306]. [243]. on average 0*T customers arrive in [0. Taking the ratio yields (7.
This page is intentionally left blank .
resp . {Rt} moves according to the differential equation R = p(R). t] are Nt At = Ui (1. Zt As earlier. and that the claim sizes U1. Thus. z/i(u) = F IinffRt< 0IRo=u 1 (1.T) = FloinfTRt< OIRo=u1 denote the ruin probabilities with/initial reserve u and infinite. However .Chapter VII Premiums depending on the current reserve 1 Introduction We assume as in Chapter III that the claim arrival process {Nt} is Poisson with rate .2) tk(u. i&(u.d...i. the aggregate claims in [0. and the evolution of the reserve may be described by the equation Rt = u . are i. the premium charged is assumed to depend upon the current reserve Rt so that the premium rate is p(r) when Rt = r.At + p(R8) ds. and T(u) = inf {t > 0 : Rt < u} is the time to ruin starting from Ro = u so that '(u) = F(T(u) < oo). .6. U2.1) (other terms are accumulated claims or total claims). Thus in between jumps.T) = F(T(u) < T). 189 . with common distribution B and independent of {Nt}. finite horizon.
say e. there is positive probability. In this situation. Now return to the general model. Example 1. Hence in terms of survival probabilities. say at interest rate b.e. it seems reasonable to assume monotonicity (p(r) is u . i. we get p(r) = p + er. and the probability of absolute ruin with initial reserve u E [p/S. Assume 0(u) < 1 for some u.4 Either i. oo) is given by i (u + p/S). dividends are paid out at rate pi .2 (INTEREST) If the company charges a constant premium rate p u but invests its money at interest rate e. A basic question is thus which premium rules p(r) ensure that 'O(u) < 1. pi > p2 and p(r) = One reason could be competition. Thus at deficit x > 0 (meaning Rt = x).Vi(v) u > e(1 .i(u) = 1 for all u. but assume now that the company borrows the deficit in the bank when the reserve goes negative. Proposition 1. we can put Rt = Rt + p/S.2.'(u)) > 0 so that V'(v) < 1. or o(u) < 1 for all u. that {Rt} will reach level u before the first claim arrives.p2. Another could be the payout of dividends: here the premium paid by the policy holders is the same for all r. rather than when the reserve itself becomes negative. RESERVEDEPENDENT PREMIUMS The following examples provide some main motivation for studying the model: Example 1 . when x > p/S. Example 1. P(r) _ p + e(r . No tractable necessary and sufficient condition is known in complete generality of the model.p/S) r > p/S p5(p/5r) 0<r<p/5 Then the ruin problem for {Rt } is of the type defined above.1 Assume that the company reduces the premium rate from pi to p2 when the reserve comes above some critical value v.3 (ABSOLUTE RUIN) Consider the same situation as in Example 1. where one would try to attract new customers as soon as the business has become reasonably safe. the payout rate of interest is Sx and absolute ruin occurs when this exceeds the premium inflow p. However. but when the reserve comes above v. 1 . That is. Proof Obviously '(u) < ilb(v) when u > v. If Ro = v < u.190 CHAPTER VII.
296297): Theorem 1.2 once more.I3IB requires a more detailed analysis and that µB < oo is not always necessary for O(u) < 1 when p(r) 4 oo. we have z/i(u) <p(u . Starting from Ro = uo. then ?(u) = 1 for all u. (1.2) for r sufficiently large so that p(oo) = limr. . Then if u > no.1.3µB for all sufficiently large r.T) = P(VT > u). and hence by a geometric trials argument.1.1. In case (b). {Vt} decreases at rate p(v) when Vt = v (i. In between jumps. say V. if and only if V)(u) < 1 for all u. the probability that Rt < uo for some t is at least tp(0) = 1 (cf. In case (a). { Vt} remains at 0 until the next arrival).5 (a) If p(r) < /.. Hence ik(u) < 1 for all u by Proposition III. Here {Vt}two is a storage process which has reflection at zero and initial condition Vo = 0. that u zPp(u .6 For any T < oo. We next recall the following results.uo) < 1. Then 0(u) = P(V > u). However. Let Op(u) refer to the compound Poisson model with the same 0.b(u.2(d). Proof This follows by a simple comparison with the compound Poisson model.o(uo) = 1 so that t/'(u) = 1 for all u by Proposition 1. V = p(V)).uo) and. (b) If p(r) > /3µB + e for all sufficiently large r and some e > 0.4. (1.1. Theorem 1. and P(Rt + oo) > 0. [APQ] pp. let uo be chosen such that p(r) < p = /3µB for r > uo. (1. This is basically covered by the following result (but note that the case p(r) .+ p(r) exists.5) and the process {Vt} has a proper limit in distribution . which was proved in 11.f p(Vs) ds. let uo be chosen such that p(r) > p = 0ILB + e for r > uo.e. That is.1 and increasing in Example 1.2(d)). then l/i(u) < 1 for all u.2) we have t Vt = At . B and (constant) premium rate p. appealing to Proposition 111. one can couple the risk process and the storage process on [0. T] i n such a way that the events {r(u) <T} and {VT > u} coincide.6) . cf. obviously infu<uo z/'(u) > 0.3. hence Rt < uo also for a whole sequence of is converging to oo.4) 0 and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0. Proposition I1I. INTRODUCTION 191 decreasing in Example 1. instead of (1. In particular.
Sx} dx. for the storage process {Vt}.8) is the rate of downcrossings (the event of an arrival in [t.9) Proof We may rewrite (1. of (1. In view of the path structure of {V t }. oo).y. and the other being given by a density g(x) on (0. It is intuitively obvious and not too hard to prove that G is a mixture of two components.8 Assume that B is exponential with rate b. B(x) = e. say. yo ^ 1 + oo Q exp {. Note that it may happen that w (x) = oo for all x > 0. t + dt] if and only if Vt E [x.h.s.8) as the rate of upcrossings.y)g(y) dy. (1.s. t + dt] can be neglected so that a path of {Vt} corresponds to a downcrossing in [t.Qw(x) . Considering the cases y = 0 and 0 < y < x separately.192 CHAPTER VII.7 p(x)g(x) = tofB (x) + a f (x . say. the flow of mass from [0. say if p(r) goes to 0 at rate 1 /r or faster as r j 0. Jo AX) (1.7) Proposition 1. RESERVEDEPENDENT PREMIUMS In order to make Theorem 1. (1. Now obviously.6x and that w(x) < oo for all x > 0. we arrive at the desired interpretation of the r. x + p(x)dt]). of (1.Sx}. one having an atom at 0 of size 'yo.6w(x) . It follows in particular that 0(u) = fg(Y)dy. the l. Then w(x) is the time it takes for the reserve to reach level x provided it starts with Ro = 0 and no claims arrive. this means that the rate of upcrossings of level x must be the same as the rate of downcrossings.6 applicable. An attempt of an upcrossing occurs as result of an arrival.h. Oeax f x e'Yg (y) dy } = p) eaxa(x) . Corollary 1. where g(x) = p( ^ exp {. and is succesful if the jump size is larger than x .8) Proof In stationarity. Then the ruin probability is tp (u) = f' g(y)dy. x] to (x.8) as g(x) = p 1 {yo13e_6x +. we thus need to look more into the stationary distribution G. say when {Vt} is in state y. u Define ^x 1 w(x) Jo p(t) dt. oo) must be the same as the flow the other way.
1. INTRODUCTION
where c(x) = 1o + fo elyg(y) dy so that (x) = eaxg(x) _
193
1
p(x)
nkx).
Thus log rc(x) = log rc(0) + Jo X L dt = log rc(0) + /3w(x), p(t) c(x) = rc (0)em"lxl = Yoes"lxl, g(x) = eaxK' (x) = e6x ,Yo)3w'(x)e'6"lxl which is the same as the expression in (1.9). That 'Yo has the asserted value is u a consequence of 1 = I I G I I = yo + f g• Remark 1.9 The exponential case in Corollary 1.8 is the only one in which explicit formulas are known (or almost so; see further the notes to Section 2), and thus it becomes important to develop algorithms for computing the ruin probabilities. We next outline one possible approach based upon the integral equation (1.8) (another one is based upon numerical solution of a system of differential equations which can be derived under phasetype assumptions, see further VIII.7). A Volterra integral equation has the general form x g(x) = h(x) + f K(x, y)9(y) dy, 0 (1.10)
where g(x) is an unknown function (x > 0), h(x) is known and K(x,y) is a suitable kernel. Dividing (1.8) by p(x) and letting K(x, y) _ ,QB(x  y) _ 'YoIB(x) p(x) , h(x) p(x) we see that for fixed to, the function g(x) in (1.8) satisfies (1.10). For the purpose of explicit computation of g(x) (and thereby %(u)), the general theory of Volterra equations does not seem to lead beyond the exponential case already treated in Corollary 1.8. However, one might try instead a numerical solution. We consider the simplest possible approach based upon the most basic numerical integration procedure, the trapezoidal rule hfxN() dx = 2 [f ( xo) + 2f (xi) + 2f ( x2) + ... + 2f (XN1) + f (xN)1
p
194
CHAPTER VII. RESERVEDEPENDENT PREMIUMS
where xk = x0 + kh. Fixing h > 0, letting x0 = 0 (i.e. xk = kh) and writing 9k = 9(xk ), Kk,e = K(xk, xe), this leads to h 9N = hN + 2 {KN,09o+KN,N9N}+h{KN,191+'''+KN,N19N1},
i.e. 9 N=
hN+ ZKN ,ogo +h{KN,lgl+•••+KN,N19N1} 1  ZKNN
(
1.11
)
In the case of (1.8), the unknown yo is involved. However, (1.11) is easily seen to be linear in yo. One therefore first makes a trial solution g*(x) corresponding to yo = 1, i.e. h(x) = h*(x) = (3B(x)/p(x), and computes f o' g*(x)dx numerically (by truncation and using the gk). Then g(x) = yog*(x), and IIGII = 1 then yields f 00 g*(x)dx (1.12) 1= 1+ 'Yo from which yo and hence g(x) and z/'(u) can be computed. u
la Twostep premium functions
We now assume the premium function to be constant in two levels as in Example 1.1, p(r) _ J 1'1 r < v P2 r > v. (1.13)
We may think of the risk reserve process Rt as pieced together of two risk reserve processes R' and Rt with constant premiums p1, P2, such that Rt coincide with Rt under level v and with above level v. For an example of a sample path, Rt see Fig. 1.1.
Rt
V
Figure 1.1
1. INTRODUCTION
195
Proposition 1.10 Let V)' (u) denote the ruin probability of {Rt}, define a = inf It > 0 : Rt < v}, let pi ( u) be the probability of ruin between a and the next upcrossing of v (including ruin possibly at a), and let q(u) = 1  V" (u) Then
1  q(u) + q ( u)z,b(v) p1(v) u = 0<u<v v
0 < u < v. (1.14)
1 + pi (v )  '02 (0) pi (u) + (0, (u  v)  pi (u)) z/i(v ) v < u < oo.
Proof Let w = inf{ t > 0 1 Rt= v or Rt < 0} and let Q1 (u) = Pu(RC,, = v) be the probability of upcrossing level v before ruin given the process starts at u < v. If we for a moment consider the process under level v, Rt , only, we get Vil (u ) = 1  q, (u ) + g1(u),O1( v). Solving for ql (u), it follows that q1 (u) = q(u). With this interpretation of q(u) is follows that if u < v then the probability of ruin will be the sum of the probability of being ruined before upcrossing v, 1  q(u), and the probability of ruin given we hit v first , q(u)z'(v). Similarly, if u > v then the probability of ruin is the sum of being ruined between a and the next upcrossing of v which is pl (u), and the probability of ruin given the process hits v before ( oo, 0) again after a, (Pu(a < oo )  p1(u))''(v) = (Vi2(u  v)  p1 (u))''(v)• This yields the expression for u > v, and the one for u = v then immediately follows. u Example 1 .11 Assume that B is exponential, B(x) = e62. Then
01 (u)
_
0 e .yiu ,,2 (u) = )3 e 72u p1S P2S
1  ~ ery1u p1S 1  Q eryly P1S
where ry; = S  ,Q/p;, so that
q

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

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

u + Z =
(u + Zr ) + ( Z  Zr) = e
ET {e
(u + Zr)  f '* eE(tT )dSt] T
e
ET [
R( u)
+ Z`],
2. THE MODEL WITH INTEREST
199
where Z* =  K* eE(tT)dSt is independent of F, and distributed as Z. The last equality followed from Rt") = eEt(Zt + u), cf. Proposition 2.1, which also yields r < oo on {Z < u}. Hence H(u) = P(u + Z < 0) = P(RT + Z* < 0; r < oo) zb(u)E [P(RT + Z* < 0 I)7T, r < oo)] _ O(u)E [H(RT(")) I r(u) < oo] .
Corollary 2.4 Assume that B is exponential, B(x) = e6', and that p(x) _ p + Ex with p > 0. Then
. o€Q/E Ir, (8(p + cu);
V) (u)
aA/Epal Ee 6n1 E +^3E1 / E
1\ E E
1r
Cbp;
E El al
where 1'(x; i) = f 2°° tnletdt is the incomplete Gamma function. Proof 1 We use Corollary 1.8 and get
w(x) fo P + Etdt = g(x) = p +0x
e log(p + Ex)  e loge,
exp {  log(p + Ex)   log p  6x }
pal(p + ex)plE1e6^ J ryo)3 70 = 1 + J p) exp {Ow(x)  Sx} dx x r^ = 1+ ' /E (p + Ex)01'leax dx + 0
f J
= 1+
a
Epo/ E
f yI/ E 1e 6(Y P)/E dy
P (
1+ OEA/E 1e6 P /Er
60/e po/ e
,;,3 )
E E
lp(u) = to foo a exp {w(x)  bx} AX)
acO/E" 1 ePE l
Yo
50 1epolE
(
+ cu); 0)
5(p
E E
3a/ (5 .2) follows by elementary algebra.x) dx e.b P/E dx /' P/ ' (p/  x)p/e 150/f I' (/3/E) (6P1'E. i. of Z is IogH[a] = f ytc(y)dy = e fa (0.3 is also valid if {Rt} is obtained by adding interest to a more general process {Wt} with stationary independent increments. 13/E). RESERVEDEPENDENT PREMIUMS u from which (2.a) . RT(u) has an exponential distribution with rate (S) and hence E [H(RT(u))I r(u) < oo] L Pe6'r (P/C . r (j3/E) In particular. From ic(a) = .e. /^ u Example 2 . As an example.pa. The process {St} corresponds to {Wt} so that c(a) or2a2/2 . Proof 2 We use Theorem 2.g.5 The analysis leading to Theorem 2./3 log(b + a)] = log ePa/f (a + a ) e which shows that Z is distributed as p/E . then {Rt} is the diffusion with drift function p+Ex and constant variance a2.V.3/E) By the memoryless property of the exponential distribution.13 /E) r (. .f. with density x(3/e1aQ/e fV (x) _ e 6X ' x > 0.pa.2y +µ ) dy .01'E) + (p/E)al aO l febP/E } IF (0 /0 jF From this (2. 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. 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 .3. and the c. where V is Gamma(b.V < x)]0 + f P(V > p/E ) + eby fv (p/E .2) follows by elementary algebra.200 CHAPTER VII.
[129] and Harrison [185]. that the analysis does not seem to carry over to general phasetype distributions. Gerber [157] p. as in the proof of Proposition 2. Emanuel et at. It must be noted. it is also used as basis for a diffusion approximation by these authors. or to nonlinear premium rules p(•). Z is normal (p/E.4 is classical. Paulsen [281]. [282]. not even Erlang(3) or H3.Y*p* W*(u) < ery*u = 0. for a martingale proof. 134 (the time scale there is discrete but the argument is easily adapted to the continuous case).e.2 is a special case of a perpetuity. [129]. [283]. 3 The local adjustment coefficient. of the form Ei° p"X" with the X„ i.p*. THE LOCAL ADJUSTMENT COEFFICIENT _ Q2a2 pa 4e E 201 I.i.3. Q2/2E). and since RT = 0 by the continuity of Brownian motion. [357]. The solution is in terms of Bessel functions for an Erlang(2) B and in terms of confluent hypergeometric functions for a H2 B (a mixture of two exponentials). and recall Lundberg 's inequality . Delbaen & Haezendonck [104].8.d.3 is from Harrison [185].3) was derived by Emanuel et at. see e.g. Further studies of the model with interest can be found in Boogaert & Crijns [71]. it follows that the ruin probability is Cu) H(u) H(0) 11 Notes and references Theorem 2.1) . se e. A r. Some of these references also go into a stochastic interest rate. Corollary 2. Paulsen & Gjessing [286] and Sundt & Teugels [356]. Paulsen & Gjessing [286] found some remarkable explicit formulas for 0(u) beyond the exponential case in Corollary 1..v. Gerber [155]. write y* for the solution of the Lundberg equation f3(B[ry *] ..g. Goldie & Griibel [167]. Logarithmic asymptotics For the classical risk model with constant premium rule p(x) . however. write Vi* (u) for the ruin probability etc. The formula (2.
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 .'y ( x)) = 0 where r.2) such that p(x) < c(. 1) and for a given E > 0.*(u) .4) we assume existence of y(x) for all x.1 Assume that for some 0 < 5o < oo. x>0 (3. Then lim sup u>oo u and e E''p(r) + 0. (3.1 ). let p* be a in (3.i)eex.1) .e. log ?i(u) < < 00 JO .1. Letting first E * 0 and next ry * T 5o yields the first statement of the theorem. choose uo such that p( x) > p* when x > u0E. c(. Let y* < So.3) When trying to extend these results to the model of this chapter where p(x) depends on x. (3. The intuitive idea behind introducing local adjustment coefficients is that the classical risk model with premium rate p* = p(x) serves as a 'local approximation ' at level x for the general model when the reserve is close to x. as solution of the equation n(x.w (u) J dt > c(3)eeu v 1 p(u+ t) . and (for simplicity) that inf p(x) > (3µs .1. then log u (u) In the proof as well as in the remaining part of the section . For the last asssertion . x * oo. i. If 60 s f 6o.C*ef*".E). as will hold under the steepness assumption of Theorem 3. The steepness assumption and p(x) + oo ensure 'y(x) * So. e(1o+e)2 (x ) u > 00. choose c(. (x.log '(u)/u < ry*(1 . the function y(x) of the reserve x obtained by for a fixed x to define y(x) as the adjustment coefficient of the classical risk model with p* = p(x). which in turn by Lundberg's inequality can be bounded by ery*(1E)" Hence limsup„.. and that p(x) * oo. it holds that f3[s] T oo. RESERVEDEPENDENT PREMIUMS and the CramerLundberg approximation V. a first step is the following: Theorem 3 . When u > uo.e. we will use the local adjustment coefficient 'y(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) . Proof of Theorem 3.>o 7(x) > 0. oo for all E > 0.5) which implies inf. i.202 CHAPTER VII.ap(x). B(x) > C(2)e(ao+f)x for all x. a) = f3(B[a] ..
let 0e (u) be evaluated for the process only with 3 replaced by /0/e and U. UJU > x cannot have a much heavier tail than the claim U itself. ruin will occur if the claim is at least u + v. Theorem 3 . Condition 3.13 below holds.13 is a technical condition on the claim size distribution B. 3) = (1 . However.v.2 is also an approximation under appropriate conditions. u Obviously. (3. Theorem 3.ea°/(ecf1)). (u) = O(u/e).0 are the same. For e > 0. the asymptotics u * oo and c . and hence '(u) > c(4)eeuc( 2)e(do+e)u The truth of this for all e > 0 implies lim inf log V. THE LOCAL ADJUSTMENT COEFFICIENT 203 where c.3 Assume that either (a) p(r) is a non decreasing function of r. which essentially says that an overshoot r.1 only presents a first step. 2 Assume that p(x) is a nondecreasing function of x and let I(u) = fo ry(x)dx. (3. .3 to be reasonably precise and use e` (u) as approximation to 0 (u). The form of the result is superficially similar to the CramerLundberg approximation. 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. one can then assume that e = 1 is small enough for Theorem 3.4)eE" Given such an arrival. then Rte) = CRtie for all t so that V). Bucklew [81]).3.(u) > so. and in particular.g. Then .7) CIO Remarks: 1. the result is not very informative if bo = oo.2).e. the limit is not u + oo but the slow Markov walk limit in large deviations theory (see e.6) The second main result to be derived states that the bound in Theorem 3. I.' (u) < eI("). Therefore the probability that a claim arrives ( before the reserve has reached level u + v is at least c(. by cU2. or (b) Condition 3. noting that in many cases the constant C is close to 1. The rest of this section deals with tail estimates involving the local adjustment coefficient. If p(x) = pis constant .. {Rte)} defined as in (1. 2. 3. The first main result in this direction is the following version of Lundberg's inequality: Theorem 3 . Then limelog l/ie (u) = I(u).
2.6).8 in terms of I(u) when the claims are exponential: Example 3 .bu}.bx}]o + b /' oo exp low (x) .bx} dx oo exp low(x) bx dx 70 Ju r oo = b J exp low (x) .204 CHAPTER VII. rather than eI(u)).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).exp {/33w(u) .8.3. say. 5.(iw(x) .bx} dx = 1 + J0 dodx(x) exp {.4 Consider again the exponential case B(x) = eax as in Corollary 1.bx} dx .3. First. we consider some simple examples. As typical in large deviations theory. 3a Examples Before giving the proofs of Theorems 3. it is formally needed only for Theorem 3. RESERVEDEPENDENT PREMIUMS 4. we get = 1+ J" AX) exp {(3w (x) . Then y(x) = b . and r j 1 'Yo v(x)dx = bu  a J0 p(x)ldx = Integrating by parts. One would expect the behaviour in 2) to be important for the quantitative performance of the Lundberg inequality (3.bx} dx fo 00 1 + [exp {/(3w(x) .(3/p(x). the logaritmic form of (3. 3. However.bx} dx 1+0.(x) dx. u .1 + b f e. J0 ^oo g(x ) dx f AX) lexp IOW (X ) bx + b u 1 exp low(x) . we show how to rewrite the explicit solution for ti(u) in Corollary 1.
1.I ( u) fool. and (3.fory(x+u)dxdy ( 3. Similarly.2. For Theorem 3. Choosing yo. u .e. and (3. (u) = I(u) + AE .8) 7(x)dxdy 11 We next give a direct derivations of Theorems 3. 3. BE * 0.10) where AE = e log 000 e. in the definition of AE converges to 0. IE(u) = I(u)/e. (X) = µ(x)..fa 7(x+u)dx/Edy o The analogue of (3. the integral is bounded by 1 eventually and hence lim sup AE < lim sup a log 1 = 0.9 ) 11000 eI(v)dy f000 e.2(X) = ev2(x) so that 7e(x) = 7(x)/e. In particular.ev 0 O /E) J0 70 70 Yo This implies lim inf A. (3. It is well known that (see Theorem XI.. (3. 70 > 0 such that 7(x) < 7o for y < yo. oo) with drift µ(x) and variance a2 (x) > 0 at x. 0.f y(x)dxd y If 7(x) is increasing .10 or Karlin & Taylor [222] pp.2.fo 7(x) dx /E dy > av 'yo /Edy = E (1 .9) yields e log . 191195) that 1P (U) = fu0 eI(v)dy = eI(u) follo e. THE LOCAL ADJUSTMENT COEFFICIENT and hence 205 f°° eI(v )dy .3.1/8 . applying the inequality 7(x + u) > 7(x) yields immediately the conclusion of Theorem 3. > lime log e = 0 and AE * 0. 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. note first that the appropriate slow Markov walk assumption amounts to u. Be = e log U000 e.0.I ( v )dy fo +u) dxdy . we get r 00 e.fo 7(x)dx/Edy f .BE.5) is infx>o 7(x) > 0 which implies that f °O . ry(x /b I u o e f0 °° e  e.7) follows.3.5 Assume that {Rt} is a diffusion on [0.3 in the particularly simple case of diffusions: Example 3.
e.+1 (u) = o( 1). Further. RESERVEDEPENDENT PREMIUMS The analogue of Example 3. .5) and 7* = 5 .. 0.206 CHAPTER VII. > . .7) follows just as in Example We next investigate what the upper bound / approximation aI (°) looks like in the case p(x) = a + bx (interest) subject to various forms of the tail B(x) of B. _ . that the interchange of the slow Markov walk oo is not justified and in fact.. the slow Markov walk assumption means 5E = b/c. the results are suggestive in their form and much more explicit than anything else in the literature.5./3 1 AX dx. G. As in Example 3. + Gq(u) + o(G9(u))• Gi (u) It should be noted . (u) representing the first few terms in the asymptotic expansion of I(u) as u + oo. lim sup Af < lim sup c log(1 ..10) holds if we redefine AE as AE = flog (j °° efo 7(x)dx/edy _ E/5 I and similarly for B. I(u ) = G1(u) + .0) = 0. ) Note that this expression shows up also in the explicit formula for lk(u) in the form given in Example 3. we have 5 > 7o and get lim inf AE > lime log e .5. this leads to (3. however . Then the solution of the Lundberg equation is y* = b .5 for risk processes with exponential claims is as follows: Example 3 . Thus 7e(x) _7(x)/e and (3. E+o e*O By (3.6 Assume that B is exponential with rate S. 7(x) is typically not explicit.6) exactly as in Example 3. Ignoring 1/5 in the formula there. 0 Now (3. G.7o C 15 I I. the slow Markov limit a * 0 and the limit u walk approximation deteriorates as x becomes large. Of course.5.0. G.(u) oo.4.Q/p*..6/p* so that u 1 I (U) = bu .0/e. . so our approach is to determine standard functions Gl (u). . Nevertheless .1 3. I. .
7 Assume that B(x) .1 =$ f cse8 Sn f e"B(x)dx = e8 Jo s eIB ( 1 . say 1 is the upper limit and B(x) . . Here B[s] is defined for all s and B[s] . if the phase generator is irreducible ( Proposition VIII.y/s)dy sn 1 1 f ' evy'7ldy = cse8r(T7) as s T oc. e.c3 logu a= 1 J 0 a + bx 1/ ( c4ul 1/° a > 1 where c3 = c2 /b. ry* loge*+ g7loglogp*.1) leads to (S7T N Ocp a.3.4) or gamma distributions.:. 1) and 17 = k + 1 if B is the convolution of k uniforms on (0.8). More precisely.x)n1.s)C' f "o o as s T S. u(logu + r7loglogu). 1. c4 = c2b 1/'/(1 .g.11) with a > 0. 2. For example.1/a).c2 Su a dx ) Su a<1 Su .3cse7*I7(77) .1) leads to . .C2p* C2 = (3clr( a))11'..1/k). the typical case is a = 1 which holds . It follows from (3. THE LOCAL ADJUSTMENT COEFFICIENT Example 3 .ry*°p*. and hence (3.clxale5x 207 (3. x T 1.11) that b[s] * co as s f S and hence 7* T S as p* + oo. 77 = 1 if B is degenerate at 1. fu I(u) Su . u Example 3 . y = 2 if B is uniform on (0. in the phasetype case . B[s] = 1 + s exB(x)dx = 1 +c1SF(a) ('+o(')) (S .cs(1 .Y . This covers mixtures or convolutions of exponentials or.8 Assume next that B has bounded support. I(u) Pt. (3. Hence (3. phasetype distributions (Example 1. more generally.12) with y > 1.
of the increment in a small time interval [0. h].13) We get b[s] .css 2%rc7eC782/2. g.u is a nondecreasing function of u.1) . of U1 + v .sp(u).Ul up to the first claim (here ru (•) denotes the solution of i = p (r) starting from ru(0) = u)..14) for the m . 3b Proof of Theorem 3. RESERVEDEPENDENT PREMIUMS Example 3 . (3.r^.u . ec78)2/2c7 dx C7 .Ote7o( u)(u.f.10 Assume that p(x) is a nondecreasing function of x.•. I (u) c8u log u 0 where c8 = 2/c7. 7 * . (3.(t))dt. one could also have considered the increment ru (T1) .12). This leads to an alternative local adjustment coefficient 7o(u) defined as solution of 1 = Ee''o(u)(vi+u .B[7o (u)] . (b) 'y(x) <'Yo(x)• Proof That 7(x) is nondecreasing follows easily by inspection of (3. assume that B(x) CO x2/2c7. Then: (a) y(x) and 7o(x) are also nondecreasing functions of x.1 Cgs o"O 0 esxex2/2c7 dx = cgsec782/2 f .(T1)) > Ee7o(u)(ul+vr»(Ti)).r„(Ti).3 (B[s] .f.15) Proposition 3.g. Hence for u<V.9 As a case intermediate between (3.4) is the formula h logEues ( Rhu) . (3. x f oo .4). this is only possible if 7o(v) 2 7o(u)• . .e7o ( u)(ul+u r.2 We first remark that the definition (3.208 CHAPTER VII. h 10. The assumption implies that ru(t) . 1 0 3e. By convexity of the m .log p*.4) of the local adjustment coefficient is not the only possible one: whereas the motivation for (3. 1 = E.c8 log .11) and (3.ru(TI)) .
(3.3.u[70(u)] fo eyo(x)dx . Also. Assume (3. note that the assumption implies that ru(t) . The case n = 0 is clear since here To = 0 so that ik(°)(u) = 0.1) . fa 7o(y)dy < u7o(u) < xyo (u) for x > u.16) Proof Define 411(n)(u) = P('r(u) < on) as the ruin probability after at most n claims (on = TI + • • • + Tn).11 Assume that p(x) is a nondecreasing function of x.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. Hence „/.10(b): Theorem 3. it is easily seen that fu x7o(y)dy < xyo (u). we obtain „I. We shall show by induction that (' Y'(n) (u) < e fo 'Yo(x)dx (3.4) considered as function of 7 is convex and 0 for y = 0.7o (u)p(u)• Since (3.u > tp(u). Separating after whether ruin occurs at the first claim or not.(n+l) (u) 1 . THE LOCAL ADJUSTMENT COEFFICIENT For (b).(n+1) (u) efo Yo(x)dxI^"Q exyo( I u u)Fu(dx )+ J . this is only possible if yo(u) > 7(u).17) shown for n and let Fu(x) = P(U1 + u .es'Yo(u)Fu(dx)} o0 e fo yo( x)dx j.e70(u)(U1P(u)T1) 209 0 + 7o(u)p(u)' 0 <_ 00['Yo( u)] . Hence 1 = EeYo(u)(U1+uru(T1)) < E.Fu(u ) + J ^(n)(u . u We prove Theorem 3.17) from which the theorem follows by letting n + oo. the case of 7 then follows immediately by Proposition 3.ru(T1) < x).2 in terms of 7o. Then (u) < efo Yo(x)dx.
RESERVEDEPENDENT PREMIUMS where the last identity immediately follows from (3.. in accordance with the notation i/iE (u).e (u) = v'. we have chosen to work with y(u) as the fundamental local adjustment coefficient. and.n) pn niE (u /n) n n_1 n.n so that n.x/n. {RtE)} (starting from u = un.11 be reasonably tight something like the slow Markov walk conditions in Theorem 3. given downcrossing occurs. C*e where the first equality follows by an easy scaling argument and the approximation by (3..210 CHAPTER VII.n. define uk. (3. However. ryk. Also. W O .nbe C*. Let Ck.n (starting from u/n) without that 2u/n is upcrossed before ruin. P k.n <Z auk}l.n = ku. Proof For ruin to occur. resp. the value of {R(E)} at the time of downcrossing is < unl.E ( u/n) ^•e. we used also Proposition 3. y* evaluated for p* = Pk. 3.2.. let Op*.3 The idea of the proof is to bound { R( f) } above and below in a small interval [x .n = sup p(x). (un2. the probability that ruin occurs in the CramerLundberg model with p* = pn.15).n) must first downcross unl. in .n inf n uk1. Y*u /E.2. To this end...10(a) for some of the inequalities. op*.3). 0.E (u/n).12 lim sup4^o f log O.E (u) denote the ruin probability for the classical model with 0 replaced by .3/e and U. pk n = uk_l.2)..n. Further. and here it is easily seen that yo(u) .E (u/ n) Y'E (un .3).: y(u). (u) < I(u).n. For these reasons.11 is sharper than the one given by Theorem 3.10(b ) that the bound provided by Theorem 3. (u). for either of Theorems 3.n u k}1.n) > k =1 II v ^k n. yo(u) appears more difficult to evaluate than y(u).E (u/n) Now as e . Lemma 3.n AX). 0.3 is required.I.n. The probability of this is at least n n. 0 It follows from Proposition 3. x + x/n] by two classical risk processes with a constant p and appeal to the classical results (3. 3c Proof of Theorem 3. by €U=.
n <X<Uk. also ryk. 11 Theorem 3.i.n = sup ?'(x).19) .n. i.! (u/n) n n m 7k. k=1 k=1 n u _ nE7 k.log Ck. we need the following condition: Condition 3.E (urn) < \ *I.13 There exists a r.n + 0(1). 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. 40 Combining with the upper bound of Lemma 3. in obvious notation one has tC (x) = y(x)/e.a( u)z.nu/en(1 + where o(1) refers to the limit e . v..n .E (u/n) OP +^p•. ne7k. y > 0 it holds that F(U>x +yIU>x) B(x + y) < F (V > y).nu /En) o(1)). . B(x) (3.n cE (2u/n) Ck ne7k. It follows that n log V'C (u) k =1 log Ypk. *p•.nu /fn( 1 Ck  e. /' (u/n) 'T nk.. 3 now follows easily in case (a). In case (b).18) (ii) the family of claim overshoot distributions is stochastically dominated by V. (3. 211 Clearly.. since ry' is an increasing function of p'. uk_1.3.F (2u/n). (u) CIO < Letting n 4 oo and using a Riemann sum approximation completes the proof. THE LOCAL ADJUSTMENT COEFFICIENT particular.. for all x .12 completes the proof. 0 with n and u fixed. so that Theorem 3.E (u/n) Op•. Indeed .e.nk=1 limsupelogv).7k.2 gives 7PE (u) < eIi"i/f = lim inf Clog 0E (u) > I (u).
E (u/n . u/n) < oo) EV). (u/n .E (2u/n .v..nu /EnE [e71.5) and the standard formula for b(0).V) = e71 nu/Eno(l) (using (3. u /en 0(i) _n so that E2 < e2ryl nu/En0(1).1 n. u/n) < oo] . v ) = inf { t > 0 : R(c ) < v R) = u } . Write EO.EV) = El + E2.18) for the last equality). Ei + E2 < e71. u/n)) . (R. (3.. ) (u u /n)) .QEU 1 .R<) (u v).( .^(E) (u.n < ery1. we first note that the number of downcrossings of 2u/n starting from RoE) = 2u/n is bounded by a geometric r. where El is the contribution from the event that the process does not reach level 2u/n before ruin and E2 is the rest. RESERVEDEPENDENT PREMIUMS To complete the proof. v ) = v . For E2.^'' = E [ . V < u/En] + P(V > u/En) (u/En . Then Y'E (u) ^(E) (u. N with EN < 1 = infx>2u/nA(x) = 0(1).n V.E (0) cf. infx>2u /n P(x) . let v < u and define T(E) (u.nu/En0(1) . Then the standard Lundberg inequality yields El < E?.. The probability of ruin in between two downcrossings is bounded by Epp .E(E) (u.212 CHAPTER VII.EV) = e.eV) • P (T(E) (u. u/n)) I T(E) (u.2y 1 ' . u/n) < oo) .of:>2 in n(x). P (T(E) (u. u/n) < oo] E [OE (u/n . T(E) (u. T() (u. .EV) = EiI 1 .. u /n) < oo] l = = < E [OE (u/n . (u/n .
l o JJJ o . [89].J r(Rs)p(R. Whereas the result of [122] is given in terms of an action integral which does not look very explicit. they also discuss simulation based upon 'local exponential change of measure' for which the likelihood ratio is ( /'t /'t Ns Lt = exp S . the results are from Asmussen & Nielsen [39]. T) is maximized over T by taking T as the time for (3.21) (the initial condition is r(0) = u in both cases). whereas the most probable path leading to ruin is the solution of r(x) _ k (x.u/n) < oo) CI  > u n n ryi n' i=1 Another Riemann sum approximation completes the proof.g.1. Typically.3EU) (3. THE LOCAL ADJUSTMENT COEFFICIENT Hence lim inf e log Ali. the approximation (3.4) and the prime meaning differentiation w.r.7(x)) (3. it might be possible to show that the limits e .7) for ruin probabilities in the presence of an upper barrier b appears in Cottrell et al. u/n) < oo { 40 )I U nryl n+liminfelogP (T(')(u. 0 and b T 00 are interchangeable in the setting of [89].T) = P „(info<t <T Rt < 0) via related large deviations techniques.t. Bucklew [81]). one can in fact arrive at the optimal path by showing that the approximation for 0(u. 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.)ds + Y(R2.3. (u) 40 213 lim inf e log(Ei +E2) + logP (r(`) (u.21) to pass from u to 0. u Notes and references With the exception of Theorem 3.7) then comes out (at least heuristically) by analytical manipulations with the action integral.13.) = exp . Similarly. s) as in (3.J y(Rs)dR. Comparing these references with the present work shows that in the slow Markov walk setup.20) (with ic(x. s). the risk process itself is close to the solution of the differential equation r(x) _ r (x. Djehiche [122] gives an approximation for tp(u. .)Ui } . where the key mathematical tool is the deep WentzellFreidlin theory of slow Markov walks (see e . 0 ) (= p(x) .=1 J An approximation similar to (3.
.3.g. RESERVEDEPENDENT PREMIUMS the simplest being to require b[s] to be defined for all s > 0 (thus excluding . . see XI. For different types of applications of large deviations to ruin probabilities .214 CHAPTER VII. e. however. We should like. to point out as a maybe much more important fact that the present approach is far more elementary and selfcontained than that using large deviations theory. the exponential distribution ).
A distribution B on (0. on Eo = E U {A} where A is some extra state which is absorbing. that is.1) is 'Here as usual . we write Pv for the case where Jo has distribution v so that Pv = KER viPi• 215 . A proper knowledge of phasetype distributions seems therefore a must for anyone working in an applied probability area like risk theory. 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. if v = (vi)iEE is a probability distribution. This implies in particular that the intensity matrix for { it } can be written in blockpartitioned form as T 0 0 . 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. and not in other cases. if a problem can be solved explicitly when the relevant distributions are exponentials. F (Jt = A eventually) = 1 for all i E E 1 and where all states i E E are transient. refers to the case Jo = i. Note that since (1. More precisely. Typically. 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<. P. We often write p for the number of elements of E.Chapter VIII Matrixanalytic methods 1 Definition and basic properties of phasetype distributions Phasetype distributions are the computational vehicle of much of modern applied probability.
j.T)) if B is the Padistribution of the absorption time C = inf{t > 0 : it = A}. E = {i. (1. Here are some important special cases: Example 1 . and the phasetype distribution is the lifetime of a particle with constant failure rate /3.e. k}. T is a subintensity matrix2.1 The phase diagram of a phasetype distribution with 3 phases. i. a.0 = t11. MATRIXANALYTIC METHODS the intensity matrix of a nonterminating Markov process. the exit rates ti and the transition rates (intensities) tij: tj 3 aj ai i ti tk tjk FkJ ak Figure 1.1 Suppose that p = 1 and write . T) (or sometimes just (a.e.216 CHAPTER VIII. i. tij > 0 for i 54 j and EjEE tij < 0 .3. that is. the ith component ti gives the intensity in state i for leaving E and going to the absorbing state A. an exponential distribution with rate parameter . t1 = /3. 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. A convenient graphical representation is the phase diagram in terms of the entrance probabilities ai. Equivalently. C is the lifetime sup It > 0 : Jt E E} of {Jt}. B(t) = Fa(^ < t ). The initial vector a is written as a row vector. In particular.2) The interpretation of the column vector t is as the exit rate vector. Thus the phasetype distributions with p = 1 is exactly the class of exponential distributions. and we have t = Te. Then a = a1 = 1. 0 2this means that tii < 0. We now say that B is of phasetype with representation (E.
p}..3 The hyperexponential distribution HP with p parallel channels is defined as a mixture of p exponential distributions with rates 51. 0 S 6 Example 1. 0 0 0 0 S 6 . a = (1 0 0 ... 0 SP 0 and the phase diagram is (p = 2) . . ....1.2 corresponding to E = {1. . 0 0 0 T= t= 0 ••• S S 0 0 0 0 0 0 . 0 ••• 0 0 Sp1 0 0 t= 0 0 00 •.x i=1 Thus E _ Si 0 T 0 S2 0 0 ... the EP distribution may be represented by the phase diagram (p = 3) Figure 1.. . 00)) S s o . . 6. so that the density is P E ai6ie6. ..1)!e Since this corresponds to a convolution of p exponential densities with the same rate S. PHASETYPE DISTRIBUTIONS 217 Example 1.... 0 •..2 The Erlang distribution EP with p phases is defined Gamma distribution with integer parameter p and density bp XP1 6x (p.
