This action might not be possible to undo. Are you sure you want to continue?
https://www.scribd.com/doc/41948462/RuinProbAsmussen
11/06/2011
text
original
Ruin Probabilities
Seren Asmussen
World Scientific
Ruin Probabilities
ADVANCED SERIES ON STATISTICAL SCIENCE & APPLIED PROBABILITY
Editor: Ole E. BarndorffNielsen
Published Vol. 1: Random Walks of Infinitely Many Particles by P. Revesz Vol. 2: Ruin Probabilities by S. Asmussen Vol. 3: Essentials of Stochastic Finance : Facts, Models, Theory by Albert N. Shiryaev Vol. 4: Principles of Statistical Inference from a NeoFisherian Perspective by L. Pace and A. Salvan Vol. 5: Local Stereology by Eva B. Vedel Jensen Vol. 6: Elementary Stochastic Calculus  With Finance in View by T. Mikosch Vol. 7: Stochastic Methods in Hydrology: Rain, Landforms and Floods eds. O. E. Barndorff Nielsen et al. Vol. 8: Statistical Experiments and Decisions : Asymptotic Theory by A. N. Shiryaev and V. G. Spokoiny
Ruin P robabilities
Soren Asmussen
Mathematical Statistics Centre for Mathematical Sciences Lund University
Sweden
World Scientific
Singapore • NewJersey • London • Hong Kong
Published by World Scientific Publishing Co. Pte. Ltd. P O Box 128, Fatter Road , Singapore 912805 USA office: Suite 1B, 1060 Main Street, River Edge, NJ 07661 UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE
Library of Congress CataloginginPublication Data Asmussen, Soren
Ruin probabilities / Soren Asmussen. p. cm.  (Advanced series on statistical science and applied probability ; vol. 2) Includes bibliographical references and index. ISBN 9810222939 (alk. paper) 1. InsuranceMathematics. 2. Risk. I. Tide. II. Advanced series on statistical science & applied probability ; vol. 2. HG8781 .A83 2000 368'.01dc2l 00038176
British Library CataloguinginPublication Data A catalogue record for this book is available from the British Library.
First published 2000 Reprinted 2001
Copyright ® 2000 by World Scientific Publishing Co. Pte. Ltd. All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means, electronic or mechanical, including photocopying, recording or any information storage and retrieval system now known or to be invented, without written permission from the Publisher.
For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy is not required from the publisher.
Printed by Fulsland Offset Printing (S) Pte Ltd, Singapore
Contents
Preface I ix
Introduction 1 1 The risk process . . . . . . . . . . . . . .. . . . .. .. . . . . 1 2 Claim size distributions .. . . . . . . . .. . . . . . . . . . . . 5 3 The arrival process . . . . . . . . . . . . . . . . . . . . . . . . 11 4 A summary of main results and methods . . . . .. . . . . . . 13 5 Conventions . .. . .. .. . . . . . . . . . . . . . . . . . . . . 19
II Some general tools and results 23 1 Martingales . .. . .. .. . . . . . .. . . . . . . . . . . . . . 24 2 Likelihood ratios and change of measure . . .. . . . . . .. . 26 3 Duality with other applied probability models . . .. . . . . . 30 4 Random walks in discrete or continuous time . . . . . . . . . . 33 5 Markov additive processes . . . . . . . .. . . . . . . . . . . . 39 6 The ladder height distribution . . . .. . .. .. . . . . . . . . 47
III The compound Poisson model 57 1 Introduction . . . . . . . . .. .. .. . .. .. . . . . . . 58 . . . . . . . . . . . . . . . 61 3 Special cases of the PollaczeckKhinchine formula . . . . . . . 62 4 Change of measure via exponential families . . . .... . .. . 67 5 Lundberg conjugation . .. . . . . . . . . . . . . . . . . . . . . 69 6 Further topics related to the adjustment coefficient .. . . . . 75 7 Various approximations for the ruin probability . . . . . . . . 79 8 Comparing the risks of different claim size distributions . . . . 83 9 Sensitivity estimates . . . . . . . . . . . . . . . . . . . . . . . 10 Estimation of the adjustment coefficient . . . . . . . . . . . . 86 93 2 The PollaczeckKhinchine formula
v
vi
CONTENTS
IV The probability of ruin within finite time 97 1 Exponential claims . . . . . . . . . . . . . . . . . . . . . . . . 98 2 The ruin probability with no initial reserve . . . . . . . . . . . 103 3 Laplace transforms . . . . . . . . . . . . . . . . . . . . . . . . 108 4 When does ruin occur? . . . . . . . . . . . . . . . . . . . . . . 110 5 Diffusion approximations . . . . . . . . . . . . .. . . .. . . . 117 6 Corrected diffusion approximations . . . . . . . . . . .. . . . 121 7 How does ruin occur ? . . .. . . . . . . . . . . . . . . . . . . . 127 V Renewal arrivals 131 1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . 131 2 Exponential claims. The compound Poisson model with negative claims . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 3 Change of measure via exponential families . . . . . . . . . . . 137 4 The duality with queueing theory .. .. .. . . . .. . . . . . 141 VI Risk theory in a Markovian environment 145 1 Model and examples . . . . . . . . . . . .. . .. . . . . . . . 145 2 The ladder height distribution . . . . . . . . . .. . . . . . . . 152 3 Change of measure via exponential families ........... 160 4 Comparisons with the compound Poisson model ........ 168 5 The Markovian arrival process . . . . . . .. .. . . ... . . . 173 6 Risk theory in a periodic environment .. . . . .. . . . . . . . 176 7 Dual queueing models .... ... ................ 185 VII Premiums depending on the current reserve 189 1 Introduction . . . . . . . . . . . . . . . . . . . .. . . . . . . . 189 2 The model with interest . . . . . .. . . . . . . . . . .. . . . 196 3 The local adjustment coefficient. Logarithmic asymptotics . . 201 VIII Matrixanalytic methods 215 1 Definition and basic properties of phasetype distributions .. 215 2 Renewal theory . . . . . . . . . . . . . . . . . . . . . . . . . . 223 3 The compound Poisson model . . . . . . . . . .. . . . . . . . 227 4 The renewal model . . . . . . . . . . . . . . . .. . . . . . . . 229 5 Markovmodulated input . . .. . . . . . . . . . . . . . . . . . 234 6 Matrixexponential distributions . . . . . . . . . . . .. . . . 240 7 Reservedependent premiums . . . . .. . . . .. . . . . . . . 244
. . . .. . . . .. . . . . . . . . . . 306 4 The distribution of the aggregate claims . . 271 6 Reservedependent premiums . . 344 AS Complements on phasetype distributions . . . . . . . . . . . . 290 5 Regenerative simulation . .. .. . . . . . 281 2 Simulation via the PollaczeckKhinchine formula . . .. . . .CONTENTS vii IX Ruin probabilities in the presence of heavy tails 251 1 Subexponential distributions . . . . . . . . . 292 6 Sensitivity analysis .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . 259 3 The renewal model . . . . . . . . . 287 4 Importance sampling for the finite horizon case . . . . . . . . . . . .. . . . . . . . . . 304 3 Large deviations . . 279 X Simulation methodology 281 1 Generalities . . . 336 A3 Matrixexponentials . . . . . . . . . . .. . . . . . 261 4 Models with dependent input . . . . . . .. . . . . . . . . . . . . . . . . . .. . . .. 316 5 Principles for premium calculation . . . . . . . .. . . . . . . . . 297 2 Further applications of martingales .. . . . . . . . . . . . . . . . . . . 264 5 Finitehorizon ruin probabilities . . .. . . . . . .. . . . . .. . . . . . The twobarrier ruin problem . . . 326 Appendix 331 Al Renewal theory . . .. . . . . . . 350 Bibliography Index 363 383 . . . . . . . 294 XI Miscellaneous topics 297 1 The ruin problem for Bernoulli random walk and Brownian motion. . . . . . . . .. . 340 A4 Some linear algebra . . . . . . . . . . . 323 6 Reinsurance . 331 A2 WienerHopf factorization . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . 285 3 Importance sampling via Lundberg conjugation . . .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . 251 2 The compound Poisson model . .
This page is intentionally left blank .
it would not be fair not to say that the practical relevance of the area has been questioned repeatedly. University of Copenhagen. the idea was close to expand these to a short book on the subject. 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!). and other projects absorbed my interest. and has been an active area of research from the days of Lundberg all the way up to today. However. Since I was to produce some handouts for the students anyway. it is not by intention. 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.Preface The most important to say about the history of this book is: it took too long time to write it! In 1991. this applies to longrange dependence which is intensely studied in the neighboring ix . Thus. As an excuse: many of these projects were related to the book. Let me take this opportunity to thank above all my publisher World Scientific Publishing Co. In particular. But the pace was much slower than expected. and the series editor Ole BarndorffNielsen for their patience. I was invited to give a course on ruin probabilities at the Laboratory of Insurance Mathematics. and my belief was that this could be done rather quickly. One reason for writing this book is a feeling that the area has in the recent years achieved a considerable mathematical maturity. It has obviously not been possible to cover all subareas. and the result is now that the book is much more related to my own research than the initial outline. the book is basically mathematical in its flavour. Risk theory in general and ruin probablities in particular is traditionally considered as part of insurance mathematics. The course was never realized. which has in particular removed one of the standard criticisms of the area. that it can only say something about very simple models and questions. Apart from these remarks. I have deliberately stayed away from discussing the practical relevance of the theory. if the formulations occasionally give a different impression.
111. [381]). http:// www.g.2. For a second reading.45. Chapters IIIVII introduce some of the main models and give a first derivation of some of their properties.5. VII. e. 111. see e.g. see also Schmidli [325] and the references in Asmussen & Taksar [52]. Good luck! I have tried to be fairly exhaustive in citing references close to the text. read Chapter I. for which I apologize to the reader and the authors of the many papers who ought to have been on the list. see in particular Michna [259].4a.2. The rest is up to your specific interests. VII. Willinger et al.maths .13 and IX. In addition. the standard stochastic control setting of diffusion models has been considered.3.13. Concerning ruin probabilities. Hojgaard & Taksar [206]. A book like this can be organized in many ways.15. One is by model. X.2 more properly).13 and XI. Asmussen. Finally. Another interesting area which is not covered is dynamic control. IX. The main motivation comes from statistical data for network traffic (e.se Lund February 2000 Soren Asmussen . VI.se/matstat / staff/asmus and I am therefore grateful to get relevant material sent by email to asmusfmaths . another by method. Here is a suggestion on how to get started with the book. It is obvious that such a system involves a number of inconsistencies and omissions. In the classical setting of CramerLundberg models. the first part of 11.14. IV. for the effects on tail probabilities. it has not been possible to incorporate more numerical examples than the few there are.6 (to understand the PollaczeckKhinchine formula in 111.lth. incorporate 11. IV. The book does not go into the broader aspects of the interface between insurance mathematics and mathematical finance.lth. I intend to keep a list of misprints and remarks posted on my web page. The present book is in between these two possibilities. 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).g. Hojgaard & Taksar [35] and Paulsen & Gjessing [284]. some papers not cited in the text but judged to be of interest are included in the Bibliography.x PREFACE field of queueing theory. some basic discussion can be found in the books by Biihlmann [82] and Gerber [157]. Resnick & Samorodnitsky [303] and references therein.1. For a brief orientation. More recently. an area which is becoming increasingly important. VIII. IV.13. I regret that due to time constraints.89.
5 from Asmussen [21] with permission from CRC Press.4 from Asmussen. Lund September 2001 Soren Asmussen Acknowledgements Many of the figures .6 by my 1999 simulation class in Lund.1 by Bjarne Hojgaard and the table in Example 111. Section VIII. . 111 . 5. not least the more complicated ones. Section VII . of which there are not many at this stage .3 are reprinted from Asmussen & Rubinstein [46] and parts of VIII. many of which were pointed out by Hanspeter Schmidli . 3 is reprinted from Asmussen & Nielsen [39] and parts of IX. Fig. 5 from Asmussen & Kliippelberg [36] with the permission from Elsevier Science .6. More substantial remarks.1 and X. A number of other figures were supplied by Christian Geisler Asmussen . as well as some additional references continue to be at the web page. Schmidli & Schmidt [47] with the permission from Applied Probability Trust . IV. supported by Center for Mathematical Physics and Stochastics (MaPhySto). Parts of X.2 by Rafal Kulik . 1 is almost identical to Section 2 of Asmussen [26] and reprinted with permission of Blackwell Publishers.6 is reprinted from Asmussen & Schmidt [49] and parts of IX.PREFACE xi The second printing differs from the first only by minor corrections. Parts of II. Fig. were produced by Lone Juul Hansen .8 . Aarhus.
This page is intentionally left blank .
They are the main topics of study of the present book. results and topics to be studied in the rest of the book. (1.3) sup St. we introduce some general notation and terminology. respectively. The probability O(u) of ultimate ruin is the probability that the reserve ever drops below zero. as defined in broad terms . Letting T(u) = inf {t > 0 : Rt < 0} = inf It > 0 : St > u}.i(u. For mathematical purposes.Rt. We denote throughout the initial reserve by u = Ro. MT = sup St.T) = P inf Rt < 0 I .2) (O<t<T Ro=ul.4) O<t<oo O<t<T 1 . (1. it is frequently more convenient to work with the claim surplus process {St}t>0 defined by St = u .1) We also refer to t/) ( u) and 0(u. M = (1. t/i(u) = P (infRt < 0) = P (infR t < 0 t>0 t>0 The probability of ruin before time T is t. is a model for the time evolution of the reserves of an insurance company. (1.Chapter I Introduction 1 The risk process In this chapter . A risk reserve process { Rt}t>o. and give a very brief summary of some of the models. T) as ruin probabilities with infinite horizon and finite horizon .
1. T3.7) k=1 k=1 The sample paths of {Rt} and {St} and the connection between the two processes are illustrated in Fig. St = E Uk . the time of arrival of the nth claim is an = T1 + • • • + Tn.. the number Nt of arrivals in [0. (1. However. say. (1. per unit time. the following setup will cover the vast majority of the book: • There are only finitely many claims in finite time intervals. That is.T) = F (MT > u) = P(r(u) < T). .b(u) = P (r(u) < oo) = P(M > u). Putting things together.pt. respectively. We denote the interarrival times of claims by T2. (1. we see that Nt Nt Rt = u + pt .2 CHAPTER I.E Uk.i(u. • Premiums flow in at rate p. the ruin probabilities can then alternatively be written as . and T1 is the time of the first claim. 1. INTRODUCTION be the time to ruin and the maxima with infinite and finite horizon. Thus. Figure 1. t] is finite. and Nt = min {n > 0 : 0rn+1 > t} = max {n > 0: Un < t}• The size of the nth claim is denoted by Un.1 .5) i..6) Sofar we have not imposed any assumptions on the risk reserve process.
. . not discuss whether this actually corresponds to practice. and the basic ruin probabilities are derived in XI. VII. but as an approximation to the risk process rather than as a model of intrinsic merit. and hence O(u) < 1 for all sufficiently large u. 1. However.b(u) = 1 for all u. 1. a basic references is Gerber [127]. rl= pP P It is sometimes stated in the theoretical literature that the typical values of the safety loading 77 are relatively small.20%.(. allowing a countable infinity of jumps on Fig. however. The models we consider will typically have the property that there exists a constant p such that Nt a E Uk k=1 p.8) holds.e. one may well argue that Brownian motion in itself could be a reasonable model. say 10% .1. If 77 < 0. • General Levy processes (defined as continuous time processes with stationary independent increments) where the jump component has infinite Levy measure. we shall. We study this case in Ch. on Fig. immaterial. and hence . It would appear obvious.s.s. and in fact: Proposition 1.1 Assume that (1.. If 77 > 0.8) The interpretation of p is as the average amount of claim per unit time. t * oo. • Brownian motion or more general diffusions. (1.1. then M = oo a.1 the slope of {Rt} should depend also on the level). however. for example. though many results are straightforward to generalize from the compound Poisson model. Thus.1. 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. For the purpose of studying ruin probabilities this distinction is. then M < oo a. one could well replace Rt by Rtnr(u) or RtA. 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. We shall not deal with this case either. since any modeling involves some approximative assumptions.) V 0. of course. Some main examples of models not incorporated in the above setup are: • Models with a premium depending on the reserve (i. that the insurance company should try to ensure 77 > 0. We shall discuss Brownian motion somewhat in Chapter IV.
i. The simplest example is 3(t) = V where V is a r .s.8) is given by ^t p = EU • lim it (3(s) ds t.10) hold with p constant. not all models considered in the literature have this feature: Example 1. If u oo. tb(u) = 1 for all u holds also when rl = 0.2 (Cox PROCESSES) Here {Nt} is a Poisson process with random rate /3(t) (say) at time t.T) = i. If U1.i(u. Nt)}.11) .Tp).s. zP(u .3 Assume p 54 1 and define Rt = Rt1p. . Then the connection between the ruin probabilities for the given risk process {Rt} and those ^(u). St In concrete models. (1. U2. 0 We shall only encounter a few instances of a Cox process..Q (say) and U1. U2. if {(3(t)} is nonergodic. Thus p may well be random for such processes. However. and independent of {(0(t). then similarly limSt/t < 0. Proposition 1. 0(u. where {Nt} is a Poisson process with rate . and here (1. . M < oo a.. and independent of {Nt}. with the most notable special case being V having a Gamma distribution.. corresponding to the Pdlya process. t t p  p' t ^ oo.T) for {Rt} is given by V)(u) = t/i (u). it is not too difficult to show that p as defined by (1. then this limit is > 0 which implies St a$ oo and hence M = oo a. are i.d.10) Again.s. The simplest concrete example (to be studied in Chapter III) is the compound Poisson model. If 77 < 0. k=1 (1. This case is referred to as the mixed Poisson process. Here it is easy to see that p = . (1.8). are i.6EU (on the average.v.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.10) is a property which we will typically encounter.. and that . namely that M = oo a. INTRODUCTION Proof It follows from (1. namely.oo t 0 J (provided the limit exists).b(u) < 1 for all u when rl > 0.i.d. rl > 0.. However. (1.8) that F N.4 CHAPTER I. _ St __ k =1 Uk pt a4. this needs to be verified in each separate case. . in connection with risk processes in a Markovian or periodic environment (Chapter VI). we obtain typically a somewhat stronger conclusion.
The term risk theory is often interpreted in a broader sense than as just to comprise the study of ruin probabilities. Gerber [159]) has a rather different flavour. see also Chapter XI. Segerdahl [334] and Philipson [289]. Taylor [364]. Embrechts et al. the claim arrivals are Poisson or renewal at the same time). in particular. Rolski. U2. and in fact p < 1 is the fundamental assumption of queueing theory ensuring steadystate behaviour (existence of a limiting stationary distribution).. we shall be able to identify p with the traffic intensity of an associated queue. [134]. Note that when p = 1. Heilmann [191]. which is feasible since in most cases the process { Rt } has a similar structure as {Rt} (for example. Daykin et al. [101]. We roughly classify these into two groups . Insurance: Mathematics and Economics. Gerber [157]. Some of the main general ideas were laid down by Lundberg [250]. Buhlmann [82]. Pentikainen & Pesonen [101]. Notes and references The study of ruin probabilities.2. Some main later textbooks are (in alphabetical order) Buhlmann [82]. but in probability and applied probability as a whole. Some early surveys are given in Cramer [91]. [76]. Daykin.. Since { Rt } has premium rate 1. De Vylder [110]. the assumption > 0 is equivalent to p < 1. the recent survey by Grandell [173] and references therein. Besides in standard journals in probability and applied probability. Note that life insurance (e. Cox processes are treated extensively in Grandell [171].. The Swedish school was pioneering not only in risk theory.g. Sundt [354]. Mitteilungen der Verein der Schweizerischen Versicherungsmathematiker and the Scandinavian Actuarial Journal. lighttailed distributions (sometimes the term . often referred to as collective risk theory or just risk theory. An idea of the additional topics and problems one may incorporate under risk theory can be obtained from the survey paper [273] by Norberg. For mixed Poisson processes and Polya processes. Hipp & Michel [198]. the research literature is often published in journals like Astin Bulletin . was largely initiated in Sweden in the first half of the century. the role of the result is to justify to take p = 1. see e .. Schmidt & Teugels [307] and Seal [326]. Straub [353]. some main texts (typically incorporating some ruin theory but emphasizing the topic to a varying degree) are Bowers et al.g. many results and methods in random walk theory originate from there and the area was ahead of related ones like queueing theory. in a number of models. another important early Swedish work is Tacklind [373]. [330]. CLAIM SIZE DISTRIBUTIONS 5 The proof is trivial. and we do not get near to the topic anywhere in this book. Grandell [171]. 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. while the first mathematically substantial results were given in Lundberg [251] and Cramer [91]. In the even more general area of nonlife insurance mathematics. Schmidli.
g. for the compound Poisson model with exponential claim sizes the ruin probability . the m. regularly varying (see below) or even regularly varying with infinite variance.2 and /LB is the mean of B.3) .B(x) satisfies B(x) = O(e8x) for some s > 0. then the conditional distribution of X . Example 2 .8. 2a Lighttailed distributions Example 2. P B[s]= (8Is ) . In contrast. On the more heuristical side. Here lighttailed means that the tail B(x) = 1 .x given X > x is again exponential with rate b (this is essentially equivalent to the failure rate being constant). 6 has density r(p)xPleax b(x) P and m. B is heavytailed if b[s] = oo for all s > 0.1) The parameter 6 is referred to as the rate or the intensity. In particular. i.g. B[s] is finite for some s > 0. the exponential distribution is by far the simplest to deal with in risk theory as well. INTRODUCTION 'Cramertype conditions' is used). but different more restrictive definitions are often used: subexponential.f. and heavytailed distributions. a simple stopping time argument shows that this implies that the conditional distribution of the overshoot ST(u) . For example in the compound Poisson model. As in a number of other applied probability areas.1 (THE EXPONENTIAL DISTRIBUTION) Here the density is b(x) = beax (2. s<8. and can also be interpreted as the (constant) failure rate b(x)/B(x). a fact which turns out to contain considerable information. where B(bo.2) = 0. (2.u at the time of ruin given r(u) is again exponential u with rate 8.f.O(u) can be found in closed form.6 CHAPTER I.2 (THE GAMMA DISTRIBUTION) The gamma distribution with parameters p.e. The crucial feature is the lack of memory: if X is exponential with rate 6. Equivalently. if 1 °O AB Jbos x B(dx) > 0. 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'.
> 1 for p < 1 and = 1 for p = 1 (the exponential case). An appealing feature is its simple connection to the Poisson process: B(x) = P(Xi + • • • + XP > x) is the probability of at most p .3 (THE HYPEREXPONENTIAL DISTRIBUTION) This is defined as a finite mixture of exponential distributions. In particular.1 Poisson events in [0. CLAIM SIZE DISTRIBUTIONS 7 The mean EX is p/b and the variance Var X is p/b2. x] so that B(x) = r` e.v.d. and exponential with rate d. An important property of the hyperexponential distribution is that its s. p). B(x) = r(p) Asymptotically. X2. p) °° where r (x.1) (or the 1/pth root if p < 1).. p. 2. . we develop computationally tractable results mainly for the Erlang case (p = 1. i = 1.c..) VarX1 (EX )2 p is < 1 for p > 1.. among others. then X v Xl + • • • + X.. 0. one has r(bx.ate (b2 ): L• i=o In the present text. by Grandell & Segerdahl [175] and Thorin [369]. u Example 2 . .y i=1 where >i ai = 1. if p is integer and X has the gamma distribution p.c. u . p) = J tPletdt. are i. JP 1 B(x) r(p ) XP ie ax In the sense of the theory of infinitely divisible distributions. or just the Erlang(p) distribution.). The exact form of the tail B(x) is given by the incomplete Gamma function r(x.2. Ruin probabilities for the general case has been studied.. In particular. the Gamma density (2. the squared coefficient of variation (s.i. P b(x) = r` aibiea.. is > 1. 0 < ai < 1. ..v. where X1. This special case is referred to as the Erlang distribution with p stages.2) can be considered as the pth power of the exponential density (2.