5 Let B be phasetype with representation (E. MATRIXANALYTIC METHODS Figure 1.1)"n! aT"e. a. T).e. j E E.g.t2 yt bP.4 For example. The basic analytical properties of phase type distributions are given by the following result . p:.g. see A. and is defined as the class of phasetype distributions with a phase diagram of the following form: 1 617 ti t2 2 b2. dp.3 0 Example 1 .f. i. B[s] = f0°O esxB (dx) is a(sI T)lt (d) the nth moment f0°O xnB(dx) is (. Proof Let P8 = (p ^) be the sstep EA x EA transition matrix for {Jt } and P8 the sstep E x Etransition matrix for {Jt} . 36) yields s d.aeTxe. Then for i .3 . the Erlang distribution is a special case of a Coxian distribution. (b) the density is b(x ) = B'(x) = aeTxt. [APQ ] p.4 (COXIAN DISTRIBUTIONS) This class of distributions is popular in much of the applied literature. ds^ = ds' = ttlaj + tikpkj.d. the backwards equation for {Jt} (e.1 tP1 1 Figure 1. (c) the m. Theorem 1 . Then: (a) the c. .218 CHAPTER VIII. Recall that the matrixexponential eK is defined by the standard series expansion Eo K"/n! 3. the restriction of P8 to E. E t ikp kj = kEE kEE 3For a number of additional important properties of matrixexponentials and discussion of computational aspects .f is B (x) = 1 .
s) is the m .f. of the initial sojourn in state i.5) Indeed .p.f. d8 P8 = TP8. in which case the time to absorption is 0 with m .tii tii .g. we i w.g.jEE B'(x) _ cx Pxe = aeTxTe = aeTxt (since T and eTx commute). ti/ . j#i jEE tijhj + his = ti. define hi = Eie8S.T) 1t..1 ) n +l n ! a (s I + T ) . Part (d) follows by differentiating the m. the rule (A. i.n1t = (1)nn!aTn1Te (1)nn! aTne.tii and have an additional time to absorption either go to state j which has m . 1.5) ki = 1 + tii L jj:Ai tii (1. Alternatively.s I .tii is the rate of the exponential holding time of state i and hence (tii)/(tii .6) .s j# tii i (1. we arrive once more at the stated expression for B[s]. = aPxe.5) as hi(tii + s) = ti  t ij hj. the solution is P8 = eT8.e. h = (T + sI)1t. After that.f. this means in vector notation that (T + sI)h = t. For (c). for n = 1 we may put ki = Ei( and get as in (1. PHASETYPE DISTRIBUTIONS 219 That is. (1) n+1n!aT . and (b) then follows from 1: aipF. Then h tit ti + ti3 h j .g.12) for integrating matrixexponentials yields B[s] = J esxaeTxt dx = a ( f°°e(81+T)dx ) t a(sI . tij / .T) 't = (. B(n)[0] = _ Alternatively. and since obviously P° = I.p. i.g. hj . Since 1 . or w. (Jx E E) = this proves (a).n lt .1. Rewriting ( 1..f. d" dsn a (.tii we go to A.B(x) = 1'a (( > x) = P. . and since b[s] = ah.
s. 0 Example 1.h. Consider for example 3 9 a= (2 2). are idempotent.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.6 Though typically the evaluation of matrixexponentials is most conveniently carried out on a computer. there are some examples where it is appealing to write T on diagonal form. MATRIXANALYTIC METHODS which is solved as above to get k = aTle. another the case p = 2 where explicit diagonalization formulas are always available. T= 2 111 so that 2 2 Then (cf. This implies that we can compute the nth moment as (1)"n! aT "e 1"n! 1 1 22 9 9 10 70 7 1 10 10 1 9 +6. One obvious instance is the hyperexponential distribution. see the Appendix. making the problem trivial. we get the density as 9 9 6 (1 1) 10 7 1 0 10 2 aeTyt = e x . Example A3."n! ( ( l 2 2 ) 17 9 0 \ 1 / 10 10 32 n! 35 6" +n!353 Similarly.220 CHAPTER VIII.
i.7 If B is phasetype with representation (v. 5 and serves at this stage to introduce Kronecker notation and calculus (see A.4b for definitions and basic rules): Proposition 1. .11aDD. 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). • The phasetype distribution B is zeromodified. (1. T) is then defined to be oo on a set of probability 1. then the matrix m.1.7) Proof According to (A.4. < 1.e 11BIJ = 1laDD < 1.29) and Proposition A4. and in fact one also most often there allows a to have a component ao at A. where the initial vector a is substochastic. a random variable U having a defective phasetype distribution with representation (a. or one just lets U be undefined on this additional set.e a mixture of a phasetype distribution with representation (a/llall.T) with weight hall and an atom at zero with weight 1 . hail = E=EE a.T).f. i. 0 Sometimes it is relevant also to consider phasetype distributions.hall. 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. B[Q] of B is f3[Q] = J e'1zB(dx) _ (v (9 I)(T ® Q)1(t ® I). This is the traditional choice in the literature. There are two ways to interpret this: • The phasetype distribution B is defective.g.
B[s] = p(s)/q(s) to be phasetype: the density b(x) should be strictly positive for x > 0 and the root of q(s) with the smallest real part should be unique (not necessarily simple. h can be chosen with strictly positive component.Ce7'. one has k = 0. MATRIXANALYTIC METHODS la Asymptotic exponentiality Writing T on the Jordan canonical form. 0 Of course. T). i is real and positive. See in particular the notes to Section 6. let . Neuts. and we have eTx . Rolski. Lipsky [247]. not only in the tail but in the whole distribution. v.4c). All material of the present section is standard. cf. the result follows (with C = (ah)(ve)).222 CHAPTER VIII. but in many practical cases. assume that T is irreducible . Then the tail B(x) is asymptotically exponential.. Using B(x) = aeTxe . B(x) . distributions with a rational m. see his book [269] (a historical important intermediate step is Jensen [214]).f.hve7x. h be the corresponding left and right eigenvectors normalized by vh = 1 and define C = ah • ve .F. Schmidt & Teugels [307] and Wolff [384].g. the conditions of Proposition 1. Example A5. let v. here k = p1). O'Cinneide [276] gave a necessary and sufficient for a distribution B with a rational m. The Erlang distribution gives an example where k > 0 (in fact.. the Erlang case).8). Notes and references The idea behind using phasetype distributions goes back to Erlang. No satisfying . it is easily seen that the asymptotic form of the tail of a general phasetype distribution has the form B(x) _ Cxkenx.1 of the Appendix. the text is essentially identical to Section 2 of Asmussen [26].f. Other expositions of the basic theory of phasetype distributions can be found in [APQ]. (1.8 Let B be phasetype with representation (a.8 are far from necessary ( a mixture of phasetype distributions with the respective T(') irreducible has obviously an asymptotically exponential tail. (or Laplace transform) are often used where one would now work instead with phasetype distributions. cf.8) Proof By PerronFrobenius theory (A.q be the eigenvalue of largest real part of T.g. x * oo. we give a criterion for asymptotical exponentiality of a phasetype distribution B. but the relevant T is not irreducible. In Proposition A5. In older literature. Schmidli. Here is a sufficient condition: Proposition 1. 1. 77 > 0 and k = 0. but todays interest in the topic was largely initiated by M. where C. . 2.
as the lifetimes of items (say electrical bulbs) which are replaced upon failure.. be i. U1<t < U1+U2. however. JtJt1) Then { 0<t<U1 . the renewals form a Poisson process and we have u(x) = 0. but is in part repeated below.. the problem has an algorithmically tractable solution if B is phasetype: Theorem 2.i. we refer to U as the renewal measure.. if U is absolutely continuous on (0. If B is exponential with rate 0.g. we denote the density by u(x) and refer to u as the renewal density. The explicit calculation of the renewal density (or the renewal measure) is often thought of as infeasible for other distributions.. ...1 of the Appendix. n=O We may think of the U.1 Consider a renewal process with interarrivals which are phasetype with representation (cr.d. (2.+UnEA). . with common distribution B and define4 U(A) = E# {n = 0. RENEWAL THEORY 223 algorithm for finding a phase representation of a distribution B (which is known to be phasetype and for which the m. and U(A) is then the expected number of replacements (renewals) in A. Jt={Jt?ul}.2.. is Markov and has two types of jumps . .. or the density is available ) is.f. U2. +UnEA} 00 = EEI(U1 +..1. 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.t. A related important unsolved problem deals with minimal representations: given a phasetype distribution .r. Lebesgue measure.1) Proof Let {Jtk)} be the governing phase process for Uk and define {Jt} by piecing the { J(k) } together..: U1 + . Then the renewal density exists and is given by u(x) = ae(T+ta)xt. Let U1. + U0 is 0 . known... 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 +. oo) w..T). For this reason. but nevertheless.
1. that is. (b) £(t) has a limiting distribution as t * oo.T) where v = aT1 /µB.1 Corollary 2. is the first k with Uk = 00. The renewal density at x is now just the rate of jumps of the second type. the lifetime of the renewal process. and let µB = aTle be the mean of B. define the excess life e(t) at time t as the time until the next renewal following t. Then the lifetime is zeromodified phase type with representation (a.3 Consider a renewal process with interarrivals which are phasetype with representation (a.1) follows by the law of total probability.224 CHAPTER VIII. the density is veTxt = B(x)/µB. u Returning to nonterminating renewal processes . Equivalently. IIafl < 1.e. T).1) remains valid for that case. as the time of the last renewal. i. 2.e.T). This is defined as U1 + .. which is ti in state i.. Then: (a) the excess life t(t) at time t is phasetype with representation ( vt. which is phase type with representation (v.2 Consider a terminating renewal process with interarrivals which are defective phasetype with representation (a. Proof Just note that { it } is a governing phase process for the lifetime. and the jumps of the first type are governed by T. Hence the intensity matrix is T + ta. this is welldefined. fi(t) U2 U1 . the phasetype assumptions also yield the distribution of a further quantity of fundamental importance in later parts of this chapter . . + Uit_1 where s.U1 U3 U2 U3 U4 Figure 2. However. see Fig. . Hence ( 2. i. Corollary 2. MATRIXANALYTIC METHODS of the last type from i to j occurs at rate tiaj .IIBII which is > 0 in the defective case. and the distribution of Jx is ae ( T+t«)x. since Uk = oo with probability 1 . B is defective . and hence ( 2. u The argument goes through without change if the renewal process is terminating.T + ta).T) where vt = ae (T+ta)t .
) ( t2 ) .T) where vt is the distribution of it which is obviously given by the expression in (a). The time of the next renewal after t is the time of the next jump of the second type.e. we first compute the stationary distribution of Q. u Example 2 . cf. 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).6. i. (2. 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 (2.4 Consider a nonterminating renewal process with two phases.2): aT1 e = AB = 1 µB µB a + aT'Tea aT1(T + ta) µB PB a + aea a + a µB µB =0.e. Hence in (b) it is immediate that v exists and is the stationary limiting distribution of it. T1 and eTx commute. = qz ql (x1 xz) = ql + qz ql + q ' and the nonzero eigenvalue A = ql . According to Example A3. hence e(t) is phasetype with representation (vt.2. Al. RENEWAL THEORY 225 Proof Consider again the process { Jt } in the proof of Theorem 2.q2. we get B(x) aeTxe aT1eTxTe µB µB PB = veTxt. (ii) First check the asserted identity for the density: since T. The renewal density is then aeQtt = (al a2) ( 7i 7"2.2) v(T + ta) = 0.1. Next appeal to the standard fact from renewal theory that the limiting distribution of e(x) has density B(x)/µB.
. MATRIXANALYTIC METHODS e. Then Q= 0 55 )+(1o)=( j ad ). A = 25.a27r1) (t1 .a27rl) (tl .e2bt) 13 Example 2 .4 yields the renewal density as u(t) = 5152 e. and Example 2.6 Let B be hyperexponential.52a1.t2) . Hence 7r = (1/2 1/2). t1B 0 Example 2 .5 Let B be Erlang(2).226 CHAPTER VIII.tl) 7r2t2 + eat (a17r2 . Then _ Q Hence 51 0 0 52 + 51 52 _ 5152 51a2 ) (al a2) 52a1 62a1 Slat + 52a1 51a2 51a2+52a1 A = 51a2 . and Example 2.t2) 1 + eat (a17r2 .`t (al a2) + C 11 172 ir12 / \ t 2 ) r1 (7r1 7r2) ( t2 7rltl + J + eAt (al a2) ( 71(t2 .(biaz + aza.52) 25152 51x2+5251 51a2+5251 Notes and references Renewal theory for phasetype distributions is treated in Neuts [268] and Kao [221]. The present treatment is somewhat more probabilistic.4 yields the renewal density as u(t) = 2 (1 . )t (51 .
The stars represent the ladder points ST+(k). T) where a+ is given by a+ = . with 0 denoting the Poisson intensity. T). G+(.(u) = a+e(T+tQ+)u Note in particular that p = IIG+II = a+e. Corollary 3. Within ladder steps. Now just observe that the initial vector of {mx} is a+ and that the lifelength is M. Considering the first. {St} the claim surplus process. and if there is a subsequent ladder step starting in j whic occurs w.3. i. Proof The result follows immediately by combining the PollaczeckKhinchine formula by general results on phasetype distributions: for (a).f3aT1. and rewriting in matrix form yields the phase generator of {my} as T + ta+. T). we shall.2. Thus the total rate is tip + tia+.1 on the next page. Then each claim (jump) corresponds to one (finite) sample path of the Markov process. itself phasetype with the same phase generator T and the initial vector a+ being the distribution of the upcrossing Markov process at time ST+_. which occurs at rate ti. Corollary 2. a+j.p. For (b). 3.1 Assume that the claim size distribution B is phasetype with representation (a. the Markov processes representing ladder steps can be pieced together to one {my}. marked by thin and thick lines on the figure. and M is zeromodified phasetype with representation (a+. add a more selfcontained explanation of why of the phasetype structure is preserved. cf. 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. Here we have taken the terminating Markov process underlying B with two states. B the claim size distribution. Since the results is so basic. . The essence is contained in Fig. represent the maximum M as the lifetime of a terminating renewal process and use Corollary 2. however. Next.3. We asssume that B is phasetype with representation (a. T(0) < oo) the ladder height distribution and M = supt>o St.i. (b) V. r(u) the time of ruin with initial reserve u.e. Then: (a) G+ is defective phasetype with representation (a+.) = F(ST(o) E •. we see that the ladder height Sr+ is just the residual lifetime of the Markov process corresponding to the claim causing upcrossing of level 0. use the phasetype representation of Bo. 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. T + to+).
3e3x + . T = (3 .M {mx} ST+(2)  S . see Corollary 2. 0 Example 3.1 This derivation is a complete proof except for the identification of a+ with . Figure 3.. This is in fact a simple consequence of the form of the excess distribution B0.2 Assume that . MATRIXANALYTIC METHODS t .QaT1.t t d kkt S.228 CHAPTER VIII...3.7)diag so that a+ = QaT 1 = 3 ( 3 2 2) 0 3 9 2 14 7 2 11 2 T+ta+ = 3 0 07/+( 7I \ 2 14 .1 .Q = 3 and b(x) = . 7e7x 2 2 Thus b is hyperexponential (a mixture of exponential distributions) with a (2 2 ).
T). the duality result given in Corollary 11. see Shin [340]. We assume p = PB/µA < 1 and that B is phasetype with representation (a. his derivation of +'(u) is different.and Markovmodulated models. 0(8) (u) (recall that z/i(u) refers to the zerodelayed case and iY(8) (u) to the stationary case). this was obtained in Section 3. Fig. For an attempt. see Stanford & Stroinski [351] . and the argument for the renewal case starts in just the same way (cf. 3.6.1 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+.1): Proposition 4. 3. T) for some vector a+ = (a+. For the compound Poisson model. we encounter similar expressions for the ruin probabilities in the renewal. if we define {mz} just as for the Poisson case (cf. The result carries over to B being matrixexponential. (a) G+ is of phasetype with representation (a+. THE RENEWAL MODEL This is the same matrix as is Example 1.1 In the zerodelayed case. with A denoting the interarrival distribution and B the service time distribution. the discussion around Fig.6). 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 .^(u) = a+e( T+ta+)ue = 24eu + 1 e6u 35 35 0 Notes and references Corollary 3.4.T).4.j). cf. It is notable that the phasetype assumption does not seem to simplify the computation of finite horizon ruin probabilities substantially. where a+ is the (defective) . That is. but there the vector a+ is not explicit but needs to be calculated (typically by an iteration). In the next sections. 4 The renewal model We consider the renewal model in the notation of Chapter V.2 are taken from Gerber [157]. see Section 6. but that such a simple and general solution exists does not appear to have been well known to the risk theoretic community. The parameters of Example 3. For further more or less explicit computations of ruin probabilities. We shall derive phasetype representations of the ruin probabilities V) (u).1 can be found in Neuts [269] (in the setting of M/G/1 queues.
T). Then . In fact. 4. Then {m. the Palm distribution of the claim size is just B.*} from {St+y . The key difference from the Poisson case is that it is more difficult to evaluate a+. MATRIXANALYTIC METHODS (b) The maximum claim surplus M is the lifetime of {mx}.1.1).T). where B0 is the stationary excess life distribution corresponding to B.*'} is Markov with the same transition intensities as {mx}. obviously mo = m. Hence by Theorem 11. B0 is phasetype with representation (aT1/µa.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.T)• Proposition 4. CHAPTER VIII.Sy} in the same way as {mx} is defined from {St}.6. where u w(a +) = aA[T + to+) = a J0 e(T+t+)1A(dy). it follows by integrating y out that the distribution a+ u of mo is given by the final expression in (4.5. Proof Obviously.4 Consider the renewal model with interarrival distribution A and the claim size distribution B being of phasetype with representation (a. (c) {mx } is a (terminating) Markov process on E.3 a+ satisfies a+ = V(a+). the form in which we derive a+ for the renewal model is as the unique solution of a fixpoint problem a+ = cp(a+). We have now almost collected all pieces of the main result of this section: Theorem 4 . Fig. Since the conditional distribution of my given T1 = y is ae4y. Also. Nevertheless. G(') = pBo. which for numerical purposes can be solved by iteration. (4. with intensity matrix Q given by Q = T + to+. the calculation of the first ladder height is simple in the stationary case: Proposition 4.3.1) Proof We condition upon T1 = y and define {m. But by Corollary 2. where a(8) = aT1/PA. but with initial distribution a rather than a+.230 distribution of mo.
1) and a(8) _ aT.3) (defined on the domain of subprobability vectors .1). a+l ) = cp (a+°)) . the maximum claim surplus for the stationary case has a similar representation as in Proposition 4. .2..^(8)(u) = a ( 8)e(T+ta +) xe. It remains to prove convergence of the iteration scheme (4. THE RENEWAL MODEL 231 .4.. and that this is given by Proposition 4.1 by noting that the distribution of mo is a+.3). thus . The term tf3 in cp(i3) represents feedback with rate vector t and feedback probability vector (3. i. The second follows in a similar way by noting that only the first ladder step has a different distribution in the stationary case. a+ can be computed by iteration of (4.1/pA.•. only with initial distribution a(*) for mo.3) Proof The first expression in (4. I {mx} .^(u) = a+e ( T+ta+)xe.M..1 .1(b).2 ) follows from Proposition 4. a+) > 0 = a+o) implies a+) _ (a+) > W (a+)) = a+) . (4.0) is an increasing function of /3. (4.0. .. by a+ = lim a +n) where a+°) . Hence ^p(. In particular .e. i y ^ T1= y `•r Figure 4.2) where a+ satisfies (4. a+2) = ^p (a+l)) . Furthermore .
Proof Suppose first Qh = sh. F[s] being interpreted in the sense of the analytical continuation of the m.T)1t. we use an argument similar to the proof of Proposition VI. Then (4. 0 = a+) < a+ yields a+) _ (a+0)) (a+) = a+ (n and by induction that a(n) < a+ for all n .4) whenever EeR(S)U < oo. Similarly.5) Since s $ sp(T).ST.1. which links together the phasetype setting and the classical complex plane approach to the renewal model (see further the notes). this implies that ahA[s] # 0.. the normalization is equivalent to F(s) = 1.} can contain at most n .4) makes sense and provides an analytic continuation of F[•] as long as s ¢ sp(T). the corresponding right eigenvector may be taken as (sI . Then s is an eigenvalue of Q = T + ta+ if and only if 1 =. Then e4'h = e82h and hence sh = Qh = (T + taA[Q])h = Th + A[s]tah.T1. 0 0 We next give an alternative algorithm. Assume the assertion shown for n . so to complete the proof it suffices to show that &+ < a+) for all n. s ¢ sp(T). n) &+n) T a+. and let &+". Then each subexcursion of {St+Tl . However.5) yields h = (sI .232 CHAPTER VIII.5 Let s be some complex number with k(s) > 0. To prove the converse inequality. both quantities are just 0 . and hence we may assume that h has been normalized such that ahA[s] = 1.f.g. Fn ). For n = 0.1 arrivals (n arrivals are excluded because of the initial arrival at time T1 ). Thus by (4. Theorem 4.T)'t • A[s] (4. limn4oo a ) < a+. a+ ) exists . let F be the distribution of U1 . In that case.) = P(mTl = i. with B[s]. 7+ ]. (4.P[s] = A[s]B[s]. (4. Then F[s] = a(sI . It follows that n1) so that on Fn the feedback to {mz} after each ladder step cannot exceed &+ a+ n) < a f ^ e(T+ t&+ 1))YA(dy) o < a is e(T+t«+1')YA(dy) _ w (a+1 )) = a+n). To this end.T)It. MATRIXANALYTIC METHODS and (by induction ) that { a+ n) } is an increasing sequence such that limn.4).2. Obviously.4. . Let Fn = {T1 + • • • + Tn+1 > r+}be the event that {my} has at most n arrivals in [T1. Thus .
Then G+ is phase. the matrix Q in Theorem 2.. . in turn. yd satisfying R(ryi) > 0. Pd with corresponding eigenvectors hl. .5..T) = 1 ata+ = a+. Let d denote the number of phases. hd.. and define hi = (piI . 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. This immediately implies that Q has the form CD1 and the last assertion on the diagonal form . .p1i . This gives d roots 'y..6 Suppose u < 0.type with representation (a+..' that the equation F(s) = 1 has d distinct roots p1.. Given T has been computed.6) i=1 i=1 Proof Appealing to Theorem 4. Pd in the domain ER(s) > 0 . 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 )... Corollary 4. the classical algorithm starts by looking for roots in the complex plane of the equation f3[y]A[ry] = 1. Q has diagonal form d d Q = dpivi®hi = dpihivi. . and hence by the WienerHopf factorization identity (A. and the solution is .T)It.. T) with a+ = a(QT)/at. t(ry) > 0. 0). . . . (4. we have IG_ [s] I < 1 .lt we get Qh = (T + to+)h = T(sI ...1 has the d distinct eigenvalues .T)lt + t = s(sI .5(c) means that a+(sI T)1t = 1. pdhd.. Q = CD1 where C is the matrix with columns hl. In older literature . D that with columns p1 hl.4. Further. 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 6.6.. letting vi be the left eigenvector of Q corresponding to pi and normalised by vihi = 1 . Hence with h = (sI T). hd. Notes and references Results like those of the present section have a long history.. THE RENEWAL MODEL 233 Suppose next F(s) = 1.T)lt = sh. As in Corollary 4.9) we have G+[s] = 1 which according to Theorem 1. Since R(s) > 0 and G _ is concentrated on (oo. we get at a(Q .. W v M(d) in the notation of Chapter V)...
In risk theory. where R is an unknown matrix. The distribution of W comes out from the approach but in a rather complicated form . the intensity matrix is A and the stationary row vector is ir . For surveys . Neuts and his students. similar discussion appears in Kemperman [227] and much of the queueing literature like Cohen [88]. Here phase. The solutions are based upon iterations schemes like in Theorem 4..234 then in transform terms CHAPTER VIII.exponential form of the distribution was found by Sengupta [335] and the phasetype form by the author [18]. It turns out that subject to the phase. whereas the approach was introduced in queueing theory by Smith [350]. Numerical examples appear in Asmussen & Rolski [43].type assumption .type assumptions are basic. see Neuts [269].4.. e. Asmussen & O'Cinneide [ 41] for a short self. is phasetype.F. 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.. In queueing theory. but the models solved are basically Markov chains and processes with countably many states ( for example queue length processes ). T('). the fixpoint problems look like R=Ao+RAI+R2A2+ .) d (see. For further explicit computations of ruin probabilities in the phasetype renewal case . That is . 5 Markovmodulated input We consider a risk process {St } in a Markovian environment in the notation of Chapter VI. with representation say (a(' ). the ruin probability can be found in matrixexponential form just as for the renewal model. involving . which contains somewhat stronger results concerning the fixpoint problem and the iteration scheme. The exposition here is based upon [18]. We assume that each B. The number of elements of El=> is denoted by q. E(t)). starting around in 1975.g. see Dickson & Hipp [118]. the background Markov process with p states is {Jt}. MATRIXANALYTIC METHODS d F 1 + a J e°" ip(u) du = Ee°w = 11(t.contained derivation). an alternative approach (the matrixgeometric method ) has been developed largely by M. [270] and Latouche & Ramaswami [241]. [119]. and the distribution of an arrival claim is B. a pioneering paper in this direction is Tacklind [373]. and appears already in some early work by Wallace [377]. The matrix. The arrival rate in background state i is a.
However. The connection between the two models is a fluid representation of the Markovmodulated risk process given in Fig. and the one E(•) for B.4. Diagonalization Consider a process {(It. has states o. the analysis involves new features like an equivalence with first passage problems for Markovian fluids and the use of martingales (these ideas also apply to phasetype renewal models though we have not given the details).Vt)} obtained by time reversing the I component. The two environmental states are denoted o. (a) 0 0 ♦ o ° tl ♦ • 0 0 o } o o (b) 0 } ♦ • 0 o f o Figure 5.2. p = ql = Q2 = 2. the phase space E(°) for B. We start in Section 5a with an algorithm involving roots in a similar manner as Corollary 4. 5. 5a Calculations via fluid models. states . Section 5b then gives a representation along the lines of Theorem 4.6. The stationary distribution is obtained by finding the maximum of the Vcomponent of the version of {(It. 5.1. for which the relevant fixpoint problem and iteration scheme has already been studied in VI.1 In Fig.5. say with slope r(i) on intervals where It = i. MARKOVMODULATED INPUT 235 some parameters like the ones T or a+ for the renewal model which need to be determined by similar algorithms. 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. •. Vt)}t>o such that {It} is a Markov process with a finite state space F and {Vt} has piecewiese linear paths. The key unknown is the matrix K. O. This calculation in a special case gives also the ruin probabilities for the Markovmodulated risk process with phasetype claims.1.
whereas Ee8s' = oo for all t and all s > so where so < oo. Recall that in the phasetype case. Eli) + Proposition 5. F is the disjoint union of E and the Eli). t. o. 2. Bi[s] = a(i)(T(i) + sI)it('). a) of {It}. consider the vector a satisfying (A + (13i(Bi[ s] .92a(2) 0 0 T(2) 0 0 0 f33a(3) 0 0 T(3) with the four blocks denoted by Ei„ i. 5. MATRIXANALYTIC METHODS 4.236 CHAPTER VIII. corresponding to the partitioning + Epp). 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. i E E. Let E denote the matrix .1))diag ) a = sa and the eigenvector b = . resp.1(b) {(It . This implies that in the fluid context. 4}. 4.31a(l) (/3i)diag . a) = 1. < oo for all s. a) : i E E. 5. j = 1. of E into components indexed by E. in the fluid model Eel'. Second. Thus F = {o. r(i) _ 1. 4. The fluid model on Fig .A 0 Or 1A/ _ t(i) 0 t(2) 0 0 0 0 0 t(3) 0 T1 0 0 0 . F = E U { (i.(Ni)diag r(i.1))diag + sII = 0 (5.1) if and only if s is an eigenvalue of E.1 A complex number s satisfies 'A+ (f3i(Bi[s] . '31a(1) 0 0 f32a(2) 0 0 AI = t(1) 0 0 0 t(2) 0 0 0 t(3) 0 T1 0 0 0 0 T(2) 0 '33a(3) 0 0 T(3) The reasons for using the fluid representation are twofold. a E E(i) } . The intensity matrix for { It} is (taking p = 3 for simplicity) I A .1(a). In the general formulation .Vt)} is then obtained by changing the vertical jumps to segments with slope 1. •. If s is such a number. V. we have more martingales at our disposal. A claim in state i can then be represented by an E()valued Markov process as on Fig. First.
T('))1t(i) .(sI .sI + E12 (sI . it follows that Ell E12 ( E 21 E22) (d) = s 1 d I .sI 0 0 t(3) 0 0 = 0. For the assertions on the eigenvectors.1).A . where c.E22 . Noting that E11c + E12d = se by definition.E12E22 E21 I .Nla(1) 0 0 T 1. with Eii replaced by Eii .sI 0 0 0 T(2) . c = a.sI 0 0 0 T(3) . MARKOVMODULATED INPUT 237 indexed by E. it follows that if Qla(1) 0 0 . 0 .E22)1 E21) a = 0.sI)1t)) iag I = 0 which is the same as (5. iEE (a> of 0* 1 AI.32a(2) (/3i)diag . Then E21c+E22d = E21a .sI.sI+ ((3ia(i)(T(i) .A .sI) (sI . t(1) 0 0 then also 0 t(2) 0 . d = (sI E22)1E21a = E ai(sI .sI ()3i)diag . assume that a is chosen as asserted which means (Ell .E21a + sd = sd. resp .5. Then (up to a constant) c = a. and let d = (sI . d correspond to the partitioning of b into components Proof Using the wellknown determinant identity Ell E12 E21 E22 E22 I ' I Ell .E22)1 E21a E21a . E(1) + + E(P).E22)1 E21a.
.4 that {e"1b(v) is a martingale . pi(u.2 Assume that E = Or 'Al has q = ql + + qp distinct eigenvalues si.. j) pi( u .v) = = p i( u .v.v) = (j. a) = Pi (Vw(u. For u. MATRIXANALYTIC METHODS Theorem 5. a)). .j)c v . Then we get V)i (u) as sum of two exponential terms where the rates s1. j.sv)b(v) = 0. Thus 0(u) = esu/d = pe7 ° as u should be.v) = v) I. a).Q. Example 5 ..4 Assume that E has two states and that B1.v)=inf{t >0:Vtu orVt=. e89uc(e)) (d(1) ./' u = e' (esiuc ( 1) . it follows by Proposition II. j. define w(u. a)). j. < 0 and let b(v) = I d(„)) be the right eigenvector corresponding to s. ..a)d^ ).( u.a Solving for the pi(u.. We can take a = c = 1 and get d = (s + b)16 = 5/(3 = 1/p. To determine 0 (u).. B2 are both exponential with rates 51 i b2. s2 are the negative eigenvalues of Al +01 A1 E _ A 2 b1 0 52 A2 +32 0 . j. the result u follows. v.O. sq with $2s..v}. . v > 0.. Proof Writing Or'Alb( v) = svb( v) as (AI ..v) = Optional stopping at time w (u. a )d(a + e8 °vpi (u .5.. Here E has one state only.j. w(u)=inf{t >O:Vtu}. c j.. a) and noting that i1 (u) = >I j. u) Iw(u. w(u. q. Then . a) = (j.v) = j). Letting v ^ oo and using Rsv < 0 yields e8'u = Epi(u.. v) yields C{V) = e8 .238 CHAPTER VIII.. Example 5 . v. v. Iw(u..3 Consider the Poisson model with exponential claims with rate 5.upi(u. v = 1. d("))1 e. . j. .pi(u.. we first look for the negative eigenvalue s of E = I 0 I which is s = ry with yy = b . I' i( V P2 (w (u) < oo.
xxej • a 00 oo el . MARKOVMODULATED INPUT 239 5b Computations via K Recall the definition of the matrix K from VI.b (u) = Pi(M > u) = 9(i)euue. is 0 /3 f R(i .k. oo)) j)ye.2) the l.s. the Pidistribution of M is phasetype with representation (E(1) + + E(P).( 2. 9('). U) where t(j) + t(j)O(j j = k uja.y = to B k7 j # k In particular. 8^')IT(j)) where e 3^') =. Proof We must show that G+ (i.2. (y. according to VI. j.33(e = 0 a(j))(K ®T ( j))(ej (9 I). (5. j. 0 Theorem 5 . dx)Bj(y . we get the following phasetype representation for the ladder heights (see the Appendix for the definition of the Kronecker product 0 and the Kronecker sum ®): Proposition 5.h. In terms of K.6 For i E E.3) . i. j.x) 00 f ° (') (j) eT (yy)edx . •) is phasetype with representation (E(i).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.3j eye.5.5 G+(i. (') a T( However .
For j = k. Starting from Jo = i.g. Then b*[0] is rational if and only b(x) is matrixexponential.. a) to (k.. i. that the density b(x) can be written as aeTxt for some row vector a. which occurs w. we have sofar concentrated on a claim size distribution B of phasetype. and a new ladder step of type k must start in phase y. i. An alternative characterization is that such a distribution is matrixexponential.. +aii10+anI then a matrixexponential representation is given by b(x) = aeTxt where a = (b1 b2 . equivalently.5).k y. This yields the asserted form of uja.. intensity matrix U. Bk7 . a) is obviously chosen according to e(`). some square matrix T and some column vector t (the triple (a.e. we have the additional possibility of a phase change from a to ry within the ladder step.. the current ladder step of type j must terminate. However.2. For a transition from (j. 6 Matrixexponential distributions When deriving explicit or algorithmically tractable expressions for the ruin probability. a m.240 CHAPTER VIII. Piecing together these phase processes yields a terminating Markov process with state space EiEE E('). Associated with each ladder step is a phase process. Furthermore. +bn0i1 0n +a10n1 +. u Notes and references Section 5a is based upon Asmussen [21] and Section 5b upon Asmussen [17]. t = (0 0 .2) ..y) to occur when j # k.e. 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.p. Numerical illustrations are given in Asmussen & Rolski [43]. 0 1)'. .. (6. the ratio between two polynomials (for the form of the density. if b* [0] = b1 +b20+b302 +. T.f. say. bn1 bn).. the initial value of (i. and lifelength M.) which is rational. oo) and b* [0] = f °O eBxb(x) dx the Laplace transform. with phase space EU> whenever the corresponding arrival occurs in environmental state j (the ladder step is of type j). t) is the representation of the matrixexponential distribution/density): Proposition 6. see Example 1. which occurs at rate t^^7.1 Let b(x) be an integrable function on [0. MATRIXANALYTIC METHODS Proof We decompose M in the familiar way as sum of ladder steps . and it just remains to check that U has the asserted form. which occurs at rate t(i).
Namely.2). where c = 1 + 1/47r 2.6. . Writing b(x) = c(e( 2ni1 ) y/2 . s) is given by 27r i . u Remark 6. but as follows from Proposition 6. (6. bn of the denominator to be distinct and expand the r..1 0 0 )3 = (111).2).. T= 0 0 1 . u giving b(x) = E 1 ciebiz/bY. 1 .(6.3) that we can take 0 1 0 0 a= (1 + 47r2 0 0). (6.. MATRIXEXPONENTIAL DISTRIBUTIONS 241 T = 0 1 0 0 0 . For a proof. 0 0 .4) 0 0 1 c This representation is complex. personal communication).3) 0 0 0 0 0 . matrixexponentiality implies a rational transform.3 A set of necessary and sufficient conditions for a distribution to be phasetype are given in O'Cinneide [276].T)1 is so..1. One of his elementary criteria.e(tai1)x/2 + e'T) it follows that a matrixexponential representation ()3. s = c/ 2 . of (6.47r2 3 . T. cannot be phasetype. namely to asssume the roots 6l.an_3 an _ 4 .. a2 a1 Proof If b(x) = aeTxt...1) as E 1 c. see Asmussen & Bladt [29] (the representation (6. (6. Thus. .2 A remarkable feature of Proposition 6. t= 0 .47x2 3 1 0 . we can always obtain a real one (a. . since 1 + 4ir2 03 + 302 + (3 + 47x2)0 + 1 + 47r2 it follows by (6.1 is that it gives an explicit Laplace tranform inversion which may appear more appealing than the first attempt to invert b* [0] one would do. Example 6 . 0 1 an an1 an _2 . shows that the distribution B with density b(x) = c(1 cos(21r x))ex. 0 0 0 0 1 0 0 . The converse follows from the last statement of the theorem. . S.s.. .3) was suggested by Colm O'Cinneide..h.1 0 . S = f c/2 0 21ri ./(0 + bi). t). b(x) > 0 for x > 0. then b*[0] = a(0I T)1t which is rational since each element of (01 .
Then (cf. (6. T. then: Proposition 6. Corollary 111.5 (6. For the second algorithm. we use a representation (a. and present two algorithms for calculating '(u) in that setting. t) a phasetype representation with a the initial vector.6) in Section 3 seems to use the probabilistic interpretation of phasetype distribution in an essential way. MATRIXANALYTIC METHODS Example 6 .1)2 + 6).4) the Laplace transform of the ruin probability is /g(e)PO 0*[e] _ /' eeu^G(u)dU = 0 9(/3a0p(9)ap (9)/q(9)) . and that the minimal number of phases in a phasetype representation increases to 0o as 5 . We recall (see Section 3. and can use this to invert by the method of Proposition 6. we have represented ti* [0] as ratio between polynomials (note that 0 must necessarily be a root of the numerator and cancels). then 5(u) = a+e(T+t+)uTle where a+ = /3aT1. t) of b(x).6) The remarkable fact is. 0.5) Thus. we take as starting point a representation of b* [0] as p( O)/q(9) where p. As for the role of matrixexponential distributions in ruin probability calculations. T. we shall only consider the compound Poisson model with arrival rate 0 and a matrixexponential claim size distribution B.242 CHAPTER VIII. For the first. leading to matrix calculus in high dimensions when b is small. that despite that the proof of (6.4 This example shows why it is sometimes useful to work with matrixexponential distributions instead of phasetype distributions: for dimension reasons . T the phase generator and t = Te. 7 + 155ex b(x) Then it is known from O'Cinneide [276] that b is phasetype when 6 > 0. Consider the distribution with density = 15 ((2e2x . .6) holds true also in the matrixexponential case. (6.1 shows that a matrixexponential representation can always be u obtained in dimension only 3 independently of J. q are polynomials without common roots. recall that t = Te) that if B is phasetype and (a. But since 15(1 +6)02 + 1205 0 + 2255 + 105 b* [9] _ (7 + 155)03 + (1355 + 63)92 + (161 + 3455)9 + 2256 + 105 Proposition 6.1 to get i (u) = f3esus.3.
U =. From the general matrix identity ([331] p.T .5 ).to+)1T .T)1T 2 = and 1 = AB IT2 + 82T . (6.T .a+(9I . the assertion is equivalent to a+(BI . Now. since (91T)1T .T)1 + 1 ib* (91.1BVA1.to+)1 = (BI .T)1 + (6I . we get (91.'t. Then in Laplace transform formulation . we get b+ = 0aT1(9I T). MATRIXEXPONENTIAL DISTRIBUTIONS 243 Proof Write b* = a(9I . but we shall give an algebraic proof.1t du = .T)1T1t.1t = b* .T)1 (91.T).T)1t.t.6).7) 9( cf.A .1UB(B + BVA1UB). (91.b* (6. (6.1 + b+ = b++ 1 .T)1 so that b* b** b** a+(9I .1 = ^(T1 + ( 91T)1).1 = A1 .B=land V=a+. xb(x) dx = aT2t.1 + 82 (9I .T)1 J0 00 b(x) dx = f aT1t.6b* .to+)1T . this can be verified by analytic continuation from the phasetype domain to the matrixexponential domain .1t = f3a (0I T)1T1t .T)1t ( l .6.T . Presumably.T)1ta+(OI . .T)1t)1a +(9I . b+ = a +(BI . b+ = a+(9I . with A = 91T. 519) (A + UBV ).
Lipsky [247] and Asmussen & O'Cinneide [41].1t = /3a (9I .s.h.1. See Fig. For expositions on the general theory of matrixexponential distributions. 0 Notes and references As noted in the references to section 4.244 CHAPTER VIII. 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. From this it is straightforward to check that b**/(b+ . MATRIXANALYTIC METHODS .3 is taken).1. . a key early paper is Cox [90] (from where the distribution in Example 6.7). In Corollary VII. of (6. VII. to piece together the phases at downcrossing times of {Rt} (upcrossing times of {St}) to a Markov process {mx} with state space E. 3. The proof of Proposition 6. see the Notes to VII. the ruin probability(u) was found in explicit form for the case of B being exponential.1.8 a(T1 + (01.T)1)t = 8 (1 .8. A key tool is identifying poles and zeroes of transforms via WienerHopf factorization. see Asmussen & Bladt [29].b*). We present here first a computational approach for the general phasetype case (Section 7a) and next (Section 7b) a set of formulas covering the case of a twostep premium rule. but the argument of [286] does not apply in any reasonable generality). 7 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate 0. /3aT1(0I . some key early references using distributions with a rational transform for applied probability calculations are Tacklind [373] (ruin probabilities) and Smith [350] (queueing theory).1.82b*.3a (1 0 T 2 + 1 T 102 (9I + 02 1 T)1) t P + 7.la. 7.5 is similar to arguments used in [29] for formulas in renewal 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. a. which is selfexplanatory given Fig. (for some remarkable explicit formulas due to Paulsen & Gjessing [286].1) is the same as the r.T)1T.T)1T2t . T). cf. Much of the flavor of this classical approach and many examples are in Cohen [88].
O<. Note that in general >iEE Vi (U) < 1. Define further vi(u) as the probability that the risk process starting from RD = u downcrosses level u for the first time in phase i. t) is the vector of state probabilities for mt. we obtain V)(u) = P(m„ E E) = v(u)P(0. the A(t) and hence Vi(u) is available by solving differential equations: Proposition 7. In fact. Proof The first statement is clear by definition. the definition of {m8} depends on the initial reserve u = Ro.1 A(0) = v(u) and A'(t) = A(t)(T + tv(u . t + s) . Let P(tl. i.tl < t2 < u.t)). Since v(u) = (vi(u))iEE is the (defective) initial probability vector for {m8}. Ai(t) = P(mt = i). P(tl.t2) be the matrix with ijth element P (mt2 =j I mtl = i). Given the v(t) have been computed. t2) = exp where Q(t) = ds [P(t. though still Markov.1 The difference from the case p(r) = p is that {m2}. Figure 7. Also. is no longer timehomogeneous.7.u)e = A(u)e (7.1z I. 0 < t < u. By general results on timeinhomogeneous Markov processes. >iEE Vi (U) is the ruin probability for a risk process with initial reserve 0 and premium function p(u + •). in contrast to Section 3.e. RESERVEDEPENDENT PREMIUMS 245 Rt l0 u .I] I 80 { tq f Q(v) dvl t1 1 .1) where A(t) = v(u)P(0.
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. 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 interpretation of Q(t) as the intensity matrix of {my} at time t shows that Q(t) is made up of two terms: obviously.t) for the second. two things can happen: either the current jump continues from u + p(u)dt to u. A'(t) = A(t)Q(t) = A(t)(T + tv(u .t)). The intensity of a jump from i to j is tij for jumps of the first type and tivj(u .(u) p ( u) = . the probability that level u + p(u)dt is downcrossed for the first time in phase j is vj (u + p(u)dt).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. vi. dt]. Proposition 7. {mx} has jumps of two types.2 For i E E. the probability that level u is downcrossed for the first time in phase i is ai. In the first case.t and being followed by a downcrossing. MATRIXANALYTIC METHODS However.246 CHAPTER VIII. Hence Q(t) _ T + tv(u . 0 < t < u. Given this occurs. Given A'.(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 . 0 Thus. whereas in the second case the probability is p(u)dt • tjvi(u).Qdt) vi(u) + vi'(u)p(u)dt + p(u) dt E{tji+tjvi(u)}. the probability of which is 1 . from a computational point of view the remaining problem is to evaluate the v(t). Thus.3dt. or it stops between level u + p(u)dt and u. we get vi(u) = aidt + (1 . Given A. jEE . (7. given A.t).
<oo.) is the tail of a (defective) random variable so that P(Bv) + 0 as v 4 oo.) P"(AnBv) = P(AnB. Now since both P(A n Bv) 3 0 and P"(A n Bv) .^ 0. To deal with this. From Section 3. starting from v"(v) = . supRt>v l t<7 I where o. (v) is given by the r.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.7.)P"(AnB. (u) on both side and dividing by dt yields the asserted differential u equation. then P(A n Bv) _ P"(A n BV'). When solving the differential equation in Proposition 7. after a certain level v. Then pv(r) p(r) r < v p r>v ' and (no matter how p is chosen) we have: Lemma 7. say. say. of (7. Then P(B. V . . RESERVEDEPENDENT PREMIUMS 247 Subtracting v. vi (U) = lim v= (u).0 as v + 00 we have P(A) P"(A) = P(AnBv)+P(AnBv) P"(AnB...s.i7rT1/p. This yields v. F" etc.4) backwards for {va (t)}v>t>o..h. Rt . (u) for any values of u and v such that u < v. Since the processes Rt and Rt coincide under level B. and similarly P"(Bv) . Thus. consider a modification of the original process {Rt} by linearizing the process with some rate p. denotes the time of downcrossing level u . we have p(r) = p = vi (u) 0aTe. we can first for a given v solve (7. P u which implies that v.) + 0 as v + oo.5). we face the difficulty that no boundary conditions is immediately available.3 For any fixed u > 0. refer to the modified process. Let p" (t).2..
. say.1. the evaluation of Vi(u) requires q(u) = 1 .10 that in addition to the O'(•).6) We may think of process Rt as pieced together of two standard risk processes RI and Rte with constant premiums p1. < 0}). (7. Corollary 3. Therefore u pl(vvueTa t 1.e. p(r) P. The algorithm based upon numerical solution of a Volterra integral equation (Remark VII.. Then v(u) is the initial distribution of the undershoot when downcrossing level v given that the process starts at u. for u > v the distribution of v .. r<v r > v.RQ (defined for or < oo only) is defective phasetype with representation (v(u). cf. typically the complexity in n is 0(n2) for integral equations but 0(n) for integral equations. v = u.x) dx f v(u)eT xt dx . which is available since the z/i'(.1a.248 CHAPTER VIII. However.v v(u)eTat 1 1 .1. the probability of ruin between a and the next upcrossing of v. The precision depends on the particular quadrature rule being employed.7) f o (the integral is the contribution from {R. T). let v(u) = a+2ieiT +ta+>)(uv).9.) are so.V" M 0 . To evaluate p1(u). where. Recall that q(w) is the probability of upcrossing level v before ruin given the process starts at w < v. > 0} and the last term the contribu tion from {R.7) equals 01 (v .1. 3u etc. p2. The trapezoidal rule used in [288] gives a precision of 0(n 3). such that Rt coincide with RI under level v and with Rt above level v. where v = inf It > 0 : Rt < v}. as well p1(u).. 2/n. MATRIXANALYTIC METHODS Next consider a sequence of solutions obtained from a sequence of initial values {v.zp1(u)/(1 . 7b Twostep premium rules We now assume the premium function to be constant in two levels as in VII.z51(v)).q(v dx +( ) ) = ( ) ( q( )) vueTva (7. We recall from Propositon VII. assuming u > v for the moment. 1/n. The f iin in (7. Let ii'( u) = a+'ie(T+ta +^)"e denote the ruin probability for R't where a+ = a+i) = laT1/pi. (u)}. 2u. numerically implemented in Schock Petersen [288]) and the present one based upon differential equations require both discretization along a discrete grid 0.. i. 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. while the fourthorder RungeKutta method implemented in [30] gives 0(n5). 0 < u < v.
From Example 3.e6u 35 . Since µB = 5/21.1 from which we see that pl (u) = 1 + 1 249  1 .x) dx 1 ^(v) ( 1 .and A2 = 3 . p2 < 3. all quantities involved in the computation of b(u) have been found in matrix form.2.4) can be written as (Y(u) ®a+)e(T+t°+>)°1 (T ® (T .8) equals v v(u)eTxta+2) e(T+ta +))( vx)edx which using Kronecker calculus (see A.^1(v) 1 .2. Then one gets X20 20 21 f 1ea1(u v) + 1 3 3 ^ A 2(u e .v(u)eTve).8) The integral in (7.7. 01(u) _ 24 u + 35 e6u 1 35 e 4(u) _ 35 . (7.e6v Let Al = 3 + 2V'2.be the eigenvalues of T + to( 2 ).21 = ? yields 0(u) = 1.4 Let {Rt } be as in Example 3. I.v(u)eTVe . so we consider the nontrivial case example p2 = 4 and p1 = 1..e. Example 7.24ev .v(u ) eTV e J v(u)eTxtz/)l (v .24e.01 (v .v) + (2^ + 3v2 ea'(u " .u .v) 1eai(u v) + 7 7 1 e\2(u v) 1 3 ^') eA2 (u. B is hyperexponential corresponding to 3 0 3 a(2 2)' T= ( 0 7 t.to+))11 {e{T®(Ttoy+ ))}„ .(7 The arrival rate is (i = 3.jl (t ®e) Thus.2V"2. RESERVEDEPENDENT PREMIUMS 1 .x) dx} V 1 1(v) f V v(u) eTxt.
1 V2 = 4e5"+6 35e6v .21(35e6v .24e5v .24es" . 21 3 In particular.250 CHAPTER VIII.1)' ?.+ it (3 4'I 1 ea2(uv e1\2(u") 7 + ( 32 +4.1.24es" ./2) ea 1(u .2 35e6v ..7) we see that we can write pi (u) = v(u)V2 where V2 depends only on v. The analysis and the example are from Asmussen & Bladt . ) e sv + ( 2v/2./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. MATRIXANALYTIC METHODS From (7. 192esv + 8 P1 . pi (u) = p12(u)/p1 l(u) where p1i(u) p12(u) 35e6v .24e5v . Notes and references [30].1 Thus.v)esv + 7 4_ 2. and one gets 12e5" .
B(x). III.f. x 4 oo.N(µ. and instead we shall work within the class S of subexponential distributions . a2)) with density 1 e(logyFh) 2/2az . For example. Some main cases where this lighttail criterion are violated are (a) distributions with a regularly varying tail. B(x) = ex0 with 0<0<1. (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. For further examples. For the definition .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) = 1.4. x 2iror2 (c) the Weibull distribution with decreasing failure rate . B[s] = f e8x B(dx) is finite for some s > 0 in the lighttailed case and infinite for all s > 0 in the heavytailed case.46 and at numerous later occasions require a light tail. see I. L(tx)/L(x) 4 1.g.2b. we require that B is concentrated on (0. for all t > 0. oo ) and say then that B is subexponential (B E S) if 251 . the exponential change of measure techniques discussed in II. A rough distinction between light and heavy tails is that the m. B(x) = L(x)/x" where a > 0 and L(x) is slowly varying.
1 Let B be any distribution on (0. Thus . X2) > u x)/B(x) = 2. given X1 + X2 > x. one can check that x x where U is uniform on (0. P(max(Xi. Proof By the inclusionexclusion formula. then P(X1>xI X1+X2>x)* 2. 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.F(X1 > x. X2 with distribution B.v. if X1 + X2 is large .'s.1) then means P(X1 +X2 > x) 2P(Xi > x). . Then: (a) P(max(Xi. X2) > x}. We later show: Proposition 1. (1. 1).v. then (with high probau bility) so are both of X1. X2) > x) is P(X1 > x) + P(X2 > x) . x 3 00. HEAVY TAILS B*2\ 2. oo).1(b) is oo.252 CHAPTER IX.p. oo). we have {max(Xi.2B(x). To capture the intuition behind this definition.3 Consider the standard exponential distribution. X2) > x} C {X1 + X2 > x}. Thus the liminf in Proposition 1.2. proving (a).2 If B E S. that is. in the subexponential case the only way X1 + X2 can get large is by one of the Xi becoming large.B(x)2 . In contrast. the behaviour in the lighttailed case is illustrated in the following example: Example 1. X2 but none of them exceeds x. (b) liminf BB(() ) > 2. X2 > x) = 2B(x) . 1/2 'typical' (with distribution B) and w. B(x) ax. the r. X2) > x) ^' 2B(x). Then X1 +X2 has an Erlang(2) distribution with density yeY so that B*2(x) xex. In terms of r.'s X1. P(Xi <yI Xi+X2>x) 1B(y). 1/2 it has the distribution of X1I X1 > x. X1 is w. Since B is concentrated on (0. the distribution of independent r. That is.v.p. B(x) Here B*2 is the convolution square. and thus the lim inf in (b) is at least lim inf P(max(Xi. As contrast to Proposition 1. note first the following fact: Proposition 1. That is.
yo] as X + 00. Hence lim sup a+oo B*2(x) 2B((1 . we get BZ(x)) > 1 + B(y) + B(B()y) (B(x) . then the overshoot X .S)x + B(Sx)2 < lim sup B(x) xaoo B(x) lim sup 2L((1 x^oo .z B(x) . then either one of the Xi exceeds (1 . Let 0 < 5 < 1/2. 253 Proof Assume B(x) = L(x)/xa with L slowly varying and a > 0.5 If B E S. then B(B(x)y) * 1 uniformly in y E [0.4 Any B with a regularly varying tail is subexponential.xIX > x converges in distribution tooo.B*(n ) B(dz) (1.B*n(x .y)/B(x) > 1 since y > 0. If lim sup B(x .1(b) we get B*2(x)/B(x) * 2.'s: if X . y] and (y. 1 < B(x ) B( x) Y) < B( 0). The uniformity now follows from what has been shown for y = yo and the obvious inequality y E [0. we therefore get lim sup B*2(x)/B(x) > 1+B(y)+ 1 . B( 0 . We now turn to the mathematical theory of subexponential distributions. x].2) B(x) B(x ) B(x) Jo with n = 1 and splitting the integral into two corresponding to the intervals [0. If X1 + X2 > x. SUBEXPONENTIAL DISTRIBUTIONS Here is the simplest example of subexponentiality: Proposition 1.B(y) = 2. we get limsupB*2(x)/B(x) < 2. Using the identity B*(n+1)(x) = 1+ + 1)(x) 1+ 2 1 .B(y)) .6)x)/((1 .] Proof Consider first a fixed y.B E S.S)x. Finally lim inf B(x . [In terms of r.v. This follows since the probability of the overshoot to exceed y is B (x + y)/B(x ) which has limit 1. or they both exceed Sx. and combining with Proposition u 1. Proposition 1. a contradiction.1.yo].5)x)' + 0 _ 2 L(x)l xa (16) Letting S 10.y)/B(x) > 1.
Proof For 0 < 5 < e.6 If B E 8. 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 . so assume the proposition has been shown for n. we have by Proposition 1.z) B(dz) (n + O(e)) ^x JO B(x) (n + 0(0) I B (x) . x oo. then for any n B*n(x)/B(x) * n.z) B(dz) _y B(x) 111 Lx B .5 and dominated convergence.z) B(dz).z) B(dz) 2B(x) o rv 2 0 2 using Proposition 1.1) for all large n so that B(n) > cle6n for all n. HEAVY TAILS Corollary 1. Given e > 0. B(x) \Jo _ B(x . Then by (1.y) sup v>o B(v) B(x) which converges to 0 by Proposition 1.z) B(x) Here the second integral can be bounded by B*n(y) B(x) .z ) (x ) = 1 + (^ B(x . Proof We use induction.2.5 that B(n) > e6B(n . The case n = 2 is just the definition. and this immediately yields the desired conclusions. O The following result is extremely important and is often taken as definition of the class S.5 and the induction hypothesis.B(x .B*2 (x) B(x) (x .nI < e for x > y. b[c] = oo for all e > 0. choose y such that IB*n(x)/B(x) . Proposition 1. B*(n+1) (x I xy + Jxx y) W.254 CHAPTER IX. 0 Proof of Proposition 1. The first integral is y B(x .7 If B E S.. its intuitive content is the same as discussed in the case n = 2 above. then e"R(x) * oo.2). This implies B(x) > c2e5x for all x.
Proof Define 5 > 0 by (1+5)2 = 1+e. 0 Proposition 1.z) B(dz ) + sup < 1 + sup f x<T B ( x) x>T 0 B(x .y)Ai(dy) = (x)o(1) (1. then there exists a constant K = KE such that B*n(x) < K(1 + e)nB(x) for all n and x.(al + a2)B(x).y)Ai(dy) v) f o . A2 be distributions on (0.5 easily yields P(X1 + X2 > x.y)B(dy) = B(x)ov (1)• v (1. e > 0. an = supx>o B*n(x)/B(x). Combining these estimates and letting a 4.z) B(dz) x . X2 be independent r.y))/B(x). choose T such that (B(x)B*2(x))/B(x) < 1 + b for x > T and let A = 1/B(T). 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) .B(x . Proof Let X1. Then Al * A2(x) = P(X1 + X2 > x).1.X2 > xv) < A1(xv)A2(x v) .i).2). Xi <= v Ai (x .z) B(dz) < 1 + A + an(1 + d) .z) B(x) < 1 + A + an sup f x B(x .'s such that Xi has distribution Ai. For any fixed 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. Proposition 1.ajB(x)Ai(v) = ajB( x)(1+o„(1)) (j = 3 .ala2B(x)2 which can be neglected. 0 Lemma 1.0 completes the proof.z) B(x . Since P(X1+X2 > x. a2 with a1 + a2 > 0. it follows that it is necessary and sufficient for the assertion to be true that JX_VA (x . Then Al * A2 (x) .X1 > xv.4) .8 If B E S. oo) such that Ai (x) _ aiB(x) for some B E S and some constants al.3) Using the necessity part in the case Al = A2 = B yields f xv B(x . Then by (1.v.9 Let A1. an+1 fX B*n( *n(x .
B1 * B2 E S does not hold in full generality (but once B1 * B2 E S has been shown.2A(x).Bl (x) + B2 (x) follows precisely as in the proof of Proposition 1. HEAVY TAILS 'VV B(x .h. B1 * B2 (x) . u Corollary 1.5) Here approximately the last term is B(x)o„(1) by ( 1.5) becomes x B(x .2.11 Let B E S and let A be any distribution with a ligther tail.12 Assume that Bi(x) = Li(x)lxa.y)B(dy). it should hold that B1 * B2 E S and B1 * B2 (x) .v)B(v) + _'U Aq(x . u It is tempting to conjecture that S is closed under convolution.(x) is decreasing for x > x0 with limit 0 at oo. Then B E S provided fo "O exA(x) b(x) dx < oo. the l. i = 1.3) follows if CHAPTER LX.4).256 Now (1. Then A * B E S and A * B(x) . Recall that the failure rate A(x) of a distribution B with density b is A(x) = b(x)/B(x) Proposition 1. That is. B2 E S.Bl (x) + B2 (x) when B1. That is. Then L = L1 + L2 is slowly varying and B1 * B2(x) sim L(x)/x«. V (1.aiB(v)) = B(x)o„(1). We next give a classical sufficient (and close to necessary) condition for subexponentiality due to Pitman [290]. L2 are slowly varying. whereas the two first yield B(x)(Ai(v) . Proof Taking Al = A2 = A.13 Let B have density b and failure rate A(x) such that .9). f " By a change of variables. L2 slowly varying. . with a > 0 and L1. of (1. However. it is easy to see that if L1. a1 = a2 = a yields A*2(x) . then so is L = L1 + L2.v)Ai(v) . Hence Corollary 1.B(x) Proof Take Al = A. u Corollary 1. A2 = B so that a1 = 0.s. A(x) = o(B(x)). In the regularly varying case.Ai(x . then A E S.10 The class S is closed under tailequivalence. a2 = 1.y)Ai(dy) = B(x)o„(1).2aB(x) . if q(x) aB(x) for some B E S and some constant a > 0.
1)xcl1 . Thus B*2(x )/ B(x) . L(x) y° (a . Thus. x . we can use the same domination for the second integral but now the integrand has limit 0 .16 For L(x) slowly varying and a > 1. Jo For y < x/2.1 B(x) eA( x)A(xv )A(y)A(y) dy f B(x . By (1. Thus.A(y)\(y) dy + fox/ 2 eA(x )..14 Consider the DFR Weibull case B(x) = ex0 with 0 <.A(xy)A ( y).y) < A (y) for y < x/2. Then B(x) = eA(x).y ) b(y)dy = B (x) o ox _ J = ox/2 eA( x)A(xy ). f ' L(y) dy . The middle bound shows that it converges to b(y) for any fixed y since \ (x .(x . subexponentiality has alrady been proved in Corollary 1. the u lognormal distribution is subexponential. Further. 0 A(x) . elementary but tedious calculations (which we omit) show that A(x) is ultimately decreasing.y) < yA(x . Define A(x) = fo .(y) dy.12. Example 1. the first integral has limit 1 .y) dy.1 has limit 1 + 0.A(y)a(y ) = ev'(y) b(y). Since ) (x .y) y\(y)• The rightmost bound shows that the integrand in the first integral is bounded by ey"(v).15 In the lognormal distribution.A(x .1. Goldie & Teugels [66]): Proposition 1.2). proving B E S. Thus A(x) is everywhere decreasing. To illustrate how Proposition 1.U) /or) v 2x This yields easily that ex. and exa(x)b(x) = (3x01e(10)x9 is integrable.. we first quote Karamata's theorem (Bingham.`(x)b(x) is integrable.y) * 0. Thus by dominated convergence .3 < 1. the DFR Weibull distriu bution is subexponential. Example 1. In the regularly varying case.e009xv)2/2a2/(x 2irv2) logx ( ) 't ((logx . replace B by a tail equivalent distribution with a failure rate which is everywhere decreasing). SUBEXPONENTIAL DISTRIBUTIONS 257 Proof We may assume that A(x) is everywhere decreasing (otherwise. B*2(x) .13 works in this setting. an integrable function by assumption. a(x) = ax01. Then b(x) = Ox0lexp.
1)X(x)/x > y) = P(X > x[1 + y/(a . 'y(x) = EXix>.1/A(x) and P(X ixil'Y (x) > y) * e'. then B(x) .y(x)B(x). However. We conclude with a property of subexponential distributions which is often extremely important: under some mild smoothness assumptions. Then 7(x) .4 is necessary in full generality.1) and P(X (. Proof ( a): Using Karamata's theorem.E(X . Thus exa(x)b(x) .13 may present a problem in some cases so that the direct proof in Proposition 1.1)] I X > x) L(x[1 + y/(a .L(x)/x" and )t(x) . HEAVY TAILS Proposition 1.17 If B has a density of the form b(x) = aL(x)/x°+1 with L(x) slowly varying and a > 1. f O B(y) dy .18 (a) If B has a density of the form b(x) = aL(x)/xa with L(x) slowly varying and a > 1.258 From this we get CHAPTER IX.1)]) xa L(x) (x[1 + y/(a .)/Y(x) > y) (1 + y/(a . More precisely. (c) Under the assumptions of either ( a) or (b). the overshoot properly normalized has a limit which is Pareto if B is regularly varying and exponential for distributions like the lognormal or Weibull.a/x.x)+ _ 1 °° P(X > x) P(X>x )J L PX >y)dy 1 x L(y)/ydy L(x)/((a1)x'1) x )l ° J °° ( ()l a x a1 Further P ((a . (1 + y/(a . we get (1.ea b(x) is integrable.1))a .6) EX(x) . the monotonicity condition in Proposition 1. yo] .1)])a 1 1 . let X W = X . Then: Proposition 1.1))^ ' (b) Assume that for any yo )t(x + y/A(x)) 1 A(x) uniformly for y E (0.xjX > x. . then 7(x) x/(a .
EK G(u)..(x). P The proof is based upon the following lemma (stated slightly more generally than needed at present). be i. i.n0 1•P(K= n)•n = EK.2 Let Y1.1/. and that for each Proof Recall from Section 1 that G*n (u) z > 1 there is a D < oo such that G*n(u) < G(u)Dzn for all u.nn.A(x) I X > x) = exp {A(x) .1 If Bo E S. cf.15. with EzK < oo for some z > 1..7) is referred to as 1/A(x) being selfneglecting.+YK> u) = ^•P(K = n)G* n(u ) . with common distribution G E S and let K be an independent integervalued r.2. THE COMPOUND POISSON MODEL 259 We omit the proof of (c) and that EX (x) . 0 G(u) L G(u) .. St > u}. It is trivially verified to hold for the Weibull. Notes and references A good general reference for subexponential distribution is Embrechts. Then P(Y1 + • • • + YK > u) . Theorem 2 . nG(u). The remaining statement (1. Let St = Ei ` Ui .14. then Vi(u) P Bo(u). Kliippelberg & Mikosch [134].and lognormal distributions . Y2. We get p(yl+. . We assume p = /3µB < 1 and are interested in the ruin probability V)(u) = P(M > u) = P(r(u) < oo). 1. u a oo.f yl 0 0 = exp {y (1 + 0(1))} 0 fY A( x + u /A( x)) a(x) du } The property (1. d.A(x + y/A(x))} =a(x) a(x + x) dx = ex p ex P .t be the claim surplus at time t and M = sups>0 St. Recall that B0 denotes the stationary excess distribution.. r(u) = inf it > 0.8) in (b) then follows from P (A(x)X (x) > y) = F(X > x + y/. 2 The compound Poisson model Consider the compound Poisson model with arrival intensity /3 and claim size distribution B. . Examples 1. Bo(x) = f0 B(y) dy / µB. Lemma 2.v.
a]. Bo ¢ S. .2. In general: Proposition 2.p) and EzK < oo whenever pz < 1. mathematically one must note that there exist (quite intricate) examples where B E S. lognormal . For some numerical studies. u The condition Bo E S is for all practical purposes equivalent to B E S. Since EK = p/(1.13)..x400 PBB(x) PB Leta+oo. However.p)p'. The approximation in Theorem 2. Bo E S is immediate in the regularly varying case. The PollaczeckKhinchine formula states that (in the setup of Lemma 2.1)xa1' vxe(109x11)2/202 2 +° /2 µB = eµ Bo(x) eµ+O2/2(log x)2 27r' = µB = F(1/0 ) Bo(x 1 ) . (2. as well as examples where B ¢ S.3 If B E S. _ B(x^sx Bo(x) µ8 I aoB(y )dy = (^) .x^ ) B(x) _ f or ( lox .1) In particular . see Abate.?(xµ 8 (x). Note that in these examples .µ J ) . the result follows immediately from Lemma 2.1 is notoriously not very accurate.2) M = Yl + • • • +YK where the Yt have distribution Bo and K is geometric with parameter p. P(K = k) = (1.µB(01 . r(1/Q) xlQexp B(x) = ex' From this . in our three main examples (regular variation . The tail of Bo is easily expressed in terms of the tail of B and the function y(x) in Proposition 1. Weibull) one has Bo(x ( B(x) . and for the lognormal and Weibull cases it can be verified using Pitman 's criterion (Proposition 1. x 4 00. then Bo(x)/B(x) + 00. Borovkov [73] and Pakes [280]. Proof Since B(x + y)/B(x) * 1 uniformly in y E [0. HEAVY TAILS u using dominated convergence with >2 P(K = n) Dz" as majorant.. u x+a Notes and references Theorem 2.260 CHAPTER IX. Proof of Theorem 2. The problem is a very slow rate of convergence as u ^ oo.1 is essentially due to von Bahr [56].18. Bo is more heavytailed than B .1. we have fx B(y)dy = a B0 (x) > lim inf lim inf x+oo B(x) . Bo E S. See also Embrechts & Veraverbeeke [136].
Snd) = Xl +. (3. The main result is: Theorem 3 . also a second order term is introduced but unfortunately it does not present a great improvement..1) this end . In [1]. Then K M=EY.g. i.. let t9+ = i9(0) be the first ascending ladder epoch of {Snd> }. THE RENEWAL MODEL 261 Choudhury & Whitt [1]. Based upon ideas of Hogan [200]. This shows that even the approximation is asymptotically correct in the tail.. + Xn. Let U= be the ith claim . Thus G+ is the ascending ladder height distribution (which is defective because of PB < PA)... there is a representation of M similar to the PollaczeckKhinchine formula.1 when u is small or moderately large.+ E A.1. M = sup s$ . p = iB /µA < 1.9+ < oo) = P(S. Kalashnikov [219] and Asmussen & Binswanger [27]. Asmussen & Binswanger [27] suggested an approximation which is substantially better than Theorem 2.1 gives 1010.e. in [219] p. 195 there are numerical examples where tp(u) is of order 105 but Theorem 2. 1 Assume that (a) the stationary excess distribution Bo of B is subexponential and that (b) B itself satisfies B(x . We assume positive safety loading. . Somewhat related work is in Omey & Willekens [278].] The proof is based upon the observation that also in the renewal setting. Then l/i(u) 1 P P [Note that (b) in particular holds if B E S. To Bo(u) u + 00. E.y)/B (x) > 1 uniformly on compact y internals.3. Define further 0 = IIG+II = P(r9+ < oo). t9(u) = inf {n : Snd> > u} . i=1 .Ti. [279]. 3 The renewal model We consider the renewal model with claim size distribution B and interarrival distribution A as in Chapter V. one may have to go out to values of 1/'(u) which are unrealistically small before the fit is reasonable.} Then ik(u) = F ( M > u) = P(i9 (u) < oo).. G+ (A) = P(Sq+ E A.. T+ < oo) where r+ = T1 + • • • + T. T1 the ith interarrival time and Xi = U. {n= 0.y + as usual denotes the first ascending ladder epoch of the continuous time claim surplus process {St}.
G_(A) = P(S.PBBo(x). A(dy) = 1.Y2.i. oo) = F(S. (b) and does not rely on the structure Xi = Ui . As for the compound Poisson model. 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.3) and we will prove it in this form (in the next Section. x * oo.y+ given r+ < oo).FI(u).y) dy = 1 Pi (X) oo IPG_ I .y_ E A) the descending ladder height distribution (IIG II = 1 because of PB < P A) and let PG_ be the mean of G_. 0 The lemma implies that (3.. FI (x) _ fz ° F(y) dy. Proof By dominated convergence and (b).Ti). and hence FI(x) .9)9'' and Y1.2 F(x) .1 IPG_ I / F(x . we will use the fact that the proof of (3. u a 00. with distribution G+/9 (the distribution of S. Lemma 3 .. are independent of K and i..B(x).1) is equivalent to P(M > u) " . whereas for large y . B(x) _ J O° B(B(x)y) A(dy) f 1 .1) holds for a general random walk satisfying the analogues of (a).FI(x) /IPG_I. A.oo. x > 0. the contribution from the interval (.d)) E A) denote the pre19+ occupation measure and let and U_ = Eo G'_" be the renewal measure corresponding to G_. Lemma 3 .3 G+ (x) . cf. (3.d. Let F denote the distribution of the Xi and F1 the integrated tail. Proof Let R+(A) = E E'+ ' I(S. Let further 19_ _ inf {n > 0: S^d^ < 0} be the first descending ladder epoch.N. U_ (dy) is close to Lebesgue measure on (. P(K = k) = (1 .g+ > x.2).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_. The heuristics is now that because of (b). x + oo.262 CHAPTER IX. Write G+( x) = G+ ( x. HEAVY TAILS where K is geometric with parameter 9. 0] normalized by IPG_ I so that we should have to G+(x) . d+ < oo). Then 0 0 F( x . 0] to the integral is O(F(x)) = o(FI(x)).(.
F(Y= > x) FI(x)/(OIp _ 1).y) U. 0] x+00 FI(x) 00 + lim up 1 x) E F(x + n) U_ (n .9)IpG_ I Differentiating the WienerHopf factorization identity (A.F[s] = (1 .1)/F(n) < 1 + e for n > N (this is possible by (b) and Lemma 3.3. Similarly. Given e.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. We then get lim sup G+(x) xro0 Fj(x) < lim sup X)00 o F(x .y) U_ (dy) 00 FI (x) < lim sup F(x) U(N.I n=N (1 E)2 r00 F(x + y) dy + e) lim sup .G+[s]) .1. If G_ is nonlattice.O[s])(1 .UG_ I x.=1 BIp G_ I (1. we can assume that the span is 1 and then the same conclusion holds since then U(n .1.e) z lim inf G+(x)  FI (x) Ip G_ I Letting a 10.0)0k k I(u) A. and that U_(n .1.9) 1 .2). and in the last that FI is asymptotically proportional to Bo E S. choose N such that F(n . then by Blackwell 's renewal theorem U_ (n .(dy) fN FI ( x) + lim sup ZY00 N F(x . In the lattice case. u Proof of Theorem 3. > (1 . n] + 1/I µG_ I. n] F1 ( n=N _1 1+e E F(x+n) 0 + limsup xr00 FI(x) FAG.1.3. n] is just the probability of a renewal at n.1. By Lemma 3.1 I .2) yields 00 F F I (u) P(M > u) _ E(1 . (3. Hence using dominated convergence precisely as for the compound Poisson model. THE RENEWAL MODEL 263 We now make this precise. n] < (1 + e)/1µc_ I for n > N. the proof is complete.
a.a. In view of the `one large claim' heuristics it seems reasonable to expect that similar results as for the compound Poisson and renewal models should hold in great generality even when allowing for such dependence. 10(0) But since P(M > u . . u)) > P(w(u) < oo)(i lp (0))• On the other hand.(u)+n .So( u)_1 < a) < P (w(u) < oo)j/i(0) < 0(0) P(M E (u . Therefore by Lemma 3. P(M > u.(1 .IIG+II)µc_ = (1 . S+q(u) ..a. with roots in von Bahr [56] and Pakes [280]. u). FJ(u) UBBO(U) PBo(u) N = (10)Ipc_I JUA .Sty(u)_I < a} we have w(u) < oo.SS(u)}n=o.4 on the joint distribution of (S. Sty(u) . 4 Models with dependent input We now generalize one step further and consider risk processes with dependent interclaim times.u)) = o(P (M > u)) = o(FI(u)). S+9(u) . and {Su.1 is due to Embrechts & Veraverbeke [136].Se(u)_1 < a) = o(Fj(u)).AB iP We conclude by a lemma needed in the next section: Lemma 3 . HEAVY TAILS µF = (1 . Notes and references Theorem 3.So(u)) are available. we have P(M E (u .1)6+[0] .2. on the set {M > u. see Asmussen & Kliippelberg [36].l. u)). Proof Let w(u) = inf {n : Sid) E (u . must attain a maximum > 0 so that P(M > u.0)ua_ .yiui_1...a.a) N P(M > u).264 and letting s = 0 yields CHAPTER IX.4 For any a < oo. allowing also for possible dependence between the arrival process and the claim sizes. Note that substantially sharper statements than Lemma 3. Mn < u}. Then P(M E (u .
. The zerodelayed case corresponds to Xo = Xl = 0 and we write then F0.1 based upon a regenerative assumption.. M. E0. 0o(u) etc..1 = max k=0. examples and counterexamples.. G(x) (4. .i. MODELS WITH DEPENDENT INPUT 265 Various criteria for this to be true were recently given by Asmussen.. such that {SXo+t  SXo}0<t< X 1Xo . 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 <. We give here one of them. {SX1+t .1 where the filled circles symbolize a regeneration in the path. M* = max S.. and the distribution of {Sxk+t . Define S.Sxi}0<t<x2Xl . (corresponding to the filled circles on Fig.. assume pp.. Thus the assumption . (viewed as random elements of the space of Dfunctions with finite lifelengths) are i. < 0 and EoX < oo where X = X2 . We return to this point in Example 4.1) is assumed. We let F* denote the Podistribution of Si. 4.n n=0.d. 4. {Sn}n=o.1..F*(X) = P0(Si > x) . 2..Sxk}o<t<xk+1xk is the same for all k = 1.4 below..1) ..... and apply it to the Markovmodulated model of Chapter VI.. Schmidli & Schmidt [47]. see [47].1 Note that no specific sample path structure of {St} (like in Fig. Figure 4.1 except for the first one) is a random walk..1. M = sup St. +1.. = Sx.X1 is the generic cycle.X2 < ... The idea is now to observe that in the zerodelayed case. See Fig. 4. For further approaches.. t>0 S.4. Theorem 4.
HEAVY TAILS for some G such that both G E S and Go E S makes (3. u p 00. Then '00 (u) = Fo(M > u) . the assumption means that Mix) and Sl are not too far away.2) to show F(M* > u) > 1.3) hold. Fo(Si > X).2) Imposing suitable conditions on the behaviour of {St} within a cycle will then ensure that M and M* are sufficiently close to be tail equivalent.3) applicable so that F(M* > u) 141 F*(u).4) liminf u>oo F(M > u) .266 CHAPTER IX..S. The one we focus on is Fo (Mix) > x) .1) and (4.* i o<t<xn+1x.1 Assume that (4..2 Theorem 4. N N Xi=0 N Figure 4. (4.Sxn = sup Sxn+t . Sxn +t . (4. See Fig. 4. (4.2. it suffices by (4. Proof Since M > M*... jF11 F* (U).3) where Mnx) = sup o<t<xn +1 X. Since clearly M(x) > Sl .
( u)1 > a) 00 1: Po(Mn<u. Then by Lemma 3..4).E) Po ( n max St u.Sn+1Sn>aV(uSn*)) n=1 00 > (1E)EPo(Mn<u. S. Let a > 0 be fixed.e)Po (M > u. To this end.1 = limti00 St/t.: S.5) which follows since Po (M > u.Mn +1 >aV(u n=1 00 S.Po (M* > u.e)Po (M > U). Po(M* > u) . We shall use the estimate Po(M > u) Miu^+ 1 < a) = o(Po (M > u)) (4. )) > (1 .a.(u) . MW O(u)+1 < a) IN ( U n=1 A1.S. MODELS WITH DEPENDENT INPUT Define 79* (u) = inf {n = 1 . /3(u) = inf{n=1. Under suitable conditions .e)Po (MMX> > x).Sn 0<t<x„+j ( 1 .2.. 2.(1 . u))/P(M* = 0) = o(Po(M* > u)).a. Given e > 0. choose a such that Po(Si > x ) > (1 . u)} < P(M* E (u . 0 yields (4.6) 1 p pBo(u) u where B is the Palm distribution of claims and p ..1 can be rewritten as 00 (U) (4.. assume the path structure Nt St = EUit+Zt i=1 . Letting first u + oo and next e . x > a. .. Theorem 4. E (u . M^xu)+l > a) . Mn+l > a V (u ...4.+Mn+1>u} 267 (note that {M> u} = {3(u) < oo}).: Sn > u} .4.
a4' 0.4).6 below. The same is true for Sl. (ii) EozNX < oo for some z > 1. Corollary 4.I u J Po(Sl > x) dx 1 EoNxB(x) dx EoX(1 . Assume further that (i) both B and Bo are subexponential. X and N. Mix) < > UE + i=1 o<t<x Thus Theorem 4. are Fmeasurable and NX Po J:U=>x i=1 (iv) Po sup Zt > x / (0:5t<x o(B(x)) Then (4.} and satisfying Zt/t N.3PB.7). and ENX Ui . (iii) For some o field Y. i=1 (4.p) Ju P Bo(u) 1p 0 .v. since the tail of Zx is lighter than B(x) by (iv).2 Assume that {St} is regenerative and satisfies (4.268 with {Zt} continuous. the proof of Lemma 4. cf. oX (see Proposition A1.8) x Write . Then the Palm distribution of claims is B(x) = E N Eo 0 I( U1 < x) . we get 00 (u) 1 IPF.1 is in force. independent of {> CHAPTER IX. HEAVY TAILS N` U.'s order EoNx • B(x). and the rest is just rewriting of constants: since p = 1+tlim St = 1+ . Proof It is easily seen that the r.X both have tails of sup Zt. and also for Mix) since Nx FNX U.Q = EoNx/EoX.6) holds with p = .
in particular lighttailed. Thus we conclude that (4. note that the asymptotics of i/io( u) is the same irrespective of whether the Brownian term Zt u in St is present or not. MODELS WITH DEPENDENT INPUT 269 Example 4 . and for some constants ci < oo such that cl + • • • + c. (i) holds.t} is standard compound Poisson and {Zt} an independent Brownian motion with mean zero and variance constant a2.. More precisely. . .6) u holds. . X2 = 1. and taking F = o.e. The number N. Again .9). we assume that B E S.6) holds.t + EN'I Ui where {>N`1 Ui .3 As a first quick application. 1) is Poisson with rate /3 = fo /3(s) ds so that (ii) holds. Bo E S. The key step of the proof is the following lemma.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). i=1 B = >2 7riaiBi i=1 and we assume p = 014 B = Ep ri/3ipB.0 (thus (iv) is trivial). Taking again Xo = Xi = 0.6) holds. Assume that B E S. Theorem 4. The arrival rate is /3i and the claim size distribution Bi when Jt = i. Zt .(NX). Example 4 . We now return to the Markovmodulated risk model of Chapter VI with background Markov process {Jt} with p < oo states and stationary distribution 7r. X3 = 2. < 1... We consider the case where one or more of the claim size distributions Bi are heavytailed. then (iv) holds since the distribution of supo<t<i Z(t) is the same as that of I Zl 1. Bo E S. The regenerative assumption is satisfied if we take Xo = Xi = 0. . consider the periodic model of VI. > 0.. 3 The average arrival rate / and the Palm distribution B of the claim sizes are given by P P Q = ir i/i. i. Then (4. In particular.5 Consider the Markovmodulated risk model with claim size distributions satisfying (4. X2 = 1. of claims arriving in [0.4 Assume that St = Zt . X3 = 2. 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 .. we conclude just as in Example 4.3 that (4.4. (iii) is obvious.
Then P P(Yx > x) . we can define the regenerations points as the times of returns to i. i =1 P(Yo > x I ^ ) < CG(x)zN1+ +Np for some C = C(z) < oo.X i=1 j=1 where conditionally upon F the Xi. 1. . oo) and define p Ni Yx = EEX'i .F) = P(Yo > X+x I •^) G (x +x)>2ciNi i=1 .. Let {Fi}t=1 P be a family of distributions on [0.2.. are independent with distribution Fi for Xij. Thus dominated convergence yields ( P(Yo>x P(Yo>x . . 6 Let (N1..F) < CG(x)zn'1+. u Proof of Theorem 4..v. Assume EzN1+"'+Np < oo for some z > 1 and all i. as x a oo.ciG(x). and F a aalgebra such that (N1.}P. X > 0 a r. For lighttailed distributions.. i1 = E\ G(x) In the general case.270 CHAPTER IX. i=1 Proof Consider first the case X = 0. cp with cl + > 0 it holds that Fi(x) . Markovmodulation typically decreases the adjustment coefficient y and thereby changes the order of magnitude of the ruin .. ."+Np .. An easy conditioning argument then yields the result when Jo is u random. i=1 P(Yx > x ^) < P(Y0 > x I.Fmeasurable. HEAVY TAILS Lemma 4 . . NP ) be a random vector in {0.G( x ) > ciNi . oo) such that G E S and some c1.c'(x) where c = ciENi . . 2 . If Jo = i. and the rest of the argument is then just as the proof of Corollary 4.. .5. and that for some + cp distribution G on [0. It follows by a slight extension of results from Section 1 that P P(Yo > x I Y) G( x) ci Ni.^•) G(x) P ^ E ciNi = C. . P P P(YX and > x I. NP ) and X are . The same dominated convergence argument completes the proof.