INTRODUCTION Example 2 .f. a. there exists a xo < oo such that B(x) = 0 for x > xo. Important special cases are the exponential.g.d.8 CHAPTER I. the restriction T of the intensity matrix of the Markov process to E and the row vector a = (ai)iEE of initial probabilities. We give a more comprehensive treatment in VIII. 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.6 (DISTRIBUTIONS WITH BOUNDED SUPPORT) This example (i. The density and c. of which one is absorbing and the rest transient. B(x) > 0 for x < xo) is of course a trivial instance of a lighttailed distribution.4 (PHASETYPE DISTRIBUTIONS) A phasetype distribution is the distribution of the absorption time in a Markov process with finitely many states. it is notable from a practical point of view because of reinsurance: if excessofloss reinsurance has been arranged with retention level xo.(2. B(x) = aeTxe where t = Te and e = (1 .e. which is slightly smaller but more amenable to probabilistic reasoning..f. .. are b(x) = aeTxt.8) j=1 j=1 j=1 where the parameters in (2. The couple (a. Equivalent characterizations are that the density b(x) has one of the forms q b(x) j=1 = cjxienbx. (or. The parameters of a phasetype distribution is the set E of transient states. This class of distributions is popular in older literature on both risk theory and queues. a rational Laplace transform) if B[s] _ p(s)/q(s) with p(s) and q(s) being polynomials of finite degree. See XI.6. Example 2 .6. then the claim size which is relevant from the point of view of the insurance company itself is U A xo rather than U u (the excess (U . the Erlang and the hyperexponential distributions.5 (DISTRIBUTIONS WITH RATIONAL TRANSFORMS) A distribution B has a rational m. However.8) are realvalued. 1)' is the column vector with 1 at all entries.7) are possibly complexvalued but the parameters in (2.xo)+ is covered by the reinsurer). resp. equivalently. q2 q3 (2. T) or sometimes the triple (E. but the current trend in applied probability is to restrict attention to the class of phasetype distributions. Example 2 .7) q1 b(x) = cjxieWWx + djxi cos(ajx)ea'x + > ejxi sin(bjx)e`ix . T) is called the representation.1 and defer further details to u Chapter VIII. We give some theory for matrixu exponential distribution in VIII.
(2.12) Sometimes also a location parameter a > 0 and a scale parameter A > 0 is allowed. It follows that the density is 't (1ogX . b(x) = crx''le`xr.9 (THE PARETO DISTRIBUTION) Here the essence is that the tail B(x) decreases like a power of x. the mean u is eµ+a /2 and the second moment is e2µ+2o2. a2). All moments are finite.u l b(x) = d dx or J ax lor 1 exp Asymptotically.9) which is heavytailed when 0 < r < I. Here failure rates b(x) = b(x)/B(x) play an important role. or equivalently as the distribution of a°U+µ where U .8 (THE LOGNORMAL DISTRIBUTION) The lognormal distribution with parameters a2.1). (2. the tail is B (x ) 2 x. Example 2 . the exponential distribution representing the simplest example since here b(x) is constant.p a 1 (2. x < a.2.1.7 (THE WEIBULL DISTRIBUTION) This distribution originates from reliability theory.pl = 1 W (logx . In particular. one being B(x) (1 + X)b(x) (1 + x)a+1' x > 0. Writing c = d/r.10) The loinormal distribution has moments of all orders. However. u Example 2 .N(p. 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).11) ex log logx 2r p 1 1 2 ( a ) f 1 (lox_P)2} (2. a)/A)a+1' x > a. b(x) _ A(1 + (x a The pth moment is finite if and only if p < a . CLAIM SIZE DISTRIBUTIONS 9 2b Heavytailed distributions Example 2. (2. and then b(x) = 0.N(0. There are various variants of the definition around.13) u . we obtain the Weibull distribution B(x) = eCx'. p is defined as the distribution of ev where V .
Thus. in particular. i.(1 + 2x + 2x2)e2x) p = 2 (2.12 (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) . 6 is defined as the distribution of et' where V has the gamma density (2. the loggamma distribution is a Pareto distribution. u Example 2 . (2. in particular. the loggamma distribution (with exponent 5) and a Pareto mixture of exponentials.2).15) x2 + 16x3 ) a3x/2) 3 (1 .x6+lr(p) (2.(1 + Zx + $ p = 3. x + 00. B(x) = O(xP).10 (THE LOGGAMMA DISTRIBUTION) The loggamma distribution with parameters p. (2. where Y is Pareto distributed with a = (p . satisfies L(xt)/L(x) 4 1.17) where L (x) is slowly varying.v.L( x ). examples of distributions with regularly varying tails are the Pareto distribution (2.14) The pth moment is finite if p < 5 and infinite if p > 5. Choudhury & Whitt [1] as the class of distributions of r. { s () 1s+3s29s3log(1+2s I p=3. In general.10 CHAPTER I. INTRODUCTION Example 2. For p = 1. oo) is slowly varying . another standard example is (log x)').16) 11 Example 2. A = 1 and X is standard exponential.13).1)/p. The simplest examples correspond to p small and integervalued.12) (here L (x) * 1) and ( 2. x 4 oo (any L having a limit in (0. u . the density is { 3 (1 .11 (PARETO MIXTURES OF EXPONENTIALS) This class was introduced by Abate.'s of the form YX.e. 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). The density is 8p(log x)pi b(x) .
the claim size distribution represents of course only one aspect (though a major one). Similar discussion applies to the distribution of the accumulated claims (XI. Namely. for example the lognormal distribution is subexponential (but not regularly varying).. Thus.. the subexponential class of distributions provide a convenient framework for studying large classes of heavyu tailed distributions. it will be seen that we obtain completely different results depending on whether the claim size distribution is exponentially bounded or heavytailed. and so is the Weibull distribution with 0 < r < 1. and based upon such information it seems questionable to extrapolate to tail behaviour. 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. We return to a closer study in IX.1) that any distribution with a regularly varying tail is subexponential.1. one may argue that this difficulty is not resticted to ruin probability theory alone. However. From a practical point of view... When studying ruin probabilities.13 (THE SUBEXPONENTIAL CLASS OF DISTRIBUTIONS) We say that a distribution B is subexponential if xroo lim B `2^ = 2. At least as important is the specification of the structure of the point process {Nt } of claim arrivals and its possible dependence with the claims. U2. We give some discussion on standard methods to distinguish between light and heavy tails in Section 4f. Also. though the proof of this is nontrivial. 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. this phenomenon represents one of the true controversies of the area. By far the most prominent case is the compound Poisson (CramerLundberg) model where {Nt} is Poisson and independent of the claim sizes U1. but can never be sure whether this is also so for atypical levels for which far less detailed statistical information is available.3. which each have a ( timehomogeneous) small rate of experiencing a . 3 The arrival process For the purpose of modeling a risk process . the knowledge of the claim size distribution will typically be based upon statistical data.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. THE ARRIVAL PROCESS 11 Example 2. (2.18) B(x) It can be proved (see IX.
epidemics in life insurance etc. too general and one neeed to specialize to more concrete assumptions . A more appealing way to allow for inhomogeneity is by means of an intensity . the negative binomial distribution. in just the same way as the Poisson process arises in telephone traffic (a large number of subscribers each calling with a small rate). with a common term {Nt} is a Markovmodulated Poisson process . when Jt = i. the first extension to be studied in detail was {Nt } to be renewal (the interarrival times T1 .. the periodic and the Markov modulated models also have attractive features . it is more questionable whether it provides a model with a similar intuitive content as the Poisson model. 5.i. gives rise to an arrival process which is very close to a Poisson process.(3.e. Cox processes are.8 (t) is a periodic function of t. in Chapter VII). However .. are i.. Another one is Cox processes. such that 8(t) = . IV (and. 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 . Mathematically. This applies also to the case where the claim size distribution depends on the time of the year or . but with a general not necessarily exponential distribution ). This model can be intuitively understood in some simple cases like { Jt} describing weather conditions in car insurance . not many detailed studies of the goodnessoffit of the Poisson model in insurance are available . The compound Poisson model is studied in detail in Chapters III. we study this case in VI . its basic feature is to allow more variation (bursty arrivals ) than inherent in the simple Poisson process. This model ..6. In others. To the author 's knowledge . 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. found the Poisson distribution to be inadequate and suggested various other univariate distributions as alternatives . INTRODUCTION claim . I. Nevertheless .12 CHAPTER I. radioactive decay (a huge number of atoms each splitting with a tiny rate ) and many other applications. In order to prove reasonably substantial and interesting results . and also that the ruin problem may be hard to analyze .3(t) fluctuating over time.d. has some mathematically appealing random walk features . The one we focus on (Chapter VI) is a Markovian environment : the environmental conditions are described by a finite Markov process {Jt }too. getting away from the simple Poisson process seems a crucial step in making the model more realistic. with the extension to premiums depending on the reserve. T2. Some of them have concentrated on the marginal distribution of NT (say T = one year ). Historically. e. however. where {/3 (t)}too is an arbitrary stochastic process . The point of view we take here is Markov dependent random walks in continuous time (Markov additive processes ).g. which facilitate the analysis. see 11. An obvious example is 3(t) depending on the time of the year (the season). in particular to allow for certain inhomogeneities. to be studied in Chapter V. so that .
The M/G/1 workload process { Vt } may also be seen as one of the simplest storage models.'s like V is available. this amounts to Vo having the stationary distribution of {Vt}). Similarly. ruin probabilities for risk processes with an input process which is renewal.4. time series and Gaussian processes. A SUMMARY OF MAIN RESULTS AND METHODS 13 the environment (VI. this gives only f0 O°i (u)du which is of limited . others are quite different. point processes and so on. stochastic geometry. methods or modeling ideas developed in one area often has relevance for the other one as well. More generally. 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. stochastic differential equations. R = p(R) in between jumps. it is desirable to have a set of formulas like (4. and which seems well motivated from a practical point of view as well. and a lot of information on steadystate r. In fact. The ones which appear most related to risk theory are queueing theory and dam/storage processes.v.6) . and the limit t 4 oo is the steadystate limit. reliability. queueing theory. that quite often the emphasis is on computing expected values like EV.1) permitting to translate freely between risk theory and the queueing/storage setting. It should be noted. In the setting of (4. A stochastic process {Vt } is said to be in the steady state if it is strictly stationary (in the Markov case.T) = P(VT > u).1) where V is the limit in distribution of Vt as t + oo. Thus. however. Markovmodulated or periodic can be related to queues with similar characteristics. dam/storage processes. genetics models.0 (u. with Poisson arrivals and constant release rule p(x) = 1. Mathematically. 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 . and here (4. others being branching processes. 0(u) = P(V > u). The study of the steady state is by far the most dominant topic of queueing and storage theory. Some of these have a certain resemblance in flavour and methodology. it is a recurrent theme of this book to stress this connection which is often neglected in the specialized literature on risk theory.1). (4. extreme value theory.1) holds as well provided the risk process has a premium rule depending on the reserve. interacting particle systems. A general release rule p(x) means that {Vt} decreases according to the differential equation V = p(V) in between jumps.
Here ?P(u) is explicit provided that .1 is a sample path relation should be stressed : in this way the approach also applies to models having supplementary r. The infinite horizon (steady state ) case is covered by letting T oo. • The compound Poisson model with premium rate p(x) depending on the reserve and exponential claim size distribution B. 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. Example VIII. Similarly. much of the study of finite horizon problems (often referred to as transient behaviour) in queueing theory deals with busy period analysis which has no interpretation in risk theory at all. The fact that Theorem H. Thus .3.1). The qualifier 'with just a few phases ' refers to the fact that the diagonalization has to be carried out numerically in higher dimensions.1 . • The compound Poisson model with a claim size distribution degenerate at one point. INTRODUCTION intrinsic interest . Here Vi(u) is given in terms of a matrixexponential function ( Corollary VIII. the ideal is to be able to come up with closed form solutions for the ruin probabilities 0(u). though overlapping. see Corollary VII. 3. • The compound Poisson model with constant premium rate p = 1 and B being phasetype with a just few phases . e . see Boxma & Cohen [74] and Abate & Whitt [3].6. the functions w x f d 1 exdx () .g.14 CHAPTER I.1.3.p(y) y^ Jo p(x) can be written in closed form.2). as is typically the case.8. see Corollary III. 3.3. the two areas. • The compound Poisson model with some rather special heavytailed claim size distributions. . p = 0/8 and y = 8 .'s like the environmental process {Jt} in a Markovmodulated setting.T). have to some extent a different flavour.3.v. which gives a sample path version of (4.1) 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 ). B(x) = ebx. Vi(u. Here O(u) = peryu where 3 is the arrival intensity. A prototype of the duality results in this book is Theorem 11. 4b Exact solutions Of course . which can be expanded into a sum of exponential terms by diagonalization (see..
Embrechts.u(y)/a2(y) dy} 4c Numerical methods Next to a closedform solution. T) themselves. Given this can be done. [s.Lef$er function. Ab(u). relevant references are Grubel [179].b(u)du . 191). We don't discuss Laplace transform inversion much.7. the formulas ( IV. Here are some of the main approaches: Laplace transform inversion Often. However.S(u) 1S(oo) f °D exp {.ff 2µ(y)/a2(y) dy} dx . • An astable Levy process with drift . where Furrer [150] recently computed ii(u) as an infinite series involving the Mittag. T) du dT 0 TO 00 in closed form than the ruin probabilities z/'(u). For the finite horizon ruin probability 0(u. O(u. T) can then be calculated numerically by some method for transform inversion.1) is the explicit form of the ruin probability when {Rt} is a diffusion with infinitesimal drift and variance µ(x). see VIII. say the fast Fourier transform (FFT) as implemented in Grubel [179] for infinite horizon ruin probabilities for the renewal model. Also Brownian models or certain skip free random walks lead to explicit solutions (see XI .f f 2µ(y)/a2(y) dy} dx  (4. it is easier to find the Laplace transforms = f e8 . 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 . A notable fact ( see again XI. 1). Abate & Whitt [2].4. T). but are somewhat out of the mainstream of the area .2) is the natural scale. . a2 (x): Ip (u) = where S(u) = f °O exp {. the second best alternative is a numerical procedure which allows to calculate the exact values of the ruin probabilities.1) are so complicated that they should rather be viewed as basis for numerical methods than as closedform solutions. A SUMMARY OF MAIN RESULTS AND METHODS 15 • The compound Poisson model with a two step premium rule p(x) and B being phasetype with just a few phases. Grubel & Pitts [132] and Grubel & Hermesmeier [180] (see also the Bibliographical Notes in [307] p. esuTb( u. (u.
One example where this is feasible is the renewal equation for tl'(u) (Corollary III.3. see VIII.16 CHAPTER L INTRODUCTION Matrixanalytic methods This approach is relevant when the claim size distribution is of phasetype (or matrixexponential). (4. 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. For the compound Poisson model with p = 1 and claim size distribution B with moment generating function (m. Differential. whereas for the renewal arrival model and the Markovian environment model U has to be calculated numerically. either as the iterative solution of a fixpoint problem or by finding the diagonal form in terms of the complex roots to certain transcendental equations. and in quite a few cases (Chapter VIII).Ce"u.or integral equation.3) where C = (1 .g.4) 00['Y]1)'Y = 0.1) and y > 0 is the solution of the Lundberg equation (4. u * oo. 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. 4d Approximations The CramdrLundberg approximation This is one of the most celebrated result of risk theory (and probability theory as a whole). which can equivalently be written as f3 [7] = 1 +13 .) B[s]. dt] most often leads to equations involving both differential and integral terms. and in particular the naive idea of conditioning upon process behaviour in [0. U is explicit in terms of the model parameters. 0(u) is then given in terms of a matrixexponential function euu (here U is some suitable matrix) which can be computed by diagonalization. However.f.and integral equations The idea is here to express 'O(u) or '(u.7.p)/(13B'[ry] . In the compound Poisson model with p = 1. most often it is more difficult to come up with reasonably simple equations than one may believe at a first sight. it states that i/i(u) .3) in the compound Poisson model which is an integral equation of Volterra type. T) as the solution to a differential.
in some cases the results are even more complete than for light tails.2. in such cases the evaluation of C is more cumbersome. a Markovian environment or periodically varying parameters. For example. and use the fact that first passage probabilities are more readily calculated for diffusions than for the risk process itself. when the claim size distribution is of phasetype. some further possibilities are surveyed in 111 . T) for large u are available in most of the models we discuss.6) 4e Bounds and inequalities The outstanding result in the area is Lundberg's inequality (u) < e"lu. J B dx. the claim size distribution should have an exponentially decreasing tail B(x). In particular. incorporating correction terms may change the picture dramatically. Approximations for O(u) as well as for 1(u. (4. . It has generalizations to the models with renewal arrivals.4. In the case of heavytailed distributions. u > oo. 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. In fact. other approaches are thus required. Large claims approximations In order for the CramerLundberg approximation to be valid. This list of approximations does by no means exhaust the topic. T). Diffusion approximations are easy to calculate. However. often for all u > 0 and not just for large u. for the compound Poisson model ^(u) p pu In fact . The CramerLundberg approximation is renowned not only for its mathematical beauty but also for being very precise. corrected diffusion approximations (see IV.7 and IV. but typically not very precise in their first naive implementation. See Chapter IX. the exact solution is as easy to compute as the CramerLundberg approximation at least in the first two of these three models. 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.6) are by far the best one can do in terms of finite horizon ruin probabilities '(u. However.
4f Statistical methods Any of the approaches and results above assume that the parameters of the model are completely known. How do we produce a confidence interval? And. which is a standard statistical problem since the claim sizes Ui. . obtained say by observing the risk process in [0. This is proved for the compound Poisson model in 111. However.) at various places and in various settings. one may question whether it is possible to distinguish between claim size distributions which are heavytailed or have an exponentially decaying tail. to have smaller ruin probabilities than when B is nondegenerate with the same mean m. given NT. empirical evidence shows that the general principle holds in a broad variety of settings.8. in the compound Poisson model. and to plot the empirical mean residual life 1 N . it is a general principle that adding random variation to a model increases the risk. fitting a parametric model to U1.U(k)) i =k+ i . We return to various extensions and sharpenings of Lundberg's inequality (finite horizon versions. . as a general rule. they have however to be estimated from data. e.d. it has the advantage of not involving approximations and also. 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. though not too many precise mathematical results have been obtained.k (U(`) . The standard suggestion is to observe that the mean residual life E[U . For example.. In practice. For example.i. say degenerate at m.. one expects a model with a deterministic claim size distribution B. the difficulty comes in when drawing inference about the ruin probabilities.3). . INTRODUCTION Compared to the CramerLundberg approximation (4. When comparing different risk models. UNT may be viewed as an interpolation in or smoothing of the histogram). 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. However. more importantly. UNT are i.g. lower bounds etc.. can we trust the confidence intervals for the large values of u which are of interest? In the present author's opinion. . . . of being somewhat easier to generalize beyond the compound Poisson setting.18 CHAPTER I.x U > x] = B(x) f '(yx)B(dx) typically has a finite limit (possibly 0) in the lighttailed case and goes to oo in the heavytailed case. This procedure in itself is fairly straightforward. T].
2.. UN. However. formula VI. this is a straightforward way to estimate finite horizon ruin probabilities. 5 Conventions Numbering and reference system The basic principles are just as in the author's earlier book Applied Probability and Queues (Wiley 1987.. The problem is entirely analogous to estimating steadystate characteristics by simulation in queueing/storage theory.d. 4g Simulation The development of modern computers have made simulation a popular experimental tool in all branches of applied probability and statistics. and of course the method is relevant in risk theory as well. Thus Proposition 4. and in fact methods from that area can often be used in risk theory as well .3) and Section VI. Klnppelberg & Mikosch [134]. but is not very satisfying. Simulation may be used just to get some vague insight in the process under study: simulate one or several sample paths. CONVENTIONS 19 as function of U(k)..e. See further Embrechts. because it appears to require an infinitely long simulation.5. . < U(N) are the order statistics based upon N i. and look at them to see whether they exhibit the expected behaviour or some surprises come up. claims U1. Truncation to a finite horizon has been used. The infinite horizon case presents a difficulty. A main problem is that ruin is typically a rare event (i.v's) which can be generated by simulation. (or a functional of the expectation of a set of r.3) or Section 3 of Chapter VI are referred to as Proposition VI. good methods exist in a number of models and are based upon representing the ruin probability zb(u) as expected value of a r. to observe whether one or the other limiting behaviour is apparent in the tail.i.(5. and also discuss how to develop methods which are efficient in terms of producing a small variance for a fixed simulation budget.2. having small probability) and that therefore naive simulation is expensive or even infeasible in terms of computer time. Still.3 (or just VI.. We look at a variety of such methods in Chapter X. The chapter number is specified only when it is not the current one.3). formula (5. in all other chapters than VI where we just write . in this book referred to as [APQ]). respectively. where U(1) < . .4. For example.v. reference [14]. the more typical situation is to perform a Monte Carlo experiment to estimate probabilities (or expectations or distributions) which are not analytically available..
IIGII the total mass (variation ) of a (signed ) measure G . if B(x) .29) refer to the Appendix. formula (5.f. cumulant generating function. random variable s. with probability Mathematical notation P probability. (A.o.4.Used in asymptotic relations to indicate that the ratio between two expressions is 1 in the limit.p.B(x) = P(X > x) of B. B[s] the m. b[s] is defined always if Rs < 0 and sometimes in a larger strip (for example. as for typical claim size distributions. right hand side (of equation) r.e. EX2/(EX)2.g.d. i. n i oo.g.2.f. E. h + 0. moment generating function. i.v. n! 27r nn+1/2en. A different type of asymptotics: less precise. and for a defective probability distribution IIGII < 1. squared coefficient of variation. E expectation.The same symbol B is used for a probability measure B(dx) = P(X E dx) and its c. left hand side (of equation) m. for a probability distribution IIGII = 1.g. cumulative distribution function P(X < x) c.r. References like Proposition A.i.g. say a heuristic approxi1 + h + h2/2. The Laplace transform is b[s]. In particular.v. B(dy).d. . mation. If.ceax. with respect to w. r.g. log E[s] where b[s] is the m.c.s.f. B(x) = P(X < x) = fx. w. independent identically distributed i.f.t. infinitely often l.h. see under b[s] below. B is concentrated on [0.20 CHAPTER L INTRODUCTION Proposition 4. then for 1s < 5).h.d. oo). (moment generating function) fm e82B(dx) of the distribution B.3) or Section 3. Abbreviations c. or a more precise one like eh .s.f. .f. B(x) the tail 1 .
Usually. the ith unit row vector is e'i.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). . xa.5. Thus..e. . matrices have uppercase Roman or Greek letters like T. F o r a given set x1. Matrices and vectors are denoted by bold letters. (the dimension is usually clear from the context and left unspecified in the notation). (xi)diag denotes the diagonal matrix with the xi on the diagonal (xi)row denotes the row vector with the xi as components (xi).e. intensity interpretation. a2) the normal distribution with mean p and variance oa2. i. 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. 7r.. Usually. 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. Xt_ the left limit limstt X8f i. I(A) the indicator function of the event A. the ith entry is 1 and all other 0. N(it. an example or a remark. Unless otherwise stated. 0 marks the end of a proof. and column vectors have lowercase Roman letters like t. CONVENTIONS {6B the mean EX = f xB(dx) of B ABA' the nth moment EXn = f x"B(dx) of B. E[X. all stochastic processes considered in this book are assumed to have sample paths in this space. the processes we consider are piecewise continuous. the value just before t. 21 D [0. i. oo) the space of Rvalued functions which are rightcontionuous and have left limits. . In the Frenchinspired literature.A] means E[XI(A)]. Notation like f3i and 3(t) in Chapter VI has a similar . often the term 'cadlag' (continues a droite avec limites a gauche) is used for the Dproperty. of numbers.e. a. R(s) the real part of a complex number s. though slightly more complicated. only have finitely many jumps in each finite interval. row vectors have lowercase Greek letters like a. A.
5. cf. . VI. e. Notation like BE and B(t) in Chapter VI has a similar.1. J the rate parameter of B for the exponential case B(x) = eby. cf. ry The adjustment coefficient. interpretation.g. p the net amount /3pB of claims per unit time. 111. cf.22 CHAPTER L INTRODUCTION B the claim size distribution.5. I. I. though slightly more complicated. FL. or quantities with a similar time average interpretation.1. EL the probability measure and its corresponding expectation corresponding to the exponential change of measure given by Lundberg conjugation. 'q the safety loading .
When encountered for the first time in connection with the compound Poisson model in Chapter III. More precisely. fundamental ( at least in the author' s opinion) and the probability involved is rather simple and intuitive. The topic is. used in Chapter VI on risk processes in a Markovian (or periodic) environment. Sections 4. the level of the exposition is. 23 . Sections 4. Due to the generality of the theory. The general theory is. however. All results are proved elsewhere . the relevance for the mainstream of exposition is the following: The martingale approach in Section 1 is essentially only used here. 5) are. a parallel selfcontained treatment is given of the facts needed there. strictly speaking. 5 on random walks and Markov additive processes can be skipped until reading Chapter VI on the Markovian environment model. in part. somewhat more advanced than in the rest of the book. The reader should therefore observe that it is possible to skip most of the chapter. in particular at a first reading of the book. however. The likelihood ratio approach in Section 2 is basic for most of the models under study.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 duality results in Section 3 (and. not crucial for the rest of the book. however. in most cases via likelihood ratio arguments.
The more general Theorem 6. and in the limit (1.1 is basic for the study of the compound Poisson model in Chapter III.u denote the overshoot. and the ruin probabilities are ip(u) = P (T(u) < oo).0. (b) St a$ oo on {T(u) = oo}.5 can be skipped. the time to ruin r(u) is inf It > 0 : St > u}. T(u) > T] . As usual. T(u) < T] + E [eryST . V) (u. the second term converges to 0 by (b) and dominated convergence (e7ST < eryu on {r(u) > T}). SOME GENERAL TOOLS AND RESULTS The ladder height formula in Theorem 6.QµB < 1. T) = P(T(u) < T). StUit. 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. T(u) < oo] + 0 = eryuE [e7Vu).2 Consider the compound Poisson model with Poisson arrival rate . 1 Martingales We consider the claim surplus process {St} of a general risk process.2) takes the form 1 = E [e'ys().. Thus N.)AT = E [e7ST(°).1 Assume that (a) for some ry > 0. {e'YS° }t>0 is a martingale. (1. Let e(u) = ST(u) . claim size distribution B and p = . f1 . We get 1 = Ee7So = E e'Y S(.T(u) < cc] = e7uE {e7f(u) I T(u) < cc] z/.2) As T > oo. using r(u) A T invokes no problems because r(u) A T is bounded by T). Our first result is a representation formula for O(u) obtained by using the martingale optional stopping theorem . however.24 CHAPTER II.(u). Example 1 . Proposition 1.
the conditions of Proposition 1. Since {St} has stationary independent increments.Ft = a(S" : v < t).6a for details) that typically a solution to the Lundberg equation K(y) = 0 exists. condition (a) of Proposition 1.i.a = a . 1. the conditional distribution of the overshoot e(u) = U . and p =.r" where y = S . the martingale property now follows just as in Example 1.= e"(') where K(a) = . Thus.5 For the compound Poisson model with B exponential. it follows that E [e7st+v I J] = e"rstE [e7(st+vSt) I Ft] = e7StEe"rs° = elst where . of the normal distribution.3 Assume that {Rt} is Brownian motion with variance constant o. u Corollary 1.1) shows that Eels. By standard formulas for the m.1 is satisfied.1) . Thus.u + x is again just exponential with rate S. and thus Ee7s° = 1. and thus by the memoryless property of the exponential distribution . From this it is readily seen (see III. and (b) follows from p < 1 and the law of large numbers (see Proposition III. From this it is immediately seen that the solution to the Lundberg equation ic(y) = 0 is y = 2p/a2. Proof Since c(a) = /3 (B[a] . A simple calculation (see Proposition III. Eeas° = e"(°) where n(a) = a2a2/2 . Thus 00 E [e'rt (") I T(u) < oo] = I e5e  dx = f 5edx .4 (LUNDBERG ' S INEQUALITY ) tion 1 .g.2 and drift p > 0. the ruin probability is O(u) = pe. Under the conditions of Proposi Proof Just note that C(u) > 0.1.a it is immediately seen that y = S .Q. MARTINGALES 25 where {Nt} is a Poisson process with rate . with common distribution B (and independent of {Nt}). Now at the time r(u) of ruin {St} upcrosses level u by making a jump . u Corollary 1.3/6 < 1.d.Q(B[a] ./3. O(u ) < e7". and thus Ee'rs° = 1.1.x. are i.. Then {St } is Brownian motion with variance constant o2 and drift p < 0.ap.Q and the U. The available information on this jump is that the distribution given r(u) = t and S.1) . Since {St} has stationary independent increments.2.1. Example 1 .1 are satisfied.2(c)).f.a. B(x) _ edx.(„)_ = x is that of a claim size U given U > u .
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. P correspond to the claim surplus process of two compound Poisson risk processes with Poisson rates /3. Embrechts. then z/'(u) = e7" where 'y = 21A/a2. on (DE. Thus the sets S = I tlim +oot t =. . which we equip with the natural filtration {. and by the law of large numbers for the Poisson process . Example 2 . Proof Just note that ^(u) = 0 by continuity of Brownian motion. (2. The number Nt F) of jumps > e before time t is a (measurable) r. then S and S are disjoint .v. F). and F. Grandell & Schmidli [131].1) 'though not always: it is not difficult to construct a counterexample say in terms of transient Markov processes. However. the parameters of the two processes can be reconstructed from a single infinite path. [172]. hence so is Nt = limfyo N2`i. A]. F(S) = P(S) = 1. and is further exploited in his book [157].3 below). More recent references are Dassios & Embrechts [98]. P are then singular (concentrated on two disjoint measurable sets). Two such processes may be represented by probability measures F.26 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Corollary 1. oo).Ft}too and the Borel afield F. A somewhat similar u argument gives singularity when B $ B. cf. Theorem 2. Delbaen & Haezendonck [103] and Schmidli [320]. A E Ft.F). P on (DE. as shown by the following example this setup is too restrictive: typically'.1 Let F.6 N S = { lim Nt I t +00 t gJ are both in F.6 If {Rt} is Brownian motion with variance constant a2 and drift p > 0. But if a $ ^ .e. we look for a process {Lt} (the likelihood ratio process) such that P(A) = E[Lt. and in analogy with the theory of measures on finite dimensional spaces one could study conditions for the RadonNikodym derivative dP/dP to exist. B. I.. 0 and claim size distributions B. u Notes and references The first use of martingales in risk theory is due to Gerber [156]. Grandell [171]. The interesting concept is therefore to look for absolute continuity only on finite time intervals (possibly random.
ELt = 1 follows by taking A = DE in (2.Pt}adapted process {Lt}t>o (2.r. A E F8. Proof Under the assumptions of (i). F the Borel o•field and P a given probability measure on (DE. Lt < 0] can only be nonnegative if P(A) = 0. A E Ft . under the assumptions of (ii) we have for A E rg and s < t that A E Ft as well and hence E[L8. Then Ft (A) = E[Lt. we have E [ LTIFT]1 = LT on {T < T}. Hence the family {Pt} is t>o consistent and hence extendable to a probability measure F on (DE. then {Lt} is a nonnegative martingale w. Hence E [_ .F such that (2. A E F.e.2. G C {T < oo}. (i) If {Lt}t> o is a nonnegative martingale w. the restriction of P to (DE. Then P(A) = E[Lt. This proves (i).t. A]. If r is a stopping time and G E PT.1) holds.1) holds. The truth of this for all A E Y.t.3 Let {Lt}. then there exists a unique probability measure Pon . implies that E[LtI.F).r. G l ] = E [_I(G)E[LTIFT] ] = E { _I(G)Lr ] = P(G).t.Y) such that P(A) = Pt(A). F) such that ELt = 1.1) and nonnegativity by letting A = {Lt < 0}. P) such that LLt = 1. Lets < t. 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 . ({. define P by Pt (A) = E[Lt. A]. then { 1 P(G) = EG .F8] = L8 and the martingale property. ({Ft} . A] = E[Lt.r. Then Lt > 0 and ELt = 1 ensure that Pt is a probability measure on (DE.2(i). By the martingale property. _. Conversely. 1 J (2. G ] = E [LT .Tt) is absolutely continuous w. if for some probability measure P and some {. Finally. . that the restriction of P to (DE.Pt)) The following result gives the connection to martingales.Ft}. . LIKELIHOOD RATIOS AND CHANGE OF MEASURE 27 (i.2) Proof Assume first G C {T < T} for some fixed deterministic T < oo..2 Let {Ft}t>o be the natural filtration on DE. (ii) Conversely. using the martingale property in the fourth step. Ft). P be as in Proposition 2. Proposition 2.A] = EE[LtI(A)IF8] = EI(A)E[LtIFB] = EI(A)L8 = PS(A).
first in the Markov case and next (Sections 4.t. Xt = (Jt. the natural filtration {. Consider a (timehomogeneous) Markov process {Xt} with state space E.. is nonnegative and has Ey Lt = 1 for all x. and this problem will now be studied. First we ask when the Markov property is preserved.r. For the definition. (2.r. Xt = St.1: Corollary 2. t. 1 Since everything is non negative. T(u) < oo].4 Under condition (a) of Proposition 1. A change of measure is performed by finding a process {Lt} which is a martingale w. r(u) < oo] occuring there than with the (conditional) expectation E[e'r{(u ) Jr(u) < oo] in (1. we have F(G) = V )(u). . St).2) by noting that 1 = ersr(„) = e1'ue7Ou). Now just rewrite the r. where {Rt} is the risk reserve process..h. applying (2. say. {St} = {u .s.Ft} . is Markov w. and letting T * oo.3 we obtain a likelihood ratio representation of the ruin probability V) (u) parallel to the martingale representation (1. we need the concept of a multiplicative functional.O(u) = eryuE[e'YC(u). SOME GENERAL TOOLS AND RESULTS In the general case . in continuous time (the discrete time case is parallel but slightly simpler).1.3) to G of{r < T} we get 1111 F(Gn {r <T}) = E[ 1 . The crucial step is to obtain information on the process evolving according to F.Rt} the claim surplus process and {Jt} a process of supplementary variables possibly needed to make the process Markovian.1) is that it seems in general easier to deal with the (unconditional) expectation E[eryVu). In the context of ruin probabilities.2) follows by monotone convergence.28 CHAPTER II.4) Proof Letting G = {r(u) < oo}. Lr(u) 11 The advantage of (2. we assume for simplicity that {Xt} has Dpaths. of (2. u From Theorem 2.4) compared to (1. 5) for processes with some randomwalklike structure.1) in Proposition 1.Gn {r <T} . each F. The problem is thus to investigate which characteristics of {Xt} and {Lt} ensure a given set of properties of the changed probability measure. (2. To this end.1).t. Rt) or Xt = (Jt. both sides are increasing in T. one would typically have Xt = Rt.
7).6) for any . t and let Px be the probability measure given by t. since Zt • (Y8 o Ot ) is . By definition of Px. Proof Since both sides of (2. ({Xt+u} 0<u<8) Theorem 2.6) implies (2.(Xtitl) with all t(i) < t + s.5) are Tt+e measurable.r. 0 . (2. The precise meaning of this is the following: being .F8measurable r. this in turn means Ex[Lt+8Zt(V8 oet)] = Ex[LtZtExt[L8Y8]].7). Y8.2.'s of the form Zt • (Y8 o 0t) comprises all r. A].Ftmeasurable.5) is equivalent to Ex[Lt+8Vt+8] = E8[Lt • (L8 o 91)Vt+8] for any . o 9t = V.Ft]..7) for any Ftmeasurable Zt and any .'s of the form fl' f. Lt has the form Lt = 'Pt ({x }0<u<t) for some mapping cot : DE[O. oo). The converse follows since the class of r.Y8f t < s. let {Lt} be a nonnegative martingale with Ex Lt = 1 for all x.Pt+8measurable r.s.. since Ext [L8Y8] = E[(Y8 o et)(L8 o 8t)I. where Ot is the shift operator. the Markov property can be written E.Ft] = Ex.v.t. Vt+e. Ex[Lt+8Zt(Y8 o et)] = Ex[LtZt(Y8 o 0t)(L8 o Bt)]. Then the family {Px}xEE defines a timehomogeneous Markov process if and only if {Lt} is a multiplicative functional.v.Ft }. and then L.8) which is the same as (2.(A) = Ex [Lt.5 Let {Xt} be Markov w. which is the same as Ex[Zt(Y8 o Bt)] = E8[ZtEx. or. t.v. t] * [0. (2. nonnegative and Lt+8 = Lt•(Lso9t) (2. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 29 on DE and define {Lt} to be a multiplicative functional if {Lt} is adapted to {.5) Pxa. Similarly. Indeed.v. (2. o 9tI. the natural filtration {Ft} on DE. s.Ft+8measurable. Zt. (2.v. for all x.T9measurable Y8.[Y. which in turn is the same as Ex[Lt+8Zt • (V8 o Bt)] = Ex[Lt • (L8 o 91)Z1 • (Y8 o et)] (2.YB] for any Ftmeasurable r.
6 For {u . aN where or* = T UN_k+l.. SOME GENERAL TOOLS AND RESULTS to define a timehomogeneous Markov process.5 can be found in Kuchler & Sorensen [240]. CN. . {Vt} .. The corresponding claim sizes are Ul. T] in the following setup: The risk process {Rt}o<t<T has arrivals at epochs or. . 0 < vl < . u Notes and references The results of the present section are essentially known in a very general Markov process formulation. . . t] = LtExt L8 = Lt. . we shall establish a general connection between ruin probabilities and certain stochastic processes which occurs for example as models for queueing and storage. < aN < T. In between jumps. The result is a sample path relation. Ro = u (say). Indeed.Ft] = LtE[L8 o 9t I. t.e. and just after time or* {Vt} makes an upwards jump of size UU = UN _k+l. and the time to ruin is 7(u) = inf {t > 0: Rt < 0}. with a proof somewhat different from the present one. the premium rate is p(r) when the reserve is r (i. UN. A more elementary version along the lines of Theorem 2. reflection at zero and initiar condition Vo = 0. More precisely . (using the Markov property in the second step) so that the martingale property is automatic.. In between jumps.1) The initial condition is arbitrary. and thus for the moment no parametric assumptions (on say the structure of the arrival process) are needed. 3 Duality with other applied probability models In this section.At where At = k: vk <t U.. Thus R = Ro + f p(R8) ds .. see Dynkin [128] and Kunita [239]. then Remark 2.. the arrival epochs are Qi. We work on a finite time interval [0. R = p(R)). The formulation has applications to virtually all the risk models studied in this book... it xEE suffices to assume that {Lt} is a multiplicative functional with Ex Lt = 1 for all x.30 CHAPTER H. The storage process {Vt }o<t<T is essentially defined by time reversion... (3. } E[Lt+B I..
3) (u) Proof Let rt' denote the solution of R = p(R) subject to r0 = u. V)(u. DUALITY WITH OTHER APPLIED PROBABILITY MODELS 31 decreases at rate p(r) when Vt = r (i.1) we have Vt = At  f P(Vs)ds where A= U= AT .___ . That is. The sample path relation between these two processes is illustrated in Fig.. (3._.x.11 4. .T) = inf Rt < 0 P (O<t<T P(r(u) < T) be the ruin probability.AT_t. (3. Theorem 3.__.1 Define r(u) = inf It > 0: Rt < 0} (r(u) = oo if Rt > 0 for all t < T) and let ii(u. Then rt°) > rt°) for all t when u > v.2) k: ok <t and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0..__._: 1} 0 011 =T01N ^N3 To 0 011 014 01N Figure 3.. {Vt} remains at 0 until the next arrival).1..____•_.e. In particular. instead of (3..3.1 The events {T(u) < T} and {VT > u} coincide. 3. V = p(V)). :.....T) = P(VT > u). Note that these definitions make {Rt} rightcontinuous (as standard) and {Vt} leftcontinuous.
Then the time reversibility of the Poisson process ensures that {At } and {At } have the same distribution (for finitedimensional distributions. we have r(u) > T.Ul < roil  Ul = RQ„ Va1V_1 < RQ2. this represents a model for storage.U1 = Rol. u A basic example is when {Rt} is the risk reserve process corresponding to claims arriving at Poisson rate . we have RQ„ < 0 so that indeed r(u) < T. and then '0 (u) = P(V > u). We get: Corollary 3. Historically. and so on. Suppose next VT < u (this situation corresponds to the broken path of {Rt} in Fig. if nothing else n = N). and a general premium rule p(r) when the reserve is r. Hence if n satisfies VVN_n+1 = 0 (such a n exists.2 Consider the compound Poisson risk model with a general premium rule p(r).and left continuity is immaterial because the probability of a Poisson arrival at any fixed time t is zero). Theorem 3. Proof Let T ^ oo in (3.U1 > roil . If VaN > 0. u Notes and references Some main reference on storage processes are Harrison & Resnick [187] and Brockwell. Hence RQ„ > 0 for all n < N. say V.i. Then similarly VVN = r0. Historically. one may feel that the interaction between the different areas has been surprisingly limited even up to today. and in between rainfalls water is released at rate p(r) when Vt (the content) is r. see Siegmund [344].d. Resnick & Tweedie [79]. Corollary 3. . Then the storage process {Vt} has a proper limit in distribution. 3.1 and its proof is from Asmussen & Schock Petersen [50].T l . the distinction between right. 3. The arrival epochs correspond to rainfalls. the connection between risk theory and other applied probability areas appears first to have been noted by Prabhu [293] in a queueing context. = r(VT) .3 and being i.32 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Suppose first VT > u (this situation corresponds to the solid path of {Rt} in Fig. say of water in a dam though other interpretations like the amount of goods stored are also possible.3).2 from Harrison & Resnick [188]. Then Vo. Nevertheless.1 with Ro = u = ul). Some further relevant more general references are Asmussen [21] and Asmussen & Sigman [51]. if and only if O(u) < 1 for all u. Thus we may think of {Vt} as having compound Poisson input and being defined for all t < oo.1 with Ro = u = u2). and since ruin can only occur at the times of claims. with distribution B. The results can be viewed as special cases of Siegmund duality. we can repeat the argument and get VoN_1 > Ra2 and so on.
.e.. and is reset to 0 once the r.N (4.1 in terms of Lindley processes .1) Thus {Wn}n=o...1.. W1. . Then the events {r(u) < N} and {WN > u} coincide.... = Xo + Y1 + • • • + Y. Let further N be fixed and let Wo.1. .1... Theorem 4. has a proper limit W in distribution as n + oo. .... WN = YN .. In particular. i.. 1} is often referred to as simple random walk or Bernoulli random walk).1.Y1 according to Wo = 0. if Wo = 0 then (Z1+•••+Zn) WN = Zl+•••+ZN..s.4. ...2) (for a rigorous proof. Z2.2) satisfies the same recursion as in (4. R valued sequence Z1. . Here F is a general probability distribution on R (the special case of F being concentrated on {1. of (4.1)).w. where the Yi are i ... Z2. . the Lindley process Wo.... . .. W1. 0 Corollary 4.... with common distribution F (say).2).. (b) 1/i(u) = P(•r(u) < oo) > 0 as u * oo. can be viewed as the reflected version of the random walk with increments Z1. (c) The Lindley process {WN} generated by Zl = Y1... For discrete time random walks .. Z2 = YN_1 i .. . Proof By (4. Xo = 0.d..1. Most often. WN be the Lindley process generated by Z1 = YN.. I.. 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... is defined by assigning Wo some arbitrary value > 0 and letting Wn+1 = (Wn + Zn+1)+• (4. just verify that the r .1 Let r(u) = inf In: u + Y1 + • • • + Yn < 0}... evolves as a random walk with increments Z1i Z2.. Z2 = Y2. W2.2 The following assertions are equivalent: (a) 0(u) = P(r(u) < oo) < 1 for all u > 0... as long as the random walk only takes nonnegative values..d.. hits (oo.. ZN = .. For a given i. 0).YNn+1) n=0. .. generated by Z1. Z2 .1.min n=0. n=0. N min (Y1 + • • • + YNn) n=0. there is an analogue of Theorem 3.h. N min (Y1 + + Yn).N From this the result immediately follows..Yl min (YN .i. {Wn}n=0...
YN in Theorem 4. The converse follows from general random walk theory since it is standard that lim sup (Y1 + • • + Yn) = oo when Y1 + • • • + Yn 74 oo. (e). The following result gives the necessary and sufficient condition for {Ln} to define a new random walk: .3 The i.. 176) but appears to be rather intractable. the condition 00 F(YI+•••+ Yn<0)<00 n=1 is known to be necessary and sufficient ((APQ] p.1 have the same distribution for n = 0. assumption on the Z1..l.5 does not necessarily lead to a random walk: if. doubly u infinite (n'= 0. there is a more general version of Corollary 4..2... YN). a sufficient condition for (e) is that EY is welldefined and > 0. the Y1. (e)Yi+•••+Yn 74 . either M = oo a.. 0 By the law of large numbers. or M < oo a. Px to Fn are equivalent (have the same null sets) so that the likelihood ratio Ln exists. Proof Since (YN. ±2.d. ±1.. For a random walk. (Yi + • • • + Yn) > oo a.=o. Next consider change of measure via likelihood ratios.s.N so that WN _P4 M = supra=0... . .. + Z.. . In general.s.. ZN or. 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..s. One then assumes Yn to be a stationary sequence.ooa. the Lindley processes in Corollary 4. Similarly. In that case . Combining these facts gives easily the equivalence of (a)(d). ... Remark 4 . .g.) and defines Zn = Yn.1..l.s .2 and Theorem 4.) sup n=0.. . N. (Z1 + • • • + Zn) = m and P(W > u) = P(M > u) = i (u ).g. Clearly. e.the result is a sample path relation as is Theorem 3...34 CHAPTER II. W v m and P(W > u) = P (m > u) = 0(u).o. Y1) has the same distribution as (Y1..1 is equivalent to WN D MN = (Z1 + . .i.1.. then the restrictions of Fx.1. a Markovian change of measure as in Theorem 2... SOME GENERAL TOOLS AND RESULTS (d) m = inf. equivalently. . (d) #.1. Thus the assertion of Theorem 4. . w. .1 is actually not necessary .
and define Ln by (4.f.4.g.nrc(a )} (4. where h (y) = g(0. The corresponding likelihood ratio is Ln = exp {a (Y1 + • • • + Yn) .s. Then the change of measure in Theorem 2. the changed increment distribution is F(x ) = E[h(Y).g.. the random walk property implies Ex f (Y1) = Eo f (Y1 ). Y) = h(Y ) a. = 1 for all n and x.. of F).5 corresponds to a new random walk with changed increment distribution F(x) = e'(a) Jr e"'F(dy) . In particular. Y1). y ).3) 1Pxa. e. for some function h with Eh(Y) = 1. h(Yn) (4. Then the change of measure in Theorem 2. Proof If (4... cf.4. Since L1 has the form g (Xo. (a) = log F(a] is the c. this means E(g(x.3) holds for n = 1. u A particular important example is exponential change of measure (h(y) = e°y'(") where r.Y2) = h(1'i)h(I'a).. implying g (x. Breiman [78] p.5 corresponds to a new random walk if and only if Ln = h(Y1) . Conversely..s. In that case. then n n Ex [f f = Ex H fi a( YY) i=1 i_1 ( Y=) h(YY) H Ef=(Y=)h(Y=) = II J fi(Y )P( d) from which the random walk property is immediate with the asserted form of F. and so onforn =3..4 Let {Ln} be a multiplicative functional of a random walk with E_L. Y ) f (Y)] = E[g(O.4). Y < x].3) holds. We get: Corollary 4. 100 ). Y) f (Y)] for all f and x.5 Consider a random walk and an a such that c(a) = log F[a] = log Ee° ' is finite. For n = 2.4) ({Ln} is the familiar Wald martingale . . (4. we get L2 = L1 (L1 o91 ) = h(Y1)g(X1. RANDOM WALKS INDISCRETE OR CONTINUOUS TIME 35 Proposition 4.