It follows from Theorem 4.p)Bo(u). The present approach via Theorem 4.4. this is applied for example to risk processes with Poisson cluster arrivals. Theorem 2. In contrast.. The main result of this section.1 is from Asmussen.2 and Example 4. for lighttailed distributions the value of the adjustment coefficient y is given by a delicate interaction between all B.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. For further studies of perturbations like in Corollary 4... cf.. ) form a general stationary sequence and the U. > 0) matter for determining the order of magnitude of the ruin probabilities in the heavytailed case. see Schlegel [316]. An improvement was given in Asmussen & Hojgaard [33]. VI.e. Theorem 5. Theorem 4. 5 Finitehorizon ruin probabilities We consider the compound Poisson model with p = /3pB < 1 and the stationary excess distribution Bo subexponential.T2. We start by reviewing some general facts which are fundamental for the analysis. Schmidli & Schmidt [47]. and independent of (T1.7. and the final reduction by Jelenkovic & Lazar [213]. IV. That paper also contains further criteria for regenerative input (in particular also a treatment of the delayed case which we have omitted here). I T(u) < oo).1.3. Floe Henriksen & Kliippelberg [31] by a lengthy argument which did not provide the constant in front of Bo(u) in final form. the discussion provides an alternative point of view to some results in Chapter IV.7). Within the class of risk processes in a Markovian environment.4.d.T2. in particular Proposition 2.6) to hold in a situation where the interclaim times (T1. Then O(u) . As usual. ). i. 5a Excursion theory for Markov processes Let until further notice {St} be an arbitrary Markov process with state space E (we write Px when So = x) and m a stationary measure. cf.pl(1 . this should be compared with the normal limit for the lighttailed case. r(u) is the time of ruin and as in IV. this then easily yields approximations for the finite horizon ruin probabilities (Corollary 5. as well as a condition for (4.5 shows that basically only the tail dominant claim size distributions (those with c. m is a (orfinite) . Combined with the approximation for O(u). FINITEHORIZON RUIN PROBABILITIES 271 probabilities for large u.4. cf.. 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). we let PN"N = P(. Notes and references Theorem 4.5.. states that under mild additional conditions. i.4. Essentially.i. 5 was first proved by Asmussen.
w(Fc) < oo ) 'In general Markov process theory.s. u For F C E.y = Qx (. r.s=j are the transition probabilities for {St}. but the example of relevance for us is the following: Proposition 5. Lebesgue measure. (note that we allow x. For the present purposes it suffices . oo). a main difficulty is to make sense to such excursions also when Px(w(F°) = 0) = 1. a familiar case is time reversion (here m is the stationary distribution). to the r. k as indicator functions. to consider only the case Px(w(F`) = 0) 0. St is distributed as y . Sw(F.). j. Say {St} is reflected Brownian motion on [0. in the terminology of general Markov process theory. .z. 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).rij = mjsji where r13. k on E.2) means ffh(a. Then (5. however . resp. for states i.2) for all bounded measurable functions h. We let QS be the corresponding distribution and Qx.1 A compound Poisson risk process {Rt} and its associated claim surplus process {St} are in classical duality w . and starting from So = y. m.s.h. {St} and {Rt} are in classical duality w.= y. HEAVY TAILS measure on E such that L for all measurable A C E and all t > 0. where we can take h.t + EI U. {Rt}.t.h. y to vary in. The equality of the l.r. the whole of R and not as usual impose the restrictions x > 0.z) dx G(dz) = ffh(y + z) k(y)dy G(dz). say. follows by the substitution y = x . Rt is distributed as x + t . an excursion in F starting from x E F is the (typically finite) piece of sample path' {St}o<t<w(F°) I So = x where w(Fc) = inf It > 0: St 0 F} .00). The simplest example is a discrete time discrete state space chain. and (5. x = 0+ and F = (0. Let G denote the distribution of ENt U.t.272 CHAPTER IX.2) with t = 1 means m. .)k(x . Thus. t.>N` Ui. Proof Starting from Ro = x.
s. But in the risk theory example (corresponding to which the sample paths are drawn). Qx y is the distribution of an excursion of {St} conditioned to start in x E F and terminate in y E F. QR and QRy are defined similarly.y as a measure on all strings of the form i0i1 .= y) Theorem 5 . and we let Qy y refer to the time reversed excursion .2 Qy.SS(F.. in = y. The theorem states that the path in (b) has the same distribution as an excursion of {Rt} conditioned to start in y < 0 and to end in x = 0. .1 The sample path in (a) is the excursion of {St} conditioned to start in x = 0 and to end in y > 0.y() = P ({SW(F`)t} 0<t<w(F °) E So = x. 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). w(0. i1. io = x.1 for the case F = (oo.(0)_ = y < 0 is the same as the distribution of w(y) where w(z) = inf It > 0 : Rt = z}.. /^s x (S1 = Z1. FINITEHORIZON RUIN PROBABILITIES 273 y E F (in discrete time. 5. S... this simply means the distribution of the path of {Rt} starting from y and stopped when 0 is hit. in E F. The theorem is illustrated in Fig . Sw(Fo)_ should be interpreted as Sw(F^)_1). [note that w(z) < oo a. z > 0. oo) = r(0) x= St y (a) Figure 5. when p = .y = Qy Q. Thus. Qx. Sn = in = y. .. .y(2p21 . in with i0. That is.5. 0]. Sn+1 E Fc) nx.. the one in (b) is the time reversed path..itt) = P Px(w(Fc) < 00..3 The distribution of r(0) given r(0) < oo. . Sw(F)1 = y) .13AB < 1] Proof of Theorem 5.2.). In particular: Corollary 5. x = 0. We can then view Qy.
in) = oo jEF^ Sxin1 ... Rn = in = x. Rn+1 E FC) TioilTili2. . i0 = y. . = in = y....ii .. in E F. .. S. in = x....ik_1EF .ik_1EF Sxin_1 ...TI( 2n2n _1 . in with 20. note first that Pt' (R l = il.. R Qy x(2p21 .. . To show Q y x (i0 i 1 . R . HEAVY TAILS E E Px (Si = 21i . 20 = y.... ..ik1EF Similarly but easier Sxin_1 . . Si1y 00 jEF° E E 5xik_ 1 ..... in = x... Silt' E Sxik_1 ..y(inin _ E SYj jEF` 00 Sxik _1 . 2p). Sn+1 E Fc) n=1 i1... 2p) when 20.. in) = Qx. Si11 S 1 .. 2n) = Qx..(F<)1 = Y) S S and Qx y( ipil . Rn+1 E F`) F (w(Fc) < 00...274 note that Fx(w(Fc) < 00...rin_1in E Txj jEFC m21 s2120 m2252221 m in Ssn n1 mjSjx Mx m2p mil min1 jEF` 1 Sinin _ 1 ...i„_iEF Similarly... i0) Q x.. t' y and Qy x are measures on all strings of the form ipi l .. MY Thus Qx(ioii ..in1 ... Si l io E mjSjx...gilt' k=1 ii . in)  Pt' (R1 = ii..... Si1y k=1 i1 .. . (Fc)1 = y) 00 CHAPTER IX. ....in E F... .J (i.. Rn = in = x.... S. 21 . Silt' E SO k=1 i1.
v. y > u.r.')distribution of Yu is Bo"). U T(O) = T (u) Y Figure 5.2 The distribution of (Y.t.'s are defined w. P(o) ). Y > u).5. We are interested in the conditional distribution of T(u) = T(0) given {T(0) < oo.')density of Y is B(y)/[.2.t. 1 w . 7(0) < oo. Bo") is also the P(u. Y > y} .r. ST(o) > y.')distribution of Z since P(Z>aIY>u) = 1 °° B(y) B(y + a) dy FLBBo(u) B (y) J°° (z) dy . the P(u. Z = Zl = ST+( 1)_ the value just before the first ladder epoch (these r.2. that is.(o) > y} = {T(0) < oo. see Fig.p. 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 .B(a) +a PBBo(u) . Z) is described in Theorem 111. That is. The formulation relevant for the present purposes states that Y has distribution Bo and that conditionally upon Y = y. that is. P(") = P(. 5. Now the P(u. Z follows the excess distribution B(Y) given by B(Y) (x) _ B(y + x)/B(y). S. Let Y = Yl = Sr+( 1) be the value of the claim surplus process just after the first ladder epoch .UBBo(u)]. the case r (O) < oo.2. To clarify the ideas we first consider the case where ruin occurs already in the first ladder segment . the distribution w.
Fig. Then. . It is straightforward that under the conditions of Proposition 1. That is . . the duration T+ (n) ...: r+ (n) < oo.p) in Pi"'')distribution. a slight rewriting may be more appealing. Bo") ). > u with high probability. r(u)/Z 4 1/(1 . i. .. Since w(z)/z a$. 2. Hence Z.. and distributed as (Y. Now Bo E S implies that the Bo ")(a) + 0 for any fixed a. The idea is now to observe that if K(u) = n. . this in principle determines the asymptotic behaviour of r(u). Zk be defined similarly as Y = Y1.18(c) Bo")(yY (u)) + P(W > y) ( 5.d.3.o be defined by w(z) = inf It > 0: Rt = z} where {Rt} is is independent of {St}.T+(2).. we get the same asymptotics as when n = 1.. Z = ZI but relative to the kth ladder segment.3 implies that the P("'1)distribution of T(u) = r(0) is that of w(Z)..1. Y1 + • • • + Yn > u} denote the number of ladder steps leading to ruin and P("'n) = P(• I r(u) < oo.p).p).. cf.. are i. and since its dominates the first n . denote the ladder epochs and let Yk. i.. In the proof. Z1). it therefore follows that T(u)/Z converges in Pi"'')probability to 1/(1 .276 CHAPTER IX.3) holds. . more precisely. HEAVY TAILS Let {w(z)}Z^. Recall the definition of the auxiliary function y(x) in Section 1. and YI. Zn).e. P(Z < a I Y > u) 3 0. Yn_1 'typical'. Z). 1/(1 .r+ (n . must be large and Z1.. Then Corollary 5. the random vectors (YI.i. in particular of Z. 5. . K(u) = n)... 4 Assume that Bo E S and that (5. However..p) then yields the final result T(u)/y(u) + W/(1 . then by the subexponential property Yn must be large. (Y.e.3) where the distribution of W is Pareto with mean one in case ( a) and exponential with mean one in case (b). Z/'y(u) * W in Pi "' ')distribution . Since the conditional distribution of Z is known (viz.p) in F(u) distribution. z ^ oo. . We now turn to the general case and will see that this conclusion also is true in P(")distribution: Theorem 5 .. let r+(1) = T(0). conditionally upon r+ (n) < oo. Then 7(u)/y(u) ^ W/(1 .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. Zn_1 'typical' which implies that the first n1 ladder segment must be short and the last long.
Bo (ri1) ®B( . . I A"(u ))II + 0. +Yn1<u.2. Lemma 5. the condition on A'(u) A A"(u) follows from Bo being subexponential (Proposition 1.n). Proof We shall use the easily proved fact that if A'(u). FINITEHORIZON RUIN PROBABILITIES 277 16 Z3 Z1 r+(1) T+(1) T+(1) Figure 5.n) (y1. Yn . .. Y„1. I A'(u)) = P(u.. then IIP( I A'(u)) Taking A'(u) = {Y. P (Yj. A"(u) _ {K(u)=n} = {Y1+ P(. . P(. Further.u) E •) . . suitably adapted)..u) E • I A'(u)) = Bo (n1) ®Bou) .u) II 0. > u}. A"(u) are events such that P(A'(u) AA"(u)) = o(F (A'(u)) (A = symmetrical difference of events).Yn1iYn ..5 Ilp(u...3 In the following..5. II ' II denotes the total variation norm between probability measures and ® product measure.Yl+ +Yf1>u}.
the discussion just before the statement of Theorem 5. see Fitzsimmons [144]). n.4.P) > y) Corollary 5. . Z' are arbitrary random vectors.P(Z' E •)II > 0 (here Y. be independent random vectors such that the conditional distribution of Zk given Y. It therefore suffices to show that the P(u'")distribution of T(u) has the asserted limit.u has distribution Bout That is.... .Bo (n1) ®Bo' 0. Y'.. Y") u etc. n .. copies of {w(z)}...6 IIPIu'n ) CHAPTER IX.. .. By Lemma 5.1. k = 1... the density of Yn is B(y)/[IBBO(u)].t. Z11). Z. P(u) since by Theorem 2.4)..p) < y). The same calculation as given above when n = 1 shows then that the marginal distribution of Zn is Bou). HEAVY TAILS ((Z1'. Y1 +. Thus F(u'n)(T(u) /7(u) > y) = F(u'n)((wl (Z1) + .. .278 Lemma 5 . +wn(Z n))l7( u ) > 1y) ^' P(u'n)(wn (Zn)/7(u) > y) 4 NW/(1 . . y > u.. . the F'distribution of r(u) is the same as the P'distribution of w1(Zl) + • • • + wn(Zn). Zn) E •) . in particular his Proposition (2. then 11P(Z E •) .1 P PBo(u) • P(W/(1 .i.. the marginal distribution of Zk is Bo for k < n. + Y" > u) Flul (K (u ) = n) _ Cu) P"F(1'i +. . n_1 < u.. and that Yk has marginal distribution B0 for k = 1.7 O (u. whereas wn(Zn) has the same limit behaviour as when n = 1 (cf. ..P) Bo(u) for n = 1.1 and Y„ . Proof Let (Y11... The first step is to observe that K(u) has a proper limit distribution w. Now use that if the conditional distribution of Z' given Y' is the same as the conditional distribution of Z given Y and JIF(Y E •) . 2. .+y 1 p"F(Yn > u) P)Pn1 P/(1 .). {wn(z)} be i.6. (Y. in our example Y = (Y1... ..1).. Notes and references Excursion theory for general Markov processes is a fairly abstract and advanced topic....2.d. Then according to Section 5a. . Zn).y(u)T) ..P(Y' E •)II * 0. . wk(Zk) has a proper limit distribution as u + oo for k < n. Proof of Theorem 5. Let {wl(z)}. and clearly Zi.r.. Zn are independent. For Theorem 5.' = y is BM.. Similarly (replace u by 0).
V.(u) = P(V > u) = f f (y) dy . u (6.1 Assume that B is subexponential and that p(x) > 00.6.. 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). Proof of Theorem 6. p(Y) and the result follows. claim size distribution B. We will show that the stationary density f (x) of {Vt} satisfies f (x) /B(x) r(x) We then get V. 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 . nontrivial and we refer to Asmussen [22]. Extensions to the Markovmodulated model of Chapter VI are in Asmussen & Hojgaard [33].(3 u u J B(y) dy . More precisely. max VB>0I Vo=0^ o<s<t J11JJJ Lemma 6 .1) The key step in the proof is the following lemma on the cycle maximum of the associated storage process {Vt}. 3. Assume for simplicity that {Vt} regenerates in state 0 . T) when T + oo with u fixed.2. the results only cover the regularly varying case.y) . = supo<t<0. Theorem 6 . . 6 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate /3. however. the probability that is exceeds u is then B(u . i. and define the cycle as a = inf{t>0: Vt=0. Corollary II. x > oo. that MQ becomes large as consequence of one big jump.e./3Ea B(u). The heuristic motivation is the usual in the heavytailed area. cf. one expects the level y form which the big jump occurs to be 0(1). that fo p(x)1 dx < oo.2 Define M. Then P(MT > u) . Asmussen & Teugels [53] studied approximations of i (u.1. The rigorous proof is. and premium rate p(x) at level x of the reserve. Then 0 (u) Qf "O ^) dy.B(u).
P(MT > u) $B(u) Ft µ(1 . Then D(u) = f(u)p(u) and.q ( u)) 1 . there exist constants c(u) 4 0 such that the limiting distribution of r(u)/c(u) given r(u) < oo is exponential.280 CHAPTER IX. D(u) = DQ(u)/µ. .q(u) Now just use that p(x) * oo implies q (x) + 0. HEAVY TAILS Define D(u) as the steadystate rate of downcrossings of {Vt} of level u and Da (u) as the expected number of downcrossings of level u during a cycle. It is also shown in that paper that typically. u Notes and references The results are from Asmussen [22]. by regenerative process theory. where also the (easier) case of p(x) having a finite limit is treated . Further the conditional distribution of the number of downcrossings of u during a cycle given Mo > u is geometric with parameter q(u) = P(Mo > u I Vo = u). Hence f (u)r(u) = D(u) = Do(u) .
.. Fox & Schrage [77].d. z) 2 = Zit NE ii ii According to standard central limit theory . Ripley [304]. We shall be brief concerning general aspects and refer to standard textbooks like Bratley. where a2 = Var(Z ).z) 4 N(0. 281 .2) is an asymptotic 95% confidence interval . replicates Zl.. ZN. The crude Monte Carlo ( CMC) method then amounts to simulating i. vrN(z . 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. and this is the form in which the result of the simulation experiment is commonly reported. topics of direct relevance for the study of ruin probabilities are treated in more depth.Chapter X Simulation methodology 1 Generalities This section gives a summary of some basic issues in simulation and Monte Carlo methods . Rubinstein [310] or Rubinstein & Melamed [311] for more detail .96s z f (1. estimating z by the empirical mean (Z1 + • • + ZN)/N and the variance of Z by the empirical variance N s2 = E(Z{  N 2. .i. 4Z). Hence 1.
an added programming effort. Typically variance reduction involves both some theoretical idea (in some cases also a mathematical calculation). The difficulty in the naive choice Z = I(T(u) < oo) is that Z can not be simulated in finite time: no finite segment of {St} can tell whether ruin will ultimately occur or not. We mention in particular ( regression adjusted) control variates and common random numbers. one can argue that unless Var(Z') is considerable smaller than Var(Z). generated at the same time as Z. Z = I inf Rt < 0 (0<t<T = I('r(u) < T). This is a classical area of the simulation literature. conditional Monte Carlo and importance sampling. variance reduction is hardly worthwhile. it is straightforward to use the CMC method to simulate the finite horizon ruin probability z = i. and many sophisticated ideas have been developed. Further. typically by modifying Z to an alternative estimator Z' with EZ' = EZ = z and (hopefully) Var(Z') < Var(Z). Sections 24 deal with alternative representations of Vi(u) allowing to overcome this difficulty. we then have EZ = EZ = z. lb Variance reduction techniques The purpose of the techniques we study is to reduce the variance on a CMC estimator Z of z. and in most cases this modest increase of N is totally unproblematic.282 CHAPTER X. 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. v. We survey two methods which are used below to study ruin probabilities.b(u. 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'. so that Z' is a candidate for a Monte Carlo estimator of z. The situation is more intricate for the infinite horizon ruin probability 0(u). writing Var(Z) = Var(E [Z I Y]) + E(Var[Z I Y]) . SIMULATION METHODOLOGY In the setting of ruin probabilities. and a longer CPU time to produce one replication. Letting Z' = E[Z I Y]. Conditional Monte Carlo Let Z be a CMC estimator and Y some other r . there are others which are widely used in other areas and potentially useful also for ruin probabilities. However. Say that Var(Z') = Var(Z)/2. Therefore.
Thus we cannot compute L = Z/z (further.3). the obvious possibility is to take F and P mutually equivalent and L = dP/dP as the likelihood ratio. Nevertheless. GENERALITIES 283 and ignoring the last term shows that Var(Z') < Var(Z) so that conditional Monte Carlo always leads to variance reduction. L such that z = EZ = E[LZ]. Then z Var(LZ) = E(LZ)2 .[E(LZ)] = E Z2 Zz .96 sis v^ N 2 1 where srs = N j(LiZi . However. (ZN. To this end. LN) from P and uses the estimator N zrs = N > L:Zj i=1 and the confidence interval zrs f 1. even if the optimal change of measure is not practical. L = z/Z (the event {Z = 0} is not a concern because P(Z = 0) = 0).v. Variance reduction may or may not be obtained: it depends on the choice of the alternative measure P. This may also be difficult to assess . . i. it may often be impossible to describe P in such a way that it is straightforward to simulate from P). . 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. and the problem is to make an efficient choice. In order to achieve (1. one would try to choose P to make large values of Z more likely. (1.. Thus. it gives a guidance: choose P such that dP/dP is as proportional to Z as possible.zrs) = 2 1 N 2 2 2 i=1 i=1 N > Lt Zi . L1).3) Thus. but tentatively.E [Z Z]2 = z2 . the argument cheats because we are simulating since z is not avaliable analytically. .e..1.zrs. it appears that we have produced an estimator with variance zero. using the CMC method one generates (Z1.z2 = 0. a crucial observation is that there is an optimal choice of P: define P by dP/dP = Z/EZ = Z/z.
The CMC method leads to a variance of oZ = z(1 . Z = I(A) and A is a rare event. z I.5 or even much smaller . u + oo.96 2 z2 z increases like z1 as z .e. 10 . SIMULATION METHODOLOGY 1c Rare events simulation The problem is to estimate z = P(A) when z is small . the optimal P is the conditional distribution given A. Again. The optimal change of measure ( as discussed above) is given by P(B) = E [ Z] i. let z(u) = P(A(u)). large sample sizes are required. However.1. but if the point estimate z is of the order 105. For each u. and let Z(u) be a Monte Carlo estimator of z(u).z) 1 > 00.B = iP(AB) = P(BIA). Z z V5 In other words . in terms of the halfwidth of the confidence interval. Two established efficiency criteria in rare events simulation are bounded relative error and logarithmic efficiency.. However. We shall focuse on importance sampling as a potential (though not the only) way to overcome this problem. we may try to make P look as much like P(•IA) as possible. In ruin probability theory. To introduce these.0. I. say of the order 103 or less. assume that the rare event A = A(u) depends on a parameter u (say A = {r(u) < oo}).z) 1001.e. assume that the A(u) are rare in the sense that z(u) * 0. 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. it does not help telling whether z is of the magnitude 104. as is the case of typical interest.. the issue is not so much that the precision is good as that relative precision is bad: oZ z(1 .96 2Z ( 1 .1. if z is small. This leads to the equation 1.e. A = {T(u) < T} or A = {r(u) < oo} and the rare events assumption amount to u being large. N . a confidence interval of width 10 4 may look small. and further it is usually not practicable to simulate from P(•IA). We then . Another way to illustrate the problem is in terms of the sample size N needed to acquire a given relative precision . i.96oz /(zV) = 0.100 . An example where this works out nicely is given in Section 3. say 10%.284 CHAPTER X. Thus.z) which tends to zero as z ^ 0.
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. are i. This allows Var(Z(u)) to decrease slightly slower than z(u)2. 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.e. where X1.1.log z(u) of (1.2. where Z = I(M > u) may be generated as follows: 1. this means that the sample size N = NE(u) required to obtain a given fixed relative precision (say a =10%) remains bounded. logarithmic efficiency is almost as good as bounded relative error.. let Z +.d. The term logarithmic comes from the equivalent form . Var(Z(u)) hm sup U+00 z (u) 2E < oo (1. it is appealing to combine with some variance reduction method . it is not efficient for large u . 3. X2. . where M = X1 + • • • + XK. However. We shall here present an algorithm developed by Asmussen & Binswanger [ 271. F(K = k) = (1 . According to the above discussion.4). 2 Simulation via the PollaczeckKhinchine formula For the compound Poisson model. Notes and references For surveys on rare events simulation.0.(2. O (u) = z = EZ. which gives a logarithmically efficient estimator .X1 + + XK.p)pk. . with common density bo(x) = B(x)/µB and K is geometric with parameter p. see Asmussen & Rubinstein [45] and Heidelberger [190]. and in practice.p)pk..4) for any e > 0.. P(K = k) = (1 . The algorithm gives a solution to the infinite horizon problem . Therefore . i. the PollaczeckKhinchine formula III. If M > u. Thus.log Var(Z(u)) lim inf > 2 u+oo . .. Generate X1. XK from the density bo(x). Otherwise. the mathematical definition puts certain restrictions on this growth rate. 2. so that NE (u) may go to infinity. Generate K as geometric.1) may be written as V) (u) = P(M > u).i. let Z +. Let M . but as a CMC method .
For the simulation...X1 ..p)Bo(x)...286 CHAPTER X. asymptotically it presents no improvement : the variance is of the same order of magnitude F(x).Xl . Z(1) (u) has a smaller variance than Zl (x). K > 2] = P2p(Xl > x) = P2Bo(x) (here we used that by positivity of the X.. assume in the following that Bo(x) ...X(K1)) .. Xl > x.. we generate only X1.XK_1). note first that To check the formula for the P(X(n) > x I X(1). However. y < 0).. XK1. XK...X(n_1)) Bo(X(„_l) V X) Bo(X(n1)) . . and that Bo(y) = 1. just note that EZ(1)(u ) 2 > E[Bo (x .. compute Y = u . To see this..b(u) = P (Xl +•••+XK>u) = EF[Xl + . .L(x)/x`' with a > 0 and L(x) slowly varying. conditional probability.X(2). Theorem IX. So.1) V)(u) . SIMULATION METHODOLOGY when the claim size distribution B (and hence Bo) has a regularly varying tail. As a conditional Monte Carlo estimator ..2.. form the order statistics X(1) < X(2) < . Then (cf.p/(l ..SK1)2. Thus.X(2). Z(1)(u) is defined as 0)... ... and considering only the remaining ones.. and let Z(2)(u) = _ P (SK B0((u > u I X(l).. X1 + + XK_ 1 > x when X1 > x. .... < X(K) throw away the largest one X(K).S( K_1)) V X(K1)) / Bo(X(K 1)) where S(K_l) = X(1) + X(2) + • • • + X(K_1). and the problem is to produce an estimator Z(u) with a variance going to zero not slower (in the logarithmic sense ) than Bo(u)2..XK1] = EBo(uX1 .+XK > uIXl. A first obvious idea is to use conditional Monte Carlo: write i.XK_1 and let Z( 1)(u) = Bo (Y) (if K = 0. The idea of [27] is to avoid this problem by discarding the largest X. we thus generate K and X1i . 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.
1) . However .Khinchine formula and importance sampling . Binswanger and HOjgaard of [28] give a general survey of rare events simulation for heavy tailed distributions .. 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.S (n1)) V X (. . 111. X(2).. IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 287 We then get P(S" > x I X( 1). and simulate from FL. The algorithm is sofar the only efficient one which has been developed for the heavytailed case. Notes and references The proof of Theorem 2. BL by I3L = /3B[y]. X . . BL instead of 0.S(n_1) I X(1). in the renewal or Markov.5). Thus. X(2). X(n1)) Bo((x . X (. Then the algorithm given by { Z (2) (u) } is logarithmically efficient. 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. using 13L.. X(n1)) P(X(TZ) + S(. and that paper contains one more logarithmically efficient algorithm for the compound Poisson model using the Pollaczeck.y. use the the CramerLundberg approximation so that z(u) = '(u) = e7"ELe7E(") where ^(u) = ST(") . the continuoustime process {St} is simulated by considering it at the discrete epochs {Qk} corresponding to claim arrivals. that is. Compute y > 0 as solution of the Lundberg equation 0 = K(y) = )3(B[y] . the algorithm for generating Z = Z(u) is: 1. l)) BO(X(n1)) Theorem 2 .. For practical purposes._1) > P(X(n) > _ X X(1). Asmussen . . 1 Assume that Bo (x) = L(x)/x° with L(x) slowly varying. BL(dx) = e7sB(dx)/B[y]. .. and define )3L. l)) . X(2). for the purpose of recording Z(u) = erysr(u). ..u is the representation 0(u) = e7sr(u) overshoot (cf. and we refer to [27]. B. 3 Importance sampling via Lundberg conjugation We consider again the compound Poisson model and assume the conditions of Ce7".1 is elementary but lengty.3.modulated model P(r+ < oo) and G+ are not explicit ). .
Let S . so that changing the measure to FL is close to the optimal scheme for importance sampling . 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. to deal with the infinite horizon problem . B) is not logarithmically efficient when (/3. namely ELery£("). The answer is no.2ryu _ z (u)2/C2. There are various intuitive reasons that this should be a good algorithm. and avoid simulating the known part e7".d. and we have: Theorem 3... b) # (/3L.i. SIMULATION METHODOLOGY 3. BL). Let FL (dx) = 'For the renewal model. Generate T as being exponential with parameter . Otherwise. . X2. and the change of measure F r FL corresponds to B > BL. We may expect a small variance since we have used our knowledge of the form of 0(u) to isolate what is really unknown. P'[y] < oo for some ry > 0.288 2.F. 4. the results of IV. let Z F e_'s. We formulate this in a slightly more general random walk setting '. BL could improve the variance of the estimator . Let X1.1) (simulated with parameters ^3. cf. let S. with distribution F. be i.. u It is tempting to ask whether choosing importance sampling parameters .T. return to 3. and assume that µF < 0 and that F[y] = 1. r(u) < oo) and FL (both measures restricted to.. one must restrict attention to the case 4µB > 1. b different from .3. Ti. The algorithm generalizes easily to the renewal model . Let Sf0 CHAPTER X.1 The estimator Z(u) = e'rs* "u) (simulated from FL) has bounded relative error.. Xi = U.. = X1 + .S+U .l3 and U from B. + X. A > AL as in Chapter V. It resolves the infinite horizon problem since FL(. More precisely. the discussion at the end of Section 1b. . In fact: Theorem 3.QL..Q.2 The estimator (3. In detail . The proof is given below as a corollary to Theorem 3.r(u) < oo) = 1. M(u) = inf {n : S„ > u}.(u)) are asymptotically coincide on {r(u)} < oo. Proof Just note that EZ(u)2 < e . If S > u.7 tell that P(.
Since ELM(u)/u + 1/ELXi. where e' = EL Iog dFL (Xi) > 0 by the information inequality.2) dF Theorem 3..2) (simulated with distribution F of the X3 has bounded relative error when . . For the second.3. EFZ(u)2 = EeW2(FIF) = Ep [W2(FIFL)W2(FLIF)] = EL [W2 ( FIFL)w(FLIF)] = ELexp {Kl+. Since K1. Proof The first statement is proved exactly as Theorem 3 .d. write W(F IF) _ F(XI). where Kl og (X) (j) 2 ) = log dFL (Xi) . K2. When F # FL. Here ELK. By the chain rule for RadonNikodym derivatives..2ryELXi. = c'. EFZ(u)2 EFZ(u)2 lim sup z(u)2eeU = lim cop C2e2..2 > 0. 1.. + KM(u))} = exp {ELM(u)(E . are i.+KM(u)}.yu+elu u +oo etry' 1 > lim up C2e2. IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 289 e7yF(dx). is not logarithmically efficient.i. it thus follows that for 0 < e < e'/ELXi.yu = G. F(XM(u)).2ryELXi)} ... let F be an importance sampling distribution equivalent to F and M(u) dF Z(u) _ I (Xi) . The importance sampling estimator is then Z( u) = e'rSM( )...2'X1 .3 The estimator (3. . More generally. (3. Jensen's inequality and Wald's identity yield EpZ(u)2 > exp {EL(K1 + .P = FL..
3'eO'x f f P (U" . u Notes and references The importance sampling method was suggested by Siegmund [343] for discrete time random walks and further studied by Asmussen [ 13] in the setting of compound Poisson risk models . U' . with the present (shorter and more elementary) proof taken from Asmussen & Rubinstein [45].e. so that one would expect the change of measure F 4 FL to produce close to optimal results.i. The results of IV.2. T".3"eQ x 0 J eQ zB (z) dz x (x > 0) and /3' = /3". SIMULATION METHODOLOGY u Proof of Theorem 3.4 indicate that we can expect a major difference according to whether y < 1/r. then the estimator Z(u) = e7Sr(°)I(r(u) < yu) (simulated with parameters /3L. In fact: Proposition 4. T) with T < oo. BL) has bounded relative error. /3'eQ'YR'( x + y) dy = . 4 Importance sampling for the finite horizon case The problem is to produce efficient simulation estimators for '0 (u. . the references in Asmussen & Rubinstein [45] and Heidelberger [190]. /3".B'=B". U' .1 is from Lehtonen & Nyrhinen [244].4. from 3' P(U'T'>x) ^ = ^ eQ'zB (z) dz. The extension to the Markovian environment model is straightforward and was suggested in Asmussen [ 16].'(y). Then according to Theorem 3. optimality is discussed in a heavy traffic limit y 10 rather than when u + oo. see e.'(y) or y > 1/r. The optimality result Theorem 3. Next. Consider compound Poisson risk process with intensities /3'. In [13]. CHAPTER X. As in IV.290 which completes the proof.T' = U" .T". we conclude by differentiation that Bo(x)=B' (x)forallx > 0. Further discussion is in Lehtonen & Nyrhinen [245].T".T' has a left exponential tail with rate /3' and U" .1 If y > 1/ic'('y). U". yu) is close to zk(u).T" > x) J /3"e0 yB (x + y) dy = .g. claim size distributions B'.T" has a left exponential tail with rate /3'. This immediately yields / = 3". generic interarrival times T' . First by the memoryless distribution of the exponential distribution . all that needs to be shown is that if U' .3. we write T = yu. The queueing literature on related algorithms is extensive . The easy case is y > 1/k'(y) where O(u.T' D U" . then /3' B' = B". B" and generic claim sizes U'.
2 The estimator (4. IMPORTANCE SAMPLING FOR THE FINITE HORIZON CASE 291 Proof The assumption y > 1/n'(y) ensures that 1fi(u.O(u. Remark 4 . yu) in the sense that . yu) is of order of magnitude a71. 3 Theorem IV.1).(u) * 1 (Theorem IV. T(u) < yu] .yk(ay) determines the order of magnitude of z'(u.'(7). Bounding u ELZ(u)2 above by a7u. (4. (4.log Var(Z(u)) l im of .4. the result follows as in the proof of Theorem 3.(ay).4.3) and we have: Theorem 4.3) (simulated with parameters /gay. and that ryy > ry.1) which is all that is needed here can be showed much easier .2) Since the definition of ay is equivalent to Eay r(u) . 7y (4.log Var(Z(u)) _ .yy> 2 . The corresponding estimator is Z(u) = eavS' ( u)+T(u)K (ay)I(T( u) < yu). yu) = eayu Eay Leay^(u)+r(u)K(ay).1). yu)/z.8). one would expect that the change of measure F Pay is in some sense optimal.4. Proof Since ryy > y.yu.4. lim inf uoo 27yu . T( u) < yu] e2ryyuEay le 2ay^(u).8 has a stronger conclusion than (4.to g x ( u ) u u so that (1.1) so that z(u) = zP(u.log 4')u) 4 u (Theorem IV. T(u) < yu] e Hence by (4. we have ic(ay ) > 0 and get Eay Z(u)2 = Eay [e  2aySr( u)+2r(u )r. and in fact. that ryy = ay . .1. Bay) is logarithmically efficient. We recall that ay is defined as the solution of a'(a) = 1/y.5) follows. We next consider the case y < 1/r. Further . Let Qy2 = .
Qyu1/2 < T(u) C yu e. the algorithm in Section 3 produces simulation estimates for the tail P(W > u) of the GI/G/1 waiting time W). yu . related discussion is given in a heavy traffic limit q J.4. 0 rather than when u 3 oo.7y x(u) > hm inf u+Oo U .o . (5.3. '%(u) = P I info Rt < 0) = P(VV > u).3: for many risk processes {Rt }.u1/2 < r(u) < yu Le ] l = e.1): then by Proposition A1.4).a yu +l/ur' (av)Ei`av reav^(u)+(T(++)(U) yu .1) where the identity for Vi(u) requires that Vt has a limit in distribution V. One main example is {Vt} being regenerative (see A. In most of the simulation literature (say in queueing applications). Hence lira inf log ryyu + vyu 1/2 tc(ay) . there exists a dual process { V t} such that i.yu1/2 <1 T(u) < yu l r > e7vu +avul/ 2r. we believe that there are examples also in risk theory where (5. Then z(u) = Eay Z(u) > Eay avS'(u)+T( u)k(av 1 ). yu .uaoo U That lim sup < follows similarly but easier as when estimating En. the object of interest is {Vt} rather than {Rt}.a vt(u).2).T) = P O<t<T inf Rt < 0 = P(VT > u).(av)Eav l e.292 CHAPTER X. zi(u) = INV. However.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..1) (see Proposition IV.o.1) is used to study Voo by simulating {Rt} (for example. In Asmussen [13].4. N(0. > u) = E f I(VV > u) dt 0 (5. 5 Regenerative simulation Our starting point is the duality representations in 11.1) may be useful.b(u.yu)/(uyu1/2) . SIMULATION METHODOLOGY Vara„ (r(u))/u so that (T(u) . and (5.2) . Z (u)2 above.
Simulate a zerodelayed version of {V t } until a large number N of cycles have been completed. To derive confidence intervals . > u) (and more general expectations Eg(V. a standard transformation technique (sometimes called the delta method) yields 1 V 2 (h (Zi. record Zi'i = (Z1'). the regenerative estimator z%(u) is consistent. For details .. Thus the method provides one answer on to how to avoid to simulate { Rt} for an infinitely long time period.d. )). + Z2N)) . Z2 a4* z2.t(u)) 4 N(0. + Z1N>) .+Z(N) z 1. provides estimates for F ( V. z2) z2/z1 yields Vh = (z2/z2 1/zl). Zl the LLN yields Z1 a$' Z(1) +. oh) for h : R2 ^ R and Ch = VhEVh. The method of regenerative simulation. j = 1. EZ2'i = z2 = E Thus.. Therefore . Z1 = (Zl1i +.3) . Z2'> the time during the cycle where { Vt} exceeds u and zj = EZJ'). 02) (5. i (^(u) .. which we survey below . Then (Z1z1i Z2z2 ) 4 N2(0...h (zl. letting J0 'o I (Vt > u) dt .. let E denote the 2 x 2 covariance matrix of Z('). Thus. . is the cycle length. Taking h(zl. and Z2'>) where Zi'i = w. Z2 = N (X21' + . Vh = (8h/8z1 8h/ 8z2). REGENERATIVE SIMULATION 293 where w is the generic cycle for {Vt}. (u) ?2 = E fo I(Vt > u) dt = 0( u ) zl Ew as N > oo. Z2 .E). consider first the case of independent cycles .. . For the ith cycle.. z2)) > N(O.. Then Z(1). Z(N) are i . i. 2.5... EZ1'i = z1 = Ew.
z^ i=1 so a2 can be estimated by 2 2 = 72 S11+ 12 S22 . Rubinstein [310] and Rubinstein & Melamed [311]. 6 Sensitivity analysis We return to the problem of 111 . There is potential also for combining with some variance reduction method. with distribution depending on a parameter (. We here consider simulation algorithms which have the potential of applying to substantially more complex situations. Notes and references The literature on regenerative simulation is extensive.Z) ^Z(=) .96s/v"N. However . Let X have a density f (x. say risk processes with a complicated structure of the point process of claim arrivals and heavy tailed claims .g. () dx so that differentiation yields zS d( fco(x)f(x. Z of the form Z = ^p(X) where X is a r .2. SIMULATION METHODOLOGY 2 Eli = Z2 z1 + 2 E22 .0 . In 111. The regenerative method is not likely to be efficient for large u but rather a brute force one. the expectation z = EZ of a single r.v.5) Z1 Z1 Z1 and the 95% confidence interval is z1 (u) ± 1. see e. () depending on C. in some situations it may be the only one resolving the infinite horizon problem . Before going into the complications of ruin probabilities .294 where 01 2 CHAPTER X.2 E1 2 z1 z1 Z2 The natural estimator for E is the empirical covariance matrix N S = N 1 12 (ZW . () dx = E[SZ] f(X.g S12 (5.9. Here are the ideas of the two main appfoaches in today 's simulation literature: The score function ( SF) method . to evaluate the sensitivity z/i( (u ) = (d/d() 0(u) where ( is some parameter governing the risk process . 9. Then z(() = f cp(x) f (x. asymptotic estimates were derived using the renewal equation for z /i(u). () dx f Ax) (dl d()f (x' () f ( z. v.C)dx = f w(x) d( f ( x. consider an extremely simple example .
The likelihood ratio up to r(u) for two Poisson processes with rates /3. ()) is 0 for C < Co and 1 for C > Co so that the sample path derivative cp'(h(U. giving h( (U.t.6. () is an unbiased Monte Carlo estimator of zS. however . In the setting of ruin probabilities . r(u) = Tl + • • • +TM(u)). For example . A related difficulty occurs in situations involving the Poisson number Nt of claims: also here the sample path derivative w. Thus. for some Co = (o(U). () = .() d( is the score function familiar from statistics .1 Consider the sensitivity tka(u) w. say W(x) = I(x > xo) and assume that h(U.r. Let M(u) be the number of claims up to the time r(u) of ruin (thus. SENSITIVITY ANALYSIS where 295 S = (d/d()f (X. C). Infinitesimal perturbation analysis (IPA) uses sample path derivatives. () = (8/8()h (u. The following example demonstrates how the SF method handles this situation. So assume that a r. SZ is an unbiased Monte Carlo estimator of z(. () where U is uniform(0. = E [`d (h(U. Example 6 . Then z(() = Ecp(h(U. For the SF method. To see this. The derivations of these two estimators is heuristic in that both use an interchange of expectation and differentiation that needs to be justified. () = d log f (X. Thus . ()) is 0 w . /3o is M(u) Oe (3T: < oo) . this phenomenon is particularly unpleasant since indicators occur widely in the CMC estimators . () Thus. ( where h( (u.t.r. C) f(X. For IPA there are. this is usually unproblematic and involves some application of dominated convergence . () _ (eSx. () = log U/(2. zc = E [d co(h(U. one can take h (U. () is increasing in C. IPA will estimate zS by 0 which is obviously not correct. the Poisson rate /3 in the compound Poisson model. ()) h((U. ()) d( hc(U. cp(h(U.1). ()).log U/(. p. cp' (h(U. one. 11 /3oeOoT. if f (x. nonpathological examples where sample path derivatives fail to produce estimators with the correct expectation. Then . just take cp as an indicator function .v. with density f (x. I(r(u) . C)). /3 is 0. () can be generated as h(U.
r. a relevant reference is VazquezAbad [374]. Example 6. differentiating w. However. whereas for the SF method we refer to Rubinstein & Shapiro [312]. ) we have VarL(ZQ(u)) ZO(u)2 O(u2)e2 u2e2ryu yu .3L. We then arrive at the estimator ZZ(u) = (M(u) . the risk process should be simulated with parameters .r.T(u)) e7uerVu) for ?P. We recall (Proposition 111. for different models and for the sensitivities w. since ELZp(u)2 < (M(U) _T(u) \ 1 2 a2ryu = O(u2)e27u. change the measure to FL as when simulating tp(u). 0 Notes and references A survey of IPA and references is given by Glasserman [161] (see also Suri [358] for a tutorial).9 . 4) that V5. j3 and letting flo = 0. the estimation of z(ip(u) is subject to the same problem concerning relative precision as in rare events simulation . Thus. BL). different parameters.t. SIMULATION METHODOLOGY Taking expectation.t. .0(1) so that in fact the estimator Zf(u) has bounded relative error. we get 1 M(u) 00(u) = E (_Ti)I(T(U)<) E [(M(u) .T(u)) I(T(u) < co) ] . To resolve the infinite horizon problem . There have been much work on resolving the difficulties associated with IPA pointed out above.1 is from Asmussen & Rubinstein [46] who also work out a number of similar sensitivity estimators.296 CHAPTER X. In the setting of ruin probabilities.3 (u) is of the order of magnitude ue7u. in part for different measures of risk than ruin probabilities.3 (u) (to generate Zp (u).
d.. a) = r(u)). Besides its intrinsic interest . That is.. Consider first a Bernoulli random walk.(u) = 0 ) = 0) or it is trivial to translate from one setup to the other. where X1. R„ = u+X.. Y'a(U) = P(T (u) = r+(a)) = 1 . in the Bernoulli random walk example below.(u) is defined as the probability of being ruined (starting from u) before the reserve reaches level a > u. as e. '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 ).. T(u. wherel T(u) = inf {t > 0 : Rt < 0} .Chapter XI Miscellaneous topics 1 The ruin problem for Bernoulli random walk and Brownian motion.1. and {1.i. 297 . X2.. are i. T+(a) = inf It > 0 : Rt > al...1}valued .+• • •+X. defined as Ro = u (with u E {0.P(•r(u. with P(Xk = 1) = 9.. }). either this makes no difference (P(R.. The twobarrier ruin problem The twobarrier ruin probability 0. in most cases . Oa(U ) can also be a useful vehicle for computing t/i(u) by letting a * oo.. a) = r(u) A T+(a).g. .