The traditional formal definition is that {Xt} is Rvalued with the increments Xt(1)_t(o). the interpretation is that the rate of a jump of size x is v(dx) (if f of Ixlv (dx) = oo.t(i ..5) Note that the structure of such a process admits a complete description. but we omit the details ). {Xt} can be written as the independent sum of a pure drift {pt}. The simplest case is 3 = JhvMM < oo. say the beginning of each month or year .. with premium rate p. In discrete time. (4. In risk theory.Xn < x). or imbedded into continuous time processes . v2 = 0 and v = 3B).Ft] = Eof (X.6) More precisely. the claim surplus process for the compound Poisson risk model . An equivalent characterisation is {Xt} being Markov with state space R and E [f (Xt+e . they arise as models for the reserve or claim surplus at a discrete sequence of instants. Roughly.). . . the pure jump process is given by its Levy measure v(dx).. Xt(2)_t(l). (4. A general jump process can be thought of as limit of compound Poisson processes with drift by considering a sequence v(n) of bounded measures with v(n) T v. {Xt} is a random walk.36 CHAPTER II. Xt =Xo+pt+oBt+Mt. we are . However. the tradition in the area is to use continuous time models.7) for all e > 0. Xt (n)t(n1) being independent whenever t(O) < t(1 ) < . < t(n) and with Xt( i)_t(i_l) having distribution depending only on t(i) .1). e f x:IxJ>e} v(dx) < oo (4. corresponds to a process with stationary independent increments and u = p. say by recording the reserve or claim surplus just before or just after claims (see Chapter V for some fundamental examples). The appropriate generalization of random walks to continuous time is processes with stationary independent increments (Levy processes). SOME GENERAL TOOLS AND RESULTS Discrete time random walks have classical applications in queueing theory via the Lindley process representation of the waiting time . a Brownian component {Bt} (scaled by a variance constant) and a pure jump process {Mt}. however. In continuous time.Xt)I..e. this description needs some amendments. which corresponds to the compound Poisson case: here jumps of {Mt} occur at rate 0 and have distribution B = v/0 (in particular . i. given by a the increment distribution F(x) = P(Xn+l . see Chapter V. a positive measure on R with the properties e J x2v(dx) < oo.
jxJ v(dx) < oo. having Poisson arrivals with rate . Furthermore. oo]. Now consider reflected versions of processes with stationary independent increments. 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.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. 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) = 1.1)v(dx) (4. 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). A different interpretation is as the workload or virtual waiting time process in an M/G/1 queue.e. is easily seen to be f3pB < 1. where c(a) = ap + a2a2/2 + f 00 provided the Levy measure of the jump part {Mt} satisfies f".v.Q and distribution B of the service times of the arriving customers. [The condition for V < oo a. i. Chapter III.s. Corollary 4.6 In the compound Poisson risk model with constant premium rate p(r) .min Xt (4. T) = P(VT > u). Then the storage process {Vt} has constant release rate 1. Proposition 4. and the reflected version is then defined by means of the abstract reflection operator as in (4.10) .2). then Ee'(xtxo) = Eoeaxt = etx(a).1.3 and decreases linearly at rate 1 in between jumps. O(u. and b(u) = P(V > u). (ex .6).] Processes with a more complicated path structure like Brownian motion or jump processes with unbounded Levy measure are not covered by Section 3.8) O<t<T (assuming Wo = Xo = 0 for simplicity). has upwards jumps governed by B at the epochs of a Poisson process with rate . V E [0.7 If {Xt} has stationary independent increments as in (4. defined as a system with a single server working at a unit rate. WT = XT . cf. 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. VT + V for some r.
u Note that (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) .38 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Proof By standard formulas for the normal distribution. then the changed parameters in the representation (4. e. t.g. Xt Xo) with E2Lt = 1 for all x. Theorem 4 . This is of course no coincidence since the distribution of Xl . use the representation as limit of compound Poisson processes.Xt)I Ftl = Eof (X8)g(s. Proof For the first statement .1 that E eaMt = exp fmoo In the general case . In particular.10) is the LevyKhinchine representation of the c. . let e" (a ) = Eoeaxl.1. Xt +B . Q2 = v2. Chung [86]). By explicit calculation . 8 Assume that {Xt} has stationary independent increments and that {Lt} is a nonnegative multiplicative functional of the form Lt = g(t.6) are µ = µ + Oo2 . Then the Markov process given by Theorem 2.Xo is necessarily infinitely divisible when {Xt} has stationary independent increments. we show in the compound Poisson case ( IlvIl < oo) in Proposition III.g. X8) = Eof (X8)L8 = Eof (X8)• For the second. we use the characterization (4.1 ) v(dx) .4 o) aµ + ((a + 0 ) 2  0 2 )o 2 /2+ r w J 00 (e (a + 9)x  a 9x )v(dx) 00 a(µ + O 2) + a2a2 / 2 + J (eax .1)eexv(dx).Xt)I'Ftl = E [f(Xt+B .11) (eax . 5 has stationary independent increments as well..Xt)g(s. Eea(µt + QBt) = et{aµ +a2OZ/2}. v(dx) = e9xv (dx).5) and get E [f(Xt+B .xo)tk ( e). of an infinitely divisible distribution (see. (4.f.Xt)L8 o 0tIFt] = E [f (Xt+s . if Lt = e9(xt .
b with b(x) > 0 for all x such that b(x) > 0). is defined as a bivariate Markov process {Xt} = {(Jt. (5.2.. MARKOV ADDITIVE PROCESSES 39 Remark 4. dB/dB = b/b when B.3 =. Ei instead of P2.9 If X0 = 0.4). u 5 Markov additive processes A Markov additive processes. the corresponding claim sizes .3 and claim size distribution B. then the martingale {eex(t)tk(e)} is the continuous u time analogue of the Wald martingale (4.1) For shorthand .0 in the following. 0.o[f (S8)g(J8)]. 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 . MAP stands for the Markovian arrival process discussed below. .8..11 For an example of a likelihood ratio not covered by Theorem 4.3B[B].10 Let Xt be the claim surplus process of a compound Poisson risk process with Poisson rate .g.(3B(dx). we write Pi. where . 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. a = 0.. and let the Px refer to the claim surplus process of another compound Poisson risk process with Poisson rate. Ei. U2. Thus (since µ = p = 1. . Example 4 . b = a = 0) the changed process is the claim surplus process of another compound Poisson risk process with Poisson rate . it is then easily seen that Lt = H dB(Ui) i:o. Then we can write v(dx) _ /3eOxB(dx) = / (dx).(3 = .Ft] = Ejt.0.. Example 4 .3 and claim size distribution B # B.St)g(Jt+s)I. v(dx) _ .l3 and claim u size distribution B. <t whenever the RadonNikodym derivative dB/dB exists (e. Recalling that U1. are the arrival times and U1.5. B have densities b. abbreviated as MAP in this section2. the structure of MAP's is completely understood when E is finite: 2and only there . one reason is that in parts of the applied probability literature. corresponding to p = 1. As for processes with stationary independent increments .. B(dx) = B[9] B(dx).
a MAP is the same as a semiMarkov or Markov renewal process.jEE• On an interval [t.. . An alternative description is in terms of the transition matrix P = (piA.it = A.J1=j)= Fij (dx) Pij In simulation language. which we omit and refer to Neveu [272] or cinlar [87]. Y2. with the Y„ being interpreted as interarrival times.g. In continuous time (assuming Dpaths).Sr_1.. vi(dx) in (4.o(Ji = j..[a) = (Ei[easl.i. Jn = j.. a jump of {Jt} from i to j # i has probability qij of giving rise to a jump of {St} at the same time..6) depending on i. Y1 E dx) where Y„ = S„ . As an example. oo). v. by generating Yn according to Hij when J„_1 = i. t+s) where Jt . the distribution of which has some distribution Bij. (That a process with this description is a MAP is obvious.jEE (here pij = Pi(J1 = j)) and the probability measures Hij(dx)=P(Y1 EdxlJo=i.40 CHAPTER H. 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. consider the matrix Ft [a] with ijth element least Ei .f. a MAP is specified by the measurevalued matrix (kernel) F(dx) whose ijth element is the defective probability distribution Fij(dx) = Pi. As a generalization of the m.) If E is infinite a MAP may be much more complicated. let {Jt} be standard Brownian motion on the line. the converse requires a proof. If all Fij are concentrated on (0. SOME GENERAL TOOLS AND RESULTS In discrete time. 1 J1 ='^])iJEE = (Fij[a])i . Fn[a] = F[a]n where P[a] = P . {St} evolves like a process with stationary independent increments and the parameters pi.1 For a MAP in discrete time and with E finite. this means that the MAP can be simulated by first simulating the Markov chain {J„} and next the Y1.9 EE = (iii&ij[a])i j EE . {Jt} is specified by its intensity matrix A = (Aij)i. Proposition 5. In addition.
a= .1)) . u Proposition 5.1) } (recall that qjj = 0). MARKOV ADDITIVE PROCESSES Proof Conditioning upon (Jn.qkj + k?^j qkj Bkj [a] } = Ei [east. Sn ) yields Ei[easn+ '. Proof Let {Stt) } be a process with stationary independent increments and pa rameters pi . assume that the Markov chain/process {Jt} is ergodic. kEE Jn = k]Ek[e"Y" .ijgij(Bij[a] . Jt = k] { 1 . J1 = A which in matrix formulation is the same as Fn+1 [a] = Fn[a]F[a]. j E E) and So = 0. Jn+1 = A] = 41 Ei[ e 5„. Jt = j] is given by etK[a]. Jt = k] { xk kEE j la] .2 Let E be finite and consider a continuous time Markov additive process with parameters A.5.1 )v(dx). Then. up to o( h) terms.4c). 00 r(i) (a) = api + a2ot /2 + f (e° . vi(dx). pi. In matrix formulation . where K[a] = A+ (r. we infer that in the discrete time case the .(')(a)) diag + (). Jt = j] Ejesh'^ + E Ak j hEi [ease . By PerronFrobenius theory (see A. Then the matrix Pt[a] with ijth element Ei [east. vi(dx) (i E E). Jt = j] (1 + htc (j) (a)) j + Ak j qk j (Bk +h E Ei [east . u In the following. Bij (i. 013 . aSt h = (1 + Ajjh) Ei [east . \ diag Ft[a] = Ft[a]K. which in conjunction with Fo[a] = I implies Ft[a] = etK[a) according to the standard solution formula for systems of linear differential equations. this means that F't+h [a] = Ft[a] II+h(rc(i)(a)) +hA+h(Aijgij(Bij[a]1)) I. qij.
Eikjt = ttc'(0) + ki . Eie"sth^a) = e'Pt[a]h( a) = e.h(a)vva)etw(a). and write k = k(°).e=e°tk. The function ic(a) plays in many respects the same role as the cumulant g.4 Eie"sth(a) = h=a)et?("). (5.tK(a)h(a) J jj it L o is a martingale. Corollary 5. a. Since v(").r. its derivatives are 'asymptotic cumulants'. 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. Proof For the first assertion.etx It then follows that E feast+^(t+v)K(a)h(a) I ^tl l . Corollary 5. cf.c(a) (and h(")). Jt = j] . Jeast. The corresponding left and right eigenvectors v("). Proposition 5.4c). cf.42 CHAPTER II. as will be seen from the following results. Furthermore. we are free to impose two normalizations. u Let k(a) denote the derivative of h() w. Yrh(a ) = 1. h(") may be chosen with strictly positive components.t. of a random walk.2) where 7r = v(°) is the stationary distribution. h(") are only given up to a constants. just note that [a]h(a) = eietK (a)h(a) = etK(a)h(a).f. Then h(°) = e.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). Corollary 5. We also get an analogue of the Wald martingale for random walks: Proposition 5.5 EiSt = tK'(0) + ki . . and we shall take V(a)h(a) = 1.4.3 Ei [east.7. Proof By PerronFrobenius theory (see A. In particular. and appropriate generalizations of the Wald martingale (and the associated change of measure) can be defined in terms of .
(5. Squaring in Corollary 5.g.2ttc (0)Evkjt + 0(i).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. u The argument is slightly heuristic (e. summing and letting a = 0 yields E„ [St + 2Stkj.") }) . 8 Also for E being infinite (possibly uncountable ). subtraction yields Vary St = tic"(0) + O(1). t im v^"St = '(0) Proof The first assertion is immediate by dividing by tin Corollary 5.+ k. 43 Ei [Steast h(a) + east k^a)1 = et"(a) (kia) + tic (a)hia)) . In the same way. Corollary 5. one obtains a generalization of Wald's identity EST = Er • ES. More precisely. (5.a) + ttc (a)2hia ) Multiplying by v=.7 No matter the initial distribution v of Jo. the existence of exponential moments is assumed ) but can be made rigorous by passing to characteristic functions.3) Let a = 0 and recall that h(°) = e so that 0=°) = h(o) = 1. Remark 5 .. MARKOV ADDITIVE PROCESSES Proof By differentiation in Proposition 5.St]2 = t2/c'(0 ) 2 + 2ttc (0)vk .5.4. For the second . E=ST = tc'(0)E7.Eikjr . for a random walk: Corollary 5.4) .5. [E. Ee"st typically grows asymptotically exponential with a rate ic(a) independent of the initial condition (i.. the distribution of Jo).5 yields + W (a)k. t a oo. tam E tSt a (0). . ] = t2tc (0)2 + 2tK'(0)vk + ttc"(0) + O(1) .6 For any stopping time T with finite mean. Since it is easily seen by an asymptotic independence argument that E„ [Stkjt] u = trc'(0) E„kjt + O(1). there is typically a function h = h(") on E and a ^c(a) such that Ey a"st t" (") * h(x).e. we differentiate (5.
in particular that f is bounded. see. h(Jo) Lo is a Px martingale for each x E E.6) We shall not exploit this approach systematically.5) is a martingale can be expressed via the infinitesimal generator G of {Xt } = { (Jt. inconvenient due to the unboundedness of ea8 so we shall not aim for complete rigour but interpret C in a broader sense.(9) {Lt}t>o = .6. Then {Lt } is a multiplicative functional.44 CHAPTER IL SOME GENERAL TOOLS AND RESULTS for all x E E. 0 Proposition 5.5.6.5 defines a new MAP. St)} be a MAP and let 0 be such that h(Jt) OStt. 0) = n(a) h(i).for the present purposes. (5. Usually.e.. 1) (i. From (5. 0) = h(i )( 1 + ttc(a)).9 The condition that (5. however. where {Jt} is deterministic period motion on E = [0. xEE .s. In view of this discussion . we take the martingale property as our basic condition below (though this is automatic in the finite case).. s) = ea8h(i).4 that { h(Jt) easttK(a) L o (5. gha(i. St) } as follows. however. For t small . An example beyond the finite case occurs for periodic risk processes in VI. V. Given a function h on E. First. Jt = (s+t) mod 1 P8a. this leads to h(i) + tcha( i.f (x) tyo t provided the limit exists. We then want to determine h and x(a) such that Ejeasth (Jt) = etK(a)h(i). u forsEE). G is defined as Gf (x) = lim Exf (Xt) . Remark 5.5) is a martingale . this is. some extra conditions are imposed.10 Let {(Jt.1) one then ( at least heuristically) obtains lim Ex eaSv v a) K( v+oo nEx easttK(a)EJt east(vt)K(a) u[J = Ex easttk(a)h(Jt) It then follows as in the proof of Proposition 5. and the family {f LEE given by Theorem 2. let ha(i.3b and Remark VI.
c(0)e = tc(0)e .5..Qi < oo and Bi a probability measure.12 The expression for A means h(e) Aij = hie) Aij [1 + gij(Bij[0] i 0 j.ic(0)e = ic(0)Oh e) h(e) . if vi(dx) is compound Poisson.1) eft ea' f ij (dx) = Hij (dx) Hij [0] . We omit the details. this gives a direct verification that A is an intensity matrix: the offdiagonal elements are nonnegative because Aij > 0. MARKOV ADDITIVE PROCESSES Proof That { Lt} is a multiplicative functional follows from L8 ogt = h(Jt+s) es(St+ . Ai = µi + 0Q. in the discrete time case.7) In particular. and by A = Oh(°) K [0]Oh(e ) vi(dx) = e"xvi (dx). In the infinite case . one can directly verify that (5. 0 < qij < 1 and Bij [0] > 0. u Theorem 5. 0<b<oo. 1 + q(b . ^i = of qij Bij [0] 1 + qij ( Bij [0] . qij = r. then also vi (dx) is compound Poisson with e Ox ^i = /3iBi[0].10 is given by P = eK(e) Oh e) F[e]Oh('). Then the MAP in Proposition 5.St)sl(e) h(Jt) 45 The proof that we have a MAP is contained in the proof of Theorem 5.11 below in the finite case. Bi(dx) = Bi(dx).1) . Bi [0] Remark 5. That 0 < qij < 1 follows from the inequality qb <1.11 Consider the irreducible case with E finite..1) holds for the P. (5. 0<q<1. Here Oh(e) is the diagonal matrix with the h=e) on the diagonal.(0)j.tc(0)e = 0 . Bi. vi (dx) = f3 Bi(dx) with .7(dx) Bij [0] Bij(dx) in the continuous time case . . That the rows sum to 1 follows from Ae = Oh(e) K[O]h(B) . In particular.
Now we can write K[a] =A+A ) ( K[a + 0] . First note that the ijth element of Ft[a] is etK(e)Ej [e(a+B)st E:[east Jt = j] = Ej[Lteas' . this means that Ft[a] = etw ( e)Ohc) Ft[a + 9]oh (e) (5. (dx) of a process with stationary independent increments follows from Theorem 4.8). (dx). Yi E dx.tc(0)I. Jt = A. in continuous time ( 5. .t.tc(0)' )Ah() = Oh(o) K[a + 0]Oh() .13) for matrixexponentials . Jt = j] = hie) .K [O])Oh(e) (0) l + ( A + (tc(') (a + 0) .8) h(. Jl = j) = Ei[Lt. H1. Jl = j] :(Yi E dx. since Hij. are probability measures .11. Letting a = 0 yields the stated expression for A. it follows that indeed the normalizing constant is H1 [0].8) yields et'[a] = Ohie )et (K[a +e]K(e)I)Oh(°) By a general formula (A.tc(') (8)/ d)ag h 7 Aiiii (Bii[a + 0] . Similarly. In matrix notation .tc (') (0) corresponds to the stated parameters µ.46 CHAPTER II.e) Consider first the discrete time case .. SOME GENERAL TOOLS AND RESULTS Proof of Theorem 5. .Bay [0]) That k(') (a + 0) . v. is absolutely continuous w..8. Here the stated formula for P follows immediately by letting t = 1. v= . Ji = j) h(e) eeyK(B)p h(8) h(e) eexK ( h=e) e)Fi. This shows that F. a = 0 in (5. this implies k[a] = A 1 ) (K[a + 0] . Hence the same is true for H=j and H. Further Fib (dx ) = P=(YI E dx..r. F:j with a density proportional to eei .
Much of the pioneering was done in the sixties in papers like Keilson & Wishart [224]. oo). is slightly less general than the present setting. [261]. hardly a single comprehensive treatment. Though the literature on MAP's is extensive. [226] and Miller [260]. 6 The ladder height distribution We consider the claim surplus process {St } of a general risk process and the time 7. see also Fuh & Lai [149] and Moustakides [264]. 7+ < oo). 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). i. < x) = 11 (S. has no mass on (oo. the literature on the continuous time case tends more to deal with special cases.6. THE LADDER HEIGHT DISTRIBUTION 47 Finally note that by (5. Conditions for analogues of Corollary 5. The closest reference on exponential families of random walks on a Markov chain we know of within the more statistical oriented literature is Hoglund [203]. For the Wald identity in Corollary 5. [262] in discrete time. however.e. Notes and references The earliest paper on treatment of MAP's in the present spirit we know of is Nagaev [265].+ < x. [225].3 for an infinite E are given by Ney & Nummelin [266]. .1). 0].1) = Aij4ij(Bij[a] .(u) = inf {t > 0 : St > u} to ruin in the particular case u = 0 . however. Note that G+ is concentrated on (0. which.6.)Ajjgij(Bij[a+0] ..Bij[0]) = hjel)ijgijBij[0](Bij[a] . there is.. an extensive bibliography on aspects of the theory can be found in Asmussen [16]. Write r+ = T(0) and define the associated ladder height ST+ and ladder height distribution by G+(x) = 11 (S.7). h. and is typically defective.
ST+(1) and so on.1. oo). they have a semiMarkov structure (but in complete generality. G+ is given by the defective density g + (x) =. it follows that for g > 0 measurable. the dependence structure seems too complicated to be useful).2) . where basically only stationarity is assumed. 6.5 below. In other cases like the Markovian environment model. The main result of this section is Theorem 6. In any case. Theorem 6 . the sum of all the ladder steps (if rl > 0. g(y)R+(dy) = E f g(St)dt. On Fig. o 00 (6.00 ). define the prer+occupation measure R+ by R+(A) = E f o "o I(St E A. Also.1.2.(3B(x ) = pbo(x) on (0. i. has no mass on ( 0. by approximation with step functions . For the proof of Theorem 6. the second ladder point is ST+(2) where r+(2) is the time of the next relative maximum after r+(1) = r+. see Fig. In simple cases like the compound Poisson model. a fact which turns out to be extremely useful. To illustrate the ideas. SOME GENERAL TOOLS AND RESULTS M ST+(2) Sr. the ladder heights are i.. 0]. the second ladder height (step) is ST+(2) .i. which gives an explicit expression for G+ in a very general setting. Recall that B(x) = 1 . = ST+(1) Figure 6. there are only finitely many).T+ > t)dt = E f 0T+I(St E A) dt.e. 0 f T+ (6.1) The interpretation of R+(A ) is as the expected time {St} spends in the set A before T+. at present we concentrate on the first ladder height.. The first ladder step is precisely ST+. 1 For the compound Poisson model with p = 01LB < 1. 6.e.1.d.48 CHAPTER K. we shall first consider the compound Poisson model in the notation of Example 1. Here bo(x) _ B(x)/µB. and the maximum M is the total height of the ladder. R+ is concentrated on (oo.B(x) denotes the tail of B. i.1 The term ladder height is motivated from the shape of the process {Mt} of relative maxima. Thus.
2(a): T+ > t Figure 6.O<t<T) = P(STEA. see Fig.2.ST<St. since the distribution of the Poisson process is invariant under time reversion. .O<t<T). P(STEA. 0 < t < T) P(STEA.St<0. {St }o<t<T is constructed from {St}o<t<T by timereversion and hence.6.T+>T) = P(STEA. 0 < t < T. 0]. That is.2 R+ is the restriction of Lebesgue measure to (00.ST<St.0<t<T) = F(ST E A. St S* t a Figure 6. has the same distribution as {St}o<t<T. 6.ST_t. THE LADDER HEIGHT DISTRIBUTION Lemma 6 . 49 Proof Let T be fixed and define St = ST .2(b): r+ < t Thus. ST < ST_t.
3 Lemma 6 . E A} precisely when r+ > t. for A C (0. it follows that R+ (A) is the expected time when ST is in A and at a minimum at the same time . 6.St _)I(r+ > t). The probability of this given { Su}u<t is B(A . That is.y)R+(dy) 00 Proof A jump of {St} at time t and of size U contributes to the event IS. s.y) (here we used the fact that the probability of a jump at u t is zero in the second step. and since the jump rate is /3.3 where the bold lines correspond to minimal values.50 CHAPTER II. oo). Fig.2) in the last). .r. and (6. G+(A) = Q f 0 B(A . U + St_ E A. But since St 4 oo a.T+ > t] dt 0 T+ _ /3E f g( St) dt = 0 f g(y) R+(dy) 0 00 where g(y) = B(A .St). this is just the Lebesgue measure of A. oo). Figure 6. we get G+ (A) = f 00 /3 dt E[B(A ..3 G+ is the restriction of /3R+*B to (0. T+ > t] 0 _ /3 f E[B( A .. cf.St _). SOME GENERAL TOOLS AND RESULTS Integrating w.t dT.