We choose a = ry where ry is the Lundberg exponent.. and the other more advanced but applicable also in some other settings.(1B)u oJ 0. 7/la(a . In a general random walk setting .r(u.a(u)).e. where a is any number such that Ee°X = F[a] <oo. u + 1. The martingale is then {zuzXl+•••+X„ } = {zR° }.y] = 1. By optional stopping. Wald's exponential martingale is defined as in 11.2).. one elementary but difficult to generalize to other models. tba(2) _ (1 .1) is solution..+Xn) F[ a]n n=0..1. Conditioning upon X1 yields immediately the recursion 'a(1) = 19+00a(2).1) o If 0 = 1/ 2. and in view of the discrete nature of a Bernoulli random walk we write z = e7.o)T/la (1) + 8z/'u(3). MISCELLANEOUS TOPICS Proposition 1. u Proof 2. C1_0\a.(4. i. (1.1) = (19)4/'0(a3)+9ba(a1). = (1 .298 CHAPTER XI. then 'Oa(u) _ au a We give two proofs . = z°Va(u) + za(1  .a) = 0) + zap ( R.0)/0. the solution of F[. The Lundberg equation becomes 1=F[ry]=(19)+9z.1 For a Bernoulli random walk with 0 0 1/2.. z and the solution is z = (1 . and insertion shows that ( 1.o» = z°P (RT ( u.o)'t/1a(a .(u) I\ e = 1 oa ' ()i a = u..a) Y.. zu = EzRO = EzRT(u.. Proof 1.4) by ea(u+Xl+.2) Oa(a ..
pa( u) _ u Corollary 1. If p<0.e7u)/(e7° . and solving for 9/la(u) yields Z/)a(u) = (e 76 .u)/u.2 For a Bernoulli random walk with 9 > 1/2.u) = et(a2 /2 +aµ) the Lundberg equation is rye/2'yp = 0 with solution y = 2p.ba(u) = e2µa .1.5) . thenz1 (u)=1. (1. . u Proposition 1. {R. Then for p 0 0.1 yields 't/la(u) = (a . i1(u) = e211 . RANDOM WALK. Proof Let a+ oo in (1. {Rt} is itself a martingale and just the same calculation as in the u proof of Proposition 1.1 If p = 0. then Vi(u) = 1.zu)/(za . However. TWO BARRIERS 299 and solving for 4/la(u) yields t/ia(u) = (za .. (1. Applying optional stopping to the exponential martingale {e7R. } yields e7u = Ee7R° = e°Wa(u) + e7a(1 . BROWNIAN MOTION.3 Let {Rt} be Brownian motion starting from u and with drift p and unit variance . If p = 0.} is then itself a martingale and we get in a similar manner u = ER° = ER ra( u) = 0 • Y'a (u) + all  au Y'a( u)). If 9 = 1/2.1).2) is trivial (z = 1).1) for p # 0.1). then Proof Since 'Oa (U)  au a Eea(R°.4 For a Brownian motion with drift u > 0. Corollary 1.• a2µa e2µu .0a(u)). 1h (u) = a el u \1 If 9 < 1/ 2.
passing to even more general cases the method quickly becomes unfeasible (see. 1 .5a). CHAPTER XI.7/la(u)). Ic 5ry 'pa(u) Using y = 6 . 5). . say. (u) _ O(u) .a) < 0) + e7°P (R(u. Here is one more case where this is feasible: Example 1. this immediately yields (1.e7a (u) = 6 /0 . and hence e7u = Ee7Ro E [e7R(. however.a) = a on {r (u. letting a * oo yields the standard expression pe7u for . the paths are upwards skipfree but not downwards.300 Proof Let a * oo in (1.vi(a) Proof By the upwards skipfree property. and thus one encounters the problem of controlling the undershoot under level 0.7). 0.3. VIII. implying R(u.a) = r+ (a)} and similarly for the boundary 0. It may then be easier to first compute the onebarrier ruin probability O(u): Proposition 1. MISCELLANEOUS TOPICS u The reason that the calculations work out so smoothly for Bernoulli random walks and Brownian motion is the skipfree nature of the paths.a) = a ) + e ' ° ( 1 . valid if p < 1 (otherwise . a) I R(u a ) < 0] P (R(u . However.616).4). For most standard risk processes .7) . (1.6 If the paths of {Rt} are upwards skipfree and 7//(a) < 1.a ) < 0) + e 7aF ( R (u.5 Consider the compound Poisson model with exponential claims (with rate.a) = a) = 5 y = P (R (u. 7O(u) = 7/la(u) + (1 . Here the undershoot under 0 is exponential with rate 5.+^a(u))^(a) If 7k(a) < 1.0(a) 0 < u < a. we obtain 'Oa a7u .0 (u) (where u p =.. 7/'(u) = 1).e7a Again .
= eµuTµ2/2Po (T( u) E dT) 2 eµuTµ2/2 u T3/2 ex p u 27r p 12 T . Here {St } is Brownian motion with drift 0 (starting from 0). 10) follows then by straightforward differentiation..9) = 2P(ST > u). P(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).d. For the symmetric (drift 0) case these are easily computable by means of the reflection principle: Proposition 1.8 Let {Rt} be Brownian motion with drift . = 1 ..11 ) is the same as (1. TWO BARRIERS 301 Note thas this argument has already been used in VII. (1.T) P(MT > u) where MT = maxo<t<T St. the density dPµ / dP0 of St is eµsttµ2/2.11) VIT ) Proof For p = 0.µ so that {St} is Brownian motion with drift µ . and hence Pµ('r(u) E dT) = Eo [e µsr(.ST<u) = P(MT>u. MT > u) = P (ST > u) + P (ST > u. 0(u. Corollary 1. ( 1. T(u) E dT..µ%T (1.4) I = . we have ili(u. Then the density and c.10) Pµ (T(u) < T) !..7 For Brownian motion with drift 0.8) Proof In terms of the claim surplus process { St} = {u .8 ).1a for computing ruin probabilities for a twostep premium function. For µ # 0. BROWNIAN MOTION. of r(u) are ( U2 Pµ (T(u ) E dT) = 2^T 3/2 exp µu . Hence P(MT>u.2 . + µ2T) } .µ T I + e2µ"4) ( .f. and (1 .Rt}.)_ _( u)µ2 /2..r(u). MT > U) = P(ST > u) + P(ST > u) (1. RANDOM WALK. (i). T ) = P(T(u) < T ) = 241.1. We now return to Bernoulli random walk and Brownian motion to consider finite horizon ruin probabilities.
oo) with drift µ(x) and variance a2 (x) at x. S(x) = f x s(y)dy. as defined above as the probability of actually hitting 0. S(oo) = f c s(y)dy. is (1.10) and that the value at 0 is 0.10 Consider a diffusion process {Rt} on [0. We finally consider a general diffusion {Rt} on [0.13) with 0 as lower limit of integration. MISCELLANEOUS TOPICS which is the same as (1. and in a similar spirit as in VII. Let s(y) = ef0 ry(. as defined in (1. T are integervalued and nonnegative.h. the behaviour at the boundary 0 is more complicated and it may happen. (1.8 also applies to the case 9 54 1/2.3 we can define the local adjustment coefficient y(x) as the one 2µ(x)/a2(x) for the locally approximating Brownian motion..g. that 0(u).. but we omit the details.13) The following results gives a complete solution of the ruin problem for the diffusion subject to the assumption that S(x). If this assumption fails. Thus. whenever u.11) then follows by checking that the derivative of the r. oo).T) = P(ST = u) + 2P (ST > u). Proof The argument leading to ( 1. is zero for all u > 0 but that nevertheless Rt ^4 0 (the problem leads into the complicated area of boundary classification of diffusions.12) P(ST = v) = 0 otherwise.g. Breiman [78] or Karlin & Taylor [222] p.T+2. see e. 226).. Here {2T( (v}TT)/2) v=T.9) goes through unchanged.10). We assume that u(x) and a2 (x) are continuous with a2 (x) > 0 for x > 0. Vi(u.9 For Bernoulli random walk with 9 = 1/2. and (1.12) is the same as ( 1. u Small modifications also apply to Bernoulli random walks: Proposition 1. The expression for F ( ST = v) is just a standard formula for the u binomial distribution. The same argument as used for Corollary 1. close to x {Rt} behaves as Brownian motion with drift µ = u(x) and variance a2 = a2(x). Theorem 1.T)dx. is finite for all x > 0.. e.T2. 0 0 (1.302 CHAPTER XI.9).T (1. such that the drift µ(x) and the variance a2(x) are continuous functions of x and that a2(x) > 0 .s.
b('u) = Eu &0. TWO BARRIERS 303 for x > 0. .13) is finite for all x > 0. RANDOM WALK.ba.b = a+/3S. b = 0. Using s'/ s = 2p/a2. Wa.10.b(u) + L. the function S(x) is . BROWNIAN MOTION.b(u) be the probability that {Rt} hits b before a starting from u.ba.S(u) (1. Assume further that S (x) as defined in (1.16) yields 4b (u) = 1 . 15) i.16). 0 Proof of Theorem 1.S(u)/S(a).11 Let 0 < b < u < a and let t&0. 0 in (1.b(u) = S(a) .10. [117]. where Lq(u) = 0'22u) q "(u) + p(u)q(u) is the differential operator associated with the diffusion.17) Hence L.b('u) = Eu . The obvious boundary conditions '0a. Lemma 1. A good introduction to diffusions is in Karlin & Taylor [222]. Further references on twobarrier ruin problems include Dickson & Gray [116]. E„ q(Rdt) = q(u)+Lq(u)dt. then 0 < 2l. In view of (1. Then YIa.0(u) = 1 for all u > 0.b(u)dt. elementary calculus shows that we can rewrite L as Lq(u) d 1a2 (u)s(u)d [ s (u) ? ] .b = 0 implies that VQ b/s is constant.(u) < 1 for all u > 0 and ^ S^ Conversely. Letting b J. see in particular pp. 191195 for material related to Theorem 1.1. we can ignore the possibility of ruin or hitting the upper barrier a before dt. see Asmussen & Perry [42].e. S(oo) < oo separately u completes the proof.b(a) = 0 then yield the result. so that Y)n.b(b) = 1.14) fails. O. i. (1 . (1.S(b) Proof Recall that under mild conditions on q.e LVa.14) S(oo) < 00. then. A classical reference for further aspects of Bernoulli random walks is Feller [142]. if (1. For generalizations of Proposition 1. If (1.16) S(a) . If b < u < a.6 to Markovmodulated models .b (Rdt) = Oa. 1'. Notes and references All material of the present section is standard. and we get Wo. Letting a T oo and considering the cases S(oo) = oo.b(Rdt).
)AT .1. The emphasis is often on stationary distributions .5. They all use the fact that ( tx(a) l ( eaRt = eau + aSttx(a) < e7yu. where C_ = B(x) _ B(x) sup 2no fy° e7(Y )B(dy)' f2e7(Y2)B(dy)' C+ i/i(u.9 ) and optional stopping applied to the stopping time r(u) A T.4.(.4. 2 Further applications of martingales Consider the compound Poisson model with adjustment coefficient ry and the following versions of Lundberg 's inequality (see Theorems 111.3) < e 7yu.5): _ z/'(u) < e 7u. with the drift and the variance depending on an underlying Markov process . Markovmodulated Brownian models .(7) . 1 y < k (y). (2. yielding eau = Ee. Another basic quantity is the speed measure M . correponding to piecewise linear paths or .(a) (2.1 ) was given already in II. and here are alternative martingale proofs of the rest .5) A martingale proof of (2. Lo is a martingale (cf. (2.3.o•K(a) = Ee . yu) where W (ay) = y..6) . but by duality.6.17).aRo . much of the literature dels with the pure drift case. IV.2. Remark 11. MISCELLANEOUS TOPICS referred to as the natural scale in the general theory of diffusions (in case of integrability problems at 0. defined by the density 1/va(u)s(u) showing up in (1. 7y = ay . (2.(T(u)AT) r. one works instead with a lower limit 5 > 0 of integration in (1.1) (2. Lo I.ytc (ay).4) I.aR. 111 .2) C_e7u < t(u) < C+e _7u. See Asmussen [20] and Rogers [305] for some recent treatments and references to the vast literature. which is motivated from the study of modern ATM (asynchronous transfer mode ) technology in telecommunications. y > .13)). is currently an extremely active area of research. yu) '+/1(u) .t&(u. variance 0. equivalently. information on ruin probabilities can be obtained .304 CHAPTER XI. (2.
B(r))/B(r). RT(u)_) given r(u) < oo. (2.2. eyuk (ay) = e7yu e > eyu"(ay ) ij(u. dr) 1 = 1 I0 /o C+ C+ From this the upper inequality follows. when Rt_ = r. (B(y) . A claim leading to ruin at time t has c.)r(u)r.6).T(u)K(ay) I yu < r(u) < T] F(yu < r(u) < T) > e. Proof of ( 2.6) with = 'y that eyu .( u ) I T(U) < 00] .1. we have ic(ay ) < 0 and use the lower bound E [e7Rr („).f.1.. u Proof of (2. dr JO Zoo ) f e7'B(r + dy) B(r) Jo ^00 ^00 H(dt.yu) Y Similarly for (2. .7R.4). and the proof of the lower inequality is similar. Equivalently.(u.E [e.4): We take a = ay in (2. Rt has distribution B(r + dy)/B(r). dr) denote the conditional distribution of (T(u).6) below by 1 E Le7Rr(.d.3).yuk (ay)(u&(u.(ay)I T(u) < yu] P(r(u) < yu) (using RT(u) < 0). For (2. FURTHER APPLICATIONS OF MARTINGALES 305 (we cannot use the stopping time r(u) directly because P(r(u) = oo) > 0 and also because the conditions of the optional stopping time theorem present a problem). Let H(dt. yu))• Letting T + oo yield e_ayu > eyur4ay)(0(u)  Notes and references See II.3). dr) e 7( yr)B(dy) B(r) f oo o 0 r > H(dt.T)  V. y > r. so that i/1(uL yu) < eayu .1 . it follows easily from (2. we have tc(ay) > 0 and we can bound (2.yu))• b(u. Hence E [e7Rr (u) Jr(u) < ool ^00 H( dt.2): As noted in Proposition II.
However .means (as at other places in the book) that the ratio is one in the limit (here n * oo).(B) = log EeOX 1 is defined for sufficiently many 0.1 We will go into some more detail concerning (3. e. (3. v2 later. however . the parameter will be u rather than n). 1) but only the dominant exponential term . . ri. cle .?n typically only give the dominant term in an asymptotic expression ..2). logarithmic asymptotics . and gave sharp asymptotics for probabilities of the form P (S. which in the setting of (3. og For sequences fn. + X. Cramer considered a random walk Sn = X1 + . its generality.nn or C2e.g. The limit result (3.^ e nn 1 > x n 0o 2xn (3..306 CHAPTER XI.1) is an example of sharp asymptotics : . not quite so much in insurance risk.1) where we return to the values of 0. .gn if nioo lim 109 fn = 1 log gn (later in this section. (3.1) amounts to the weaker statement lim 1 log P I Sn > x I = 17. then P C S. we will write fn 1. Example 3. Thus . The advantage of the large deviations approach is.1). such that the cumulant generating function r.. gn with fn + 0 .2) can be rewritten as F (Sn/n > x) 1g a'fin. The last decades have seen a boom in the area and a considerable body of applications in queueing theory. in being capable of treating many models beyond simple random walks which are not easily treated by other models ./n E I) for intervals I C R. The classical result in the area is Cramer's theorem. gn 4 0. Thus.. logarithmic asymptotics is usually much easier to derive than sharp asymptotics but also less informative . nroo n n /// Note in particular that (3. large deviations results been. 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. if x > EX1. Accordingly.3na with a < 1. and that a considerable body of theory has been developed. large deviations results have usually a weaker form.the correct sharp asymptotics might as well have +.1) does not capture the \ in (3. For example.
the mean rc'(0) of the distribution of X1 exponentially tilted with 0. the sup in the definition of rc* can be evaluated by differentiation: rc*(x) = Ox .4) immediately yields (3. replacing Sn in the exponent and ignoring the indicator yields the Chernoff bound P Sn > x 1 < e°n (3.425. which is a saddlepoint equation . More precisely. we have P(nx < Sn < nx + 1. Since P nn > x) = E {e_8 ' ( 9). of P(X1 E dx) = E[e9X1K.rc(0) where 0 = 0(x) is the solution of x = rc'(0). (3.3) is put equal to x. P with mean nx and variance no.1). V > 0 e.960/) * 0. nx < Sn < nx + 1.r.sseo f which in conjunction with (3. the LegendreFenchel transform or just the Legendre transform or the large deviations rate function). if we replace Sn by nx + o / V where V is N(0.t.(e)i XI E dx]. rc*(x) = sup(Ox . and hence for large n P(Sn/n > x) > E [e. S rtn > x 1.3. 2 where o2 = o2(x) = rc"(0).(0)) e 307 (other names are the entropy. i. exponential change of measure is a key tool in large deviations methods.r. Define .9S„+n' ( 9).4) n Next. LARGE DEVIATIONS Define rc* as the convex conjugate of rc.tin f o') o e9o^y 1 1 ey2/2 dy 21r = etin 1 Bo 27rn . In fact.e. we get P(Sn/n > x) E [e9nx +nK(9)9" '.96o /] > 0.4 enn +1.2).q = rc* (x). Most often. since Sn is asymptotically normal w.
's. (iii) #c (8) = limn..o log Ee9Sn /n. e > 0 such that (i) Kn (0) = log Ee°Sn is welldefined and finite for 'y .. (ii) lim supn. to be made rigorous. and the WentzellFreidlin theory of slow Markov walks. Mogulskii's theorem which gives path asymptotics.3 For each i > 0. we shall concentrate on a result which give asymptotics under conditions similar to the GartnerEllis theorem: Theorem 3 .'(u) )Ng a"u./^ >7 < zn n for n n0.p > 7 < zn. r(u) = inf {n : Sn > u} and o(u) = P('r(u) < oo). In the application of large deviations to ruin probabilities. we introduce a change of measure for X1.. (iv) tc(ry) = 0 and r.s. .. . Xn given by Fn(dxl. . integrates to 1 by the definition of Icn).. . . however.2 (GLYNN & WHITT [163]) Let X1.. which is a version of Cramer's theorem where independence is weakened to the existence of c(O) = limn. MISCELLANEOUS TOPICS which is the same as (3. Assume that there exists 'y.3.. We shall need: Lemma 3 ..308 CHAPTER XI.v. Pn Sn1 . 1]. be a sequence of r. Sanov's theorem which give rare events asymptotics for empirical distributions.dxn) where Fn is the distribution of (X1i . Xn) and sn = x1 + • • • + xn (note that the r.e < 8 < y + e... Further main results in large deviations theory are the GartnerEllis theorem.. The substitution by V needs. Then i/.e < 8 < y + e.. X2. 260 for details.h.. and write Sn = X1 + • • • + Xn. is differentiable at ry with 0 < K'(y) < 00. commonly denoted as is the saddlepoint approximation. there exists z E (0. n Icn(0) exists and is finite for ry ... see Jensen u [215] or [APQ] p. which is of similar spirit as the dicussion in VII.. For the proof.dxn) = 05nKn(7)Fn(dx1. 1) and no such that Sn .. We further write µ = tc'(ry). that is. Ee9X n < oo for e < 0 < e.1). asymptotics for probabilities of the form P ({S[nti/n}o<t<l E r) for a suitable set r of functions on [0.
Then V.> .ne(p+ 17). is of order . We first show that lim inf„_. The rest of the argument is as before.77) follows by symmetry (note that the argument did not use µ > 0).n e(µ +o)w"(7) [Eep(B +7)Sn]1 /p [EegoX.n m µ 1 + rl . Let r7 > 0 be given and let m = m(77) = [u(1 + 77)/µ] + 1..y) . > 1 +17] m(7). u Proof of Theorem 3.m(7).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 +.077 n^oo n and by Taylor expansion and (iv ). the r .n > u ) = [ Em [em Em 1e. The corresponding claim for Pn(Sn/n < µ . This proves the existence of z < 1 and no such that Pn (Sn/n > µ.. mµ Sm > u] km e7Sm+n.2. 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. it is easy to see that the r.2. can be chosen strictly negative by taking p close enough to 1 and 0 close enough to 0.91) + o(O ) as 0 J.s. h.YS.. log zl'(u)/u > 'y. S.71 < e and jq9j < e.h.µ?7 . This establishes the first claim of the lemma . LARGE DEVIATIONS Proof Let 0 < 9 < e where a is as in Theorem 3. can be chosen strictly negative by taking 9 small enough. 0.s.. ( U) P(S.+r7) < zn for n > no.Bµ .3.ne(µ limsup 1 log Pn (Sn/n > µ + 17) < ic(9 + ry) . P n(Sn/n > {c+77) < e no(µ 309 +n)Enees n +n)elcn(B +7).+r. 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.]1/q = e.r (7) n = e.W. in particular the r.s. Since I EeqOX „ ] 1/q is bounded for large n by (ii). S. for Sn.. h. Clearly.
6) for some z < 1 and all n > n(E).+wn(7).310 ]Em I e.. we write P(T(u) = n) = Il + I2 + I3 + I4 'i/I(u) _ E00 n=1 where n(b) Lu(10/µJ Ii = 1: F(T(u) = n). For lim supu. I > IL exp `S. MISCELLANEOUS TOPICS (7).I < µl1 1+77 I M 1_ 1+277 S. logO(u)/u > ry. 14 = = E Lu(16)/aJ+1 Lu(1+6)/µJ+l = n) and n(S) is chosen such that icn('y )/n < 6 A (.n Yµ 1 + m + r ('Y) } U n \ 77 m µ µ7 1 < 1+ 77 ) Here E.3.0 log i'(u )/u < 'y. P(T(u) = n) < P(Sn > u) = En [e7S.. (iv) and Lemma 3.(Y).. this is possible by (iii)..7) so that n(b) I1 < e'Yu E en. n=1 . Pn \ > la+ 8 I < zn (3. we get lum inf z/i(u) 1 +12r7 >_ ry + 77 Letting r7 J. Sn > u] < eYu+Kn(7)pn(Sn > u) (3.log z) /2 and Sn Fn\ n >lb+S) <Zn..n(ry)/u 4 0andm/u* (1 + r7)/µ. 0 yields liminfu __.(•) goes to 1 by Lemma 3.. I2 = F(T(u) = n).. and since Ic. Obviously. 3. n=1 n=n(b)+1 00 Lu(1 +6) /µJ 13 F( T (u) = P(T(u) n).YS +^c CHAPTER XI.
S.' 1 + b) n e7u x 1 /2 1 n x n / 2x (3. u . eryu en logz/2p n nt n. C 26u `p / +1 I e6u(1+6)/µ (3.3.10) 00 I4 < E F(Sn_1 < u. > u) Lu(1+6) /µJ +l 00 )^n 'YSn+kn (7) .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). 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 .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. µ n=n(6)+1 \ 1u(16)/µ1 00 1 zn < e7u E Z n/2 < e(U xn/2 E n=n(6)+1 n=0 eYu = 1 . Sn1 C U. we get lim sup log u/00 O (U) < y + b(1 + b) U Letbl0.
4 there is an aj > 0 and a cj < oo such that Ij < c3e.312 CHAPTER XI. ryue«iu . 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.Q is so small that w = 1 .7' a"ju. I2.xl/2 to give the desired conclusion. we have rcn (a + 7) < 2n^c(7 + a) < 4narc' (7).4. For 14. cf. Then for n large.3ui where . Letting c11 = maxn<n. say n n1. we need to redefine n(b) as L. we replace the bound P(Sn > u ) < 1 used in (3..u(1+b)/rc'(7)). the typical time is u/rc'(7) just as for the compound Poisson model.8) by P(S.11 ) can be sharpened to x 4 [u(1+6)/µJ /2 1 . MISCELLANEOUS TOPICS The following corollary shows that given that ruin occurs.2.('+'Y). > u) < e"' E eIsn = ectueKn (a+'Y)Kn(7) where 0 < a < e and a is so small that r.z 1/z For I1. it holds for each b > 0 that 0(u) 1' g F(T(u) E (u(1 .(u) = I1+I2+I3+I4'^ ery( u). e'. (7 + a) < 2arc'(7). the last steps of (3.4 Under the assumptions of Theorem 3.4/3rc'(y) > 0.9) can then be sharpened to x LQuJ /2 I2 < e7u 1 . 13 = P(T (u) E (u(1 b)l^ (7). this is straightforward since the last inequality in (3.b)/i(7).. u(1 + b)/i(7)) Proof Since V. it suffices to show that for j = 1. IV. Corollary 3. For I. u . For 12. 2.
'(y) > 0.. for the ruin probability z/'h(u) of any discrete skeleton {Skh}k=0.. Let {Nt}t>0 be a possibly inhomogeneous Poisson process with arrival rate .e. whether P ( sup St > u ltg a ^" 0<t<oo // (3..2 shows that the discrete time structure is used in an essential way. criteria are given in Duffield & O'Connell [124]. and in fact. Theorem 3. V(s) with m. It is then wellknown and easy to prove that Sn has a normal distribution with mean np and a variance wn satisfying i lim wn = wz = Var(X1 ) + 2 E Cov(Xl. 2 is in force with y = 2p/wz.f. (iv) becomes existence of a limit tc(9) of tct(9) _ log Ee8S° It and a y > 0 with a(y) = 0.g.LARGE DEVIATIONS 313 Example 3 . r.. The reader not satisfied by this gap in the argument can easily construct a discrete time version of the models! The following formula (3.1. 11 Inspection of the proof of Theorem 3.3. Hence z z\ 2 z nrn(9) _ n Cn0p+BZn/ * .14) is needed in both examples .1. i. we shall give two continuous time examples and tacitly assume that this can be done..5 Assume the Xn form a stationary Gaussian sequence with mean p < 0. Obviously many of the most interesting examples have a continuous time scale.13) One would expect this to hold in considerable generality.3(s) at time s. t] is Rt = E V (Un) n: o„ <t . The problem is whether this is also the correct logarithmic asymptotics for the (larger) ruin probability O(u) of the whole process.2 then immediately yields the estimate log F( sup Skh > u) a7u (3.. and we conclude that Theorem 3 . the key condition similar to (iii). If {St}t> 0 is the claims surplus process. To verify these in concrete examples may well present considerable difficulties. but nevertheless.. Assuming that the further regularity conditions can be verified. Thus the total reward in the interval [0.v. 09(9).(O) = 9µ+02 for all 9 E R. Xk+l) k=1 00 naoo n provided the sum converges absolutely.12) k=0. An event occuring at time s is rewarded by a r..
Thus. we conclude that Cu) log e7 u (cf.'`1 U.15) . nondecreasing and with finite limits Un as s T oo ( thus. 0 Example 3 . An apparent solution to this problem is to calculate the premium rate p = p(t) at time t based upon claims statistics .9t = /3 J t (Ee8U° i8l . this is not realistic . Un(s). it contributes to St by the amount Un(t .1) . the CramerLundberg model implicitly assumes that the Poisson intensity /3 and the claim size distribution B (or at least its mean µB) are known.0 and assume there are y.noise model is the same as the one for the Cramer Lundberg model where a claim is immediately settled by the amount Un. leading to St = At(1+77) Joo t S8 ds. (3. Thus. n: o.t. e. At = . Of course . Thus by (3.Lundberg model has the larger ruin probability. We further assume that the processes {U1(s)}8>0 are i. we have S.1) ds . Of course.14). If the nth claim arrives at time Qn = s. More precisely.1) ds rt (3. 7 Given the safety loading 77. Kt (0) t (Ee9U"it8i J0 .Q„) . we have rct (9)/t 4 ic (9). Un represents the total payment for the nth claim). assuming a continuous premium inflow at unit rate. the best estimator of /3µB based upon Ft. where Ft = a(A8 : 0 < s < t). Then logEeOR° = J0 /3(s)(^8(9) . Example 3. the Cramer. Most obviously.1) ds . a differential equation in t). but that a claim is not settled immediately. . (9) < oo for 9 < 'y + C.314 where the an CHAPTER XI.2 are trivial to verify.. e > 0 such that ic('y) = 0 and that r. 0 and since EeOUn(8) + Ee°U^ as s * oo. one would take p(t) = (1 + rt)At/ t.g.. <t which is a shotnoise process. We let ic (9) = 3(EeWU° . then the payments from the company in [on. i. MISCELLANEOUS TOPICS are the event times. = U„ ( t . Since the remaining conditions of Theorem 3.. O'n +S] is a r .s). is At .d. the above discussion of discrete skeletons).14) (to see this . It is interesting and intuitively reasonable to note that the adjustment coefficient ry for the shot .It.v. if the nth claim arrives at time a.9t.6 We assume that claims arrive according to a homogeneous Poisson process with intensity 0 .. derive .
1) or . standard exponential .18) Thus (iii) of Theorem 3. we conclude that t.1) .19) with equality if and only if U is degenerate./3. Thus. typically the adaptive premium rule leads to a ruin probability which is asymptotically smaller than for the CramerLundberg model . 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 . i.(1 +i) f > i= 1 s ds = E Ui 1 . (3.i.d.e.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. Ui Nt / t 01i 315 St = Ui . the solution of /3(Eelu . the Vi = . Indeed.(1 + r7) log t (3. equivalently.16) i=1 o i=1 Let ict (a) = log Eeast . we have Nt t N.log Oi are i.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 .14) that rt _ 13 Jo _ (a [1_( i+77)log]) ds_flt = t (a) (3.b(u) IN a'Yu (cf.i. and since the remaining conditions are trivial to verify. uniform (0.d. one has y > y' (3.17) K(a) f o 1 O (a[I + (1 + 77) log u]) du )3.20) (3.3. (3. LARGE DEVIATIONS With the Qi the arrival times. rewrite first rc as te(a) _ /3E 1 1 +(1+77)aUJ eau 1 . It then follows from (3. To see this .(1 + 17)0µB = 0.
this in turn yields y > y*. and k(x) < 0. see also Nyrhinen [275] for Theorem 3. [245]. Further applications of large deviations idea in risk theory occur in Djehiche [122]. are i. see Nyrhinen [275] and Asmussen [25].7. In addition to Glynn & Whitt [163]. = P(N = n) = e(3an However. with common distribution B and independent of Nt. we are interested in estimating P(A > x) for large x. 0 < x < x0.20) is due to Tatyana Turova. [257] and Nyrhinen [275]. and since tc(s).d. . 4 The distribution of the aggregate claims We study the distribution of the aggregate claims A = ^N' U. MartinL6f [256]. assuming that the U. Further. say one year.xo. Therefore e7'U _ k(U) E [1+(1+77)y*U] . a* (s) are convex with tc'(0) < 0 .2 expressing the finite horizon ruin probabilities in terms of the distribution of A. 11 Notes and references Some standard textbooks on large deviations are Bucklew [81]. This implies n(y*) < 0.i. much of the analysis carries over to more general cases. the function k(x) = e7*x .1 . k'(0) < 0. so there exists a unique zero xo = xo(r7) > 0 such that k(x) > 0. the study is motivated from the formulas in IV. y = y* can only occur if U . at time t. the proof of (3. In particular.2.19). x > x0. Dembo & Zeitouni [105] and Shwartz & Weiss [339]. For notational simplicity. we then take t = 1 so that p.(1 + ri)y*x is convex with k(oo) = 00. MISCELLANEOUS TOPICS Next. Further. 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. rc*' (0 ) < 0.1 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. k(0) = 0. The main example is Nt being Poisson with rate fit. For Example 3. though we do not always spell this out.. using that Ek(U) = 0 because of (3. Lehtonen & Nyrhinen [244].316 CHAPTER XI.
818' where s' = sup{s : B[s] < oo}. B"' [s] lim (B".x)//3B"[9] is standard normal. i. we define the saddlepoint 9 = 9(x) by EBA = x. The exponential family generated by A is given by Pe(A E dx) = E [eeA K(9).9(Ax). 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 ."(0) = .e. K'(0) _ ic'(9) = x. no(a) = logE9e'A = rc(a + 9) . (4. The analysis largely follows Example 3.1). e9x+K(°) P(A > x) B 2ir /3 B" [9] Proof Since EBA = x.2) implies that the limiting Pedistribution of (A .1. In particular.4.[s])3/2 = 0.1) where )30 = . B"[s] = oo. A E dx] .3B"[9].1 Assume that lim8T8. Then Ee"A = e'(") where x(a) _ 0(B[a] .3e(bo[a] .3B[9] and Be is the distribution given by eox B9(dx) = B [9] B(dx). Hence P(A > x) = E e [e9A+ ic(9). For a given x. Vare(A) = s.ic(9) = . THE DISTRIBUTION OF THE AGGREGATE CLAIMS 317 4a The saddlepoint approximation We impose the Poisson assumption (4. only with 0 replaced by a9 and B by B9. A > x)] = eex+K( e)E9 [e . Then as x * oo.1). Proposition 4. This shows that the Pedistribution of A has a similar compound Poisson form as the Fdistribution.
Remark 4 .ycix °ie6x B. large x. it holds that EA = . b is logconcave.v.x') where x' = sup {x : b(x) > 0}. the distribution of A is approximately normal .3) The result to be surveyed below improve upon this and related approximations by taking into account second order terms from the Edgeworth expansion. B covers distributions with finite support or with a density not too far from ax° with a > 1. where q(x) is bounded away from 0 and oo and h (x) is convex on an interval of the form [xo. 4b The NP approximation In many cases .(D X . Var(A) _ ^3p.2) is often referred to as the Esscher approximation. i. b(x) = q(x)eh(z).4) .318 CHAPTER XI. The present proof is somewhat heuristical in the CLT steps. Thus . more generally.3) and related results u for the case of main interest . [138].e. just the same dominated convergence argument as in the proof of Theorem 2. Furthermore 00 b(x)Sdx < oo for some ( E (1.EN B(x). For a rigorous proof. The (first order) Edgeworth expansion states that if the characteristic function g(u) = Ee"`}' of a r. In fact.1 goes all the way back to Esscher [141]. leading to P(A > x) :. and (4. In particular. For details. then P(A > x) . b is gammalike. Notes and references Proposition 4.(3µB)/(0µB^))1/2 has a limiting standard normal distribution as Q ^ oo. For example. MISCELLANEOUS TOPICS It should be noted that the heavytailed asymptotics is much more straightforward. bounded with b(x) .1 yields: Proposition 4.Q{AB (4. either of the following is sufficient: A. For example. or. Y satisfies 9(u) ti eu2/2(1 + ibu3) (4. it is quite questionable to use (4. 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 . 1 . see Embrechts et al.1).l3pB. Jensen [215] and references therein. some regularity of the density b(x) of B is required. A covers the exponential distribution and phasetype distributions. 2).2 If B is subexponential and EzN < oo for some z > 1. under the Poisson assumption (4.2i and that (A .
one needs to show that 163..f. Rather than with the tail probabilities F(A > x). If the distribution of Y is close to N(0.5) follows by integration. . u5. Thus if EY = 0.. (4. as u2 u3 u4 9(u) = Ee'uY = exp {iuci . A particular case is a. so that 1(u) 3 exp { . K3 = E(Y . If this holds . Remark 4.5) may be negative and is not necessarily an increasing function of y for jyj large.. in particular. however. THE DISTRIBUTION OF THE AGGREGATE CLAIMS where b is a small parameter. and from this (4. are small. resp. the NP (normal power) approximation deals with the quantile al_E..5 (y3 . . K4 . K2 = Var (Y).3!). Let Y = (A . the density of Y is 1 °° _ eiuy f(u) du 2x _.. zl_e be the 1 .2 ^ \1 .i 3 K3 } Pt^ exp .99... s. one expects the u3 term to dominate the terms of order u4.l = EY.c2i. and so as a first approximation we obtain a1_E = EA + yle Var(A) . then P(Y < y) 4(y) .EA)/ Var(A) and let yl_E.2 2 . the standard normal distribution. defined as the the solution of P(A < yle) = 1 . . where Kl .e..5).3& (y). which is often denoted VaR (the Value at Risk)..EY)3.4.equantile in the distribution of Y.s. of (4. (4.y2)^P(y)• 319 Note as a further warning that the r. In concrete examples .: EA + zl_E Var(A) .2K3 + 4i 64 + .6) .6(1 . the CLT for Y = Y6 is usually derived via expanding the ch.h.2X2 . ylE should be close to zl_E (cf. Var(Y) = 1 as above .1). f °o 9(y) = 1 e'uye u2/2(1 + iSu3) du 27r _ cc(y) .i 6 r 1 3 so that we should take b = ic3/6 in (4. Heuristically. are the cumulants .5) is obtained by noting that by Fourier inversion.
3ni /3 .7) as 1 (3) a1E = Qµa +z1 . and assume that there exist n ) Pn_i . For example. 4c Panjer 's recursion Consider A = constants a. this yields the NP approximation 6(Z1 _E .zlE)W(zlE) 1 .E(/3PB^1 )1^2 + s(z1E .E + (yl. MISCELLANEOUS TOPICS A correction term may be computed from (4.y2)cp( y) term.E )Azl E) 4(z1E) + ( ylE .5(1 .zl E)V(zl_E) .EA)3 a1_E = EA + z1_E(Var (A))1/2 + 1 Var(A) Under the Poisson assumption (4.1) E (A . 21 .1)^ 2) µ'E Notes and references We have followed largely Sundt [354].zlE )w(zl _E) = which combined with S = EY3/6 leads to q^ 1 Y1 . Another main reference is Daykin et at. .k = /3µB^1 / (.E = z1E + S(zi_E .EA ) / Var(A).1)! n ^eQ ..320 CHAPTER XI.E .1)EY3. Note. k3 is small for large /3 but dominates 1c4.S(1 . [101]. however. We can rewrite (4.yi. b = /3 for the Poisson distribution with rate /3 since Pn = Pn1 n! n (n .. n = 1..6 (1 .(y) terms dominate the S(1 ..E)A1 l E) 1 E 4)(yl E) ^' .5) by noting that the 4.zl E)^o(zl E) .1). this holds with a = 0. as required . Using Y = (A . This leads to t( yl E) .zi. let pn Pn = (a+ = P(N = n). K5 . b such that EN 1 U%... that [101] distinguishes between the NP and Edgeworth approximations. In particular .S(1 ..6pBki) d/2. the kth cumulant of A is /3PBk' and so s.
.4. Then fo = >20 9onpn and fi = 1 E In particular.14) is therefore a + b/n.13) Namely. j = 1. (4. The expression for fo is obvious.4 Assume that B is concentrated on {0..k . 2. if go = 0. u Proof of Proposition 4. .12) we get for j > 0 that fj n a b + n p nlgj *n 00 U I n 1 *n = E a+bUi=j pn19j n=1 j i=1 CC) n Ui EE n=1 Ia +b Ul i=1 =j pn_1 . which would consist in noting that (in the case go = 0) fj = pn9jn n=1 (4... (4. (4. .4 is that the algorithm is much faster than the naive method..11) Remark 4. n = k=n1 9k(n1 )9j k • (4. the complexity (number of arithmetic operations required) is O(j3) for (4. 2. . By symmetry.. . 2.12) where g*n is the nth convolution power of g. 1. . .. E[a +bU=I >Ui =j l i=1 J (4.4.5 The crux of Proposition 4. Since the sum over i is na + b. THE DISTRIBUTION OF THE AGGREGATE CLAIMS 321 Proposition 4.10) f o = po. and calculating the gj*n recursively by 9*1 = 9j. . Hence by (4. then j (a + b!) 1ag k_1 3 gkfj.12)... the value of (4...13) but only O(j2) for Proposition 4.1. n.9).4. fj = E (a+ b k =1 )9kfi_k . (4. j = 0..14) is independent of i = 1. fj = P(A = j).} and write gj = 2 . . j1 g. . j = 1.
322
00 J
CHAPTER XI. MISCELLANEOUS TOPICS
EE (a + bk I gkg3 _ k lieni n=ik=0 (a+bk l gkE g j'`kpn = E (a+b!)9kfi_k n=0 k=0 k=0 ^I 1 E(a+b. agofj+ k Jgkfjk, k=i /
and and (4.9) follows . (4.11) is a trivial special case.
u
If the distribution B of the Ui is nonlattice , it is natural to use a discrete approximation . To this end, let U(;+, U(h) be U; rounded upwards, resp. downwards , to the nearest multiple of h and let A}h) = EN U. An obvious modification of Proposition 4.4 applies to evaluate the distribution F(h) of A(h) letting f( ) = P(A() = jh) and
g(h) gkh+
= P (U(h2 = kh) = B((k + 1)h)  B(kh ), k = 0, 1, 2, ... , = P (U4;+ = kh) = B(kh)  B (( k  1)h) = gk  l,, k = 1, 2, ... .
Then the error on the tail probabilities (which can be taken arbitrarily small by choosing h small enough ) can be evaluated by
00 00
< P(A > x ) f (h) j=Lx/hl j=Lx/hl
Further examples ( and in fact the only ones , cf. Sundt & Jewell [355]) where (4.9) holds are the binomial distribution and the negative binomial (in particular, geometric ) distribution . The geometric case is of particular importance because of the following result which immediately follows from by combining Proposition 4.4 and the PollaczeckKhinchine representation: Corollary 4.6 Consider a compound Poisson risk process with Poisson rate 0 and claim size distribution B. Then for any h > 0, the ruin probability zb(u) satisfies 00 00
f^,h) Cu) < E ff,+, j=Lu/hJ j=Lu/hJ (4.15)
f! h)
5. PRINCIPLES FOR PREMIUM CALCULATION
where f^ +, f^ h) are given by the recursions
(h) 3 (h) (h)
323
fj,+ = P 9k fjk,+ ' I = 17 2, .. .
k=1 3 (h)
(h)
=
P
(h)
f9,  (h) gk,fAk, e 1  ago, k=1
j = 1+2,
starting from fo + = 1  p, f(h) = (1  p)/(1  pgoh) and using 07
g(kh) 1 (k+1)h
=
Bo((k + 1 ) h)  Bo(kh ) =  f
AB
kh
B(x) dx, k = 0, 1, 2, ... , k = 1,2 .....
gkh+
Bo(kh )  Bo((k  1 ) h) = 9kh)1 ,
Notes and references The literature on recursive algorithms related to Panjer's recursion is extensive, see e.g. Dickson [115] and references therein.
5 Principles for premium calculation
The standard setting for discussing premium calculation in the actuarial literature does not involve stochastic processes, but only a single risk X > 0. By this we mean that X is a r.v. representing the random payment to be made (possibly 0). A premium rule is then a [0, oo)valued function H of the distribution of X, often written H(X), such that H(X) is the premium to be paid, i.e. the amount for which the company is willing to insure the given risk. The standard premium rules discussed in the literature (not necessarily the same which are used in practice!) are the following: The net premium principle H(X) = EX (also called the equivalence principle). As follows from the fluctuation theory of r.v.'s with mean, this principle will lead to ruin if many independent risks are insured. This motivates the next principle, The expected value principle H(X) = (1 + 77)EX where 77 is a specified safety loading. For 77 = 0, we are back to the net premium principle. A criticism of the expected value principle is that it does not take into account the variability of X which leads to The variance principle H(X) = EX+77Var(X). A modification (motivated from EX and Var(X) not having the same dimension) is
324
CHAPTER XI. MISCELLANEOUS TOPICS
Var(X).
The standard deviation principle H(X) = EX +rl
The principle of zero utility. Here v(x) is a given utility function, assumed to be concave and increasing with (w.lo.g) v(O) = 0; v(x) represents the utility of a capital of size x . The zero utility principle then means v(0) = Ev (H(X)  X); (5.1)
a generalization v(u) = Ev (u + H(X)  X ) takes into account the initial reserve u of the company. By Jensen 's inequality, v(H(X)  EX) > Ev(H(X)  X) = 0 so that H(X) > EX. For v(x) = x, we have equality and are back to the net premium principle. There is also an approximate argument leading to the variance principle as follows. Assuming that the Taylor approximation
v(H(X)  X) ^ 0 +v'(0)(H (X)  X) + v 0 (H(X)  X)2 ,/2
is reasonable , taking expectations leads to the quadratic v"H(X )2 + H(X) (2v'  2v"EX) + v"EX2  2v'EX = 0 (with v', v" evaluated at 0) with solution
H(X)=EXv^±V( ^ )2Var(X).
Write
( vI ) 2 \
Var(X) v^  2v^Var(X)/ I  (
, Var(X) )2
If v"/v' is small, we can ignore the last term. Taking +f then yields H(X) ,:: EX 
2v'(0) VarX;
since v"(0) < 0 by concavity, this is approximately the variance principle. The most important special case of the principle of zero utility is The exponential principle which corresponds to v(x) = (1  e6x)/a for some a > 0. Here (5.1) is equivalent to 0 = 1  e0H(X)EeaX, and we get
H(X) = 1 log Ee 0X .
a
5. PRINCIPLES FOR PREMIUM CALCULATION
325
Since m.g.f.'s are logconcave, it follows that H,, (X) = H(X) is increasing as function of a. Further, limQyo Ha (X) = EX (the net premium princiHa (X) = b (the premium ple) and, provided b = ess supX < oo, lim,, H(X) = b is called the maximal loss principle but is clearly not principle very realistic). In view of this, a is called the risk aversion The percentile principle Here one chooses a (small ) number a, say 0.05 or 0.01, and determines H(X) by P(X < H(X)) = 1  a (assuming a continuous distribution for simplicity). Some standard criteria for evaluating the merits of premium rules are 1. 77 > 0, i .e. H(X) > EX. 2. H(X) < b when b (the ess sup above ) is finite 3. H(X + c) = H(X) + c for any constant c
4. H(X + Y) = H(X) + H(Y) when X, Y are independent
5. H(X) = H(H(XIY)). For example , if X = EN U= is a random sum with the U; independent of N, this yields
H
C^
U; I = H(H(U)N)
(where, of course, H(U) is a constant). Note that H(cX) = cH(X) is not on the list! Considering the examples above, the net premium principle and the exponential principle can be seen to the only ones satisfying all five properties. The expected value principle fails to satisy, e.g., 3), whereas (at least) 4) is violated for the variance principle, the standard deviation principle, and the zero utility principle (unless it is the exponential or net premium principle). For more detail, see e.g. Gerber [157] or Sundt [354]. Proposition 5.1 Consider the compound Poisson case and assume that the premium p is calculated using the exponential principle with time horizon h > 0. That is,
N,,
Ev I P  E U;
i =1
= 0 where
v(x) = 1(1  e°x
a
Then ry = a, i.e. the adjustment coefficient 'y coincides with the risk aversion a.
326
Proof The assumption means
CHAPTER XI. MISCELLANEOUS TOPICS
0 a (1  eareo (B[a11)
l
i.e. /3(B[a]  1)  ap = 0 which is the same as saying that a solves the Lundberg u equation. Notes and references The theory exposed is standard and can be found in many texts on insurance mathematics, e.g. Gerber [157], Heilman [191] and Sundt [354]. For an extensive treatment, see Goovaerts et al. [165].
6 Reinsurance
Reinsurance means that the company (the cedent) insures a part of the risk at another insurance company (the reinsurer). Again, we start by formulation the basic concepts within the framework of a single risk X _> 0. A reinsurance arrangement is then defined in terms of a function h(x) with the property h(x) < x. Here h(x) is the amount of the claim x to be paid by the reinsurer and x  h(x) by the the amount to be paid by the cedent. The function x  h(x) is referred to as the retention function. The most common examples are the following two: Proportional reinsurance h(x) = Ox for some 0 E (0, 1). Also called quota share reinsurance. Stoploss reinsurance h(x) = (x  b)+ for some b E (0, oo), referred to as the retention limit. Note that the retention function is x A b. Concerning terminology, note that in the actuarial literature the stoploss transform of F(x) = P(X < x) (or, equivalently, of X), is defined as the function
b * E(X  b)+ =
f
(s  b)F(dx) _ f
6 00
(x) dx.
An arrangement closely related to stoploss reinsurance is excessofloss reinsurance, see below.
Stoploss reinsurance and excessofloss reinsurance have a number of nice optimality properties. The first we prove is in terms of maximal utility: Proposition 6.1 Let X be a given risk, v a given concave nondecreasing utility function and h a given retention function. Let further b be determined by E(X b)+ = Eh(X). Then for any x,
Ev(x  {X  h(X)}) < Ev(x  X A b).
6. REINSURANCE
327
Remark 6 .2 Proposition 6.1 can be interpreted as follows. Assume that the cedent charges a premium P > EX for the risk X and is willing to pay P1 < P for reinsurance. If the reinsurer applies the expected value principle with safety loading q, this implies that the cedent is looking for retention functions with Eh(X) = P2 = P1/(1 + 77). The expected utility after settling the risk is thus
Ev(u + P  P1  {X  h(X)})
where u is the initial reserve . Letting x = u + P  P1, Proposition 6.1 shows that the stoploss rule h (X) = (X  b)+ with b chosen such that E(X  b)+ u = P2 maximizes the expected utility. For the proof of Proposition 6.1, we shall need the following lemma: Lemma 6 .3 (OHLIN'S LEMMA) Let X1, X2 be two risks with the same mean, such that Fj(x) < F2 (x), x < b, Fi(x) ? F2(x), x > b for some b where Fi(x) = P(Xi < x). Then Eg(X1) < g(X2) for any convex function g. Proof Let Yi=XiAb, Zi=Xivb.
Then
P(Yl < x) _ Fi(x) <_ F2 (x) = P(Y2 < x) x < b 1=P(Y2<x) x>b so that Y1 is larger than Y2 in the sense of stochastical ordering . Similarly, P(Zl < x) _ 0 = P(Z2 < x) x < b Fi(x) > F2(x) = P(Z2 < x) x > b
so that Z2 is larger than Zl in stochastical ordering. Since by convexity, v(x) = g(x)  g(b)  g'(b)(x  b) is nonincreasing on [0, b] and nondecreasing on [b, oo), it follows that Ev(Y1) < Ev(Y2), Ev(Zi) < Ev(Z2). Using v(Yi) + v(Zi) = v(Xi), it follows that
0 < Ev(X2)  Ev(Xi) = Eg(X2)  Eg(X1),
using EX1 = EX2 in the last step. u
Proof of Proposition 6.1. It is easily seen that the asssumptions of Ohlin' s lemma hold when X1 = X A b, X2 = X  h(X); in particular, the requirement EX1
328
CHAPTER XI. MISCELLANEOUS TOPICS
= EX2 is then equivalent to E(X  b)+ = Eh(X). Now just note that v is convex. u
We now turn to the case where the risk can be written as N
X = Ui
i=1
with the Ui independent; N may be random but should then be independent of the Ui. Typically, N could be the number of claims in a given period, say a year, and the Ui the corresponding claim sizes. A reinsurance arrangement of the form h(X) as above is called global; if instead h is applied to the individual claims so that the reinsurer pays the amount EN h(Ui), the arrangement is called local (more generally, one could consider EN hi(Ui) but we shall not discuss this). The following discussion will focus on maximizing the adjustment coefficient. For a global rule with retention function h* (x) and a given premium P* charged for X  h* (X), the cedents adjustment coefficient y* is determined by
1 = Eexp {ry*[X  h*(X)  P*]},
for a local rule corresponding to h(u) and premium P for X look instead for the ry solving
J _f
(6.2) N 1 h (Ui), we
[ X_P_^
1 = Eexp
[ Ei  h(Ui)] P [U
= Eexp{ry
h(Ui)]
l (6.3) This definition of the adjustment coefficients is motivated by considering ruin at a sequence of equally spaced time points, say consecutive years, such that N is the generic number of claims in a year and P, P* the total premiums charged in a year, and referring to the results of V.3a. The following result shows that if we compare only arrangements with P = P*, a global rule if preferable to a local one. Proposition 6.4 To any local rule with retention function h(u) and any
N
J}
P > E X  N h(Ui)
4 =1
(6.4)
there is a global rule with retention function h* (x) such that
N
Eh*(X) = Eh(U1)
i=1
and 'y* > ry where ry* is evaluated with P* = P in (6.3).
b)+ = Eh(U) (and the same P) satisfies 71 > ry. Proof As in the proof of Proposition 6.h(Ui)] . Remark 6.3).b)+ with b determined by E(U .P. ry* > 0 because of (6. this implies 7* > 7.4) and u g(x) = e7x in Ohlin's lemma.6) u where C[ry] = Ee'r(u4(u)). . The arrangement used in practice is. expectations like those in (6. that 01[ry] < 0[y] where 0[y] = Ee'r(U^') .h(U) (as in the proof of Proposition 6. REINSURANCE Proof Define N 329 h* (x) = E > h(Ui) X = x .d. (6. i. X2 = U .4)..h * (X) .5 Because of the independence assumptions . and so on.P } < 1 = Eexp E[Ui. however.4).h(Ui)P JJJ l:='l {ry ] or. y = Ei [Ui . Eexp 7 [E [Ui .4.i.P I = EC [7]N. we get N 1 = Eexp ry E[Ui ii . Local reinsurance with h(u) = (u . then (6. the excess ofloss rule hl (u) = (u .h(Ui)] . u But since ry > 0.5) reduce quite a lot.P]}. as often local as global.h(Ui)] . (6. N E X .b)+ is referred to as excessofloss reinsurance and plays a particular role: Proposition 6. we get EX = EN • EU. This follows by taking Xl = U A b.4).h( UU) = EN • E[U . Then for any local retention function u .6.d. Assuming for simplicity that the Ui are i. ' ii (6. Applying the inequality Ecp(Y ) > EW(E (YIX )) (with W convex ) to W(y ) = eryy.h(U)]. appealing to (6.5) holds trivially.h(u) and any P satisfying (6. it suffices to show that Eexp {ry ii 'UiAb.P > EexP{7[X .6 Assume the Ui are i.6).
Bowers et at.330 CHAPTER XI. The original reference for Ohlin's lemma is Ohlin [277]. MISCELLANEOUS TOPICS Notes and references The theory exposed is standard and can be found in. see also Sundt [354]. The present proof is from van Dawen [99]. [76].many texts on insurance mathematics. e.g. . Heilman [191] and Sundt [354]. See further Hesselager [194] and Dickson & Waters [120].
.e.1) (here U(t) = U([0. stating that U(t+a)U (t) ^ a. Then Blackwell 's renewal theorem holds. i. If F satisfies the stronger condition of being spreadout (F*' is nonsingular w . Y2. That is... Y2. Y. when t is large. = T„ . U(A) is the expected number of renewals in A C R in a zerodelayed renewal process. Y1. . denoted by F in the following and referred to as the interarrival distribution. + U2 where U1 is a finite measure and U2(dt) = u(t)dt where 331 ... Technically. then Stone 's decomposition holds : U = U.r. .. the distribution of Yo is called the delay distribution. The renewal theorem asserts that U(dt) is close to dt/µ.T„_1).. . t]) so that U(t + a) . are independent and Y1. t +a]).U(t) is the expected number of renewals in (t. The associated renewal measure U is defined by U = u F*" where F*" is the nth convolution power of F. of interarrival times and the time Yo = To of the first arrival (that is. some condition is needed: that F is nonlattice. t 00 (A. The number max k : Tk_j < t of renewals in [0. t] is denoted by Nt. The point process is called a renewal process if Yo. If Yo = 0. Lebesgue measure for some n > 1). not concentrated on {h. Lebesgue measure dt normalized by the mean to of F. 2h. note in particular that U({0}) = 1...} for any h > 0. . The mathematical representation is either the ordered set 0 < To < T1 < . of epochs or the set Y1.. the renewal process is called zerodelayed.t. all have the same distribution.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..
stating that U(t)/t > 1/p. u u PF 4 00.332 APPENDIX u(t) has limit 1/µ as t 4 oo.R. the statements being EN(t + a) .5) 2This condition can be weakened considerably . IV).i.9.2) Z(u) = J0 u z(x)U(dx).4) If F is spread.4) that z is Lebesgue integrable with limZ.2). (A. that z(u) has a limit z(oo) (say) as u 4 oo. i. z(x) = 0.2) has the unique solution Z = U * z. resp.3) Further. 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 .e.x)F(dx). and that F has a bounded density2.out. the asymptotic behavior of Z(u) is given by the key renewal theorem: Proposition A1. (A. Equivalently. but suffices for the present purposes . IV). then Z(u) i f0 z(x)dx . then it suffices for (A. µF (A.2 Assume that Z solves the renewal equation (A. oo). Under weak regularity conditions (see [APQJ Ch. in convolution notation Z = z + F * Z. wee shall need the following less standard parallel to the key renewal theorem: Proposition A1. (A. and F(dx) a known probability measure . z(u) a known function. Note in particular that F is spreadout if F has a density f. In 111. U Z(u . (A.EN(t) .a. A weaker (and much easier to prove) statement than Blackwell's renewal theorem is the elementary renewal theorem..1 if F is nonlattice and z (u) is directly Riemann integrable (d. Both result are valid for delayed renewal processes. Then Z(u) 4 z(oo).i". see [APQ] Ch.
The distribution F of Y1.5a. Y2. T1.k+t }t>o is independent of To. Assuming that y can be chosen such that f °° Ox F(dx) = 1.i.2) by e7x to obtain Z = z +P * Z where Z(x) = e'Y'Z(x). T1. cycles. • . . However.. Here the relevant F does not have mass one (F is defective). i.x)u(x) dx = z(u( 1 .} be a renewal process. 0 PF µF 11 In risk theory.. however. Tk and {Xt }o<t<Tk • For example. F(dx) = e7xF(dx).. A regenerative process converges in distribution under very mild conditions: . asymptotic properties can easily be obtained from the key renewal equation by an exponential transformation also when F(dx) does not integrate to one.(3. refer to the zerodelayed case. is called the cycle length distribution and as before. . where the Tn are the instants where a customer enters an empty system (then cycles = busy 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. we let µ denote its mean.. and its distribution does not depend on k. To this end.. .. Yk ). multiply (A.r. A stochastic process {Xt}t>0 with a general state space E is called regenerative w.e. Note. or many queueing processes.d.. the postTk process {XT. .3) satisfied by the ruin probability for the compound Poisson model. Z(u) U = 1 u 1 u f z(u . We let FO.. 1c Regenerative processes Let {T. This program has been carried out in III. that the existence of y may fail for heavytailed F. this expression is to be interpreted as a random element of the space of all Evalued sequences with finite lifelengths. {Tn} if for any k. equivalently.. The kth cycle is defined as {XTk+t}o<t<Yk . that F is a probability measure. this covers discrete Markov chains where we can take the Tn as the instants with Xt = i for some arbitrary but fixed state i. The property of independent cycles is equivalent to the postTk process {XTk+t}t>0 being independent of To.t. The simplest case is when {Xt} has i. the present more general definition is needed to deal with say Harris recurrent Markov chains. of Yo. However. Y1 . z(x) = e7xz(x). . . a basic reason that renewal theory is relevant is the renewal equation II. Tk (or.t))u(ut) dt 0 0 J f z(oo) • 1 dt = z(OO). Eo etc. results from the case fo F(dx) = 1 can then be used to study Z and thereby Z. Hence by dominated convergence.
0<t<Yi then Zt /t a$• EU1/µ. resp . e(t )) .t. An example is Zt = fo f (X8) ds where {Xt} is regenerative w.ZT Then: (a) If E sup I ZTo+t .ZTOI < 00.e.. 2. but in fact. in total variation. then (Zt . C). are i.. {Tn}. Otherwise . and q(t) = sup It .d..tEU1/µ)/f has a limiting normal distribution with mean 0 and variance Var(Ui) + (!)2Var (Yi)_ 2EU1 Cov(U1.3 Consider a regenerative process such that the cycle length distribution is nonlattice with p < oo..v.. fi (t) = inf {Tk .r. Then it (ii..i.'s by e. and we have: holds more generally that (rl(t). then Xt ..t. (A. assume that p < 00 and define Un = ZT}1 . Then {e(t)}.0 is called cumulative w.t : t < Tk}.d.ZT }0<t<Y„+. {i7(t)} are Markov with state spaces (0. This is the case considered in [APQ] V. C(t) and ij (t) both have a limiting stationary distribution F0 given by the density F (x)/p.t. then e (t) .. for n = 1. (b) If in addition Var(Ul ) < oo.. where the distribution of X.4 Let {Zt}t^. If p = oo. under the condition of Blackwell's renewal theorem.e.334 APPENDIX Proposition A1. oo). i.i. {Tn} if the processes {ZT +t . Then {Zt}t^..r. µ 0 If F is spreadout. {Tn}.3. Then Xt Di X.6) id Cumulative processes Let {Tn} be a renewal process with i. We denote the limiting r.0 be cumulative w.oo (i.+ X.. oo).r. just the same proof as there carries over to show: Proposition A1.Tk : t < Tk} as the age. cycles (we allow a different distribution of the first cycle). P(C ( t) < a) 4 0 for any a < oo) and ij (t) * oo. is given by Eg(Xoo) = 1 E0 f Ylg (Xt)dt. Y1) le Residual and past lifetime Consider a renewal process and define e ( t) as the residual lifetime of the renewal interval straddling t. r.. [0.
v. the first statement follows. but governed by a Markov chain {Jn} (we .y) = f U(t .5 Under the condition of Blackwell's renewal theorem.U(x) (c < oo because it is easily seen that U(x + 1) . Y1 > t] 4 0. Then Eo^(t) satisfies a renewal equation with z(t) _ E[Y1 . Proof The number Nt of renewal before t satisfies Nt/t a4' p. 1) and W has distribution Fw given by dFw/dF(x) = x/pF.y)P(Yo E dy) . Hence for t large enough. ^) is given by the following four equivalent statements: (a) P (77 > x. the joint distribution of (rl.U(x) < U( 1)). U(x + 1) . Then fi(t)/t a4' 0 and. Yl > t]. V is uniform on (0. are not i. if in addition EYo < oo. Yo > 0] + f Eo^ (t . .dy )z(y) < c ^ l z(k) Eoe(t 0 0 k=o where c = sup.i. ^ > y) = 1 f +Y (z)dz. and the equivalence of (a) with (b)(d) is an easy exercise. l:) is the same as the distribution of (VW. W are independent. Since z ( k) < E[Yi .t. (c) the marginal distribution of q is FO. = z is Foz) The proof of (a) is straightforward by viewing {(r. Y1i Y2. we can bound e(t) by M(t) = max {Yk : k < 2t/p}.(t). we used: Proposition A1...d.'s with finite mean satisfies Mn/n a$• 0 (BorelCantelli). (1 V)W) where V. EC(t)/t + 0.t. use t E^(t)/t = E[Yo . and the conditional distribution of ri given l. (b) the joint distribution of (ri.^(t))} as a regenerative process. r. In the general case. and the conditional distribution of given 17 = y is the overshoot distribution R0(Y) given by FO(Y) (z) = Fo (y+z)/Fo(y).. For the second. In IV. 0 If Markov renewal theory By a Markov renewal process we understand a point process where the interarrival times Yo . the sum is o(t) so that Eo£(t)/t + 0 .d.APPENDIX 335 Theorem A1. (d) the marginal distribution of ^ is FO.6 Consider a renewal process with µ < oo. Since the maximum Mn of n i.4.i. Hence t t lt ) = f U(dy)z(t . assume first the renewal process is zerodelayed.
the Markov renewal process if for any n. Then Xt 4 Xo. Yn.... A stochastic process {Xt}t>o is called semiregenerative w.. ..g. . e. distribution ofjXt}t>o itself where Pi refers to the case Jo = i. oo).7 Consider a nonlattice semiregenerative process.. the conditional distribution of {XT„+t}t>o given Yo. . IT.. Further: Proposition A1. . 0] or (0 . J1 i ...t. These facts allow many definitions and results to be reduced to ordinary renewal.jEE is a family of distributions on (0.336 APPENDIX assume here that /the state space E is// finite) in the sense that P(Y. For example..) and (Fij )i. in [APQ]. .}. X2.i .. Jn_1.and regenerative processes.+ < x. . is given by Eg(X00) = 1 YO vjEj f g(Xt) dt µ jEE o where p = ujEEViAj. with common distribution F.} is nonlattice (it is easily seen that this definition does not depend on i).. Assume that uj = EjYo < oo for all j and that {J„} is irreducible with stationary distribution (v3)jEE.t. Notes and references Renewal theory and regenerative processes are treated. . The semiregenerative process is then regenerative w. oo). r+ < oo). = io for some arbitrary but fixed reference state io E E. Y1. T_=inf{n>0: Sn<0}. We call r+ (T_) the strict ascending (weak descending) ladder epoch and G+ (G_) the corresponding ladder height distributions. where the distribution of X. and define r+=inf{n>0: Sn>0}.r.. Jn +1=j} where J = a(JO. . Jn = i is the same as the P. . Alsmeyer [5] and Thorisson [372]. < yIJ) = Fij( y) on {Jn= i. G_(x) = P(ST_ < x. Let X1.T_ < oo). .. A2 WienerHopf factorization Let F be a distribution which is not concentrated on (oo. G+(x) = P(S. Sn = X1 + • • • + Xn the associated random walk.. be i. Jo. A Markov renewal process {Tn} contains an imbedded renewal process. the semiregenerative process is called nonlattice if {T.r. namely {Twk } where {Wk } is the sequence of instants w where Jo.d.
n=0 The basic identities are the following: Theorem A2. n=0 n=0 00 00 and the T+. m<j<n}. On {T_ > 2}.G+ * G_: (b) G_ (A) = f °° F(A .1 (a) F = G+ + G_ . (A. . Sr_ _1 is at its minimum .=EGn.7). F(A .S. F(A) is the contribution from the event {T_ = 1} = {X1 < 0}. 0]). G_.x)R+(dx). (c) G+(A) = f °. oo) (A. 0<j<m.>0. 0] and (0..T_=n} = {S. we may rewrite (a) as G_ (A) = G+(A) = F(A) + (G+ * G_)(A).APPENDIX 337 Probabilistic WienerHopf theory deals with the relation between F. In (A. 0). S. U.g. A C (0.r..8) (e. A C (0. >0.=n w=m i Figure A. 0]. A C (oo.1 .. G+. More rigorously. (d) R+ = U_. A C (oo. define w as the time where the preT_ path S1.and r_ preoccupation measures T+1 r_1 R+(A) = E E I(Sn E A).S. u . .x)R_ (dx).7) follows since G+(A) = 0 when A C (oo. n 0 R_(A) = E I(Sn E A). the renewal measures U+=>G+. oo). Proof Considering the restrictions of measures to (oc. we consider the last such time (to make w unique) so that {w=m.7) (A. (e) R_ = U+. F(A) + (G+ * G_)(A). oo).
8) is similar.. S.+ E du) E P(S. m < j <n.1. (A. .0<k<ri .. m it follows (see Fig.F(r_n_mSrEA_u).1) that P(Sj Sn. ST_ E A . ST+Edu). ST_ E A) P(T+ = m. m=1 f S mming over n = 2.u) f0m m=1 n=m+1 00 J0 OO P(S.Sn_1Edx..7) follows..1).>0. SmEdu) = P(T+=m.3 8 APPENDIX Reversing the time points 0.XnEAx) 00 f 0 f 0 00 00 1: F(A ....3. 0 < k < n. and the proof of (A. It follows that for n > 2 F (7. Sr_ E Adu) (s ee again Fig . r+ = n) n=1 n=1 0  C0 E fF(Sk< 0. 0<j<m. (b) follows from 00 G+ (A) _ E F(Sn E A.m. SnEAIS.+ E du)P(S. Aso. A.du) (G+ * G)(A)• C llecting terms. A.x)P(Sk < 0._ E A .= n. Sn1 E dx) n=1  F(A . . S.x)R+(dx)._ E A) n1 f P(r_=nw=m Sm EduSrEA) m=1 n1 F(r+=mSr+Edu).. E du) = P(T_=nm. ._ = n . clearly (Sj Sm>0. and reversing the order of summation yields P(T_ > 2.
oo). In discrete time. 0]. and G+. this holds always on the line its = 0. cf..SnEA) = P(Sn<Sk.'s. G_ are trivial. being concentrated at 0.F[s] = (1 . is based upon representing G+ as in (b). the derivation of the form of G+ for the compound Poisson model (Theorem 11.0<k<n. Since G+ is concentrated on (0. However. For example. E. we can rewrite (a) as 1 . and using timereversion as in (d) to obtain the explicit form of R+ (Lebesgue measure). In continuous time. Then for A C (oo.0+[s])(1 . which is basic for the PollaczeckKhinchine formula. 11. there is no direct analogue of Theorem A2. there are direct analogues of Theorem A2.g.SnEA) = P(Sn<Sk.Sn_k.1(a) is from Kennedy [228]. a number of related identities can be derived. the survey [15] by the author and the extensive list of references there.O<k<n. In this generality of. and similarly H_ (s) = 1 . G_ [s] are defined at the same time.SnEA) is the probability that n is a weak descending ladder point with Sn E A.O<k<n.9) whenever F[s]. such developments motivate the approach in Chapter VI on the Markovian environment model. u Remark A2. H+ (s) = 1G+[s] is defined and bounded in the halfplane Is : ERs < 0} and nonzero in Is: Rs < 01 (because IIG+lI _< 1).SnEA) = P(SnSn_ k. For (d). u Notes and references In its above discrete time version. it serves as model and motivation for a number of results and arguments in continuous time.g. see e.2 In terms of m.T+> n) = P(Sk < O. .f. P(SnEA . consider a fixed n and let Xk = Xn_k+l.1. Again. and sometimes in a larger strip.1). the analogue of a random walk is a process with stationary independent increments (a Levy process.4).1. and the proof of (e) is similar.G_ [s] is defined and bounded in the halfplane is : ERs > 01 and nonzero in Is : ERs > 0}. WienerHopf theory is only used at a few places in this book. Summing over n yields R+ (A) = U_ (A).s. if {St} is Brownian motion..6. 6+ [s]. then T+ = inf It > 0 : St = 0} is 0 a. Another main extension of the theory deals with Markov dependence. see for example Bingham [65]. The classical analytical form of the WienerHopf problem is to write 1 .0<k<n. Nevertheless. The present proof of Theorem A2. Sk = X1 + • • • + Xk = Sn .g.APPENDIX 339 and the proof of (c) is similar.G_[s]) (A.P as a product H+H_ of functions with such properties.
11) A f eAtdt = eA.13) henever A is a diagonal matrix with all diagonal elements nonzero. Here it is standard to compute matrixinverses by GaussJordan el imination with full pivoting . if m is s fficiently large. three of the c rrently most widely used ones: xample A3. however . _I 0 (A. one needs to compute matrix inverses Q1 and matrix exponentials eQt ( r just eQ ).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. JAI = max {Jjt : µ E sp(A)} and sp(A) is the set of all eigenvalues of A (the spectrum). whereas there is no similar single established a proach in the case of matrix exponentials. Here are. To circumvent this. Thus. Some fundamental properties are the following: sp(eA) = {e' : A E sp(A)} (A.1 (SCALING AND SQUARING) The difficulty in directly applying t e series expansion eQ = Eo Q"/n! arises when the elements of Q are large. 1.10) d dteAt = AeAt = eAtA (A.12) eA'AO = Ale AA (A.5 that when handling phase type distributi ons. Eo Kn/n! converges rapidly and can be evaluated without p oblems. 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). and eQ can then be computed as the mth power (by squaring if = 2). It is seen from Theorem VIII. 0 . ere A is the eigenvalue of largest absolute value. write eQ = (eK)m where = Q/m for some suitable integer m (this is the scaling step).
One then can reduce to p linear differential equations by noting that k = ZQ. and we may consider a new Markov process {Xt} which has jumps governed by P and occuring at epochs of {Nt} only (note that since pii is typically nonzero .7t) n=0 n! u °O n Pn (to see this. However . Zo = h). 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..... Here is a further method which appears quite appealing at a first sight: Example A3 . Let vi.. assume that Q is the intensity matrix for {Xt} and choose q with rt > max J%J = max qii• 1.e. i.]t)n (A. letting P = I + Q/i and truncating the series in the identity = e17t 00 Pn(.3 i (A. p different eigenvalues Aj i . Zo = a (Z = QZ. The probabilistic reason that (A. Ap. 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. In practice.3 (DIFFERENTIAL EQUATIONS) Letting Kt = eQt.4 (DIAGONALIZATION) Assume that Q has diagonal form. . condition upon the number n of Poisson events in [Olt])  Example A3. some jumps are dummy in the sense that no state transition occurs ). the procedure consists in choosing some suitable i > 0. vp be the corresponding left . construction of {Xt} by realizing the jump times as a thinning of a Poisson process {Nt } with constant intensity 77. i.2 (UNIFORMIZATION) Formally. The approach is in particular convenient if one wants eQt for many different u values of t.e. what is needed is quite often only Zt = TreQt (or eQth) with it (h) a given row (column) vector. .APPENDIX 341 Example A3.14) holds is therefore that the tstep transition matrix for {fft} is eQt = E ent (.14) E n n=0 which is easily seen to be valid as a consequence of eqt = en(Pr)t = entenpt The idea which lies behind is uniformization of a Markov process {Xt}.15) Then it is easily checked that P is a transition matrix . To this end.
we have an explicit formula for eQt once the A j. however. Qhi = vihi... vi. we can take H as the matrix with columns hl.. (A.17) eQt = E e\`thivi = E ea:thi ® vi. and writing eQt as eQt = He°tH1 = H (e\it)di.16) (A.18) contains terms which almost cancel and the loss of digits may be disasterous.. hi have been computed. under the conditions of the PerronFrobenius theorem). In view of this phenomenon alone care should be taken when using diagonalization as a general tool for computing matrixexponentials.18) Namely. i= 1 i=1 P P (A. hp. i=1 i=1 Thus. Then P P Q = > Aihivi = E Aihi (9 vi. and we need to have access to software permitting calculations with complex numbers or to perform the cumbersome translation into real and imaginary parts.. v5Q = Aivi. Example A3. the eigenvalue. Complex calculus : Typically.. and vihi ¢ 0. hp the corresponding right (column) eigenvectors. say A = (Ai)diag. i # j. say Al. There are. some cases remain where diagonalization may still be appealing. not all ai are real.. D = ) 2 2 . this last step is equivalent to finding a matrix H such that H1QH is a diagonal matrix.g H1.5 If Q= ( 411 ( q21 q12 q22 is 2 x 2. and hence A2 is so because of A2 = tr(Q). (A. Nevertheless. two serious drawbacks of this approach: u Numerical instability : If the A5 are too close.. Then vihj = 0. of largest real part is often real (say. Everything is nice and explicit here: 411+q2+D' )12_g11+q2^^ where (411422z + 4412421.342 APPENDIX (row) eigenvectors and hl. The phenomenon occurs not least when the dimension p is large. and we may adapt some normalization convention ensuring vihi = 1.
u Example A3.APPENDIX 343 Write 7r (= v1) for the left eigenvector corresponding to a1 and k (= hl) for the right eigenvector.20) ir = q2 ql qi +q 2 9l +q2 (A.19) Example A3 .e. b are any constants ensuring//Irk = 1.k1). replacing ai by A2.q. However. v2 and h2 can be computed in just the same way. where (A. k  C k2 ) =b ( A1 q 1 Q11 / where a . i. Then 7r = (ir1 7r2 ) = a (q21 Al .Q2i and after some trivial calculus one gets eQt = 7r 1 112 + eat 7r1 7r2 / (7fl 7r2) = ( 7r2 1r2 7r1 IF. 1) . Then Al = 0 and the corresponding left and right eigenvectors are the stationary probability distribution 7r and e. The other eigenvalue is A = A2 = q1 . h2 = Thus.6 A particular important case arises when Q = q1 qi ) q2 q2 J is an intensity matrix. it is easier to note that 7rh2 = 0 and v2k = 1 implies v2 = (k2 . eqt = eNlt ( ir1ki i2k1 \ ir1 k2 72 k2 + e azt 7r2k2 i2k1 7ri k2 7r1 k1 (A. l ab (g12g21 + (A1  411) 2) = 1. Of course.7 Let 3 9 2 14 7 11 2 2 .21) Here the first term is the stationary limit and the second term thus describes the rate of convergence to stationarity.
2 2 1=ab(142+(1+2)2 ) = tab. They are most often constructed by imposing some additional properties . (AA+)' = AA+. but only that dimensions match . A2 = 3/2 . e_6u A4 Some linear algebra 4a Generalized inverses A generalized inverse of a matrix A is defined as any matrix A. Generalized inverses play an important role in statistics.344 Then D= 2+ 11)' 7 T4 2 =52. and a generalized inverse may not unique.. (A.. for example AA+A = A.22) Note that in this generality it is not assumed that A is necessarily square. 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_. (A.11/2 + 5 1.23) . APPENDIX x1 3/2 .satisfying AAA = A.5 .6. (A+A)' = A+A.11/2 . A+AA+ = A+.
.. Here is a typical result on the role of such matrices in applied probability: Proposition A4.1 Let A be an irreducible intensity matrix with stationary row vector it. Rather than with generalized inverses . = 0 where m < p is the rank of A. and define D = (A . ( Q .24) = te7r . Then for some b > 0.. then there exists an orthogonal matrix C such that A = CDC' where 0 0 D = AP Here we can assume that the A . 0 01 In applied probability. Am+1 = . Assume that a unique stationary distribution w exists .25) . E.1 goes under the name fundamental matrix of the Markov chain).g.eir ).g.e. most often either an intensity matrix Q or a matrix of the form IP where P is a transition matrix. (I . one then works with Q = (Q . are ordered such that Al > 0.P). .= (I ..ew. one is also faced with singular matrices .P + e7r)1 (here ( I .P + e7r ).APPENDIX 345 A matrix A+ satisfying (A.e ® 7r)1.eir )1.1Q = Q(Q . (A..D + O(ebt). These matrices are not generalized inverses but act roughly as inverses except that 7r and e play a particular role . _ A..23) is called the MoorePenrose inverse of A.eir)1 = I . and exists and is unique (see for example Rao [300]). if A is a possibly singular covariance matrix (nonnegative definite).. lt o eAx dx = te7r + D(eAt . Am > 0. and can define /ail 0 0 0 0 0 0 A+ = C A' 0 0 0 C' .I) (A.
h. . B'(t) = e7r + DAeAt = eir + (I . Equivalently.26) 2 = 2 e7r + tD . in block notation i2h A®B= ( a11B a21 B a12B a22 B Example A4. 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. of (A. (A.3 Let 2 A= 4 3 Vf' N7 5 )' B= ( 8 ). and the columns to h.eir)eAt = eAt = A'(t).J {xe^r + D(e .s. resp.. (A. h ® it reduces to hit in standard matrix notation. ()®(6 f 6/ 7f 8^ 7 8 )=! ^)( 6 7 8 )=(6^ 7^ 8^) \ u Example A4.26) follows by integration by parts: t f t /' xeAx dx = [x {xe7r + D(eAx .I)}.346 t APPENDIX 2 xe Ax dx = eir + t(D + e7r) + D(eAt . For example.I) . the r. it follows that h ® it is the k x m matrix with ijth element hi7rj . h as 1 x m and k x 1 matrices.I) (A.s. respectively.D + D2 + O(ebt). the formulas involving O(e6t) follow by PerronFrobenius theory. Interpreting 7r.91a(2) . I.DZ(ent .I)} dx. 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. Then A(O) _ B(O) = 0. Note that h ® it has rank 1.2 Let it be a row vector with m components and h a column vector with k components.2e7r . o Finally.h. B(t) denote the l. the rows are proportional to it.e. see below. and in fact any rank 1 matrix can be written on this form.24). .27) Proof Let A(t).
5v/.28) In particular.30) eA+B = eAeB function generalizes to Kronecker notation (note that in contrast typically only holds when A and B commute): Proposition A4.3V8. and the number of such factors is precisely given by the relevant binomial coefficient.A9.5v'8 5vf9 11 A fundamental formula is (A1B1C1) ®(A2B2C2) = (A1 (9 A2)(B1 (9 B2)(C1®C2). then v1B1h1 and v2B2h2 are real numbers. it follows that e® ® e B An _ 0o oo oo Bn 7 I F n! = ` k! (I . then the Kronecker sum is defined by A(1) ®A(2) = A(1) ®Ik2 + k ®A(2). (AED B)1 = (A®I+I(9 B)l is the sum of all products of t factors.31) Indeed. Proof We shall use the binomial formula A crucial property is the fact that the functional equation for the exponential t / l (A ®B)t = I k Ak 0 B1k k=0 (A. Using (A.3v'6. such a factor is Ak (&B 1k according to (A.3vV/72f 20.31).(A. each of which is A ® I or I ® B.50 6 7 6 4f 4. C2 = h2 are column vectors.APPENDIX 347 Then A®B = 2 f 20. if Al = vi.3f 4v/.4vf.29). and v1B1h1 • v2B2h2 = v1B1h1 ® v2B2h2 = ( v1(&v2 )( B1(&B2 )( h1(&h2 ) .29) If A and B are both square (k1 = ml and k2 = m2). if A ® I occurs k times.k)! ( n0 n=0 t=0 k=0 J _ ® Ak ®Blk r ^.4 eA® B = eA ®eB. (A B)' = eA®B e! L 1=0 0 . A2 = v2 are row vectors and C1 = h1. (A. (A.
v whenever a is an eigenvalue of A and 0 is an eigenvalue be any row vectors and h. { On the other hand. Thus . Yt(2) where independent Markov processes with intensity matri{y(2) } are {Y(1) }. X ) }.32) is the intensity matrix of the bivariate continuous Markov process {Yt(1). P(t) Yt(2) }. Ps 1) = exp {sQ ( 1) } > p(2 ) = exp {sQ(2) } can therefore be rewritten as Taking s = 1 for simplicity . resp . independent Markov chains.348 APPENDIX Remark A4. we have P8 = Pal) ® p(2). (A. 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 +. 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). A special case of Proposition A4. the {Yt(2) } transitions in the {Yt(1) } component and the second transitions in the component . Yt(2 ) }. n2 n1 ) {X(2) } are independent Markov chains with transition matrices P(1).32).33) . Q(2).4 can easily be obtained by probabilistic be the sstep transition reasoning along the same lines .5 Many of the concepts and results in Kronecker calculus have p(2) is the intuitive illustrations in probabilistic terms. k any column vectors.I)(h ® k). {Yt(1). { 1't(1) }. From what has been said about matrices of {Yt( 1). Let P8f P(Sl). Then 2 0 ire At h • ve Bt kdt = (^®v)(A®B)1(e A®Ba . P8 = exp {sQ} = exp {s (Q(1) ®Q(2)) } . and Q = Q(1) ® Q (2) = Q(1) ® I + I ® Q(2) (A.6 Suppose that A and of B.s. the same time. where transition matrix of the bivariate Markov chain {X n1).3 < 0 Lemma A4 . h. in the definition (A. Let further it. p = P(1) ® {X }. first term on the r . represents ces Q( 1).