) where ak = Ti + • • • + Tk .z)B(dz) _ f I(x < z)B(dz) _ f (x). i. oo) x (0. 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. the first component representing time (the arrival time o.S8 )t> o = {St }t>o for all s > 0. The marked point process .1. THE LADDER HEIGHT DISTRIBUTION 51 Proof of Theorem 6. 4 (the points in the plane are (ak . Uk) (k = 1. Nt St =>Uk k=1 t where Nt = max{k = O.s. Lemma 6. . cf. .. 2..s > 0). U k)} k=1 a is as a marked point process M *. h]} /h (by stationarity.. we define the arrival rate as E# { k : ak E [0 .3 yields g+ (x) = . obviously.1 With r+(y) = I(y < 0) the density of R+. The points in the plane (marked by x on Fig. 6 . We call M * stationary if M* o B8 has the same distribution as M* for all s > 0. Fig.:T1 +•••+Tk <t}.Q f r+(x . The sample path structure is assumed to be as for the compound Poisson case: {St*} is generated from interclaim times Tk and claim sizes Uk according to premium 1 per unit time. 6 . In the stationary case. 0 Generalizing the setup. assuming basically stationarity in time and space. The traditional representation of the input sequence {(TT.* ) and the second the mark (the claim size Uk ). {St+8 .4) are (ak..M o 08 shifted by s is defined the obvious way. we consider the claim surplus process {St }t>o of a risk reserve process in a very general setup.6.e. as a point process on [0.. Uk) for those k for which ak .4). oo).T+ < oo). this does not depend on h).e. i. this is equivalent to the risk process {St*} being stationary in the sense of (6.
we define its Palm version M as a marked point process having the conditional distribution of M* given an arrival at time 0 . See.. and let T = T2 denote the first proper interarrival time . The two fundamental formulas connecting M* and M are Eco(M) = aE E. where TI = 0.QiBi(dx).. This more or less gives a proof that indeed (6. most often one takes h = 1). h] and the sum approximately ^o(M*)I(ul < h). k: vk E [0. h] Eco(M*) = 1 E f co(M o Bt)dt. We represent M by the sequence (Tk. Uk) k=1. o. letting h J.5) does not depend on h. where T is the first arrival time > 0 of M and h > 0 an arbitrary constant (in the literature. Assume {Jt} irreducible so that a stationary distribution 7r = (1i)iGE exists. 0. i.4 Given a stationary marked point process M*..s..g.52 CHAPTER II. Note also that (again by stationarity) the Palm distribution also represents the conditional distribution of M* o Ot given an arrival at time t. of (6. As above . Section 5) which has pure jump structure corresponding to pi = a = 0. i 1 U2 Us 1_ 0 or Q2 $ U3 *1 L 0 7 X I 11 1 Figure 6. Oh becomes the approximate probability F(ri < h) of an arrival in [0. SOME GENERAL TOOLS AND RESULTS M* U. = 0 . V(M* o eak ). e. . Sigman [348] for these and further aspects of Palm theory.2. the r.5) represents the conditional distribution of M* given vi = 0.4 Consider a finite Markov additive process (cf.e. h. vi(dx) = . Example 6 .
O fo "o F(x)dx = . let U0 be a r. we get a marked point process generated by Poisson arrivals at rate /3i and mark distribution Bi when Jt = i.5.p. This follows by noting that iP*(0) = IIG+JI = J0 "o g+(x)dx = . Note in particular that the Palm distribution of the mark size (i. First choose (Jo_.p.O for i # j. It follows that we can describe the Palm version M as follows . and by some additional arrivals which occur w. Then the ladder height distribution G+ is given by the (defective) density g+(x) = . After that.e./. let the arrivals and their marks be generated by {Jt} starting from Jo = j.OEU0.6 Under the assumptions of Theorem 6. an arrival for M* occurs before time t + dt w.6iBi + Aijgij Bij j#i iEE iEE 0 Theorem 6 . 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 . the ruin probability .. THE LADDER HEIGHT DISTRIBUTION 53 Interpreting jump times as arrival times and jump sizes as marks.oo a.OF(x). 5 Consider a general stationary claim surplus process {St }t>o. If Jt_ = i. Assume that St * .6. Jt = j is iri(3i /.s.p. Before giving the proof. qij when {Jt} jumps from i to j and have mark distribution Bij. having the Palm distribution of the claim size and F (x) = F(Uo < x) its distribution . aij for (i. the distribution of Ul) is the mixture B = E aii Bi + aij Bij J = j#i !i J.O for i = j and iriAijgij/. j) and let the initial mark Ul have distribution Bi when i = j and Bij otherwise. v. . and that p = 0EU0 < 1. A stationary marked point process M* is obtained by assigning Jo distribution Tr. the probability aij of Jt .= i.*(0) with initial reserve u = 0 is p = /3EU0. we note: Corollary 6. Jo) w.
$St_ u.s.St*_ u.e.. .o.).. { Su}0<u<t is distributed as a process {Su} . Proof of Theorem 6.. 105) shows that one can assume w. oo) x (0 . The result is notable by giving an explicit expression for a ruin in great generality and by only depending on the parameters of the model through the arrival rate 0 and the average ( in the Palm sense) claim size EU0. . oo) p(t) = P(St EA.1] here the r . A standard argument for stationary processes ([78] p. SOME GENERAL TOOLS AND RESULTS V` (0) = E E Uk k: ak E [0. Then clearly * G+ (A) = P(ST+ E A) = Consider a process { f p(t)f3dt. The last property is referred to as insensitivity in the applied probability literature. 6.Su_ <0.Su<0. in (oo. h.Mt)... the arrival times 0 < 0'1 < Q2 < .l.5.. in (0. and the kth preceding claim arrives at time t . 0).54 By (6. It follows that for A C (0..St <. moves down linearly at a unit rate in between jumps and starts from S0 = U.0<u<t) = P(St EA. 0<u<t) = P(St EA.A. which makes an upwards jump at time .. has a very simple interpretation as the average amount of claims received per unit time .o<u<t where a claim arrives at time t and has size Uo.Su< 0.0<u<tIAt) = P(St EA. . oo)). T+ = t given the event At that an arrival at t occurs . are point processes on (oo .5).5. Let p(t) be the conditional probability that ST+ E A. oo ) and the arrival times 0 > 0_1 > a_2 > .g.o. that M* and M have doubly infinite time (i. (k = St}t>o 1.St<Su.0<u<t) = P(StEA. 2. The sample path relation between { Su } and { Su } amounts to S„ = St .Q_k and has size U_ k. CHAPTER H.(left limit) when 0 < it < t and is illustrated on Fig . We then represent M by the mark (claim size ) Uo of the arrival at time 0.0<u<tIAt) = P(St EA. Now conditionally upon At . the mark at time Qk is denoted by Uk.
the support of L has right endpoint U0. t a oo. G' (A) = 3 f P(St E A. Fig.s.5.6. In Fig. 6. NIt)dt . time instants corresponding to such minimal values have been marked with bold lines in the path of { St}. Mt)dt = i3EL(A) o"o .5 where it = { St < Su. 2 therefore immediately shows that L(dy) is Lebesgue measure on (oo.. the left endpoint of the support is oo. THE LADDER HEIGHT DISTRIBUTION 55 { A Su}0<u<t U0 U0 \t tt u>0 N U_1 Figure 6. Thus. Since So = U0. A sample path inspection just as in the proof of Lemma 6 . cf. Uo]. 6. 0 < u < t } is the event that { Su } has a relative minimum at t . and we let L(dy) be the random measure L(A) = fo°° I(St E A. and since by assumption St * oo a.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.
.5 is due to Schmidt & coworkers [48].1).56 CHAPTER II.6 is Bjork & Grandell [67]. A further relevant reference related to Corollary 6.2. [147]. 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. [263] (a special case of the result appear in Proposition VI.
{Rt} and the associated claims surplus process {St} are given by Nt Nt Rt = u+t EUi. . Panjer's recursion ( Corollary XI. with common distribution B. and independent of {Nt}.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. 4. i=1 i=1 An important omission of the discussion in this chapter is the numerical evaluation of the ruin probability.d. • the premium rate is p = 1. For finite horizon ruin probabilities . i.. see Chapter IV. say. and assume that • { Nt}t>o is a Poisson process with rate j3.6) and simulation methods ( Chapter X). exact matrixexponential solutions under the assumption that B is phasetype (see further VIII.. 3). U2.4 below . Thus . are i. i. being of the form Rt = Rt+Bt + Jt where {Rt } is a compound 57 . • the claim sizes U1.e. Some possibilities are numerical Laplace transform inversion via Corollary 3. It is worth mentioning that much of the analysis of this chapter can be carried over in a straightforward way to more general Levy processes . St = uRt = EUi t. A common view of the literature is to consider such processes as perturbed compound Poisson risk processes .
Proof It was noted in Chapter I that p . we shall start by giving the basic formulas for moments. where K(k) (0) is the kth derivative of is at 0.1). and that B(k)[0] = Pak). (d) The kth cumulant of St is tf3p(k) for k > 2. e .)3t (fit' k t} = etk(8) exp {st '3t + B[s]f Finally.Rt.u . [324]. {Bt} a Brownian motion and {Rt} a pure jump process. cumulants ...g.1) = t(p .'s etc. P = PAB = 1/(1 + rl) Proposition 1.+Uk)P(Nt = k) k=O e8t k=O B[s]k . 1 Introduction For later reference.t = fltpB . Dufresne & Gerber [126]. say stable Levy motion.t = E E [ U k k=1 k=1 Nt .t = E[Ntµs] . A more formal proof goes as follows: Nt r Nt ESt = E > U k .1 (a) ESt = t(13µ$ . Write pB^) = EUn' YB = Pali = EU. (b) Var St = t. See e. We do not spell out in detail such generalizations. for (d) just note that the kth cumulant of St is tic(k) (0).t = t(p . and this immediately yields (a). we get Ee8st = 00 e8t c` Ee8 (U1+.6pBa).58 CHAPTER III. Schmidli [319].g. Furrer [150]. For (c). 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]. (c) Ee8St = et" (8) where c(s) = f3(B[s] . of the claim surplus St .1) . 0 .f.1 is the expected claim surplus per unit time. THE COMPOUND POISSON MODEL Poisson risk process.s. m.1).
(d) If 17 = 0. We return to this approach in Chapter V.Tk are i. if t = nh + v with 0 < v < h..4.. Proof We first note that for u. In particular.. lim supt. where Tk is the time between the kth and the (k . 2.2 (DRIFT AND OSCILLATION) St/ta3'p1 ast >oo.3EU01 = 1µs where rt is the safety loading. For the proof.V. and the value is then precisely v.. In this way. then St 4 co.3) is proved similarly. Indeed. we have Sok . cf. v > 0. St = oo.Tk. however.1 = .1)th claim. we need the following lemma: Lemma 1. . which is often used in the literature for obtaining information about {St} and the ruin probabilities. u + v]. then St> SnhV>Snhh. we get a discrete time random walk imbedded in the claim surplus process {St}. Obviously.Sok_l = Uk .1 is the same as if {St} was a random walk indexed by t = 0. (b) If 77 < 0. the Uk . (c) If 77 > 0. Sn+0 . rather to view {St} directly as a random walk in continuous time. (a) No matter the value of 77. meaning that the increments are stationary and independent.S„ attains its minimal value when there are no arrivals in (u. so that {Sok } is a random walk with mean EUET = EU. 1.i. The right hand inequality in (1. INTRODUCTION 59 The linear way the index t enters in the formulas in Proposition 1. then Snh . obviously 0(u) = F(maxk Sok > u).3 If nh < t < (n + 1)h. St = oo. and there are at least two ways to exploit this: Recalling that ok is the time of the kth claim. The point of view in the present chapter is. then St 00.h < St < S(n+1)h + h. then lien inft. II.d. Here is one immediate application: Proposition 1. S„+V > S„ . The connections to random walks are in fact fundamental. For example.1.
.o.5 The limiting distribution of St . 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. Snh/n a4' ESh = h(p . if P(M > 0) = 1. {Snh}n=o. Thus using Lemma 1. There is also a central limit version of Proposition 1. h.2.1.. Corollary 1. Proof The case of 17 < 0 is immediate since then M = oo by Proposition 1. Notes and references All material of the present section is standard. then {St} upcrosses level 0 a.2. {Snh}n=o. 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 . THE COMPOUND POISSON MODEL Proof of Proposition 1. lim supn_. and < 1 for all u when 77 > 0. 169) stating that lim infra.... Remark 1 . hence by induction i._. Considering the next downcrossing (which occurs w.. at least once. is a discrete time random walk. is a discretetime random walk for any h > 0..p. 1 since St 4 oo) and repeating the argument.2: Proposition 1.6 Often it is of interest to consider size fluctuations.t .3. The general case now follows either by another easy application of Lemma 1. 2h.. 0 Snh = 00. and hence by the strong law of large numbers. This contradicts u St400. it suffices to prove 4'(0) = F(M > 0) < 1.. For any fixed h. However. u 307). p. (c) are immediate consequences of (a).60 CHAPTER III. and (b). If rl > 0.s.4 The ruin probability 0(u) is 1 for all u when 77 < 0.1) as t 4 oo is normal vtwith mean zero and variance )3µsz) Proof Since {St}t>o is a Levy process (a random walk in continuous time).1. h A similar argument for lim sup proves (a). and hence it folz lows from standard central limit theory and the expression Var(St) = tf3pB (Proposition 1. it is seen that upcrossing occurs at least twice. Part (d) follows by a (slightly more intricate) general random walk result ([APQ]..3. Assuming that each risk generates claims at Poisson intensity /3 and pays premium 1 per unit time.1.. Snh u = 00 (the lemma is not needed for (d)).1). or by a general result on discrete skeletons ([APQ] p. where the size of the portfolio at time t is M(t).1(b)) that the assertion holds as t 4 oo through values of the form t = 0.
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. 11. Thus . 0 Alternatively. The following results generalizes the fact that the conditional distribution of the deficit ST(o) just after ruin given that ruin occurs (i. that r(0) < oo) is Bo: taking y = 0 shows that the conditional distribution of (minus) the surplus ST(o).34 or A. nevertheless.1 provides a representation formula for 0(u). where G+ is given n=0 by the defective density g+ (x) = 3B (x) = pbo(x) on (0.6. which we henceforth refer to as the PollaczeckKhinchine formula. but we shall be able to extract substantial information from the formula. the ladder heights are i. Fig. As a vehicle for computing tIi(u).IIG +II)EG+ .just before ruin is again B0. 1 The distribution of M is (1..1. equivalently. Summing over n. Theorem 2. p < 1.. 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}. Here bo(x) _ Proof The probability that M is attained in precisely n ladder steps and does not exceed x is G+ (x)(1 . Combined with i/i(u) = P ( M > u). and we shall here exploit the decomposition of the maximum M as sum of ladder heights. and we further get information about the joint conditional distribution of the surplus and the deficit.e. IV. n=0 (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 ). the formula for the distribution of M follows . Note that this . The expression for g+ was proved in Theorem 11.IIG+II) (the parenthesis gives the probability that there are no further ladder steps after the nth ).P) E PnBon(u) . cf.2. (2. i. The decomposition of M as a sum of ladder heights now yields: 00 Theorem 2 . This follows simply by noting that the process repeats itself after reaching a relative maximum. We assume throughout rl > 0 or. B(x)/aB. oo ). 1e.6. d. [APQ] Ch. we can rewrite the PollaczeckKhinchine formula as 00 (u) = P (M > u) = (1 .1) is not entirely satisfying because of the infinite sum of convolution powers. It is crucial to note that for the compound Poisson model.1.
the form of G+ is surprisingly insensitive to the form of {St} and holds in a certain general marked point process setup. 7r(0 ) < oo) = Q 3 Special cases of the PollaczeckKhinchine formula The model and notation is the same as in the preceding sections. As shown in Theorem 11.i. The proof of Theorem 11.5. [62]. where it requires slightly more calculation. f +b (b) the joint distribution of (ST( o). Theorem 2.5. ST(o )) given r (0) < oo is the same as the distribution of (VW. in this setting there is no decomposition of M as a sum of i.d. We assume rt > 0 throughout. ladder heights so that the results do not appear not too useful for estimating 0(u) for u>0. cf. ST(o)) is given by the following four equivalent statements: B(z) dz.62 CHAPTER III.2(a) is from Dufresne & Gerber [125]. However. Feller [143] or Wolff [384]. V is uniform on (0. Theorem 2 . (1 . In the risk theory literature. For the study of the joint distribution of the surplus ST(u)_ just before ruin and the deficit ST(„). and the conditional distribution of ST(o)_ given ST(o)_ = z is Bo z) The proof is given in IV. see for example [APQ]. Asmussen & Schmidt [49].6. see Schmidli [323] and references there. 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). Theorem A1. Again. W are independent.V)W) where V. . ST(o) > y. THE COMPOUND POISSON MODEL distribution is the same as the limiting joint distribution of the age and excess life in a renewal process governed by B. there is a general marked point process version.2 and it gives an alternative derivation of the distribution of the deficit ST(o) Notes and references The PollaczeckKhinchine formula is standard in queueing theory. (d) the marginal distribution of ST(o)_ is B0.just after ruin.1 is traditionally carried out for the imbedded discrete time random walk. (c) the marginal distribution of ST(o)_ is Bo . cf. 2 The joint distribution of (ST(o )_. cf. (a) 11 (ST(o)_ > x. 1) and W has distribution Fw given by dFyy/ dB(x) = x/µB. the PollaczeckKhinchine formula is often referred to as Beekman 's convolution formula.6. Beekman [61].
1 e ax = n1 (n . B0 is exponential with rate S and the result can now be proved from the Pollaczeck Khinchine formula by elementary calculations . the formula for P(O) holds in a more general setting. Let r ( x) be the failure rate of M at x > 0. Thus r(x) = S(1 . 3b Exponential claims Corollary 3.p.O)e(b0)x. As shown in 11.1)1 00 ( 1 . the result follows .3. 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 0(0) = p = Nl2B = 1 1 +71 Proof Just note that (recall that T+ = r(0)) 00 z/^(0) = I' (r+ < oo) = IIG+II = )3 f(x)dx =l3LB• Notes and references The fact that tp(u) only depends on B through µB is often referred to as an insensitivity property. and hence this overshoot has the same distribution as the claims themselves . Thus . however .. For a failure at x. The result can. Bon is the Erlang distribution with n phases and thus the density of M at x > 0 is (1 .p.p)pSe a ( l v)x = p( S . I. 0 . Integrating from u to oo.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.p) = S . then. hence without memory. But claims are exponential . the current ladder step must terminate which occurs at rate S and there must be no further ones which occurs w.6. use Laplace transforms.e. a further relevant reference is Bjork & Grandell [67]. also be seen probabilistically without summing infinite series .p) E pn S n x n.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. 1 . Alternatively.2 If B is exponential with rate S.
T+ <oo)+P(M> u. T+ <00) (3.+ >u. A variety of proofs are available .3). just insert the explicit form of G+.4) is similar (equivalently.2) Notes and references Corollary 3. (3.p + G+ * Z(u) = 1 .3) below. and weights 1/2 for each.3. cf.2).4) can be derived by elementary algebra from (3.s.y)f3 (y) dy.1.4) zu P(M > u .+ <u. Corollary 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.i(u) satisfies the defective renewal equation Z(u) = 1 . we use the PollaczeckKhinchine formula in Chapter IX to show that b(u) . (3.+ = y yields P(M>u. 0 Proof Write o (u) as P(M>u) = P(S. then 24 1 V. E.+ <U. (u) 35eu + 35e6u.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 .T+ <oo).h. u . and conditioning upon S. (b) use stopped martingales . 3c Some classical analytical results Recall the notation G+(u) = f^°° G+(dx).3. Then the first term on the r. The case of (3. is ?7+ ( u). II.p + f u Z(u .64 CHAPTER III.y)G+(dy) For the last identity in (3.1 p pBo(u).1) For a heavytailed B.g.S. if 3 = 3 and B is a mixture of two exponential distributions with rates 3 and 7. (Example VIII. the survival probability Z(u) = 1 . THE COMPOUND POISSON MODEL In VIII. (3. u + oo. (3.3)).y)/3B (y) dy. 2 is one of the main classical early results in the area. We mention in particular the following: (a) check that ip (u) = pe (60)u is solution of the renewal equation (3.3) Equivalently.S.y)G+(dy ) = f U V(u .
5) Proof We first find the m.. which yields the survival probability as 00 f u }t Z(u) = f f3eRtdt 0 from which (3. (3..p)s s /3 . see e.P)pB' (3./3B[s] .4) can be derived by elementary but tedious manipulations.PPB2) EM2 = PPB) + QZPBl 2(1 .4 The Laplace transform of the ruin probability is 65 fo Hence Ee8M 00 e8uiP(u)du .p) E p"Bo[s]" = 1 .f. by analytical manipulations (L'Hospital's rule) from (3. 0 Notes and references Corollary 3. [APQ] pp. We omit the details (see.Ps s(.g. numerical inversion of the Laplace transform is one of the classical approaches for computing ruin probabilities. The approach there is to condition upon the first claim occuring at time t and having size x .(3B[s] 1 .g. . In fact.3 .p)s .7) s +.7) and Corollary 3.5 can be found in virtually any queueing book.8) Proof This can be shown. e.1 Bo[s] = f oc. [APQ] pp. for example.s .7). g.p = (1 . 111112 or Feller [143].3 is standard .pBo[s] no (1 . Bo of B0 as m e8u B(du) = B[s] .s .5). Embrechts.(3 .3. 191).3 . eau B(u) du = f PB 3PB SPB 0 o (3.6) 00 = (I . Corollary 3. SPECIAL CASES OF POLLACZECKKHINCHINE Corollary 3. In view of (3. either of these sets of formulas are what many authors call the PollaczeckKhinchine formula. it is not surprising that such arguments are more cumbersome since the ladder height representation is not used.Ee8M) f ao e8' ( u)du = a8uP (M > u)du = 0 o 1 ( 1+ (1 .)3B[s]) (3. Of course.5).P)PB 2(1 ./3B[s] which is the same as (3.p)2 3(1 . Griibel & Pitts [132]. Griibel [179] and Thorin & Wikstad [370] (see also the Bibliographical Notes in [307] p. 206207). Also (3. Some relevant references are Abate & Whitt [2].5 The first two moments of M are 2 EM .
u)]k k! k0 The renewal equation (3.u) [N(k .9) follows for 0 < u < 1. For n < u < n + 1.6 If B is degenerate at p.Q (k 1 k= n  [O(k ./32(u .u) a)Qea" + (1 . eO('u) [)3(k .1)! k=1 u1 .66 CHAPTER III./3Z(u .u/p)]k ko k! Proof By replacing {St} by {Stu/p} if necessary.u + 1 )]k = QZ(u) .1 < u < n and let Z(u ) denote the r.Q) 3e.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. of (3.1).9) shown for n .u)]k k! (1 L3) 1: e_O(ku) NIN (k (k .1).9). differentiation yields Z'(u) _ /3Z(u) which together with the boundary condition Z(0) = 1/3 yields Z(u) _ (1/3) eAu so that (3.3I( 0<y<1)dy Z(y)/3I(0<u. .3+ 18+ J0 Z(uy).z/'(u) takes the form Z(u) L^J L. then p) 1: ep(k u/.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 .u) [p(k .h.Q) k=0 k! E e0( = /32(u) .)3(1 . Assume (3.u)]k1 ku+1) [/3( k . THE COMPOUND POISSON MODEL 3d Deterministic claims Corollary 3. differentiation yields Z(u) _ /3Z(u) . Z^ =eR(k.4) for Z( u) means f lhu Z(u) = 1.s.
g.4. and thus (4.(9).2).3B[9].f. we just have to multiply (4.2) (Here 9 is any such number such that r.3B = .2) shows that the solution is Ox [O]0].d.a.f.4) . 4 Change of measure via exponential families If X is a random variable with c.6 is identical to the formula for the M/D/1 waiting time distribution derived by Erlang [139]. (4.1) or equivalently.1 that c(a) = /3(B[a] .1) . in terms of the c. and define rce by (4.rc(9) = .(9) is welldefined.f.1) . We could first tentatively consider the claim surplus X = St for a single t.g. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 67 Since Z(n) = 2(n) by the induction hypothesis. See also Iversen & Staalhagen [208] for a discussion of computational aspects and further references. we set up . but will now be repeated for the sake of selfcontainedness.3) by t. (4. Formalizing this for the purpose of studying the whole process {St}.4) works as well. it follows that Z(u) = 2(u) for n<u<n+1.Qe(Bo[a] . corresponding to a compound Poisson risk process in the sense that for a suitable arrival intensity 00 and a suitable claim size distribution BB we have no(a) = rc(a + 9) . co(a) = rc(a + 9) . or equivalently BB[a] = B[^+ Repeating for t 54 1. 00 the standard definition of the exponential family {F9} generated by F is FB(dx) = e°xK(e)F(dx).a. (4. F and c.g.) The adaptation of this construction to stochastic processes with stationary independent increment as {St} has been carried out in 11. The answer is yes: inserting in (4. B9(dx) = B[9] B(dx).4. 0 Notes and references Corollary 3. say t = 1: recall from Proposition 1. K(a) = logEe'X = 109f 00 eaxF(dx) = logF[a].r.f. The question then naturally arises whether ie is the c. of F9.