. i. then IN < Ao for all A E sp(A). 4c The PerronFrobenius theorem Let A be a p x pmatrix with nonnegative elements. and if we normalize v. and the corresponding left and right eigenvectors v. h such that vh = 1. . the integrand can be written as ( 7r (9 v)( eAt ® eBt )(h ®k ) = ( 7r ®v)(eA (DBt)(h (& k). which can be found in a great number of books. = j and atk_li.1 and references there (to which we add Berman & Plemmons [63]): Theorem A4. and the corresponding left and right eigenvectors v.APPENDIX 349 Proof According to (A. (A. Now note that the eigenvalues of A ® B are of the form a +. .29). f o r each i. . h can be chosen with 3By this. .12). Then: (a) The spectral radius Ao = max{JAI : A E sp(A)} is itself a strictly positive and simple eigenvalue of A. we have AO = 1.. h = e and v = 7r (the stationary row vector).g.. We call A irreducible if the pattern of zero and nonzero elements is the same as for an irreducible transition matrix. in such that io = i. we mean that the pattern of nonzero offdiagonal elements is the same as for an irreducible intensity matrix.3 whenever a is an eigenvalue of A and 3 is an eigenvalue of B. p there should exist io. Similarly. Then the eigenvalue Ao with largest real part is simple and real.7 Let A be a p x pmatrix with nonnegative elements. see e.. (b) if in addition A is aperiodic. > 0 for k = 1. That is. E (0. then An = Aohv+O(µ") = Aoh®v+O(µ") for some u.. ao). ...34) Note that for a transition matrix. and appeal to (A. il.The PerronFrobenius theorem has an analogue for matrices B with properties similar to intensity matrices: Corollary A4. h can be chosen with strictly positive elements. A is called aperiodic if the pattern of zero and nonzero elements is the same as for an aperiodic transition matrix.8 Let B be an irreducible3 p x pmatrix with nonnegative offdiagonal elements. so that by asssumption A ® B is u invertible. . [APQ] X. . Here is the PerronFrobenius theorem. . n. . j = 1.
. one can consider A = 77I + B where rl > 0 is so large that all diagonal elements of A are strictly positive (then A is irreducible and aperiodic). we have A0 = 0. A5 Complements on phasetype distributions 5a Asymptotic exponentiality In Proposition VIII. Then for any (3. To this end. the condition is that t is small compared to Q. h such that vh = 1.(3. note that we can write the phase generator T as Q . Bi° (x) + at*x Proof Let { 4 } be the phase process associated with B(a) and (°) its lifelength. The next result gives a condition for asymptotical exponentiality. let t = (ti)iEE # 0 have nonnegative entries and define T(°) = aQ . not only in the tail but in the whole distribution.(ti)ding.1.e. h = e and v = 7r (the stationary row vector). then eBt = ea0thv + O(eµt) = eA0th ® v + O(et t) (A. Furthermore. it was shown that under mild conditions the tail of a phasetype distribution B is asymptotical exponential. Proposition A5. let {Yti°i } be a Markov process with initial distribution a and intensity .35) for some p E (oo. The content is that B is approximately exponential if the exit rates ti are small compared to the feedback intensities tij (i # j). Example A3.n t AL n=0 n! (cf. if we normalize v. I. 10) and use the formula me at e Bt = e 00 Antn = e .8.2). Corollary A4.(ti)diag where Q = T + (ti)diag is a proper intensity matrix (Qe = 0). the phasetype distribution B(a) with representation (.1 Let Q be a proper irreducible intensity matrix with stationary distribution a. but is an easy consequence of the PerronFrobenius theorem. Note that for an intensity matrix. T(°)) is asymptotically exponential with parameter t* _ r EiEE aiti as a 4 oo.8 is most often not stated explicitly in textbooks.350 APPENDIX strictly positive elements. relate the eigenvalues of B to those of B via (A. For example. the analogy of this procedure with unformization. Ao).
v/ t. Conditioning upon whether { Yt} changes state in [0.2 Pi (c(a) > x.aE where 0 < e < 1). we get dx F (Idx = j) = (1 + qij t )Sij + qij dt. a' = a .Yj(av) = j f . J^O)_ = j) Pi (v(aaV) > x. = YQ(x). Let further V be exponential with intensity V and independent of everything else. dx/ti] or not. We can think of ( ( a) as the first event in an inhomogeneous Poisson process ( Cox process ) with intensity process matrix aQ .)_ = Y(a) = 1'aS(a) = Ya(av)^ it follows that Pi ((. and this easily yields a(x)/x a' 1/t*. prove a somewhat more general result which was used in the proof of Proposition VI. Since JJ(.jEE. Hence O ((a) aa.(a) > x . t < (a). in fact . a . has a limit distribution: Proposition A5. and that Yt(a) = Yat for all t. J(()) _ = i) + at•x t tt' . it states that the state. a'/a + 1. In addition to the asymptotic exponentiality. {t Y( a) } v>0 . and write Yt = Yt(1).YQ(av) = j) Pi ( ci(a'V) > x. By the law of large numbers for Markov processes .x (1 . 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. Then {Ix} is a Markov process with to = Yo. 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}.g.1. Then a(a'V)/a (aV) a' 1. We can assume that Jta) = Yt(°). Proof Assume first ti > 0 for all i and let I. fo tY dv/t a$' t*. We shall .a' + oo (e.bij) Hence the intensity matrix of { Ix} is (qij/ti)i.9. from which the phase process is terminated .APPENDIX 351 ((1) etc.
zP)'p.. k = 1. Then P is substochastic and the vector of exit probabilities is p = e . Gnedenko & Kovalenko [164] and Glasserman & Kou [162]).5 Let B be discrete phasetype with representation (P. a).p)k1 p. 2.. P. so we shall be brief.. (c) the nth moment k 1 k"bkis 1)"n!aP"p. Penev & Turbin [238].1 and A5. and thus the parameter p of the geometric distribution u can be identified with the exit probability vector p.. Example A5. Et II I a(a^V) > x) at' . zkbk is za(I .2 do not appear to be in the literature.+ at*x • a't' L ` at t* t* J Reducing the state space of {Ix } to {i E E : t. is discrete phasetype. 1 k=1 1 0 otherwise. so is the geometric distribution. u Notes and references Propositions A5. k>1. 5b Discrete phasetype distributions The theory of discrete phasetype distributions is a close parallel of the continuous case. See also Korolyuk. the simplest discrete phasetype distribution: here E has only one element. Example A5..4 Any discrete distribution B with finite support. an easy modification of the argument yields finally the result for the case where t.. .j) and initial distribution a. = 0 for one or more i.. However.} is said to be discrete phasetype with representation (E.. let E and Pkj j=k1.. a = b = (bk)k=1. say bk = 0. ' pk 0 k>1 11 Theorem A5. these results are in the spirit of rare events theory for regenerative processes (e. . 2.Pe. > 0}.x k > K. (b) the generating function b[z] _ E' . K}. Then: (a) The point probabilities are bk = aPklp.352 rr Ia(a'V) Ei I ( > x) P APPENDIX L at (Yo (aV) .. with point probabilities bk = (1 . . A distribution B on {1.3 As the exponential distribution is the simplest continuous phasetype distribution.g.. a) if B is the lifelength of a terminating Markov chain (in discrete time) on E which has transition matrix P = (p. Indeed. Keilson [223].
T= ( 0 T(2) ) (A. (E(2).APPENDIX 353 5c Closure properties Example A5.r + 1.36) in blockpartitioned notation (where we could also write a as (a (1) 0)).a(2). . Then {Jt} has lifetime U1 + U2 . and a=1).6 (CONVOLUTIONS) Let B1. Jt t > U1 + U2. _ i E E(1) T(1) t(1)a(2) i E E(2) .a(1). a' . { Jt 2) } with lifetimes U1 ..T(1)).6 is the Erlang distribution Er which is the convolution of r exponential distributions.. U2. as is seen by minor modifications of Example A5. The discrete counterpart is the negative binomial distribution with point probabilities bk k1) (1 k = r.2 The form of these results is easily recognized if one considers two independent phase processes { Jt 1) }. r . and hence the negative binomial distribution is discrete phaseu type.. A. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2).. resp.{ 0. a. B2 be phasetype with representations (E(1).6. and piece the processes together by it = 41) 0<t<U1 U1 < t < U1 + U2 2U.7 (THE NEGATIVE BINOMIAL DISTRIBUTION) The most trivial special case of Example A5.T(2)). 11 Example A5. resp. initial distribution a and phase generator T.1 This corresponds to a convolution of r geometric distributions with the same parameter p. Then the convolution B = B1 * B2 is phasetype with representation (E. A reduced phase diagram (omitting transitions within the two blocks) is am E(1) t(1) a(2) (2) t(2) Figure A.
4 . a.p)pn1.10 (GEOMETRIC COMPOUNDS) Let B be phasetype with representation (E. Let B(") be the corresponding phasetype distribution. p at each termination. Example A5. a. i E E(1) T 0 I (A..T. resp. then C is the distribution of Ul + • • • + UN. A reduced phase diagram is 0a(1) E(1) A . Equivalently. B2 be phasetype with representations (E(1).p. In risk theory. with common distribution and N is independent of the Uk and geometrically distributed with parameter p. P(N = n) = (1 . i E E(2) 0 T(2) =IT (in blockpartitioned notation. a mixture of more than two phasetype distributions is seen to be phasetype. (E(2).3 In exactly the same way.T(2)). one obvious interpretation of the claim u size distribution B to be a mixture is several types of claims. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2).i.0)ai2).a(2). Thus. this means that a = (Oa(1) (1 . are i..O)B2 (0 < 0 < 1) is phasetype with representation (E.T(1)). if U1. To obtain a phase process for C.354 APPENDIX Example A5.E) where a(°) = fAa(a)v(da).').8 (FINITE MIXTURES) Let B1. Then it is trivial to see that B(") is u phasetype with representation (a(").0)a(2) E(2) Figure A. and o'i Oa. Example A5. we need to restart the phase process for B w.0)a(2))).d.37) (1) (1 .. and consider B(") = fA B(a) v(da) where v is a probability measure on A. T) and C = EO°_1(1 . Then the mixture B = 9B1 + (1 .a(1). U2.9 (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.p)pn1B*n. a reduced phase diagram is f a E t Figure A.
aF[T]. E). a. Note that this was exactly the structure of the lifetime of a terminating renewal u process. f2. If U1 has a different initial vector. let the initial vector be a ® v and u let the phase generator be I ® T + P ® (ta). (E(2). v. resp. a(2). Proposition VIII. Example A5. Then the minimum U1 A U2 and the maximum U1 V U2 are again phasetype.. then Jy has distribution aeTx. Thus the representation is (E(1) x E(2). T) and C = F. a. then C is the distribution of U1 + • • • + UN. if {Jt} is a phase process for U.T) where F[T] = J0 "o eTx F(dx) u is the matrix m.x)+. be the point probabilities of a discrete phasetype distribution with representation (E...APPENDIX 355 and C is phasetype with representation (E. Example A5 . let {Jtl)}. T + pta). resp... a.°_1 f„ B*?l. a(1) ® a(2 ). Corollary VIII. { 4 } as exit of {Jt}. For U1 A U2. T). Example A5 . T(1) ® T(2)). then U1 +• is phasetype with representation (E. { Jt2) } be independent with lifetimes U1.g. say with distribution F. To see this.d. v. B2 of phasetype with representations (E('). Indeed..f. we then let the governing phase process be {Jt} _ {(411 Jt2))} 2) interpreting exit of either of {4 M }. say v.9) that (U . Minor modifications of the argument show that 1. if U1. 13 (MINIMA AND MAXIMA ) Let U1. are i. U2 be random variables with distributions B1.T + pta).°.1. U2.a(1). if B is defective and N + 1 is the first n with U„ = oo. It is zeromodified phasetype with representation (E. j E F}. of F. with common distribution B and N is independent of the Uk with P(N = n) = f. but the same T..aeTx. let the phase space be E x F = {i j : i E E.2.TWWW). U2. .. let B be a continuous phasetype distribution with representation (F.7. P). cf.11 (OVERSHOOTS) The overshoot of U over x is defined as the distribution of (U . If we replace x by a r. it follows by mixing (Example A5. T + ta. cf.v. i. T(2) ). +UN 2. X independent of U. To obtain a phase representation for C . . then U1 + • • + UN is zeromodified phasetype with representation (a.T) if U is phasetype with representation (E.2.X)+ is zeromodified phasetype with representation (E. 12 (PHASETYPE COMPOUNDS ) Let fl. Equivalently.
B(bk) I < 1/n for n > k. q(n) q(n) pi(n)a .} of phasetype distributions such that Bn 3 B as n + oo. Thus the state space is E(1 ) x E(2) U E(1) U E( 2). The general case now follows easily from this... Here are the details at two somewhat different levels of abstraction: (diagonal argument . By the diagonal argument (subsequent thinnings).356 APPENDIX For U1 V U2.. and vice versa. elementary) Let {bk} be any dense sequence of continuity points for B(x). The mean of B„ is n/Sn = b and the variance is n/Sn = b2/n. and the closedness of the class of phasetype distributions under the formation of finite mixtures. and let Bn be the Erlang distribution E. we can assume that ID. however.(bk) + B(bk) for all k as n * oo. Proof Assume first that B is a onepoint distribution. Let the support of Dn be {xl(n). Now we can find first a sequence {Dm} of distributions with finite support such that D.. r # oo.8... there is a sequence {B. 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 . relies more on matrix algebra than the probabilistic interpretation exploited here).(bk) + B(bk) for all k. Hence it is immediate that Bn 4 B.(Sn) with Sn = n/b. oo).14 To a given distribution B on (0.(n) = D.n = I:pi(n)Er v ( __ ) n) ) a= 1 . Then we must find phasetype distributions Bn with B. That is.xq(n)(n)}.. i= 1 C. any distribution B on (0. Then from above. the fact that any distribution B can be approximated arbitrarily close by a distribution with finite support. the initial vector is (a(1) (& a (2) 0 0). Example A5. say degenerate at b. cf. 5d Phasetype approximation A fundamental property of phasetype distributions is denseness . see Neuts [269] (where the proof.2) } to go on (on E(2)) when { i 1) } exits. we need to allow { Jt. with weight pi(n) for xi(n). oo) can be approximated 'arbitrarily close' by a phasetype distribution B: Theorem A5..(bk)'.
u 2 (abstract topological ) The essence of the argument above is that the closure (w. 2.e. In particular.( dx) * f r f{(x)B(dx). Hence G C PET and L = PIT.15 To a given distribution B on (0 . k < n. It should be noted. Let E be the class of functions f : [0... oo). the topology for weak convergence) PET of the class PET of phasetype distributions contains all onepoint distributions. one would use the B given by some statistical fitting procedure (see below).n. Since PET is closed under the continuous operation of formation of finite mixtures.14 is fundamental and can motivate phasetype assumptions. oo) such that f (x) = O(e«x). i.APPENDIX 357 Hence we can choose r(n) in such a way that ICr( n).. For a general Bo. and that cp is known to be continuous. we can then approximate Bo by a phasetype B.i. i = 1.. f2. But To is the class G of all distributions on [0. the class CO of all discrete 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. Then ICr( n ). however. .r. E E. u Theorem A5.(x)Bf. and we can take Bn = Cr(n).n( b k ) . that this procedure should be used with care if ^p(B) is the ruin probability O(u) and u is large. oo) and any fl.. for some a < oo. then it is immediate that WI(B) = p2(B) for all distributions B on [0.t.n (bk) . oo) approximation Assume that we can compute a functional W(B) when B is phasetype. say on the claim size distribution B in risk theory. there is a sequence {Bn} of phase type distributions such that Bn Di B as n 4 oo and f ' f. k < n. replications). Corollary A5.D(bk)I < n. if information on Bo is given in terms of observations (i.B(bk )I < .d. If Cpl (B) and ^02(B) are weakly continuous.. PIT contains all finite mixtures of onepoint distributions. compute W(B) and use this quantity as an approximation to cp(B0). x 4 oo. oo) * [0...
n. if f (x ) = e°x. Bn=En z f f (x)Bn(dx) fof (x)B(dx) = ° (A. n.n(dx) + f 0 fi(x)Dn(dx). .. TO (A.. \\ 0 Corollary A5...39). and hence it is sufficient to show that we can obtain limsup n4oo fi(x)Bn(dx) < Jo 0 f fi( x)B(dx ). there is a sequence {Bn} of phase type distributions such that Bn Di B as n + oo and all moments converge. i = 1. i = 1. i=1. Now returning to the proof of (A. .16 To a given distribution B on (0 . . . for each i.2 . f° xtBn(dx ) * f °° x`B( dx).. then cc f (x)Bn ( dx) = (?!c ) e'= . and hence we may choose r(n) such that L 9l) f (x)Cr(n).  APPENDIX B implies that 00 o o 00 n.f (x)B(dx). By (A... n B=az.39) Indeed.oo J fi(x)B.38 ).(dx) > J fi(x)B(dx).358 Proof By Fatou' s lemma.... i = 1. 2.n(dx) < 1+. and the case of a general f then follows from the definition of the class E and a uniform integrability argument. . oo).14 Dn has been chosen such that 00 1 °° f fi(x)D n(dx ) < 1++ '  o \ n o f fi(x)B(dx).f (z) = f = 1 1 1 1n/ o .38) We first show that for each f E E.. liminf B... f00 fi(x)Cr. we may assume that in the proof of Theorem A5.f ' f (x)B(dx).
5e Phasetype fitting As has been mentioned a number of times already. The adjustment coefficient is a fundamental quantity.} of phasetype distributions such that Bfz + B as n * oo and Yn 4 ry where ryn = y(Bn.18 In the setting of Corollary A5./3).l3µb < 1. from a more conceptual . This is motivated in part from the fact that a number of nonphasetype distributions like the lognormal. (N. there is substantial advantage in assuming the claim sizes to be phasetype when one wants to compute ruin probabilities. the remaining results may be slightly stronger than those given in the literature. . there is a sequence {B./3) is defined as the unique solution > 0 of B[y] = l+y/j3. and therefore the following result is highly relevant as support for phasetype assumptions in risk theory: Corollary A5. The present section is a survey of some of the available approaches and software for inplementing this. lim sup ryn < 7. For practical purposes. /3) = ry for all n.. O We state without proof the following result: Corollary A5. oo) with B[y +e] < oo for some e > y = 7(B.16. . . lim inf > is proved similarly.e. one can obtain 7(Bn. but are certainly not unexpected. I. e ) and ei J...14 is classical. Notes and references Theorem A5. and in part from the fact that many of the algorithms that we describe below have been formulated within the setup of fitting distributions.3). (N or a given distribution Bo. . We shall formulate the problem in the slightly broader setting of fitting a phasetype distribution B to a given set of data (1i . If ei > 0. 0 as i * oo. Proof Let fi(x) = el'r+E. However. the problem thus arises of how to fit a phasetype distribution B to a given set of data (1.17 To a given /3 > 0 and a given distribution B on (0. . the adjustment coefficient 'y = 7(B. .> y for some sequence {ei} with ei E (0. . the loggamma or the Weibull have been argued to provide adequate descriptions of claim size distributions. then Bn['Y + ei] * B[y + ei] > 1 + 7 Q implies that 'yn < ry + ei for all sufficiently large n .APPENDIX 359 In compound Poisson risk processes with arrival intensity /3 and claim size distribution B satisfying .
Johnson & Taaffe considered a mixture of two Erlangs (with different rates ) and matched (when possible ) the first three moments . and we next describe two such approaches which also have the feature of being based upon the traditional statistical tool of like maximum likelihood.g . It seems therefore a key issue to develop methods allowing for a more general phase diagram.360 APPENDIX point of view the two sets of problems are hardly different : an equivalent representation of a set of data (1 . A method developed by Bobbio and coworkers (see e..d. g.g. Schmickler (the MEDA package. giving mass 1 /N to each S=. The likelihood function is maximized by a local linearization method allowing to use linear programming techniques.} of phasetype distribution such that Bo. for some suitable large n.'s). defined by the absence of loops in the phase diagram . at a a number of selected points . B„ The problem is that the constructions of {B„} are not economical : the number of phases grows rapidly. we have constructed a sequence { B. risk theory.f. a program package written in C for the SUN workstation or the PC is available as shareware. 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. The characteristics of all of these methods is that even the number of parameters may be low (e. one could argue that the results of the preceding section concerning phasetype approximation contains a solution to our problem : given Bo (or Be). The earliest such reference is Bux & Herzog [85] who assumed that the Erlang distributions have the same rate parameter. [317] ) has considered an extension of this setup. and in practice this sets a limitation to the usefulness (the curse of dimensionality ..g. [216] ). A number of approaches restrict the phase type distribution to a suitable class of mixtures of Erlang distributions .. and this is what matters when using phasetype distributions as computational vehicle in say renewal theory. d. we do not not want to perform matrix calculus in hundreds or thousands dimensions). [202]. the L1 distance between the c . and used a nonlinear programming approach . [70]) restrict attention to acyclic phase type distributions . . 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. cf. the number of phases required for a good fit will typically be much larger. The observation is that the statistical problem would be straightforward if the whole ( EAvalued) phase process { Jtk)} o<t<( k associated with each observa .. Asmussen & Nerman [38] implemented maximum likelihood in the full class of phasetype distributions via the EM algorithm . .g. e . (N is the empirical distribution Be. Of course. reliability or queueing theory.f. In a series of papers (e. three for a mixture of two Erlangs ). and as fitted distribution we may take B.
. eieT(n)((k.g.. Thus. it seems open whether the restriction to the acyclic case is a severe loss of generality. . . one is lead to an iterative scheme.(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.. EN where ai = N 1 I ((k) = i) tii=i iEE.T(n) (Ti ^^ 1. . the methods of [70] and [38] appear to produce almost identical results. .x)t(n) 1 and this and similar expressions are then computed by numerical solution of a set of differential equations. (n+1) _ Ea (n).T(n) (Nik IC1. In fact. E. since this is parameterdependent. Nii = = ..g..(N) = E Ea(n). The general idea of the EM algorithm ([106]) is to replace such unobserved quantities by the conditional expectation given the observations. jEEA.. it is easy to see that N (k Ea(n).. N Ti = I(J= i) dt. e.T(n) k=1 I (Jti) dt o \f a(n)eT(n )(kt(n) N f:i a(n)eT(n)xei . .. In practice. then the estimators would be of simple occurenceexposure type.APPENDIX 361 tion Sk was available.. (N) tJk Ea ( n). = j) f k=1 k =1 tE[0. (N ) (^ 54 k )+ and similarly for the cn+1) The crux is the computation of the conditional expectations.
This page is intentionally left blank .
Abate & W. 191223. Arfwedson ( 1954) Research in collective risk theory. Proc. Whitt (1994) Waitingtime tail probabilities in queues with longtail service time distributions . 363 . Queueing Systems 16. New York. Asmussen (1984) Approximations for the probability of ruin within finite time . The Mathematical Scientist 14. [3] J. G. 20. Teubner.L. Sparre Andersen (1957) On the collective theory of risk in the case of contagion between the claims . Borokov. 14. 37. Abate & W. [14] S. Abate. 253267. Asmussen ( 1982 ) Conditioned limit theorems relating a random walk to its associate. 213229. Stoch. Aktuar Tidskr. 3157 . Adv. [11) S. AT&T. Skand. [8] G. ed.Bibliography [1] J. 1985. [4] M. Skand. 345368. 1984 . Dover. In: Limit Theorems and Related Problems (A.). with applications to risk reserve processes and the GI /G/1 queue. 57. Queueing Systems 10. Appl. J. Stegun ( 1972 ) Handbook of Mathematical Functions (10th ed. New York. The case of equal risk sums. Probab. Chichester New York. [7] E. Scand. [13] S. 53100. Alsmeyer (1991) Erneuerungstheorie. pp. Arfwedson ( 1955) Research in collective risk theory. Appl. Preprint. Transactions XVth International Congress of Actuaries. B. 101116. New York. [9] G. [6] V. John Wiley & Sons.). Abramowitz & I. 587. Asmussen (1989a) Aspects of matrix WienerHopf factorisation in applied probability. [2] J. Anantharam (1988 ) How large delays build up in a GI/GI11 queue.A. Asmussen (1985) Conjugate processes and the simulation of ruin problems. [5] G. Queueing Systems 5 . Stuttgart. Aktuar Tidskr. II. [10] K. Act. Optimizations Software. 311338. Whitt ( 1992) The Fourierseries method for inverting transforms of probability distributions . [15] S. 219229. Choudhury & W. ibid. Asmussen (1987) Applied Probability and Queues. 38. Arndt (1984) On the distribution of the supremum of a random walk on a Markov chain . Whitt (1998) Explicit M/G/1 waitingtime distributions for a class of longtail service time distributions .G. [12] S. 143170.
[30] S. [25] S. 315. Asmussen (1992a) Phasetype representations in random walk and queueing problems. Methods V Problems (J. 8. 303322. Andronov. 25. Act. L. 772789. Binswanger & B. 4966. [29] S. 37. Asmussen (1995) Stationary distributions via first passage times. Asmussen & K. [19] S. Scand. 2149. Statist. 354374.). 189201. J. K. [21] S. MatrixAnalytic Methods in Stochastic Models (A. Stoch. Appl. Stoch. 313326. Asmussen. 20. [32] S. Asmussen (1989b) Risk theory in a Markovian environment. Alfa & S. T. [17] S. Chakravarty. Act. Asmussen (1998c) A probabilistic look at the WienerHopf equation. 54. [22] S. Random Processes 2. J. Bernoulli 6. Asmussen (1992c) Stationary distributions for fluid flow models and Markovmodulated reflected Brownian motion.K. [26] S. Schmidt (1995) Does Markovmodulation increase the risk? ASTIN Bull. Astin Bulletin 27. Scand. Appl. [18] S. Probab. Asmussen (1991) Ladder heights and the Markovmodulated M/G/1 queue. Asmussen & M.). Dshalalow ed. [33] S. 297318. Scand. Ann. 2943. Probab. 1989 . Ann. 106119. [20] S. Stochastic Models 11. Boca Raton. [34] S.). 313341. Advances in Queueing: Models. 555574. A. Hojgaard (1999) Approximations for finite horizon ruin prob abilities in the renewal model. Asmussen (1998b) Extreme value theory for queues via cycle maxima. Asmussen (1999) On the ruin problem for some adapted premium rules. 27. Probabilistic Analysis of Rare Events (V. SIAM Review 40. 1996. J. 137168. Ann. Asmussen & B. Kalashnikov & A. Marcel Dekker. Asmussen. Bladt (1996) Phasetype distributions and risk processes with premiums dependent on the current reserve. 1936. [27] S. Asmussen. 193226. Floe Henriksen & C. Hojgaard (2000) Rare events simulation for heavytailed distributions. Appl.M. 79102. [24] S. Asmussen & B. stationary distributions and first passage probabilities. Hojgaard (1996) Finite horizon ruin probabilities for Markovmodulated risk processes with heavy tails. [23] S. Florida. [28] S. . eds. Riga Aviation University. New York. Bladt (1996) Renewal theory and queueing algorithms for matrixexponential distributions. Asmussen (1998a) Subexponential asymptotics for stochastic processes: extremal behaviour. Binswanger (1997) Simulation of ruin probabilities for subexponential claims. Asmussen (1992b) Light traffic equivalence in single server queues. 96107. Act. Probab. Kliippelberg (1994) Large claims approximations for risk processes in a Markovian environment. Extremes 1 . eds. Frey. 69100. 2. Asmussen (2000) Matrixanalytic models and their analysis. Appl. Scand.364 BIBLIOGRAPHY [16] S. Asmussen & M. Proc. 1999 . Th. CRC Press.S. J. Rolski & V. Proc. [31] S.
Statist. [43] S. Asmussen & C. Schmidli & V.M. Math. Appl. 11251141. Asmussen & T. Stochastic Models 8. 103125. 410433. [52] S. 736755. Asmussen . Res. 419441. Asmussen & V. with applications to insurance risk Stoch. Boca Raton. Marked point processes as limits of Markovian arrival streams. Eng. Example of excessoffloss reinsurance for an insurance corporation. Advances in Queueing: Models. Kliippelberg (1996) Large deviations results for subexponential tails.). Asmussen & G. Koole (1993). 913916. Schmidt (1999) Tail approximations for nonstandard risk and queueing processes with subexponential tails. J. [42] S. [50] S. . [39] S. Asmussen & D. J. Appl. Finance and Stochastics 4. Asmussen & S.Y. 429466. Perry (1992) On cycle maxima. O'Cinneide (2000/2001) Matrixexponential distributions [Distributions with a rational Laplace transform] Encyclopedia of Statistical Sciences. 299324. J. Adv. Appl. Probab. [36] S. Opns. Asmussen & C. Adv. Hojgaard & M. Rubinstein (1999) Sensitivity analysis of insurance risk models. B. [41] S. Wiley. [40] S. [45] S. Asmussen & V. Asmussen. Read eds. Probab. Statistica Neerlandica 47. Schmidt (1995) Ladder height distributions with marks. Fitting phasetype distributions via the EM algorithm. Stoch. Management Science 45. Supplementary Volume (Kotz. Appl.A. [46] S. 115. Oper. [51] S. Prob 32 . Schock Petersen (1989) Ruin probabilities expressed in terms of storage processes. 20. [47] S. Nielsen (1995) Ruin probabilities via local adjustment coefficients . (to appear).A. Schmidt (1993) The ascending ladder height distribution for a class of dependent random walks. Asmussen & M. Olsson (1996). Taksar (2000) Optimal risk control and dividend distribution policies . Sc. Appl. 64. [48] S. Inf. 30. Appl. O'Cinneide (2000/01) On the tail of the waiting time in a Markovmodulated M/G/1 queue. Probab. Asmussen & K. 421458. Methods 8 Problems (J. Taksar (1997) Controlled diffusion models for optimal dividend payout. Sigman (1996) Monotone stochastic recursions and their duals. Rolski (1994) Risk theory in a periodic environment: Lundberg's inequality and the CramerLundberg approximation.BIBLIOGRAPHY 365 [35] S. Proc. Scand. Insurance: Mathematics and Economics 10. [38] S. O. 58. Res. 105119. [37] S. Proc. Johnson. 259274. 10. Asmussen & T. Dshalalow ed. Probab. Th. Insurance: Mathematics and Economics 20. Rubinstein (1995) Steadystate rare events simulation in queueing models and its complexity properties. 23. [49] S. first passage problems and extreme value theory for queues. 365372. Asmussen & C. [44] S. Asmussen & H. 120. CRC Press. Nerman & M. Asmussen & R. Rolski (1991) Computational methods in risk theory: a matrixalgorithmic approach. 31. 422447. H. 19. Florida.Y. Asmussen & R.). Asmussen.
Act. J. Scand. 9. [68] T. Probab. Cambridge. 1975. Insurance: Mathematics and Economics 6. Boogaert & A. Boogaert & V. Trans. J. Grandell (1988) Exponential inequalities for ruin probabilities in the Cox case. J.R.J. 610. 705766. Actuaries 21. [64] P. Ney ( 1972) Branching Processes . [65] N. C. 33. 77111. Scand. 148156.L. Bingham (1975) Fluctuation theory in continuous time.A. Appl. 1988 . [70] A. Clarendon Press. New York. 832840. [58] O. [62] J. BarndorffNielsen (1978) Information and Exponential Families in Statistical Theory. Goldie & J. Bobbio & M. 275279.H. 4148.T. [67] T. J. New York. Cambridge University Press. 190204. [63] A. Bjork & J. . Oxford. Scand. 1995. Stochastic Models 10. Adv. Bjork & J. Teugels ( 1996) Convergence rates for M /G/1 queues and ruin problems with heavy tails . Borovkov (1976) Asymptotic Methods in Queueing Theory.M. Act. BarndorffNielsen & H. Cox (1972) A low traffic approximation for queues. Appl. Crijns (1987) Upper bounds on ruin probabilities in case of negative loadings and positive interest rates. 221232. 9599. [56] B. de Waegenaere (1990) Simulation of ruin probabilities. Plemmons (1994) Nonnegative Matrices in the Mathematical Sciences.H. Wiley. [73] A. Insurance: Mathematics and Economics 9. 169186. Act. 1974. Teugels (1987) Regular Variation. Springer . Telek (1994) A bencmark for PH estimation algorithms: results for acyclic PH. Schmidli (1995) Saddlepoint approximations for the probability of ruin in finite time. SpringerVerlag. 661667. Appl. [72] P. Bloomfield & D. 11811190. [61] J. Asmussen & J. Berlin. [60] J. Probab. 129134. J. 7. [66] N. J. Beekman (1969) A ruin function approximation. Scand.366 BIBLIOGRAPHY [53] S. [71] P. Berman & R. Halsted Press. Bingham. New York. Grandell (1985) An insensitivity property of the ruin probability. Act. Soc. [54] K.L. Billingsley (1968) Convergence of Probability Measures. SIAM. Athreya & P. Wiley. [57] C. Insurance: Mathematics and Economics 4.H. Probab. [55] B. 1985 . Baker (1977) The Numerical Solution of Integral Equations. Beekman (1985) A series for infinite time ruin probabilities. [69] P. Scand. Beekman (1974) Two Stochastic Processes.A.B . [59] O. Act. J. von Bahr (1974) Ruin probabilities expressed in terms of ladder height distributions. von Bahr (1975) Asymptotic ruin probabilities when exponential moments do not exist.
J. 127130. 907920. Insurance : Mathematics and Economics 9. verve. [83] D. 3539.). Nesbitt (1986) Actuarial Mathematics . Area Commun. Cramer (1955) Collective risk theory.J. Smith ( 1983) Asymptotic analysis of a queueing system with bursty traffic. Smith ( 1986) An asymptotic analysis of a queueing system with Markovmodulated arrivals . Cox (1955) Use of complex probabilities in the theory of stochastic processes. 749763. Broeck .L. Jr. [88] J. [91] H. Hickman. Wahrscheinlichkeitsth. Croux & N. 24. Stockholm. II.W. Wiley. SpringerVerlag. Fox & L. 105119. J.A.R. 93121. Cambr. [79] P. Aut. Simulation and Estimation. 62. Probab.R. The Jubilee volume of Forsakringsbolaget Skandia. Boxma & J.) NorthHolland. H.Y . Bell. [89] M. [86] K. Adv. Soc. Schrage (1987) A Guide to Simulation. IEEE Trans. 313319. Queueing Systems 33. Stockholm. Burman & D . [82] H. Herzog (1977) The phase concept : approximations of measured data and performance analysis. Sel.L. Malgouyres (1983) Large deviations and rare events in the study of stochastic algorithms . [81] J. . 51. 14331453. [78] L.W. [87] E. Philos. NorthHolland. Nonparametric estimators for the probability of ruin. Biihlmann ( 1970) Mathematical Methods in Risk Theory. Academic Press. Bratley. Insurance : Mathematics and Economics 5. Chung (1974) A Course in Probability Theory (2nd ed. New York San Francisco London. Control AC28. Bux & U. De Vylder (1986) Ordering of risks and ruin probabilities. Goovaerts & F. [90] D. 16. 2338. M. Res.A. [76] N. J. Chandy & M. Berlin.W. Springer. [84] D.I. Illinois. Tweedie (1982) Storage processes with general release rule and additive inputs . IEEE J. [93] K.R. Bowers .. [75] O. Gerber. Skandia Jubilee Volume. Brockwell . [77] P.M. Boxma & J. Cramer (1930) On the Mathematical Theory of Risk. Fort & G. [85] W.U. Z. [80] F.L. B. New York.L. Qinlar (1972) Markov additive processes. 34. Reading. Cohen (1999) Heavytraffic analysis for the GI/G/1 queue with heavytailed service time distributions . Burman & D. Bucklew ( 1990) Large Deviation Techniques in Decision.BIBLIOGRAPHY 367 [74] O. Cohen (1998) The M/G/1 queue with heavytailed service time distribution. Itasca.C. New York. Cohen (1982) The Single Server Queue (2nd ed. Proc. [92] H. Veraverbeke (1990). Syst. 177204. Computer Performance (K. Oper.Y. Amsterdam. Geb. AddisonWesley. Breiman ( 1968) Probability. 392433. S. D. Jones & C. Appl. Cottrell. 14. Resnick & R.).J. Reiser eds. Amsterdam.C. The Society of Actuaries . Tech.
N. Rubin (1977) Maximum likelihood from incomplete data via the EM algorithm. Zeitouni (1992) Large Deviations Techniques and Applications.. Res. Insurance: Mathematics and Economics 6. [106] A. Goovaerts (1988) Recursive calculation of finitetime ruin probabilities.D. Martingales and insurance risk. Delbaen & J.B. 27. 624628. Insurance: Mathematics and Economics 7. 85116. Pentikainen & E. Stochastic Models 5. [102] P. Scand.L. Rolski (1991) Light traffic approximations in queues. 3750. 3. Res. 181217. Probab. Davidson (1946) On the ruin problem of collective risk theory under the assumption of a variable safety loading (in Swedish). Oper. Scand. 201206. [111] F. 5771. [98] A. Stockholm. Forsiikringsmatematiska Studier Tillagnade Filip Lundberg. Aktuar. De Vylder (1996) Advanced Risk Theory. Laird & D. Insurance: Mathematics and Economics 10. Tidskr. 114119. 16. 227229. English version published in Skand. 7083 (1969). van Dawen (1986) Ein einfacher Beweis fur Ohlins Lemma. [112] F. Act.368 BIBLIOGRAPHY [94] M. De Vylder & M. Goovaerts (1984) Bounds for classical ruin probabilities. [103] F. New York. De Vylder (1978) A practical solution to the problem of ultimate ruin probability. [99] R. Chapman & Hall. 17. Dempster. T. Rolski (1984) A light traffic approximation for a single server queue. Haezendonck (1985) Inversed martingales in risk theory. Insurance: Mathematics and Economics 3. Roy. [100] A. Blotter der deutschen Gesellschaft fur Versicherungsmathematik XVII. Math. 1990 . Pesonen (1994) Practical Risk Theory for Actuaries. Insurance: Mathematics and Economics 4. Steinebach (1990) On some alternative estimators of the adjustment coefficient in risk theory. Teugels ( 1990) Empirical Laplace transform and approximation of compound distributions .P. Math. Appl. 88101. Comp. [109] F. Daley & T. Springer. Csorgo & J. Haezendonck (1987) Classical risk theory in an economic environment. J. Editions de 1'Universite de Bruxelles. J. [104] F. Deheuvels & J. Appl. [107] L. Suppl.M. [108] F. Dembo & O. Steinebach ( 1991 ) On the estimation of the adjustment coefficient in risk theory via intermediate order statistics . 9 . 22. Math.. Actuarial J. Devroye (1986) Nonuniform Random Variate Generation. Delbaen & J. Dassios & P. J. Daykin. Oper. [95] M. [101] C. J. Csorgo & J. [97] D. Soc. 135159. [96] D. . Embrechts (1989). 121131. De Vylder & M. [110] F. Statist. 435436. Daley & T. [105] A. 583602. Jones and Bartlett. Boston. De Vylder (1977) A new proof of a known result in risk theory.
118. 1993 . 177192. Scand. Gerber (1988) The surpluses immediately before and at ruin. Scand.S. H. 193199. Scand. . British Actuarial J.M. 147167. Dufresne & H.BIBLIOGRAPHY 369 [113] D. 1741. Dickson (1995) A review of Panjer's recursion formula and its applications.C. 131138. with applications. Soc.M. [116] D. Scand. 191207. Astin Bulletin 21. Insurance: Mathematics and Economics 25.W. Dickson & J. 1993 . Harrison & A. Dufresne. van Doorn & G. Gray (1984) Approximations to the ruin probability in the presence of an upper absorbing barrier. J. Duffield & N. Philos. Gerber & E. Gerber (1991) Risk theory for the compound Poisson process that is perturbed by a diffusion. [123] E. Act. Dickson & C. Scand.C. Taylor ( 1975) A diffusion approximation for the ruin probability with compounding assets. J. Hipp (1999) Ruin problems for phasetype(2) risk processes. J. Scand. Dickson (1994) An upper bound for the probability of ultimate ruin.R. [122] B. Res.M. Embrechts ( 1988 ). Insurance: Mathematics and Economics 7.R. 107124. Act. Proc. Grandell & H.J. Shiu (1991) Risk theory with the Gamma process. Math.U.M. J. Insurance : Mathematics and Economics 7. [119] D. [126] F. Dynkin (1965) Markov Processes I.J.C. 229232. Act. J. 269274. Act. 4962.G.C. [127] F.M. 105115. 1984 . 1. Berlin Gottingen Heidelberg. 251262.R. 1975 . Emanuel .C.C.K. Insurance : Mathematics and Economics 22. Scand. Letters 7. [115] D. 5159. [118] D. [125] F. [124] N. J.U. Act. Springer . Waters (1999) Ruin probabilities wih compounding. Regterschot (1988) Conditional PASTA. O'Connell (1995) Large deviations and overflow probabilities for the general singleserver queue.M. Ruin estimates for large claims . J. [120] D.M. Oper. Insurance : Mathematics and Economics 11. [129] D . Insurance: Mathematics and Economics 19. [114] D. Gray (1984) Exact solutions for ruin probability in the presence of an upper absorbing barrier. 1984 . [117] D. 363374. 3745. Insurance : Mathematics and Economics 10. 4259.R. [121] D. Embrechts . Dickson & H. Waters (1996 ) Reinsurance and ruin . J. Act. 6180. Djehiche (1993) A large deviation estimate for ruin probabilities. Dickson (1992) On the distribution of the surplus prior to ruin . Act.M. Dickson & C.M. 1994 . 174186.M. Camb. 2000 . Dufresne & H. [128] E.U. J.C.B.C. [130] P. [131] P. and the amount of claim causing ruin. Hipp (1998) Ruin probabilities for Erlang (2) risk processes. Dickson & J.C. Dickson & H.C. Schmidli (1993) Finitetime Lundberg inequalities in the Cox case .