.FT. n) for a given n. Z is measurable w.(9)} (4.68 CHAPTER III. G]. THE COMPOUND POISSON MODEL Definition 4.1) and multiply from 1 to n).. But let Xk = SkT/n . (4. = exp {BST .2 For any fixed T.d. Then FB denotes the probability measure governing the compound Poisson risk process with arrival intensity.6) F(G) = Po (G) = EB [exp {BST + Ttc(0)} .5) for the density.8) follows by discrete exponential family theory. . Proposition 4. and define 09.r.FTn) = Q(SkTIn : k = 0. it suffices to consider the case where Z is measurable w. . (4. oo) governing a given compound Poisson risk process with arrival intensity.t. Then P(G) = Fo(G) = EB [exp {BST + TK(O)} . and thus (4. Let FT = o(St : t < T) denote the o•algebra spanned by the St. Ti(a)/n.10) .d.5) for the density of n i. with T taking the role of n) is the analogue of the expression exp{8(x1 + • • • + xn) .8) By standard measure theory. The following result (Proposition 4..4). EeeBSt + tk(B) = 1.2. the PBT) are mutually equivalent on. .1.7) Proof We must prove that if Z is FTmeasurable.1 Let P be the probability measure on D[0. BB by (4. in particular the expression (4. for G E FT. the corresponding expectation operator is E9. t < T. and dP(T) dP^T) That is.i.3 Let T be any stopping time and let G E FT. G C {T < oo}..7) now follows by taking Z = eBST+TK(e)I(G) u Theorem 4 . replications from Fe (replace x by xi in (4. (4.t. and PBT) the restriction of PB to FT. G]. (4.3 and claim size distribution B. (4.Tic (0)} .f. .S(k_1)Tln. The identity (4. with common c.nr. Then the Xk are i.0e and claim size distribution Be. then EBZ = E [Ze9ST _T"(9)I .9) Proof We first note that for any fixed t. v(Xi. . Xn).r.i.g.
FT]] = EB [exp { BST + Trc(9)} I(G)] . Given FT.r)rc(9)}I .1) _ 1 + a. t = T . Ee [exp { BST +Trc(9)} I(G) FT)] = 1.r is deterministic. Letting T t oo and using monotone convergence then shows u that (4. GT C_ Jr < T}. 77 Thus.7) holds. The behaviour at zero is given by the first order Taylor expansion c(a) r.1(a). subject to the basic assumption ij > 0 of a positive safety loading. so that PG = EeE0 [exp { 9ST+Trc(9)}I(G)I FT)] = Ee [exp { BST + rrrc(O)} I(G)EB [ exp {9 (ST . 5.f.1) .10). Now consider a general G.ST) + (T . according to what has just been proved. (0) + rc'(0)a = 0 + ES1 a = a (p . Then GT = G n Jr < T} satisfies GT E FT. Thus by (4. Thus. LUNDBERG CONJUGATION 69 Now assume first that G C Jr < T} for some deterministic T.7 1 Some discussion further supporting this statement is given in the next section.9) holds for G as well. c(a) is a convex function of a. (a) rc (a) (b) KL(a) 'Y 'Y Figure 5.. the typical shape of rc is as in Fig. (4. . and hence (4.1 It is seen that typically) a ry > 0 satisfying 0 = r. 5 Lundberg conjugation Being a c.g.(Y) = 13(B['Y] . Then G E FT.9) holds with G replaced by GT.5.
5.1 Consider the case of exponential claims.2) 7 Figure 5. we further note that ( 5.70 CHAPTER III.2 is B(7) = 1 + ^. Thus. e. An established terminology is to call y the adjustment coefficient but there are various alternatives around. Thus B[7] = 6/.3) cf. 5.2 s As support for memory. Fig. G = {T(u) < oo} in Theorem 4.4) ELS1 = #L(0) cf.4) yields /3L = b and that BL is again exponential with rate bL =. Equation (5. and (4. the claim surplus process has positive drift > 0. we write FL instead of F7. .1) . Fig. 5. THE COMPOUND POISSON MODEL exists .g.1) is precisely what is needed for one of the terms in the exponent .1) is known as the Lundberg equation and plays a fundamental role in risk theory .s). Taking T = r(u). u It is a crucial fact that when governed by FL. (5. Example 5 . (5.a = i(a + 7). Note that KL (a) = /L (BL [a] .QL instead of /37 and so on in the following . (5. an equivalent version illustrated in Fig.3.1(b). b[s] = 5/(b .1) (or (5.3. It is then readily seen that the nonzero solution of (5.3. Lundberg conjugation corresponds to interchanging the rates of the interarrival times and the claim sizes.1(b). the Lundberg exponent.2)) is 7 = 5/3.
V)(u) < e7u.3.Ce7u as u 4 oo.1p . (5.t.e7x)G+(dx). T = T+.6) Proof By renewal theory.5) Theorem 5 . e(u) has a limit i. To this end.+ E A} in Theorem 4.P Y j o' xeryxOB (x) dx /3k [Y] .1.3 (THE CRAMERLUNDBERG APPROXIMATION) i'(u) .8) . we can rewrite this as 0(u) = e"ELe7^(u).1 (5. we therefore have ELe7t(u) + C where C ELe7 (00) = µ+) f e7(1 .6 ). (5.7) 0 and all that is needed to check is that ( 5. take first 0 = ry.G+L)(x)) dx ry^+L) J 00 f 0 (1 .(u)} .(oo) (in the sense of weak convergence w. Then P(ST+ E A) = EL [exp { 7S?+} .4). see A . G = {S.3 takes a particular simple form. PL ) with density 1 . Letting e(u) = ST(u) .G+ L) (x) G+L) (x) IL(+) µ+L) L) where G+L) is the FL.1e. LUNDBERG CONJUGATION 71 to vanish so that Theorem 4. (5. 0 Theorem 5 . T(u) < oo] . where C . Proof Just note that e(u) > 0 in (5. Since a7' is continuous and bounded.2 (LUNDBERG'S INEQUALITY) For all u > 0. which shows that G(L) (dx) = e7xG +(dx) = e7x /3 (x) dx.ascending ladder height distribution and µ+ its mean.7) is the same as (5. ST+ E A] .5). V) (u) = P(T(u) < oo) = EL [exp {ryS.r.5.u be the overshoot and noting that PL(T(u) < oo) = 1 by (5.
this solves the problem of evaluating (5.10) VW = JI c* e° (x) dx = a (B[a] .1 .12) Example 5 . we get L where 00 (1 . of course.11) so that I 7B ['Y](B[7]1) BI [7]Q VP (7) 72 7 (using (5.1 = ^7 The accuracy of Lundberg's inequality in the exponential case thus depends on how close p is to one. From this it follows.1) (5.3 (this was found already in Example 5. Then 0(u) = pe(a_Q)u where p = /3/S. that 7 = S . THE COMPOUND POISSON MODEL In principle. (5. u .e.")G + (dx ) = 1  J0 00 3B(x) dx = 1p.4 Consider first the exponential case b(x) = Seax. Noting that SIG(L)II = 1 because of (5.8) yields +L) J0 xel'B ( x) dx (5.7). but some tedious (though elementary) calculations remain to bring the expressions on a final form.72 CHAPTER III.1)) and 7µ+L) = 'y/3 [7] 7 1/0 = /3B ['y] . or equivalently of how close the safety loading 77 is to zero. Using (5. .1 above) and that C = p.4). 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.
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)
.
)3 )PBo µB  . (/3max . That is. Letting Bo be the stationary excess life distribution. Grandell [171] pp. the premiums exceed only slightly the expected claims.80 CHAPTER III. Proposition 7.b(u) = p*e.p = (/3max 0)µB./3)PBo PB .1. the approximating risk process has ruin probability z.p .g.p + p { 1 1p ti 1 .1 As /3 f Nmax. cf.3 and Corollary 3.3 )1 } _ 1p 1 . Though of course it is based upon purely empirical grounds. 1924.(3)2 P PB 2µB 2µB Letting /3* = /3/P. Notes and references The approximation (7. numerical evidence (e. Proposition 1.Ps(/3max .7) that Ee$(Amex /j)M _ 1p _ 1p Eo [s (0max 1 .3* /S.2) was suggested by De Vylder [109].P . we shall represent this situation with a limit where /3 T fl but B is fixed. heavy traffic conditions mean that the safety loading q is positive but small. Mathematically.8µBo  Ss' u where 6 = µB/µBo = 2µa/µB 2) . and hence the ruin probability approximation is b(u) e(bAln)u. p* _ . but has an obvious interpretation also in risk theory: on the average.1.PBo [s (/3max .(bA*)". 7c The heavy traffic approximation The term heavy traffic comes from queueing theory. we have according to the PollaczeckKhinchine formula in the form (3. or equivalently that /3 is only slightly smaller than /3max = 1/µ8. [174]) shows that it may produce surprisingly good results./3)] 1 . THE COMPOUND POISSON MODEL These three equations have solutions 9/3µB2)3 30µa2)2 3µa2) (3) P+ (3) ' 0 ./3)M converges in distribution to the 2a exponential distribution with rate S = B' Proof Note first that 1 .s(/3max .
This follows since rl = 1/p .7.g. Mathematically.4) suggested by the Cramer Lundberg approximation and Proposition 6.1 1 ./3)u * v. VARIOUS APPROXIMATIONS FOR THE RUIN PROBABILITY 81 Corollary 7.3) is reasonable for g being say 1020% and u being small or moderate. as well as it is needed for the interpolation approximation to be studied in the next subsection. 2 provides the better mathematical foundation. or equivalently that 0 is small compared to µB .l3)M > (/3max . That is . the premiums are much larger than the expected claims . we shall represent this situation with a limit where 3 10 but B is fixed. . u * oo in such a way that (3max . Of course. [APQ] Ch. VIII). and hence 2µ2B 1 .ze a2unµB laB (7. It is worth noting that this is essentially the same as the approximation (2) z/i(u) Ce./3)u). 7d The light traffic approximation As for heavy traffic .Q T /3max. the term light traffic comes from queueing theory. in risk theory heavy traffic is most often argued to be the typical case rather than light traffic . These results suggest the approximation Vi(u) e6(0_. the first results of heavy traffic type seem to be due to Hadwiger [184].B AB ) 6()3max _'3) = However .0)u. light traffic is of some interest as a complement to heavy traffic .2.p _ 2rl11B PB p. light traffic conditions mean that the safety loading rl is positive and large . Notes and references Heavy traffic limit theory for queues goes back to Kingman [230]. but has an obvious interpretation also in risk theory: on the average . We return to heavy traffic from a different point of view (diffusion approximations) in Chapter IV and give further references there . then P(u) 4 e6„ Proof Write z'(u) as P((/3max .2 If .ryu . However .p. Numerical evidence shows that the fit of (7. In the setting of risk theory. while the approximation may be far off for large u. The present situation of Poisson arrivals is somewhat more elementary to deal with than the renewal case (see e . obviously Corollary 7.
e. Indeed.T > u) = J o" B(x + u)/3eax dx . Omax max m. Another way to understand that the present analysis is much simpler than in these references is the fact that in the queueing setting light traffic theory is much easier for virtual waiting times (the probability of the conditioning event {M > 0} is explicit) than for actual waiting times . The crude idea of interpolating between light and heavy traffic leads to 0 (u) C1 . For a more comprehensive treatment. and hence 00 (U) /3pBBo (u) = 0 / B(x)dx. by monotone time T of the first claim .= 1 aJ 1 a 0+ 1 = = p. Sigman [347]. i. [97].(3 10.u. 7e Interpolating between light and heavy traffic We shall now outline an idea of how the heavy and light traffic approximations can be combined. z/' (u) convergence P(U . ( 3 J O B dx.Q limIP ( u) + Q lim z/'(u) Amax &0 amax ATAm. U > u] = /3iE(U .82 CHAPTER III. THE COMPOUND POISSON MODEL Proposition 7. the Poisson case is much easier than the renewal case.3 is the same which comes out by saying that basically ruin can only occur at the F(U . see Daley & Rolski [96].3 As . u Note that heuristically the light traffic approximation in Proposition 7. Asmussen [19] and references there. Light traffic does not appear to have been studied in risk theory. (7. 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. cf. En'=2 • • • = O(/32) so that only the first terms matters. Again. ao n=1 00 n=1 (u) P) anllBBon(U) onPaBon(u) • Asymptotically.u)+.T > u). . 10 ( u The alternative expressions in (7. 0(u) /3 J B(x)dx = /3E[U .5) follow by integration by parts.
[84]. ) M. . Instead. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS 83 which is clearly useless ./3)) . we see that the following limits HT) (u'). we may ask which one carries the larger risk in the sense of larger values of the ruin probability V(') (u) for a fixed value of 0.6) is . 8 Comparing the risks of different claim size distributions Given two claim size distributions B(1). where further references can be found . available. "/Qmex Cu) CLT(u ( /3max 0) + O16 CHT( U(Qmaz . one may hope that some correction of the heavy traffic approximation has been obtained.8. (U).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. The adaptation to risk theory is new.O0 M. ^ LT Q maxQ m"^ Qlo V LT) ( CHT(v) (say). Al . that is. z/i(E) (u) = pe(QmaxQ)u.3 and use similar notation for %(B) (u) = (u). even if the safety loading is not very small. Another main queueing paper is Whitt [380]. 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). no empirical study of the fit of (7.VHT) ( ax QmQ ) h (B) ( . 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. Thus .O(E)(u) 1 (1 .x . however. _(E) (u). Notes and references In the queueing setting . with rate 1/µB = /3max. (7.3n. . the idea of interpolating between light and heavy traffic is due to Burman & Smith [83 ]. B(2).3). Substituting v = u(. Let OLT) (u) denote the light traffic approximation given by Proposition 7. to get nondegenerate limits . ^IE) exist: 1 (B) HT QmsxQ hm J e e6" 2µE/µE2)'" = e(1 6)" =  Q1Qm.Wmax f(x ) dx + pee6mQ.
6. Rather than measuring difference in size. this ordering measures difference in variability. an equivalent characterization is f f dB(') < f f dB (2) for any nondecreasing convex function f. then Bill = B(2).' 1)(u) < V)(2) (U) for all u. THE COMPOUND POISSON MODEL To this end. XI. B(') <d B(2)) if B(1)(x) < B(2)(x) for all x.ill(u) < V)(2) (U) for all u. then . Taking probabilities. B(' <. we have the convex ordering. cf. In particular (consider the convex functions x and x) the definition implies that B(1) and B(2) must have the same mean. B(2)) if f fdB(1) < f fdB(2) for any convex function f. 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. In terms of the time to ruin. then i. equivalent characterizations are f f dB(') < f f dB (2) for any nondecreasing function f. for more detail and background on which we refer to Stoyan [352] or Shaked & Shantikumar [337]. Proof According to the above characterization of stochastical ordering. we shall need various ordering properties of distributions. we can assume that 1) < St 2l for all t.s.1 If B(') <d B(2). and a particular deficit is that we cannot compare the risks of claim size distributions with the same mean: if BM <d B(2) and µB«) = /IB(2). u Of course. most often the term stoploss ordering is used instead of increasing convex ordering because for a given distribution B. B(2) and PB(1) = µB(2).2 If B(') <j. U(2) such that U(l) has distribution B(').84 CHAPTER III.1 is quite weak. In the literature on risk theory. Proposition 8. . the proof is complete. U(2) distribution B(2) and U(1) < U(2) a. one can interpret f x°° B(y) dy as the net stoploss premium in a stoploss or excessofloss reinsurance arrangement with retention limit x. Recall that B(') is said to be stochastically smaller than B(2) (in symbols. whereas (consider x2) B(2) has the larger variance. Finally. Here convex ordering is useful: Proposition 8. Proposition 8. this implies St T(l)(u) > r(2)(u) for all u so that 17(I) (U) < oo} C_ {T(2)(u) < oo}. B(') <i. A weaker concept is increasing convex ordering: B(1) is said to be smaller than B(2) (in symbols. or the existence of random variables U(l).
B(2). we have by Jensen 's inequality that E f (U) > f ( EU). B(2). say to p.(1) (. (D) (u) < O(B) (U ) for all u.3 If B(1) <.1. we have Bol) (x) f ' B(1) (y) dy < ' f' B(2) (y) dy = Bo2) (x)• µ 85 I. Then V. The problem is to specify what 'variation' means. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS Proof Since the means are equal.e.4) certainly supports this view: noting that.. Proof If f is convex. Corollary 8.6 Fix /3 at 1/1.8. Example 8. then /'(')(u) < 0(2)(u) for all u. Hence by the PollaczeckKhinchine formula . from which the result immediately follows.. and here is one more result of the same flavor: Corollary 8. Bo1) <_d Bo2) which implies the same order relation for all convolution powers.2 is the following: Proposition 8.5 If '0(1)(u) < p(2) (U) for all u and a. B. u We finally give a numerical example illustrating how differences in the claim size distribution B may lead to very different ruin probabilities even if we fix the mean p = PB.1 and µB at 1 so that the safety loading 11 is 10%. then B(1) <. A general picture that emerges from these results and numerical studies like in Example 8. This u implies that D <.6 below is that (in a rough formulation) increased variation in B increases the risk (assuming that we fix the mean). The heavy traffic approximation (7. A first attempt would of course be to identify 'variation' with variance. .3 provides another instance of this.u) = (1 _ P) E /3npnBo( 1):n(u) n=1 00 < (1. with fixed mean. larger variance is paramount to larger second moment.4 Let D refer to the distribution degenerate at 'LB .p ) E /3"µ"Bo2)* n(u) _ V(2) (u) n=1 = Corollary 8. and consider the following claim size distributions: B1: the standard exponential distribution with density ay. it is seen that asymptotically in heavy traffic larger claim size variance leads to larger ruin probabilities. Proof Consider the light traffic approximation in Proposition 7. A partial converse to Proposition 8.
We return to ordering of ruin probabilities in a special problem in VI. i.000. Let ua denote the a fractile of the ruin function.01%.lA. Kluppelberg [234]. van Heerwarden [189].4.e. 0. THE COMPOUND POISSON MODEL B2: the hyperexponential distribution with density 0.01%. B3 the comparison is as expected from the intutition concerning the variability of these distributions.1358.9A2e'2r where A.. the behaviour of which is governed by a parameter 9. 1%.4142. B. sensitivity analysis (or pertubation analysis) deals with the calculation of the derivative (the gradient in higher dimensions) of a performance measure s(O) of a stochastic or deterministic system. = 0. One then obtains the following table: U005 U0.0' U0. 1/)(u.86 CHAPTER III. A2 = 3. all distributions have mean 1. However. 11 Notes and references Further relevant references are Goovaerts et al. Note to make the figures comparable.) = a. For B1i B2. in comparison to B2 the effect on the ua does not show before a = 0.e'\1x + 0. with the hyperexponential distribution being more variable than the exponential distribution and the Erlang distribution less. and this is presumably a consequence of a heavier tail rather than larger variance. which appears to be smaller than the range of interest in insurance risk (certainly not in queueing applications!). 9 Sensitivity estimates In a broad setting. 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). A standard example from queueing theory is . we have 0r3 = 2 < or2 = 1 < 02 = 10 < 04 = 00 so that in this sense B4 is the most variable. B4: the Pareto distribution with density 3/(1 + 2x)5/2. Pellerey [287] and (for the convex ordering) Makowski [ 252].001 u0. [166]. and consider a = 5%. In terms of variances o2. 0.1%. B3: the Erlang distribution with density 4xe2x.
Similar conclusions will be found below. 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. the distribution of %3 0 is approximatively normal N(0„ Q/t). Then a p ao = 00 Qa/. Then the arrival rate /3(P) for { R(P) } is )31p. Thus at p = 1. a0 as ao 80 19P . In particular . Then if t is large . t]. Let R(P) = Rtli. Assume for example that 8 is known. a2/t). Proof This is an easy time transformation argument in a similar way as in Proposition 1. if = a e(6A)u.19P a/ .9.2 Consider a risk process { Rt} with a general premium rate p. where the partial derivatives are evaluated at p = 1.e. Thus. For example. SENSITIVITY ESTIMATES 87 a queueing network. the premium rate p and the claim size distribution B. increasing in u.1. In the present setting. say estimated from data. and s(9) the expected sojourn time of a customer in the network. i. the standard deviation on the normalized estimate ^/1' (the relative error ) is approximatively . we may be interested in a'/ap for assesing the effects of a small change in the premium.01/2u. a/3 0 . and hence the effect of changing p from 1 to 1 + Ap corresponds to changing /3 to /3/(1 + Op) /3(1 . with 0 the vector of service rates at different nodes and routing probabilities.Ap).(u) for large u. and hence a _ e(60)u + u e(60)u = ( i + which is of the order of magnitude uV. while /3 = j3 is an estimate. u Proposition 9. or we may be interested in aV)/0/3 as a measure of the uncertainty on '0 if 0 is only approximatively known. it follows that ' is approximatively normal N(0.1 Consider the case of claims which are exponential with rate 8 (the premium rate is one). where Q2 = fl ( l2 1113 / _ Ou2v)2. Then ib = Pe(613)u. Example 9.3.. obtained say in the natural way as the empirical arrival rate Nt/t in [0.
0 = t/'(u) and the CramerLundberg constant C. Proposition 9. we can rewrite the Lundberg equation as w(9+ y.t. but we shall concentrate on a special structure covering a number of important cases. for the ruin probabilities . However .3 70 = 'Ye = = 7 /3(1 we(e +'y.6 below for some discussion of this assumption). Of course. and the proofs of (9.5) are similar. Consider first the case of 8/8/3: . so that heuristically we obtain '00 50ryu = Coe"u . () Proof According to (9. mathematically a proof is needed basically to show that two limits (u * oo and the differentiation as limit of finite differences) are interchangeable. but must look for approximations for the sensitivities 0. THE COMPOUND POISSON MODEL As a consequence.4). Differentiating w. 9.3) follows by straightforward algebra.10) below. (9.3.r. u Now consider the ruin probability 0 = 0 (u) itself.uypCe7u urypO. x > 0 (9.1 or Proposition 9. 4) (9 . (. In the case of the claim size distribution B.()YC = 1 +y/ /3 \ Q2 From this (9. (9. Similar notation for partial derivatived are used below. we cannot expect in general to find explicit expressions like in Example 9.(/3 + y)we(9 + 7.((dx ) = exp {Ox + (t(x) . namely that of a twoparameter exponential family of the form Bo.^)] 1(/3+y)we(9+'y.()^ 1 . 3) ( 9 . /3 yields w e(e + Y.w(6. and write yp = 8y/8/3 and so on .3.()wC(e.88 CHAPTER III.()(0 +'0) ' (9 . it suffices to fix the premium at p = 1 and consider only the effects of changing . Viei '0(.2) (see Remark 9. (3+'y)PC (0+7.g.6) As will be seen below. Consider first the adjustment coefficient y as function of 3. 5) (Q+'Y)[we(0+7. The most intuitive approach is to rely on the accuracy of the CramerLundberg approximation . () = log(1 + y//3).Owe (9. various parametric families of claim size distributions could be considered.3 or/and B. ()} p(dx) . this intuition is indeed correct.w(O. ^) . e.
x). BarndorffNielsen [58]). SENSITIVITY ESTIMATES Proposition 9.w(9. ()} .2 of the Appendix ).4t (U)e°`U = which are wellknown and easy to show (see e.w(9.8) Letting cp = e0/e/3 and differentiating (9. z2(U) = e7" J u b(u .x) F(dx ) f u J C F(dx) = C as u 4 oo. 0(u) = /3 Ju"O B(x) dx + f 0 0(u .x)B(x) dx + J U W(u .3) for z/'(u). Z(u)/u a C//3PF where PF is the mean of F. 11 For the following.4 As u oo.x)B(x)dx. Z= zl + z2 where zl (u) = e7u J m B(x)dx.3(x) dx. By dominated convergence. we multiply by e7" and let Z(u) = elt" cp(u).8) (Section 5). PF = (1 . () .12)).8).9) (9. we get p(u) = J "O B(x) dx + J U O(u . But from the proof of Theorem 5. () . F(dx) = e'yy/3B(x)dx.QB(x) dx. it holds that 89 a ue ryu a/3 Q(1 P) 7C2 Proof We shall use the renewal equation (3.w(O. we note the formulas Ee. Hence by a variant of the key renewal theorem (Proposition A1.3 (see in particular (5. ()] exp {w (O + y.g. u 0 Proceeding in a similar way as in the proof of the CramerLundberg approximation based upon (9.p)/C'y.St (U) Ee. Be.10) (9. ()} . and alsoo zl(u) + 0 because of B['y ] < oo. u 0 Then Z = z + F * Z and F is a proper probability distribution .x). z2(u) _ 1 ^ e'ri`i7i( u . () . (9.we(9 . O} (9. Combining these estimates . the proof is complete.([a] = exp {w(9 + a. () .9.C).11) Ee. w((9 + a. Further write de = [we (9 +'y. () exp {w(9 + a.(e"U = = wS(O.
()} 1z(dy) = f [t(y) . ()](e7v .wc (O.lB(x) dx = e7uzl(u) + e7°zz(u) + V(u T where zl (u) = .we (0. By dominated convergence and (9. 8^ ue7u.e7x/3 f 00 [t(y) .8) that cp(u) .1) B(dy) 'f '[t(y) .9)(9.x).5 Assume that (9.90 CHAPTER III.6C do 89 1p 8( 1p Proof By straightforward differentiation.wc(9.w(e.2) holds.11).wc(O. ()]B(dy) dx. ()]B(dy) dx x 0 0C T ON O . 01 (i+) do = +'Y. this implies Z = z + F * Z.w(0.QB(x)dx. C)] (1 + 7 ) Proposition 9. ^)} [wc (0 + 7. THE COMPOUND POISSON MODEL [we(e+7. 0 x Multiplying by e7" and letting Z(u) = e"uV(u).w (9. z = zl + z2. 8 8() 8( (9. ())B(dy) dx. ()]e7vB(dy) 'fCd 7 c . )}B(dy)• Letting cp it thus follows from (9. oo z2 (u) f C . Then as u > oo. F(dx) = e7x. C) .x)f3 f ^[t(y) .w( (0. u Z2(U) = e7° f u ^/i(u . ^) . 2 z 07P N ue7u (3C de .w( (0.6e7u f "o f[t(y) .12) f exp {O y + (t(y) .
18) (05 + 57 _'3_y .12) takes the form y) alp a . we (9.15) (9. We get w( (0. < = a.13) (9. w(e.3ary tog('Finally.a/35a&y' ' (9. () ='I'(t.a log S = log r(c) . U 7µF from which the second assertion of (9.) log(9) = %F(a) logs where %1 = F'/]F is the Digamma function.Sry a/32 + a/37 + /37 . Z(u) /3C 91 o c'o e11(t (y) .. and the proof of the first one u is similar.(log r(a) a log S)} • r(a) 1.6 Consider the gamma density b (x ) = Sa xa.Y)a+1 ' (9. SENSITIVITY ESTIMATES as u 3 oo. ())B(dy) < oo.14) de = d( 7!3 76 = 7e = log ( \ ( \5a_ / \SSry ) 72 .1 .17) (9.1edz = 1 exp {Sx + a log x . that C = a. .yu/3C2do u86 89 1p' az/) = 8z/.12) follows.rye) S 5rya. It follows after some elementary calculus that p = a)3/5 and. ue_Yu 'C2d( 8a 8( 1 p . () = log r(a) . and also zj (u) 4 0 because of f Hence. Example 9./35' a/i'y + aryl 625ry. ( 9. a /(S . () = C/9 = a/S.16) (9..C log(9). t(x) = logx.ry) 5a1 cry (5 ..2) holds with p(dx) = xldx. 9 = S.QS 1 . by inserting in the above formulas.pa+1 . Here (9.w((9.9.
([Y] .1 16 +ry c C22ry 2( = + 70 We (e.1 = eXP {c(C .2 .92 CHAPTER III. 9 = ."62 .22.2 log (0. for a < a* = z (. Be.log c = 2 In particular.3Ee. t(x) _ . C) . C = . () = Cc . ()} = exp {c (C .l3 of Section 6a needed for the existence of ry becomes e^Q > 1+62 / 2.2a) } Thus the condition B[a*] > 1 + a* /. THE COMPOUND POISSON MODEL Example 9.9) (() .7 Consider the inverse Gaussian density ( b(x) Zx37 exp This has the form (9. Straightforward but tedious calculations . w(e.S[a] = exp {w (9 + a. further yield .21og 2. C) = B = Yc = de = do = .w(9.2) with µ(dx) = 2x3zrdx.CZ try)} 1 C C2 try . which we omit in part .3.
B are assumed to be completely unknown.12) takes the form a = a 93 ar.g. Notes and references The general area of sensitivity analysis (gradient estimation) is currently receiving considerable interest in queueing theory. we can just fix k .1) .cue_7u)3C P Remark 9. ESTIMATION OF THE ADJUSTMENT COEFFICIENT Finally. let NT 16T = ^T . in which case we can just let t(x) = 0. and we estimate y by means of the empirical solution ryT to the Lundberg equation. 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). then ryT < 0. the exponent of the density in an exponential family has the form 01 tl (x) + • • • + 9ktk (x). That it is no restriction to assume k < 2 follows since if k > 2. kT (a) = /T (BT [a] . Also. the exponent is either Ox. BT [a]= NT ^` e"U. thus. or Ct(x). Finally if k = 1. in u which case the extension just described applies.+UNT) > 1. Note that if NT = 0. In general. 10 Estimation of the adjustment coefficient We consider a nonparametric setup where /3. (9. then BT and hence ryT is undefined.oo. the results presented here are new. Thus. Thus. Van Wouve et al.. queueing networks) are typically much more complicated than the one considered here.a. we have assumed k = 2 and ti (x) = x.. However . to our knowledge. if 1 PT = /3TNT(U1+.7 and references there. by the LLN both F (NT = 0) and F (PT > 1) converge to 0 as T . ae t 1lEY u S _ . Comparatively less work seems to have been done in risk theory. [379] consider a special problem related to reinsurance. for which we refer to X.8 The specific form of (9. However. sj=1 and let YT be defined by IKT('ryT) = 0. and hence explicit or asymptotic estimates are in general not possible.2 of the parameters. . To this end.3C2de 1p' z a = c .2) is motivated as follows.10.. the models there (e. the main tool is simulation.
(27)/K'(7)2.i3)(B[7] 1) + (3(BT[7]  . For the proof.'s. N ( n[7]. 1) r. B [7]2 (10. (10.1 As T 4 oo.B[7]2 n Hence ( 10. vfoVFB[2y].B[7]2 V2 .2) rT(7) N N (0. B[27] .3) Proof Since Var(eryU) = we have B[7].a BT[7] I B[7] I + .'Y . THE COMPOUND POISSON MODEL Theorem 10 . Hence KT(7) = (F' + (OT a(B[7l 0))((BT [7] . B[2'Y]  /3T ) . we need a lemma.1) . More generally.B[7]2 }) ( T 0 .If .7 + (. . Lemma 10 .1) 'YT .1)2 + E[27] .94 CHAPTER III.)vl+ N CO.: N 0.Q and Anscombe 's theorem.v.2 As T * oo. since NT /T . V2 are independent N (0. If furthermore B[27] < oo.3T . 7T a4' 7. then (10.B[7]) + B [7] .: N ()3.B[7]) 0+ Iv/o(b[y]./^ B[27] .. 16T where V1.b[Yp'V21 T { (E[7] .T y .1) .3/ T). a2 where a2 = /3r. it is easy to see that we can write \ V1 1 l _ .2) follows from NT/T a4' .
Let 0 < E < ry.E) < 0 < kT(7 + E) for all sufficiently large T .(ry .E ) < 4T(7T) < (7 +0' which implies 'T(ry4) a$' r. ESTIMATION OF THE ADJUSTMENT COEFFICIENT which is the same as (10.(ry + e) and hence KT(7 .. it follows that 7T7 KT(7T) . I. where ryT is some point between ryT and ry. °7IT) . If ryT E (7  we have KT(7 .10.e.c'(7) N (0' T (2(7) / N (0. By the law of large numbers. Then r.3). BT[a] 3 B[a]. Now write KT(7T)  kT(7) = 4T(7T)( 7T 7).'T(a) = 1 E Uie°U' a$' EUe "u = B'[a]. and the truth of this for all e > 0 implies ryT at 'y.1 can be used to obtain error bounds on the ruin probabilities when the parameters .4) and Lemma 10. Theorem 10.'(y). OT a 95 u 4 /3.KT(7) kT(7) K'(7) . first note that e7TU N (e7U u2e27Uo'2/T) 7 . 0 are estimated from data . Proof of Theorem 10.2. Combining ( 10.E) < 4T(7T) < 4T(7 + E). lcT(a ) 4 /c(a).4) + E).1 By the law of large numbers.e. 7T E (y .e) < 0 < r.Q. 6"Y (10. To this end . y + E) eventually. NT BT [a] Hence r. NT i =1 n'(a) for all a so that for all sufficiently large T K7 .
Wn) are i . For this reason . Further work on estimation of y with different methods can be found in Csorgo & Steinebach [94]. satisfies b(. 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. t]}.1 : Vt = 0. 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. A major restriction of the approach is the condition B[2ry] < oo which may be quite restrictive.. if B is exponential with rate 8 so that ry = 8 . Notes and references Theorem 10. Csorgo & Teugels [95]. with a tail of the form P(Y > y) .96 CHAPTER III. Griibel & Embrechts [292].f. ft.. Frees [146]. various alternatives have been developed.Q.. For example .96 if a = 2. wn = inf{t > W.g. 6 < 2. i .e. Herkenrath [192].1 is from Grandell [170]. it means 2 (8 .. = 1.d.i.ueryuU ".3 or equivalently p > 1/2 or 11 < 100%.g...info< „< t S. and the known fact that the Y„ = max Vt tE[W„1. Mammitzsch [253] and Pitts.T = 3TKT ( 21T)IKT (^T)2 is the empirical estimate of vy and fc. > 0 for some t E [Wn_ 1. This approach in fact applies also for many models more general than the compound Poisson one.) = a (e. .0) < 5. Asmussen [23]) can then be used to produce an estimate of ry. Deheuvels & Steinebach [102].T VIT where r7ry.e. Wn). Letting Wo = 0.5%). Hipp [196].. [197]. Embrechts & Mikosch [133]. i. V.C1e"a ( see e. the nth busy cycle is then [Wn1. U2.. Vt = St .
The safety loading is q = 1/p . 0.1 where p = 13µB. T) = P( /r(u) <T) \ = PI inf Rt <OIRo=u1 /\0<t<T PI sup St>ul 0<t<T Only the compound Poisson case is treated. the Poisson intensity is 0 and the claim size distribution is B with m. it is assumed that i > 0 and that the adjustment coefficient (Lundberg exponent) y.Chapter IV The probability of ruin within finite time This chapter is concerned with the finite time ruin probabilities 0(u. Unless otherwise stated.f. 97 . exists. 'y) where c(a) attains it minimum value. defined as solution of c(ry) = 0 where ic(s) _ /3(B[s] . The notation is essentially as in Chapter III. B[•] and mean AB. In particular.1 (the role of ryy will be explained in Section 4b).s.g. See Fig.1) . Further let 'Yo be the unique point in (0. generalizations to other models are either discussed in the Notes and References or in relevant chapters. the premium rate is 1.
7.1 The claims surplus is {St}. EL refer to the exponentially tilted process 3 with arrival intensity S and exponential claims with rate / (thus .t. FL and independent of T(u).(U) < 00] = ELT(u)ke'YS. 1 FL. 2 that E [T(u)k.(.u is the overshoot.1) (1.2) Proof Let as in Example 111. Var[T(u) I T(u) < 00] = VarL T( U) . E[T(u) I T(u) < 00 ] = ELT (U).5 .) = e7u ELe'Y^(u) ELT(U)k = e'Yu b ELT(u)k = O(u)ELT(u)k. the conditional mean and variance of the time to ruin are given by E[r(u) I T (u) < oo] Var [T ( u) I T( u) < oo] /3u+1 J )3 _ 2/3Su+/3+S (S)3)3 (1.9). we have for k = 1.(4.1 In the compound Poisson model with exponential claims with rate S and safety loading 77 > 0. using that the overshoot l. By the likelihood identity III. PROBABILITY OF RUIN IN FINITE TIME Figure 0. . 1 Exponential claims Proposition 1.r. In particular. PL = 6/0 = 1/p > 1).. (u) is exponential with rate 0 w. the time of ruin is T(u) and ^(u) = ST(t&) .98 CHAPTER IV.
the Laplace transform of the time to ruin is given by Eea7( u) = E [eaT (u). T(u) < oo] fora > r.B = a. .s.0) .6a = 0 with solution 0 (the .1)T(u)) = VarLe(u) + (PL .1//32 (6/)3 1)2 26(/3u + 1)/(6 .1 /3u + 1 u + 1 //3 = 6/3 6/01 For (1. is V1rLSr( u) +VarL ((PL .V/ is as asserted. the 1.(yo) = 2V .1) .1)T(u) are independent with QL the same mean .2) is aLELT( u) ."(ry) = 26//32.h. which leads to the quadratic 02 + (/3 . Wald's second moment identity yields 2 EL (Sr(u) .1)) ELST(u) ELT(u) (PL .1)ELT(u)./3 . 0 Proposition 1.s. EXPONENTIAL CLAIMS For (1 . we have by Wald's identity that (note that ELSt = t(pL .2). 1).1)T(u))2 = UL where = s./3) . where = eBu I 1 .3) B = 0(a) = + (6/3a)2+4a6 2 and hence that the value of ic(yo) Proof It is readily checked that yo = 6 . Since Sr (u) and (PL .6 + a)0 .1)2VarLT(u) + 2 Ca 1I VarLT(u).h.1 (6)3)2 which is the same as the r.I (1.h.6.12 Thus the l. u + ELe(u) _ PL . This means that /3(6/(6 . Let 0 > yo be determined by ^c(0 ) = a.s.2 In the compound Poisson model with exponential claims with rate 6 and safety loading rl > 0.(PL . of (1.1.
9ST(u) +T(u)!c(0)} .3) we have E [e«T(u ). T2 . Y2. are the lengths of of the ladder segments 2.. .1 .1 where Y1.. T(u) < oo] = e. the result follows.4. .Y1 Y2 Figure 1.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.. PROBABILITY OF RUIN IN FINITE TIME sign of the square root is + because 0 > 0).T+ Ti a t U T I 1 a i F.4) The interpretation of this that T(u) can be written as the independent sum of T(0) plus a r.0. Y(u) belonging to a convolution semigroup .'s with rate 5. and M(u)+1 is the index of the ladder segment corresponding to T(u).. Note that it follows from Proposition 1.. Fig. (1. are the ladder heights which form a terminating sequence of exponential r. But by the fundamental likelihood ratio identity ( Theorem 111. Cf... T(u) < oo] = EB [exp {aT(u ) . Using 5 = 6 .3.100 CHAPTER IV.v.v. M(u) T(u) = T + E Tk k=1 where T = T(0) is the length of the first ladder segment . St Ti F. 1. More precisely.3 that we can write EeaT( u) = eeuEe 017(o)..
6(u) = Vfl/j l(Su. U2.i. 1).cos (u/. . ..T the service times of the customers awaiting service .d. T) to be evaluated by numerical integration: Proposition 1. Note that the case 6 # 1 is easily reduced to the case S = 1 via the formula V. Proof We use the formula . For j = 0.. including the customer being currently served). [4]) 00 (x/2)2n+3 Ij (x) OnI(n+j)! . If QT = N > 0. EN has an Erlang distribution with parameters (N.6.i (u. T. UN. .1.T. Corollary 11.T. Then V(u.1 )!. Since U1 . where U1.T is the residual service time of the customer being currently served and U2 .1(u.. let (cf.4.T are conditionally i. Let {Qt} be the queue length process of the queue (number in system.3 sin0 + 29) f3(0) = 1+/32/cos9.v. 2. cf.. the conditional distribution of VT given QT = N is that of EN where the r.1..T) 1 I fl(O)h(0) fdO where (1.e.k + 1). .. the following formula is convenient by allowing t.0.I ex cos B cos j O dB fo " . then VT = U1. .3 Assume that claims are exponential with rate b = 1. EXPONENTIAL CLAIMS 101 For numerical purposes . Hence 00 F(VT > u ) P(QT = N)P(EN > u) N=1 00 N1 k F(QT = N) eu N=1 k=1 °O u k! k Ee k=0 1t P(QT . i.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.. UN. density xN lex/(N .T.ST).6) fl(9) f2(0) = = fexp {2iTcos9(1+/3)T+u(/cos91) cos (uisin9) .T + • • • + UN. and exponential with rate S = 1.
cos (( k + 1)0)] f3(9) Hence the integral expression in (1.102 CHAPTER IV.k + 1) = 1 k +1 + bj j=00 j=00 00 j=kk+1 j=k1 By Euler 's formulas.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 . Then (see Prabhu [294] pp. in particular equations (1.ie .8 ) yields F(QT > k + 1) . let I _ j (x) = Ij (x).1 00 ok+lR 00 j=k1 +1)/2e . f3(0) .112 l 1( k +1)/2 [ 31/ 2 cos(kO) .44).3(k +1)/ 2ei(k + l)6 (. similar formulas are in [APQ] pp.1)] L _112 /(k+1)/2 [.(31/2 cos (( k + 2)9) .38). (1.i(k +1)e R [/3( klal/2e:0 (01 /2 e . k k2 + $k+1 E bj 00 t j . 912. 00 E '3j/2 cos(je) j=k+1 00 _ j=k+1 ^j/ zeij = .31 /2eie L 1)] 1 I/31/2eie .cos((k + 2)9)] d9.)3k+1 = e(1+0)T e201/2Tcos 7r 0 e )3(k +l)/2 [31/ 2 cos ( kO) .(31/2eie . and define tj = e(1+R)Taj/2Ij(2vT T). 8789) 00 E aj j= 00 = 1.1 R [.13(k +l)/2ei(k +1)9 R E .)3k +1 tj g'(QT >. PROBABILITY OF RUIN IN FINITE TIME denote the modified Bessel function of order j.
E Fk. We first prove two classical formulas which are remarkable by showing that the ruin probabilities can be reconstructed from the distributions of the St. T) in terms of F(. Related formulas are in Takacs [359]. 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. The rest of the proof is easy algebra. 2 The ruin probability with no initial reserve In this section . Ui < x I / (note that P(St < x ) = F(x + t. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Since P(QOO > k + 1) = flk+1. The first formula.. equivalently.3 was given in Asmussen [12] (as pointed out by BarndorffNielsen & Schmidli [59]. oo (u)31/2e^e)k = )3k z cos(k9) = R k. from the accumulated claim distribution N. there are several misprints in the formula there. the numerical examples in [12] are correct). . however. k! k=O k0 i/z Co Uk ate" o'/z e . is numerically unstable for large T. F(x. however. we are concerned with describing the distribution of the ruin time T(0) in the case where the initial reserve is u = 0. or. We allow a general claim size distribution B and recall that we have the explicit formula z/i(0) _ P(7(0) Goo) = p.2.e = e' COS a cos(uf31/2 sin 0). T). t) = P . going back to Cramer. Seal [327] gives a different numerical integration fomula for 1 .T) which. expresses V)(0.7) that _ [^ au ak+l (30 k L.0(u. t )). 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). u Notes and references Proposition 1. it follows as in (1.
resp.T))dv E^T I(M(v. 2. See Fig. PROBABILITY OF RUIN IN FINITE TIME Theorem 2 .(6. Proof For any v E [0.t)= {Stv) < SM.(0.S„ 0 <t<Tv STS„+St_T+v Tv<t<T as the event that IS. co ).3) with A = (0. f T lStv)} 0<t<T by a 'cyclic translation'. [v. 1 1 . ") } is at a minimum at time t.i. T) = P(Tr(0) > T) = P(M(0.b (0. Stv^ _ Define M(v.T)dx. and the third from the obvious fact (exchangeability properties of the Poisson process) that has the same distribution as St = { Si0)} so that P(M(v.T)) does not {Stv)} depend on v.1 In formulas. Then 1 . T T o where the second equality follows from II. v].T))dv.0<w<t} St+v .104 CHAPTER IV.1. meaning that we interchange the two segments of the arrival process of {St}o<t<_T corresponding to the intervals [0. T].(.T].T) T F(x. . we define a new claim surplus process St StM NJ Figure 2.T)) 1 fT P(M(v.
T Theorem 2 .xdx. T)). then i fT I(M(v.Sv.T)f(I z /)(0. It follows that if M(0 . T) = M(0. We claim that if M(0.. t). Hence T TE f I( M(v. 0<t<v} = {ST < St . letting w = inf It > 0 : St_ = mino<w<T Sw}. cf. this integral is 0 if STv) .ST on M (0.v<t<T} = {ST<StSv. 0<t <Tv}n{ST<ST Sv+St T+v. in which case there is a last time o where St downcrosses level u. If ST < 0.T) and Sv < 0 on M(0. v)) dv = ST T T o (note that the Lebesgue measure of the v for which {St} is at a minimum at v is exactly . THE RUIN PROBABILITY WITH NO INITIAL RESERVE 105 Now consider the evaluation of fo I(M(v. T)) dv = TEST = T fP(ST < x) dx T T NT 1 f P(ST < x) dx = 1 f P Ui T . then M(v. T) occurs. Proof The event {ST < u} = { Ei T Ui < u + T j can occur in two ways: either ruin does not occur in [0. T)) dv. v) = M(0. v < t < T} n M(0. ST > 0. It is then clear from the cyclical nature of the problem that this holds irrespective of whether M(0.T) occurs or not as long as ST < 0.Tt))f(u+t.2. v). Fig 2.T) = F(u+T. Indeed. For example. or it occurs.2 10(u. we can take v E (w E. T) occurs. T.. T T o i =1 Let f (•. v). T) as {ST<St+ vS. there exist v such that M(v. v).T) occurs. Obviously. .2.T)) dv f T I(M(0. w) for some small E.t)dt. where the last equality follows from ST < St on M(0. T]. we can write M(v. v<t<T}n{ST<STSv+St. t) denote the density of F(•.
define St = ST . t + dt] occurs if and only if St E [u.ST_ t_ and let A(z.T) = . {St > .(0. Then P(T(0) E • I T(0) < oo) = P(T_ (Z) E •). 0 < t <T . which is independent of St and has the stationary excess distribution B0. which occurs w. {S t > z. PROBABILITY OF RUIN IN FINITE TIME u Q II T Figure 2. Hence P(ST<u) = 1 . Let Z be a r. Proposition 2. 2.T)+ J0 T (1V. E [t.Tt))P(StE[u.T) = C(z.v. For a fixed T > 0.3 Define r_ (z) = inf It > 0 : St = z}.p.2. 0 < t <T.t). u + dt] and there is no upcrossing of level u after time t.2.u+dt]). O(T . z > 0.T) = {St < 0. The following representation of T(0) will be used in the next section.106 CHAPTER IV.2 . 0 < t < T. u which is the same as the assertion of the theorem.2 Here o. Proof of Theorem 111.b(u.z. ST_ _ z}. The proof is combined with the proof of Theorem 111. ST_ _ z} . C*(z. ST_ _ z}.