Probab. [148] C. Tidsskr. Appl. Probab. Appl. 8190. Veraverbeke (1982) Estimates for the probability of ruin with special emphasis on the possibility of large claims . Walkup (1967) Association of random variables. Adv. Probab. Statistica Neerlandica 47. Fields 75. [139] A. Mikosch (1997) Modelling Extremal Events for Finance and Insurance.D. Insurance : Mathematics and Economics 1. 14661474. Konig. with applications. Fuh (1997) Corrected ruin probabilities for ruin probabilities in Markov random walks. [142] W. Furrer (1998) Risk processes perturbed by astable Levy motion. Erlang 2. Jensen. Sci. [143] W. Villasenor (1988) Ruin estimates for large claims. 5974.D.J. Insurance: Mathematics and Economics 10. Insurance: Mathematics and Economics 7. J. Frenz & V. 26. 3339. Teugels (1985). A bootstrap procedure for estimating the adjustment coefficient. [136] P. Embrechts & N. C. .K. first passage times and moments of ladder variables in Markov random walks.M. Appl. Embrechts. 29. New York. [145] P. Embrechts & J.W. [146] E. M. 35. New York. [141] F. R. Act. 131137. Appl. 404422.D. 623637.W. Springer. Feller (1971) An Introduction to Probability Theory and its Applications II (2nd ed. Th. Trans. On the excursions of Markov processes in classical duality. Wiley. 175195. F. Embrechts & H. Probab. Schmidli (1994) Ruin estimation for a general insurance risk model.K. 5572. Appl. [138] P. [149] C. Maejima & J. 616624. Schmidt (1982) Queues and Point Processes. 5975. [134] P. Fitzsimmons (1987). Frees (1986) Nonparametric estimation of the probability of ruin. Heidelberg. Pitts (1993) Some applications of the fast Fourier transform in insurance mathematics . Arndt & V. [144] P. Proschan & D. J. Math. Adv. [140] J. Probab. 38. Fuh & T.). 181190. Franken. Esary. [137] P. Reprinted 1948 as 'The theory of probabilities and telephone conversations' in The Life and work of A. Griibel & S. Danish Academy Tech.).A. J. Esscher (1932) On the probability function in the collective theory of risk. 1998 . Approximations for compound Poisson and Polya processes. [133] P. Schmidt (1992) An insensitivity property of ladder height distributions. Embrechts & T. Scand. Wiley. Ann.370 BIBLIOGRAPHY [132] P. [135] P. Wiley.L. Erlang (1909) Sandsynlighedsregning og telefonsamtaler .L. 566580. Kliippelberg & T. Skand. 695712. Rel. [150] H. 159178. Mikosch (1991). Nyt Tidsskrift for Matematik B20. Lai (1998) Wald's equations. 269274. Statist. D. Astin Bulletin 16. Embrechts. Feller (1966) An Introduction to Probability Theory and its Applications I (3nd ed. Embrechts. Adv. Probab. 17. Akt. U. J. 29. [147] M.
3752. Galambos & J. Appl. Griibel (1996) Perpetuities with thin tails. 1523. Life Insurance Mathematics. Scand. Furrer (1996) A note on the convergence of the infinitetime ruin probabilities when the weak approximation is astable Levy motion. 6370. A. Unpublished.M. 131156. Vers. 1977. [159] H. Gerber (1973) Martingales in risk theory.U. Gerber (1979) An Introduction to Mathematical Risk Theory. F. Probab. 105115. 97114. Ann. Gerber (1970) An extension of the renewal equation and its application in the collective theory of risk. Schmidli (1994) Exponential inequalities for ruin probabilities of risk processes perturbed by diffusion. Kovalenko (1989) Introduction to Queueing Theory (2nd ed. S.U. Whitt (1994) Logarithmic asymptotics for steadystate tail probabilities in a singleserver queue. Goovaerts. Math. Act.J. NorthHolland. 205216.G.E.). [163] P.. Amsterdam. Amsterdam. Haezendonck (1984) Insurance Premiums. Furrer.U. Insurance: Mathematics and Economics 20. 5.). Vers. J. Skand. Act. Basel.U. Goovaerts. Huebner Foundation Monographs. Z. [168] J. [158] H. [162] P.71. contained in the authors PhD thesis. Grandell (1978) A remark on 'A class of approximations of ruin probabilities'. Scand.U. Bauwelinckx (1990) Effective Actuarial Methods. Dordrecht. Insurance: Mathematics and Economics 15. ETH Zurich. R. 424445. Gnedenko & I.V. Glasserman (1991) Gradient Estimation via Perturbation Analysis. [155] H. Michna & A. Appl.S.BIBLIOGRAPHY 371 [151] H. Glynn & W. Gerber (1986). Appl. London. 28. de Vylder & J. Kluwer.U. . [166] M. van Heerwarden & T. 205210. [153] H. Grandell (1977) A class of approximations of ruin probabilities. Insurance : Mathematics and Economics 7. 1978 . [161] P. Ver. J. Weron (1997) Stable Levy motion approximation in collective risk theory.U. [169] J. [152] H.. J. J. Goldie & R. Birkhauser . 31A. [165] M. [167] C. Ver. Mitt. Gerber (1981) On the probability of ruin in the presence of a linear dividend barrier. Probab. [160] H. 73.N. 463480. [157] H. Adv. [154] H. In Studies in Applied Probability (J. [164] B. Boston. Scand. Schweiz. Gani. Schweiz. Kaas.W. NorthHolland. Furrer & H. Probab. Kou (1995) Limits of first passage times to rare sets in regenerative processes. Mitt. Gerber (1988) Mathematical fun with ruin theory.J. Math. Springer. University of Pennsylvania. eds. 7778. 2336. [156] H. Aktuarietidskrift 1970 . Suppl. Act. Gerber (1971) Der Einfluss von Zins auf die Ruinwahrscheinlichkeit. 1981 . Glasserman & S.
J. Informatica 2. SpringerVerlag. Astin Bulletin 29. Theory and Applications. Berlin. 400409. 332. E. 566576. Res. Appl. [188] J. 14. 3 . Th. Winter School on Stochastic Analysis and Appl . [173] J. 14. Hermesmeier (1999) Computation of compound distributions I: aliasing errors and exponential tilting. 7590. Probability Theory and Mathematical Statistics (I. Liet. Harrison & A. 5766. Harrison & S. Griibel & R. Ibragimov. Harrison & S. Aktuar Tidskr. 6779. Gyllenberg & D. Korolyuk (1969) On the joint distribution of a process with stationary increments and its maximum . Archiv v. KTH. [183] M.243255. [177] B. Gut (1988 ) Stopped Random Walks . Res. Appl. [171] J. Grigelionis ( 1993) Twosided Lundberg inequalities in a Markovian environment. Stoch. Segerdahl (1971) A comparison of some approximations of ruin probabilities. Appl. Matem. Resnick (1976 ) The stationary distribution and first exit probabilities of a storage process with general release rule . Opns . AkademieVerlag. Applied Statistics 40. 143158. Probab. M. Grandell (1997) Mixed Poisson Processes. 1.). Gusak & V. SpringerVerlag.V. Lemoine (1977) Limit theorems for periodic queues.O. [185] J. M. Opns . Grandell (1979) Empirical bounds for ruin probabilities. Stoch. Appl. 355365. [184] H . [179] R. 347358. Statistical Algorithm 265. Harrison (1977) Ruin problems with compounding assets . [175] J. Rink. Chapman & Hall.. [189] A. 167176. van Heerwarden (1991) Ordering of Risks: Theory and Actuarial Applications. ed. mathematische Wirtschaft . [181] D. Tinbergen Institute Research Series 20. . Silvestrov (2000) Cram€rLundberg approximations for nonlinearly perturbed risk processes . Math. Grandell (1999) Simple approximations of ruin functions. Insurance: Mathematics and Economics 26.A.I.131135. 197214. 8. Grandell & C. 33. [182] A. Grandell ( 1992) Finite time ruin probabilities and martingales . Proc. Proc . Amsterdam. Proc. New York. Math. Resnick (1977) The recurrence classification of risk and storage processes .and Sozialforschung 6. Skand. Grnbel (1991) G/G/1 via FFT. [174] J. Hadwiger (1940) Uber die Wahrscheinlichkeit des Ruins bei einer grossen Zahl von Geschiiften . M. [176] B. Grandell (1990) Aspects of Risk Theory. Grigelionis ( 1992) On Lundberg inequalities in a Markovian environment. 3041. [172] J.I.J.S. Preprint.372 BIBLIOGRAPHY [170] J. M. Probab. [178] B. [186] J. [187] J. [180] R. 54. 5 . Grigelionis (1996) Lundbergtype stochastic processes.
Prob. S. Act. [198] C. Willekens (1986) Estimates for the probability of ruin starting with a large initial reserve. [196] C. [199] R. Huskova. 8095. Klugman (1984) Loss Distributions. 123151. 6. Appl. Heidelberger (1995) Fast simulation of rare events in queueing and reliability models. Versicherungswirtschaft. eds. ACM TOMACS 6. Astin Bulletin 19. 29. Insurance: Mathematics and Economics 5. [200] M. Michel (1990) Risikotheorie: Stochastische Modelle and Statistische Methoden.L. Hoglund (1974) Central limit theorems and statistical inference for Markov chains. Nachrichten 30. [192] U. Hipp (1989b) Estimators and bootstrap confidence intervals for ruin probabilities. Proc.BIBLIOGRAPHY 373 [190] P. 1990 . 3. J. Hipp (1989a) Efficient estimators for ruin probabilities. Chalmers University of Technology and the University of Goteborg. Math. Haggstrom. 305313. [208] V. Iglehart (1969) Diffusion approximations in collective risk theory. 378389. Herkenrath (1986) On the estimation of the adjustment coefficient in risk theory by means of stochastic approximation procedures. The ruin problem for finite Markov chains. [195] B. New York. Hoglund (1990) An asymptotic expression for the probability of ruin within finite time. Probab.). Mathematical Statistics. Taksar (2000) Optimal proportional reinsurance policies for diffusion models. [207] D. [201] L. Hipp & R. Geb. 18.R. Hoglund (1991). Electronic Letters 35. Probab. 23. Karlsruhe.A. Staalhagen (1999) Waiting time distributions in M/D/1 queueing systems. J. 25. [202] 0.. 285293. Hesselager (1990) Some results on optimal reinsurance in terms of the adjustment coefficient. [191] W.B. 12981310. Scand.a program for fitting phasetype distributions. Ann. 11631174. Verlag Versicherungswirtschaft e.M. Nerman (1992) EMPHT . 19. Hogan (1986) Comment on corrected diffusion approximations in certain random walk problems. Statist. Asmussen & 0. on Asymptotic Statistics (P. J. Hermann (1965) Bin Approximationssatz fiir Verteilungen stationarer zufalliger Punktfolgen. Insurance: Mathematics and Economics 5. 5770. Ann. Horvath & E.V. Hojgaard & M. Act. Karlsruhe. 377381. Mandl & M. Heilmann (1987) Grundbegriffe der Risikotheorie. verve. Hogg & S. Probab. 4385. Studies in Statistical Quality Control and Reliability 1992 :4. Iversen & L. Wahrscheinlichkeitsth. [193] U. Scand. J. 166180. 8996. Wiley. . [197] C. Ann. [194] 0.L. [205] T. [204] T. 285292. Z. 1998 . No. [206] B. Foruth Prague Symp. Hill (1975) A simple general approach to inference about the tail of a distribution. 259268.V. [203] T. Appl.
P. ibid. Proc. N. Act. Proc. J. Copenhagen. 116129. 61.G. [220] V. Technometrics 30. University of Chicago Press. Goovaerts (1986) General bound on ruin probabilities. 187193. [219] V. [224] J. 63. Academic Press. Taaffe (1989/90) Matching moments to phase distributions. 239256.M. 37. Jagerman (1985) Certain Volterra integral equations arising in queueing. Soc.H. [228] J. Appl. 1996. 325247.B.J. Probab. Keilson (1966) A limit theorem for passage times in ergodic regenerative processes. 123133. Keilson & D. 4151. [227] J.M. J. [222] S.A. J. Kemperman (1961) The Passage Problem for a Markov Chain. Clarendon Press. Jensen (1953) A Distribution Model Applicable to Economics.G. Lazar (1998) Subexponential asymptotics of a Markovmodulated random walk with a queueing application. 6. Proc. Reinhard (1985) Probabilities de ruine pour une classe de modeles de risque semiMarkoviens. [217] R. Oxford. [225] J. Kalashnikov (1996) Twosided bounds of ruin probabilities. [214] A. 866870. Wishart (1964). Scand. Jelenkovic & A. Kluwer. [223] J. [211] J. 60.C. Soc. Kaas & M. 561563. Karlin & H. [216] M. Manuscript. 11. Ann. Act. [221] E. Astin Bull. Princeton. Jensen (1995) Saddle Point Approximations. [213] P. Kennedy (1994) Understanding the WienerHopf factorization for the simple random walk. Insurance: Mathematics and Economics 5. Probab. Taylor (1981) A Second Course in Stochastic Processes. Kalashnikov (1999) Bounds for ruin probabilities in the presence of large claims and their comparison. Munksgaard.M. Wishart (1964) A central limit theorem for processes defined on a finite Markov chain. Cambridge Philos. Statist. Wishart (1964) Boundary problems for additive processes defined on a finite Markov chain. Soc. J. [210] D. 15. Stochastic Models 1.J. [218] V. ibid.M. Keilson & D. 31. 165167. Johnson & M. 283306. Keilson & D. NEC Research Institute. Math. Jagerman (1991) Analytical and numerical solution of Volterra integral equations with applications to queues.L. 35.M. Kalashnikov (1997) Geometric Sums: Bounds for Rare Event with Applications. 6. Cambridge Philos. [212] J. Chicago. Astin Bull. 8793. N. 711743. Appl. New York. Addenda to for processes defined on a finite Markov chain. 259281.G. 118. Janssen (1980) Some transient results on the M/SM/1 special semiMarkov model in risk and queueing theories. [215] J. Stochastic Models 5.374 BIBLIOGRAPHY [209] D. 173190. Cambridge Philos. 547567. 3. [226] J. . Kao (1988) Computing the phasetype renewal and related functions.R. Janssen & J. Amer.
132141. Klnppelberg (1989) Estimation of ruin probabilities by means of hazard rates. 4477. J. 60 . Lindley (1952) The theory of queues with a single server. 25. 390397. Ramaswami (1999) Introduction to MatrixAnalytic Methods in Stochastic Modelling. Kliippelberg & U.J. Korolyuk.F.F. Probab. Lucantoni (1991) New results on the single server queue with a batch Markovian arrival process. Insurance : Mathematics and Economics 12. Macmillan. J. Appl. SpringerVerlag. 359361. [230] J. Camb. Scand. Kingman (1961) A convexity property of positive matrices . 1995 . Insurance: Mathematics and Economics 8. [241] G.a Linear Algebraic Approach. Act. Nyrhinen (1992b) On asymptotically efficient simulation of ruin probabilities in a Markovian environment.C. 146. Lehtonen & H. Penev & A.C. Cambr. . [237] C. J. 277289. [248] D. Klnppelberg & T. Soc. Stochastic Models 7. SpringerVerlag. B24. Appl. Stadtmiiller (1998) Ruin probabilities in the presence of heavytails and interest rates . Sorensen (1997) Exponential Families of Stochastic Processes. Philos. Scand. Soc. Adv. Klnppelberg (1993) Asymptotic ordering of risks and ruin probabilities. Seminaire de Probabilties X . Turbin (1973) An asymptotic expansion for the absorption time of a Markov chain distribution. (235] C. Kunita (1976) Absolute continuity of Markov processes. Probab. Kingman (1962) On queues in heavy traffic. Bernoulli 1 . Oxford 12. Lecture Notes in Mathematics 511. [245] T. Probab. I. Proc. Actuarial J. [246] D. [238] V. Prob.J. Nyrhinen (1992a) Simulating levelcrossing probabilities by importance sampling. J. Lemoine (1989) Waiting time and workload in queues with periodic Poisson input. Mikosch (1995) Modelling delay in claim settlement. 4958. 858874. 283284.F. J. 259264. 18. [231] J. SIAM. New York. Soc. Proc. Act. Math. Cybernetica 4. Statist. J. 889900. Kingman (1964) A martingale inequality in the theory of queues.S.C. [234] C. 133135 (in Russian). [247] L. [236] C.F. Appl. [233] C. [243] A. 24.BIBLIOGRAPHY 375 [229] J. Quart. Klnppelberg & T. 383392. 279285. [232] C. [239] H. Lipsky (1992) Queueing Theory .. Roy. Kuchler & M. Kliippelberg (1988) Subexponential distributions and integrated tails.P. Quart. [242] A. Philos. 125147. 1998 . Latouche & V. [240] U. Mikosch (1995) Explosive Poisson shot noise with applications to risk retention. J. 154168. 6075. 48. 26. Scand. Lehtonen & H. Lemoine (1981) On queues with periodic Poisson input. [244] T. Appl.
Moustakides ( 1999) Extension of Wald 's first lemma to Markov processes. MeierHellstern & M . 36. Res. Anal.D. Soc. Meier ( 1984). [257] A. [264] G. with an application .K. J. Miller ( 1962 ) Absorption probabilities for sums of random variables defined on a finite Markov chain . Appl. 370377. Probab. 11 . Adv. Probab. 15.376 BIBLIOGRAPHY [249] D. Scand. 268285. [251] F. Nummelin (1987) Markov additive processes I. . [253] V. [254] V . 2939. [256] A. Act. 32. [259] Z . Proc. Act. [260] H. 58 .V. Stockholm. Cambridge Philos. Lundberg (1903) I Approximerad Framstdllning av Sannolikhetsfunktionen. Europ. [261] H. 29 . Opns. Mammitzsch ( 1986 ) A note on the adjustment coefficient in ruin theory. 4859. Englunds Boktryckeri AB. 429448. Neuts ( 1990) A single server queue with server vacations and a class of nonrenewal arrival processes . Ann. Lucantoni . J. Ney & E. Lundberg (1926) Forsdkringsteknisk Riskutjdmning. Malinovskii ( 1994) Corrected normal approximation for the probability of ruin within finite time. [263] M . K. Proc. II Aterforsdkring av Kollektivrisker . Nagaev (1957) Some limit theorems for stationary Markov chains. 286298. Math. Uppsala. 1261 (0?)1270. Ann. 561592. Appl.V. Th. Scand. . 834841. [255] V. Insurance : Mathematics and Economics 5. 31. Makowski ( 1994) On an elementary characterization of the increasing convex order . Statist. J. 223235. Act.K. J.S. 147149. Almqvist & Wiksell. Stoch. Appl. J. 3. [262] H . Miyazawa & V. Eigenvalue properties and limit theorems . Soc. 378406. MartinL6f (1983) Entropy estimates for ruin probabilities . Appl. 161174. 763776. 1994.D. A fitting algorithm for Markovmodulated Poisson processes having two arrival rates . 58. Appl. 124147. 676705. Hoist eds. J. MartinL3f (1986) Entropy.). [258] K . Michna ( 1998) Selfsimilar processes in collective risk theory. Malinovskii ( 1996) Approximation and upper bounds on probabilities of large deviations of ruin within finite time. Probab.F. Math. [265] S. Miller ( 1961 ) A convexity property in the theory of random variables defined on a finite Markov chain. Ann.D. Probab. Probab . 1996 . [266] P.M. Gut & J. Probab. Probability and Mathematical Statistics (A. J. Schmidt ( 1993) On ladder height distributions of general risk processes .S. [250] F. 22. [252] A. Miller (1962 ) A matrix factorization problem in the theory of random variables defined on a finite Markov chain . F. 2. Appl. Cambridge Philos. Scand. a useful concept in risk theory. 1986 .
F. Appl. 135148. [280] A. Insurance: Mathematics and Economics 20.K. 733746. [275] H. [277] J. Nyrhinen (1999) Large deviations for the time of ruin. Appl. 21. Probab. Baltimore . Stoch. London. Pakes (1975) On the tail of waiting time distributions. Neuts ( 1989) Structured Stochastic Matrices of the M/G/1 Type and their Applications. IEICE Trans. Probab. Hamburg. Stoch. 30. J. [273] R. 46. Stoch. 123144. Insurance: Mathematics and Economics 22. Paulsen & H. Neveu ( 1961 ) Une generalisation des processus a accroisances independantes. [281] J. Paulsen (1998) Ruin theory with compounding assets . Willekens (1987) Second order behaviour of distributions subordinate to a distribution with finite mean . Abh. Astin Bulletin 5 . Omey & E. Norberg ( 1990) Risk theory and its statistics environment . 157. [283] J. Proc. Appl. 16. Proc. Naval Research Logistics Quarterly 25. Adv. Gjessing (1997a) Optimal choice of dividend barriers for a risk process with stochastic return on investments. [282] J . Appl. O'Cinneide (1990) Characterization of phasetype distributions. [271] M. Nyrhinen ( 1998 ) Rough descriptions of ruin for a general class of surplus processes. [269] M .K. Probab.F. Appl. 12. 327361. Neuts ( 1978) Renewal processes of phase type. Stoch. Stoch. Marcel Dekker. 75. 215223. Johns Hopkins University Press. 311342. E75B . Statistics 21. Commun.G. [274] H . . 36.F. 273299. 10081026. 249266. 339353. Paulsen (1993) Risk theory in a stochastic economic environment. [276] C. Stochastic Models 3 . 445454. Neuts (1992) Models based on the Markovian arrival process. [279] E. Models 6.A. [278] E. [270] M . 71. Neuts ( 1981 ) MatrixGeometric Solutions in Stochastic Models. [272] J . J. Appl. Paulsen ( 1998) Sharp conditions for certain ruin in a risk process with stochastic return on investments. Ohlin (1969) On a class of measures for dispersion with application to optimal insurance. Neuts (1977) A versatile Markovian point process. Probab. Willekens (1986) Second order behaviour of the tail of a subordinated probability distributions. [285] J. Math. [284] J. 12551265. Proc. Gjessing (1997b) Present value distributions with applications to ruin theory and stochastic equations. Appl. 764779. J. [268] M .BIBLIOGRAPHY 377 [267] M . Omey & E.a survey. New York. 555564. Proc. Paulsen & H.F. 316. Appl.F. Sem.
Austr. [295] N. [305] C. Math. Prabhu & Zhu (1989) Markovmodulated queueing systems. Wiley. 29A. Berlin. 235243. [299] C. Schock Petersen (1989) Calculation of ruin probabilities when the premium depends on the current reserve. Prabhu (1965) Queues and Inventories. Probab. Astin Bulletin XIV. Adv. 46. [293] N. Res. Appl. Gjessing (1997c) Ruin theory with stochastic return on investments.H. New York. [300] C. [304] B.R. Math. Wiley. J. Appl. [291] S. Statist. [297] R. [290] E. Rao (1965 ) Linear Statistical Inference and Its Applications. Appl. 757764. 222261. Ann. 215246. Prabhu (1980) Stochastic Storage Processes. Oper. 691707.A. 965985. 12.378 BIBLIOGRAPHY [286] J. Math. New York. and Dams. [292] S. Resnick & G.G. Pellerey (1995) On the preservation of some orderings of risks under convolution.L. Res. Act. J. 29. 390413. Reinhard (1984) On a class of semiMarkov risk models obtained as classical risk models in a markovian environment.M. Ramaswami (1980) The N/G/1 queue and its detailed analysis.J. Griibel & P. On the ruin problem of collective risk theory. 537555. Insurance Risk. [288] S. Probab.J. Insurance: Mathematics and Economics 3. Regterschot & J.U. Paulsen & H.M. 28. 32. [302] J. . 2343. R. Probab. [298] V. [306] T. de Smit (1986) The queue M/G/1 with Markovmodulated arrivals and services. Tidskr. 19. Philipson (1968) A review of the collective theory of risk. Ann. Ann.M. Ripley (1987) Stochastic Simulation . 61. Probab. 117133. Pyke (1959) The supremum and infimum of the Poisson process. Embrechts (1996) Confidence bounds for the adjustment coefficient. J.K. [296] N. Prabhu (1961).U. Aktuar.U. 465483. Rolski (1987) Approximation of periodic queues. 820827. Ramsay (1984) The asymptotic ruin problem when the healthy and sick periods form an alternating renewal process. Soc. 147159. Springer.G. Stat.K. [294] N. Pitts (1994) Nonparametric estimation of compound distributions with applications in insurance. 4. Heidelberg. 11. Appl. Insurance: Mathematics and Economics 16. Opns. Appl. 139143. Pitts. Skand. 1989 . Statist. Samorodnitsky (1997) Performance degradation in a single server exponential queueing model with long range dependence.U. 337347. [289] C. [303] S.. Pitman (1980) Subexponential distribution functions. Queues. Adv. Math. Wiley. 4568. Probab . Inst. [301] G. Ann. Math. 2330. 30. Queueing Systems 5. 568576. [287] F. Scand. Rogers (1994) Fluid models in queueing theory and WienerHopf factorisation of Markov chains. Adv.
Stoch. Wiley. Math. [319] H. 795829. Insurance: Mathematics and Economics 16.J. 431447 [316] S. 121133. 4857. Act. 155188. Statist.L. [309] H. [314] T. Rubinstein & A. 417421. Schmidli ( 2001 ) Optimal proportional reinsurance policies in a dynamic setting . Adv. [325] H . Comp. [310] R. [311] R. [322] H. Singapore. Ann. 227244. 7. Science Culture Technology Publishing . Wiley. [324] H . V. Appl. Proc. Schmidli ( 1997a) Estimation of the Lundberg coefficient for a Markov modulated risk model . Schmidli. Teugels (1999) Stochastic Processes for Insurance and Finance. Schmickler ( 1992 ) MEDA : mixed Erlang distributions as phasetype representations of empirical distribution functions . 131156. Rubinstein & B. Ross (1974) Bounds on the delay distribution in GI/G/1 queues. 2001 . Astin Bulletin 29. 5. Probab . Schmidli ( 1997b) An extension to the renewal theorem and an application in risk theory. Rudemo (1973) Point processes generated by transitions of a Markov chain. 5. Act. 145165. [312] R. Melamed (1998) Modern Simulation and Modelling. Schmidli ( 1995 ) CramerLundberg approximations for ruin probabilities of risk processes perturbed by a diffusion . [326] H. [320] H. Lecture Notes of the 8th Summer School on Probability and Mathematical Statistics ( Varna). Ryden (1994) Parameter estimation for Markov modulated Poisson processes. Shapiro (1993) Discrete Event Systems: Sensitivity Analysis and Stochastic Optimization via the Score Function Method. [321] H. Schmidli ( 1999b) Perturbed risk processes: a review . 93104. Schmidt & J. J. Insurance: Mathematics and Economics 22.BIBLIOGRAPHY 379 [307] T. 1997.M. Scand. H.Y.Y. [308] S. [317] L. 4068. Appi. . Statist. [313] M. Stochastic Models 10. 5. Schmidli ( 1994) Diffusion approximations for a risk process with the possibility of borrowing and interest. 365388. 262286. Th. 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. [318] H. Probab. Rubinstein (1981) Simulation and the Monte Carlo Method. Schlegel ( 1998) Ruin probabilities in perturbed risk models. Operationsforsch. Probab. Wiley. Stochastic Models 6. 21. Schmidli ( 1996) Martingales and insurance risk. Wiley. Wiley. 687699.Y. [315] T. Stochastic Models 10 . Scand. Schmidli ( 1999a) On the distribution of the surplus prior and at ruin. Ryden (1996) An EM algorithm for estimation in Markovmodulated Poisson processes. 11. 135149. Appl. [323] H. Seal ( 1969) The Stochastic Theory of a Risk Business. Data Anal. J. J. Rolski.
Chapman & Hall. 1987.380 BIBLIOGRAPHY [327] H. Risk teory and the single server queue . Sigman (1992) Light traffic for workload in queues. Mitteil. 159180. Aktuar Tidsskr. 72.L. Versich. [328] H. Aktuar Tidsskr. 157166. Sengupta ( 1989) Markov processes whose steadystate distribution is matrixgeometric with an application to the GI/PH/1 queue. [337] M. [333] C: O. [345] D. Wiley. 701719. Segerdahl ( 1942) Uber einige Risikotheoretische Fagestellungen . [343] D. Shtatland ( 1966) On the distribution of the maximum of a process with independent increments. 4. Seal (1978) Survival Probabilities. Versich. [329] H.S.W. Sengupta (1990) The semiMarkov queue: theory and applications. Mitt. Insurance: Mathematics and Economics 8. 4. Scand. [338] E. Seal (1974) The numerical calculation of U(w. Adv. Math. Seal ( 1972). .L. 72. Verein Schweiz. SpringerVerlag. Ann. Almqvist & Wiksell. the probability of nonruin in an interval (0. 171178.L. Shantikumar (1993) Stochastic Orders and Their Applications. Grenander).F. 1974. J. Adv. Queueing Systems 11. New York. Verein Schweiz Versich. Academic Press. Stochastic Models 6. [332] C : O. In Probability and statistics . Segerdahl (1955) When does ruin occur in the collective theory of risk? Skand. 279299. Appl. 1955. Act. Mitt. . Ann. 21. Act. 61. Seber (1984) Multivariate Observations. Siegmund (1976) Importance sampling in the Monte Carlo study of sequential tests.the Harald Cramer volume (ed. [342] D. Th. Verein Schweiz. J. 2236. Seal (1972) Numerical calculcation of the probability of ruin in the Poisson/Exponential case.W. 383413. t).S. Segerdahl (1959) A survey of results in the collective theory of risk. Shaked & J. [339] A. 914924. Siegmund (1976) The equivalence of absorbing and reflecting barrier problems for stochastically monotone Markov processes . Probab. Appl. 191197.G. 77100. Shwartz & A. Siegmund (1975) The time until ruin in collective risk theory. Wiley. [346) D .A. t). [340] E. Weiss (1995) Large Deviations for Performance Analysis. Appl. Chapman and Hall. Stockholm. Probab. Skand. 243249. [336] B.S. L. Shin (1989) Ruin probability by operational calculus. Shin (1987) Convolution of uniform distributions and ruin probability. [347] K. 75. [335] B . Scand. 673684. [348] K . Siegmund ( 1985) Sequential Analysis . [334] C: O. Math. Probab . [341] E. 11. Math. Probab. 429442. 121139. [330] H. Sigman ( 1994 ) Stationary Marked Point Processes: An Intuitive Approach. 4383. 11. Statist. [331] G. pp. [344] D. 483487. Siegmund (1979) Corrected diffusion approximations in certain random walk problems.
[361] G.C. Act.. Soc 49. [352] D.). [363] G.L. [353] E. Act. 65102. Suri ( 1989) Perturbation analysis : the state of the art and research issues explained via the GI/G/1 queue. New York.S. Insurance: Mathematics and Economics 16. [365] J. Proceeedings of the IEEE 77. 1980 . Teugels ( 1995 ) Ruin estimates under interest force . . Scand. Probab. Scand. Actuarial J. J.V. Jewell ( 1981 ) Further results on recursive evaluation of compound distributions . Karlsruhe.and largedeviations probabilities in actuarial risk theory. 2739. NorthHolland. 1976.C. Act. 118. Verlag Versicherungswirtschaft e. Scand.. 725741. Springer. Appl. [366] O. Sundt & W. [362] G. [354] B. Slud & C. 114137.BIBLIOGRAPHY 381 [349] E. Sundt & J. 8199. [367] O. [351] D . Taylor (1979) Probability of ruin under inflationary conditions or under experience ratings . Proc.J. Thorin (1986) Ruin probabilities when the claim amounts are gamma distributed. Stoyan (1983) Comparison Methods for Queues and Other Stochastic Models (D.). [358] R. 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. Teugels (1997) The adjustment coefficient in ruin estimates under interest force . Hoesman ( 1989) Moderate. Wiley. 449461. 21. 1982 . 163175. Scand. 235254. New York. J. Smith ( 1953) Distribution of queueing times. [369] O. John Wiley & Sons. Insurance: Mathematics and Economics 19. [355] B . [350] W. Stanford & K. Scand. J. Adv. [360] G. Insurance: Mathematics and Economics 1. 149162. 5776. Astin Bulletin 12. Taylor (1979) Probability of ruin with variable premium rate. Act.L.C. Sundt & J.L.L. Stroinski ( 1994) Recursive method for computing finitetime ruin probabilities for phasedistributed claim sizes. Act. Astin Bulletin 16 . Cambridge Philos. Straub (1988) NonLife Insurance Mathematics. Ruin probabilities prepared for numerical calculations. [359] L. [368] O. Teugels (1982) Estimation of ruin probabilities . Taylor (1986) Claims reserving in NonLife Insurance.C. 722. Daley ed. J.J. Thorin (1982) Probabilities of ruin. 197208. Takhcs (1967) Combinatorial Methods in the Theory of Stochastic Processes. Sundt ( 1993) An Introduction to NonLife Insurance Mathematics (3rd ed.C. J. Taylor (1976 ) Use of differential and integral inequalities to bound ruin and queueing probabilities . 1974. [364] G. Unpublished manuscript. Astin Bulletin 24.A. 65102. 1978. 8594. [357] B. [356] B . Scand. Thorin (1977).
[378] H. Goovaerts. 1942. Oper. Wolff (1990) Stochastic Modeling and the Theory of Queues. University of Michigan. thesis. [379] M. [375] N. Wilson (1995) Selfsimilarity in highspeed traffic: analysis and modeling of ethernet traffic measurements. Submitted. 148164. Unpublished Ph. Willmot (1994) Refinements and distributional generalizations of Lundberg's inequality. 37. [383] G. Waters (1983) Probability of ruin for a risk process with claims cost inflation. 6785. Sherman & D. Statistical Science 10. Goovaerts (1983) The influence of reinsurance limits on infinite time ruin probabilities. M. J. VazquezAbad (1999) RPA pathwise derivatives of ruin probabilities. Thorisson (2000) Coupling.W. . Van Wouve. Lin (1994) Lundberg bounds on the tails of compound distributions. Appl. Astin Bulletin VII.382 BIBLIOGRAPHY [370] 0. Wikstad (1977) Numerical evaluation of ruin probabilities for a finite period. [376] A. Stationarity and Regeneration. Act. Res. Astin Bulletin IX. Thorin & N. Whitt (1989) An interpolation approximation for the mean workload in a GI/G/1 queue. D. [374] F. Skand. PrenticeHall. Wiley. [372] H. Tidskr. [384] R. [382] G. Dordrecht Boston Lancaster. F. [380] W. Insurance : Mathematics and Economics 13. J. [377] V. [371] 0. 1983 . Wikstad (1977) Calculation of ruin probabilities when the claim distribution is lognormal. R. 231246. [381] W. 936952.). Reidel. Tacklind (1942) Sur le risque dans les jeux inequitables.E. Taqqu. 743756. SpringerVerlag. 31. Wallace (1969) The solution of quasi birth and death processes arising from multiple access computer systems. Insurance: Mathematics and Economics 15. Willinger.E. Haezendonck eds. Wald (1947) Sequential Analysis. M. Scand. 5762. J. De Vylder. 142. 137153. Willmot & X.R. In: Premium Calculation in Insurance (F. De Vylder & M. Probab. Thorin & N. Aktuar. Veraverbeke (1993) Asymptotic estimates for the probability of ruin in a Poisson model with diffusion. [373] S. 4963.
189.185187. 360 excursion 155156.121129. 301 Kronecker product. 318319 Erlang distribution 7. 332333 Volterra 192194.226.308 CramerLundberg model: see compound Poisson model cumulative process 334 dams: see storage process differential equation 16. 135. 37. 3032. 111117. 4851.281.203.217.178184. 239. 15.and sum 221.228229. 7079.269. 17. 122. 79. 205. 180182. 7179.251280 heavy traffic 76. 7879. 8283 hyperexponential distribution 7. 97129.203.4447. 248 WienerHopf 144 interest rate 190. 9496. 138139. 141144. 117128. 341.292293 Edgeworth expansion 113.200201.Index adjustment coefficient 17. 217.307312 compound Poisson model 4. 3436. 110113.135.259261.285292. 201214.150. 283.328330. 323 Coxian distribution 147. 271274.9899. 5796. 207 heavytailed distribution 6. 12 CramerLundberg approximation 1617. 302303 diffusion approximation 17.287292. 3839. 17.359 aggregate claims 103106.242. 245248.299. 40. 201 Brownian motion 3 . 1112. 278 gamma distribution 67. 86. 308.137141. 361 diffusion 3. 117127 corrected 121127 duality 1314. 5.293294. 170173.160167. 119. 162164. 226.182.314316. 91. 9396. 97. 196201 inverse Gaussian distribution 76. 1415. 316323 Bessel function 102.346349 383 . 9293. 2425. 89.86. 39. 218 Cox process 4.6779. 227229. 80 81. 3334.100.318320 change of measure 2630.301 central limit theorem 60 . 2526. 14.272. 7475.249250 integral equation 16 Lindley 143 renewal 64.249. 1819.
238. nonlinear 155. 176185.340350 multiplicative functional 2830.227230.269271. 42. 6162. 4446. 100. 179 NP approximation 318320 Palm distribution 5253. 86 periodicity 12. 203204. 171. 71.128129. 269 PerronFrobenius theory 4142. 145187. 32. 134. 3639. 149. 144.285287 queue 14 .287291 INDEX matrix equation . 9899. 6970. 106108. 108109. 44. 37.161. 227228. 157. 229 M/M/1 101 Markovmodulated 185187 periodic 187 martingale 2426.275278. 162.298299. 35. 230.139141. 295. 113114.302. 25. 5758. 185187 GI/G/1 141144 M/D/1 6667 equation 16. 134135. 35. 16. 3947.218221. 15. 16.349 350 perturbation 172173. see also sensitivity analysis phasetype distribution 8.201. 96.287. 251. 141144. 108 life insurance 5. 80. 41. 213214. 132133. 175 light traffic 8183 Lindley integral equation 143 process 3334. 112113. 138. 203 Markov additive process 12.234240. 52 53. 133. 14. 257.297299. 38. 7179.261264. 3947. 306316 Levy process 3. 176185 nonhomogeneous 60 PollaczeckKhinchine formula 6167.304 process 2830.215250. 178 modulation 12.134135.123.152160. 133. 142 likelihood ratio : see change of measure lognormal distribution 9.148.336339 .350361 Poisson process Markovmodulated 12 periodic 12. 154. 260 Lundberg conjugation 6979 . 38. 39. 267269 Panjer's recursion 320323 Pareto distribution 910.315 inequality 1718.259261.160161. 99.336339 Laplace transform 15.146148. 65.348 terminating 215216.288290. 271274. 304305 random walk 3336.384 ladder heights 4756.161164.161.240244. 44. 137139. 25. 59.180. 234 matrixexponential distribution 240244 matrixexponentials 14.108.174. 245 M/G/1 13. 7576.234. 2730. 261264.178182.339 large deviations 129.
292294. 229234. 12. 177 timereversion 14. 172173. 338 utility 324. 240. 327 . 120 statistics x. 186187 renewal process 131. 326330 Weibull distribution 9. 4950. 222. 261264 reservedependent premiums 14. 256258. 8386. 87. 253. 37. 3032. 7475. 260 reinsurance 8. see also matrixexponential distribution regenerative process 264 268. 174. 1819. 317318 semiMarkov 147.359361 stochastic control x stochastic ordering 18. 160. 162. 131144. 251280 time change 4. 233. 141144. 213.INDEX 385 waiting time 141. 251. 31. 9693. 257. 335336 sensitivity analysis 8693. 5455. 260 WienerHopf theory 144.186. 152. 238 saddlepoint method 115117.336339 workload 13. 223226.262263. 307308.154157. 11. 294296 shotnoise process 314 simulation 19. 60. 251. 333334 regular variation 10. 89. 146.279280 Rouche roots 158. 186187 virtual: see workload rational Laplace transform 8. 189214.314. 244. 332333 model 12. 331336 equation 64. 279280 subexponential distribution 11. 233234. 147. 168172 storage process 13.273274.244250. 107. 280. 191192. 281296 stable process 15. 123.
extensions of the classical compound Poisson model to allow f o r reservedependent premiums. worldscientific. the ^W A l \ i l ' ''' CramerLundberg approximation.com 2779 he 9 "789810ll22293211 .." Short Book Reviews ISBN 9810222939 mi u inn i nun I I I I I I i in u www. P'i yfliother approximations (e.T [Ail i The book is a comprehensive treatment of  I i I \ classical and modern ruin probability theory. 2 A I 11 JjVb l' i  i Yj . Markovmodulation or periodicity. Special features of the book are the emphasis on change of measure techniques. It is a comprehensive treatment of the known results on ruin probabilities. for heavytailed claim size distributions).Vol.. I 1! Ruin Probabilities .g. phasetype distributions as a computational vehicle and the connection to other applied probability areas like queueing theory. y finite horizon ruin probabilities. "This book is a must for anybody working in applied probability. exact solutions. Some i (l I JL I J r of the topics are Lundberg's inequality.Advanced Series on Statistical Science & Applied Probability .
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue reading from where you left off, or restart the preview.