1) yields P(ST(o)_ E [z.3 But by sample path inspection (cf. ST(o)_ E [z. Figure 2.3.T))f3B(z) dz dT. 7( 0) < oo) = P (C(z)) dT. T + dT] I S7(o)_ E [z. u which is the assertion of Theorem 111.1) that P(T(0) E [T.T(0)<oc) = f x F(U > y + z U > z) P(Sr(o)_ E [z. It follows by division by P(ST(o)_ E [z.1) z T . Thus P(Sr(o)_>x.T) = C*(z.ST(o) >y. T(0) < oo) B(y B(z) + z) f3B(z) dz = 3 f °^ B(y + z) dz = f3 + x v f B(z) dz.T + dT]. we therefore have P(A(z. .2. Proof of Proposition 2. A(z. z + dz]. z + dz]) = P(A(z.2. (2. z + dz].T)) = P(Cx. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Then 107 P(r(0) E [T. z + dz].T)). 2. Fig. and since {St}o<t<T. {St }o<t<T have the same distribution . r(0) < oo) = 3R(z) dz JP(C(z. z + dz].2. Hence integrating (2.3). T(0) < oo) = OR(z) dz in (2.T))dT = Off(z) dz P(T_ (z ) < oo) = 3B(z) dz.T).
a martingale proof is in Delbaen & Haezendonck [103]. Tak'ecs [359].5a).2 ga(x) = Qexr(a) f "o eyr(a)B(dy) x .T+dT]).3.(yo). PROBABILITY OF RUIN IN FINITE TIME ]P(7(0) E [T. some relevant references are Shtatland [338] and Gusak & Korolyuk [181].1 Eear( y) = eyr(a).c(r(a)) l = l er( a)se+at } u yields 1 = eyr(a)Eear(y). Notes and references For Theorems 2. cf.5 and one upon excursion theory for Markov processes (see IX. T(0) < oo) = dTP(T_(Z) E [T. Let T_ (y) be defined as Proposition 2.T + dT] T(0) < oo) dT f ' P(C(z))P(Sr( o)_ E [z.3 was noted by Asmussen & Kl(ippelberg [36].108 Hence CHAPTER IV.2. see in addition to Prabhu [293] also Seal [326]. I L Let ga(x) be the density of the measure E[ear(°). r(0) < oo. 3 Laplace transforms Throughout in this section. Proposition 2. T(0) < oo) 0 = dT f 0 P(C(z))P(Z E [z. z + dz].r(a).1. ^(0) E dx] (recall that ^(0) = Sr(o)) and write ga[b] = f OD ebxga(x) dx. In the setting of general Levy processes. z + dz]. [329].1) where a > r. Theorem 2. r(a) denotes the solution < 'Yo of the equation a = ic(r (a)) = . because of77>0. Lemma 3 . Note that T_ (y) < oo a.1 and the present proof is in the spirit of Ballot theorems.6.(3(B[r( a)] . one based upon a result of Asmussen & Schmidt [49] generalizing Theorem 11. (3.s. who instead of the present direct proof gave two arguments. Lemma 3. 2.1) . Proof Optional stopping of the martingale I er (a) 9 t.
T(0) < oo] = 20[b] = za[b] (9a[b] 9a[0])/b 1 .ic(b)/b x(b) + a eb"E[eaT(" ). Z = y] = EeaT. r(u) < oo).3.x)ga (x) dx where za(u) = f. T(u) < oo] du = Proof Define Za(u) = E [eaT(" ). y + dy]. E[ear (o) I T(0) < oo .3. Then by Proposition 2. Further by Theorem 111.x)(a) B(dy)• Lemma 3 . £(0) E dx) = /3B(x + dy) dx and hence ga(x) = f e r)/3B(x + dy) _ /3 f x e(v. the result follows after simple algebra.(v) = ev''(a).r(a) = a [B[b] B[r(a)]] .r(a) oo Q f ex(br(a))dx f00 eyr(a)B(dy) x 0 Q f evraB(dy) e(a))dx 0 Q cc ev(br (a)) .°° ga(x)dx.2 P(Z E [y.ga [b] 1 .f.1] evr(a)B(dy)[ b . LAPLACE TRANSFORMS 109 Proof Let Z be the surplus .ST(o)_ just before ruin .2. (Laplace transform) of the ruin Corollary 3.ga [b] 0 TO Using Lemma 3. time T(u): u u Here is a classical result : the double m.3. It is then easily seen that Za(u) is the solution of the renewal equation Za (u) = za (u) + fo Z. rr(0) < oo) = 1_ r(a) Proof Let b = 0.g.5 f 00 o a/r(a) . u . Corollary 3.3 ga[b] = c(b) Proof + b + a .r(a) The result now follows by inserting /3B[s] = ic( s) +/3+ s and ic(r(a)) =a. b .4 E[eaT (o).r(a) b . Hence eb"du E[eaT("). (u .
This proves the first assertion of (4. where _ 1 _ 1 1 C ML w(ry) 6B'[7J 1 . The first main result of the present section is that the value umL.UProof The assumption 11 < 0 ensures that P(T(u) < oo) = 1 and r(u) a4' oo.3).e. PROBABILITY OF RUIN IN FINITE TIME 4 When does ruin occur? For the general compound Poisson model.1 Assume 77 > 0.s. and hence a..00 St = lim .r(u) = Er(u) • ES.2. for any m T(u) u . mu ) ( 0 m < ML '(u) 1 m > rL. note that by Wald's identity u + EC(u) = ES. the known results are even less explicit than for the exponential claims case.1. (u) t. For the proof.1)Er(u) . u 1 ET(u) 1 p1 u where Pw2 = 311B)m3• 7(u) .. T(u)/u mL as u + oo. Later results then deal with more precise and refined versions of this statement.2 Assume ri < 0.(u ) = o(u) a.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 . cf. and take basically the form of approximations and inequalities. = (p .1) i. we need the following auxiliary result: Proposition 4. T(u) a. St/t 1 1/m. Then given r(u) < 00. Then as u * oo. t T(u) T(u) T(u) t m = lim = lim = lim Utioo u + Sr(u) u+oo S.s. (4. i. Theorem 4 .w ) v/. By Proposition 111.6. P = /3µB > 1.h(u.mL > E T(u) < 00 ) 40. Proposition A1.110 CHAPTER IV. for any c > 0 P( Further. That is. uoo u using e.mu D 2 4 N(0.
1).1). Notes and references Theorem 4.3).4. T (u) < 00 J 0(u) e7'PL U \ I T u) . 4a Segerdahl's normal approximation We shall now prove a classical result due to Segerdahl. implying T(u) .7 6  11 Proof of Theorem 4. the result comes out not only by the present direct proof but also from any of the results in the following subsections. though it is not easy to attribute priority to any particular author. For (4 .6µB2) Z v m (3µB2) Z.3. PL (•)+ 0.2 of [86]) and (4.^ N (o. apB ) .1 is standard. which may be viewed both as a refinement of Theorem 4. note first that ( Proposition 111.1 The l.1.mL >E By Proposition 4. and as a timedependent version of the CramerLundberg approximation. T(u) < oo f / 00) e7uE L [e_7 (t1). 4).1).N(0.5) St . T) for T which are close to the critical value umL). If Z . WHEN DOES RUIN OCCUR? and that Ee(u)/u a 0.2) follows immediately from u (4. proving (4. the same conclusion holds with t replaced by r(u).1 (by considering 0(u. and (4.r(u)/m T(u) ti µB2) Z. again Proposition A1. 1'r(U) .1) is T (u)  U mL P( T (u) < I > E. According to Anscombe' s theorem (e. of (4.g.s.h. . cf. Thus. Tu) T( u) .mL U > E.2.t/m D (2) 111 .mu m .6. Theorem 7.mu (2) '• m3/2 µB 7 . this can be rewritten as u + 1(u) .
g are continuous and bounded on [0.)mu \ h(oo)Eg (r(ul) .w2) r. we get E[ T (u) .^(T(u')).3 (SEGERDAHL [333]) Let C be the CramerLundberg constant and define wL = f3LELU2mL = f3B"[ry]mL where ML = 1/(pL1) = 1/($B'[ry]1). one has 9 (r(u)_rnu) Ef (^(u)) * E.um.r. then e(u) and r(u) are asymptotically independent in the sense that. Proof Define u' = u . S( u ) < ul/4] < ET(ul / 4) = O(ul/4). we can replace T(u) by r(u'). letting Z be a N(0. and similarly as above we get E[f(^(u)) I Fr(u.mul h(oo)Eg(Z).))I h(ul /4  ^(u)) I(6 (u') C ) f < ul /4 + f(e(u') .e(u') oo w .f ( (oo)) . P because of ^(u') .6) whenever f. (oo. oo).(u. Then h(u) 4 h(oo) = E f (6(oo)). using that ul/4 .) is readily seen to be degenerate at zero if ST(u•) > u and otherwise that of T(v) with v = u .4 (SIAM'S LEMMA) If 71 < 0. we need the following auxiliary result: Proposition 4. and thus in (4. with w2 as in (4.T ( u')] = E[ T ( ul /4 . PROBABILITY OF RUIN IN FINITE TIME Corollary 4. E9(Z) (4.T(u') given F. O .l:(oo) (recall that rt < 0). oo ).L+YWLV'U) .6).5) For the proof. resp . Hence Ef (Vu )) 9 (T(u. e'°'/b (u.a C4'(y )• ( 4.3).4).u1/4)I(S(u') > u1 /4) h(oo) + 0. Then the distribution of T(u) .VU T.v.ST( u') = u1/4 .112 CHAPTER IV.ul/4.t. Let h(u) = E f (^(u)). Then for any y. Using ( 4.
that for the fit of (4.8) Note that ay > 7o and that 7y > •y (unless for the critical value y = 1/ML). CL Fig. in practice one would trust (4. Theorem 4.9) ( 4 . 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 (. u needs to be very large).5 '(u .4.3 in terms of Edgeworth expansions . 0. also Hoglund [204]. oo ) as u * oo. For refinements of Corollary 4. y u) < . Cf. umL + ywL f) = e"P(T (u) < umL + ywL) = EL [e7V "). e7v" y < ^'(7) (4 .oo.1. however .3 ery"z/i(u . Notes and references Corollary 4 . see Asmussen [12] and Malinovskii [254].5) and solve for y = y(T).umL wI V"U u (4.7) To arrive at this . 4b Gerber's time. just substitute T = umL + ywL in (4. . define ay. yy by 1 K. T(u) < umL + ywL f.7) to be good. y u) < e 7v" .yK(ay)• (4.z/)(u . 3 is due to Segerdahl [333]. see also von Bahr [55 ] and Gut [182]. Thus . WHEN DOES RUIN OCCUR? Proof of Corollary 4.dependent version of Lundberg's inequality For y > 0. y > k'(7) .7) whenever u is large and ly(T)l moderate or small (numerical evidence presented in [12 ] indicates .4) in the last. ELe7E (") . 10) '5(u) . The precise condition for (4. The present proof is basically that of Siegmund [342].(ay) = 17 7y = ay . For practical purposes .T) Ce7"4 (T . Segerdahl 's result suggests the approximation b(u. PL(T(u ) < umL + ywL) 113 4 C4(y).
f Some urther discussion is given in XI. In view of Theorem 4.7 i. 5 is due to Gerber [156 ]. . yu < T (u) < oo 1 l e ayuEav [eT ( u)K(ay). yu 11 < T(u) < oo j < eayu +Y UK(ay) Remark 4. u Differentiating w.9): Proposition 4. we have rc(ay) < 0 and get (u) . yy is sometimes called the timedependent Lundberg exponent. yu ) = < eayuEay [eay^ ( u)+T(U)K ( ay).ay4(u)+ T(u)K(ay ). From the proof it is seen that this amounts to that a should maximize ayic(a). dy) Notes and references Theorem 4 .b (u. the bound a7y° turns out to be rather crude .5. PROBABILITY OF RUIN IN FINITE TIME Proof Consider first the case y < 1/K'(y). Numerical comparisons are in Grandell [172 ].8 below .yu ) = eayuEav [e .6 It may appear that the proof uses considerably less information on ay than is inherent in the definition (4. and hence t. Hoglund [203] treats the renewal case.3 yields easily the following sharpening of (4. An easy combination with the proof of Theorem 111.r.6. who used a martingale argument. which may be understood from Theorem 4. Then ic(ay) > 0 (see Fig . the point is that we want to select an a which produces the largest possible exponent in the inequalities.v"U. 0.8). yu) < C+(ay)e7a„ where l C+(ay) = sup f 00 eayR(xy)B( . yu) is e 'Yyu/ . a. and generalizations to more general models are given in Chapter VI.2. However. T(u) < yu] < eayu + yUr(ay) Y < eayuEav [ eT(u)K(av )L T(u) < yu} Similarly.114 CHAPTER IV. see MartinLM [257] .h(u.t. we arrive at the expression in (4. if y > 1/ic'(y).1). For a different proof. which shows that the correct rate of decay of tp(u.8).Y' (u.
2 yields EaT(u) u u r.'(y ). (4.ay y 'Yay  ay . T(u) suggests heuristically that l t/.13) . yu ) ayay e ryyu ayay 27ry/3B"[ay] u Proof In view of Stam 's lemma.ay a.6 with P replaced by Pay and FL by Pay.8 If y < 1/ic'(ry). the formula 0(u. As a motivation.e. and b(u. u 4 oo. yu ) eaauEaye . yu) = e.ayC() . it is instructive to reinspect the choice of the change of measure in the proof. (4.. and ii(u) .4.ay and get Ea e ayf (00) y _ 'Ya( ayKal lay C 1 . For any a > yo. 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. [eT(u )K( ay). not inequalities.i(u.c(&) = ic(ay) is < 0.12) < yu] Here the first expectation can be estimated similarly as in the proof of the CramerLundberg ' s approximation in Chapter III. if we want EaT(u) .. T(u) < yu] . The traditional application of the saddlepoint method is to derive approximations. (4. This idea is precisely what characterizes the saddlepoint method. Ea .: T.5. then the solution &y < ay of .^3 ]1/ Bay [lay .yu) c ay .11) ' If y > 1/ r .z. then the relevant choice is precisely a = ay where y = T/u.yyu y l ay I 21ry/3B" [ay] V fU_ u + 00. we have ryas = ay .ayuEay f eay^ ( u)+T(u)K(ay).e. the choice of ay. and in case of ruin probabilities the approach leads to the following result: Theorem 4 .(u.. then ay > 0. We thereby obtain that T is 'in the center' of the Padistribution of T(u). (0) r1 (a) ' I. Proposition 4. i.
1)3 = (jB"[ay]l (Pay .13). and the equation ic'(a) = 1/y is easily seen to have .1B[ay]1 ) y(ay . it seems tempting to apply the normal approxiyu + ul/2wV.l'B)y /(Pay .I ay &y a ^c'(ay) a (1 +.13).1)3 = y3/3B"[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 .11) follows. T(U) < yu] = eyuk (ay)E''ay (ek(ay )"1/2WV.c'(a) _ /3a/(8 . we get heuristically that Eay Ler (u)r(ay). .1.ay + ayl /BLay] . (4. Example 4.116 CHAPTER IV.3(5/(S . Writing r(u) and W2 = I3ay{.c(ay)ul/2W p 2ir = eyu(ay) dz 1 rc(ay ) 2.(j (1 .ay ) r.9 Assume that B(x) = eay. a nr=. Then ic(a) = .B[ay] /ay &y y(ay .12) is 0 entirely similar.4).1) under Pay mation (4. 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). The proof of (4.1) .ay) ay +.ay)K(ay) ay ayI&YI For the second term in (4. where V is normal(0.a.a) . i B[7ay .(ay) _ y(ay . PROBABILITY OF RUIN IN FINITE TIME ry I i .7ruw2 Inserting these estimates in (4. The difficulties in making the proof precise is in part to show (4.a)2 . V < 01 Ir 00 er(ay)"1'2"'x eyur.13) rigorously.
then { __ .8 is from Arfwedson [9].tcp) Lo {Wo ( t)}t>0 . 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.11) gives the expression '31/4 ( . A related result appears in BarndorffNielsen & Schmidli [59]..3+52 1+/351/y' sy 7 B ii[ay] 25 _ 251/2(1 + y)3/2 (5 . .5.f. yu) when y < 1/ic'('y) = p/1 .1. y) a''y" L '3 _ fl ) 51 /4(1 +1IY)3/4 \.= (s. is undefined for a > 5). 2 = Var(Si ) the variance.. 0 Notes and references Theorem 4. (5.i )( v s vc ('3 + s _2 / .. DIFFUSION APPROXIMATIONS solution ay=5 117 V 1 (the sign of the square root is .p. The mathematical result behind is Donsker's theorem for a simple random walk {Sn}n=o. 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 =.because the c. is the drift and o. c a 00. and next to note that such an approximation in particular implies that the first passage probabilities are close.ay)3 0 3/2 and (4./4 ^y for 1/i (u.1) . in discrete time: if p = ES.g.
p/c < St(p) < S((n+l)/ c + Pp/c. this is an easy consequence of (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.1 below). Letting c = a2/pp.tp). we have o {i!t s: . n/c < t < (n + 1)/c. We want an approximation of the claim surplus process itself.t} _ {W_1(t)} . for the purpose of approximating ruin probabilities the centering around the mean (the tcp term in (5.3) whenever c = cp f oo as p 1 p. + {Wo(t ) . oo)).1.. where p is the critical premium rate APBTheorem 5 . such that the claim size distribution B and the Poisson rate a are the same for all p (i. It is fairly straightforward to translate Donsker's theorem into a parallel statement for continuous time random walks (Levy processes). cf. However.1) with S.2) t>o where p = pp = p . Indeed .1 As p J.p.3) takes the form LI S(P) { a2 to2/µ2 + t LI S (P) { a2 ta2/µ2 {W0(t)}. 0 . we shall represent this assumption on 77 by a family {StP) L of claim surplus processes indexed by the premium rate p. of which a particular case is the claim surplus process (see the proof of Theorem 5.7c)..a = Snp) and the inequalities Sn )C . 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. Mathematically. and this can be obtained under the assumption that the safety loading rt is small and positive. St = EN` U= . p.tcpp) y = { WC (Sct) pct) } {Wo( t)}t>o (5.3.1)) is inconvenient. (5. a2 =/3µB2) Proof The first step is to note that { WC (St P) . Lemma 111.e.118 CHAPTER IV.z } {W_1(t )}t>o (5. and consider the limit p j p.
199. (ua2 To2 op \ IPI > IG ( T . 1.u).. TS(u)=inf{t>0: WW(t)>u}. the continuity argument above does not generalize immediately.1 . u) =PIT( (u) < x) = 1 .2 suggests the approximation u 0(u. and in fact some additional arguments are needed to justify (5. [169] or [APQ] pp.1.e. . Corollary 5 .h. see Grandell [ 168]. (5.( ^ I + e2( \ I . Corollary 5. since ti(u) has infinite horizon . DIFFUSION APPROXIMATIONS Now let Tp(u) = inf{t>0: S?)>u}. ulpI/a2). w. ulpl /a2) = e2"1µl / or2. For practical purposes .. (. any probability measure concentrated on the continuous functions.s.6) This is the same as the heavy traffic approximation derived in III.6) from Theorem 5.5. (5.r.8 or [APQ] p. 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. u). has a continuous distribution. we obtain formally the approximation V. is 1/ip (ua2 /IpI. Because of the direct argument in Chapter III. ('. 196. u) is defective when < 0.Ta2 /p2).s. and the r. we omit the details . 263) that the distribution IG(•.f I \\\ J \ (5. u) of r( (u) (often referred to as the inverse Gaussian distribution) is given by IG(x. is IG(T..h.h.1.5). ^ p2 Proof Since f 4 SUP0<t<T f (t) is continuous on D a. the continuous mapping theorem yields sup W Sz2 to lP 4 sup Wi(t)• O<t<T O<t<T a2 Since the r.T) IG(Tp2/ a2).4) Note that IG(. C.(u) ti IG(oo.7c.t.5) Note that letting T * oo in ( 5.s.2 As p j p.1 I 7= . However. 119 It is wellknown (Corollary XI.
B0 * Boo. the claim size distribution B9 and the premium rate p9 depends on 0. Michna & Weron [152] suggested an approximation by a stable Levy process rather than a Brownian motion. Assume further that 039µB6 < pe.Pe. All material of this section can be found in these references. In contrast.t. and which is much more precise. For claims with infinite variance.. on the premium rule involving interest. We conclude this section by giving a more general triangular array version of Theorem 5. The first application in risk theory is Iglehart [207]. such that the Poisson rate Oe.Po = 09µB6 . In view of the excellent fit of the CramerLundberg approximation. The proof is a straightforward combination of the proof of Theorem 5. in the next subsection we shall derive a refinement of (5.5) for the compound Poisson model which does not require much more computation.1 and Section VIII. 0) { 2 StQ2 /µ2 D { W_ i(t)}t>o t>o D 2 where p = pe = pe . PROBABILITY OF RUIN IN FINITE TIME Checks of the numerical fits of (5.00µB6 + 0. we have ^A. as an example of such a generalization we mention the paper [129] by Emanuel et al. e. that 00 4090. for more general models it may be easier to generalize the diffusion approximation than the CramerLundberg approximation. as 0 * 00 and that the U2 are uniformly integrable w. However. pt? 4 peo.6) are presented.5) combined with the fact that finite horizon ruin probabilities are so hard to deal with even for the compound Poisson model makes this approximation more appealing.5) and (5. the simplicity of (5. in Asmussen [12].6 of [APQ].3 Consider a family {Ste) } oc claim surplus processes indexed by a parameter 9.120 CHAPTER IV. Furrer.g. However. in particular for large u. and two further standard references in the area are Grandell [168]. Further relevant references in this direction are Furrer [151] and Boxma & Cohen [75]. Then as 0 _+ 90. a2 = ae = 00µa6 Notes and references Diffusion approximations of random walks via Donsker's theorem is a classical topic of probability theory. See for example Billingsley [64]. Theorem 5. the B9. (5.r. pe . The picture which emerges is that the approximations are not terribly precise.1.6) therefore does not appear to of much practical relevance for the compound Poisson model. . [169].
The setup is the exponential family of compound risk processes with parameters ( B9 constructed in III./c(9 .s and p = /3µB < 1. Determine yo > 0 by r. and we want to consider the limit 77 10 corresponding to Oo f 0. 2. Bo(dx) = B[eo]B(dx). it is more convenient here to use some value 9o < 0 and let 9 = 0 correspond to n = 0 (zero drift). claim size distribution B .90] B(dx). 77 = 1/p . whereas there we let the given 3B. this is because in the regime of the diffusion approximation . 3.Q (B[s] .1 > 0. this means the following: 1. In this setup. In terms of the given risk process with Poisson intensity . 0(0) = 0. 77 is close to zero.1) . Then EOU' = Boki[0] = Biki[eo]/E[9o] and "(s) = k(sBo)k(9o). Then r. The objective of the corrected diffusion approximation is to take this and other deficits into consideration.4.T) = Peo(r(u) < T) for 90 < 0. PB('r(u ) < oo) < 1 for 9 < 0. 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. let P9 refer to the risk process with parameters Q9 = QoB0[9] = QB[9 9o]. B9(dx) =Bale] Bo(dx) e9z keo)z = B[9 .'(yo) = 0 and let 90 = 'Yo.90) .90) and the given risk process corresponds to Poo where 90 = 'yo.9(s) = Ico ( s + 9) . 9o T 0. this idea ignores (among other things) the presence of the overshoot e(u). However . P9(r (u) < oo) = 1 for 9 > 0. Since Brownian motion is skip free. risk process with safety loading 77 > 0 correspond to 9 = 0 . . . which we have seen to play an important role for example for the CramerLundberg approximation . For each 9.6. and we are studying b(u. Let PO refer to the risk process with parameters e9oz Qo = QB[90].c(s) = .6.ao (0) _ /c(s + 9 .
(6.122 CHAPTER IV. One has (6. tu2 ) i IG (t.(y) = 0. (. write r = T(u). bl IG(t81.3) this implies (take u = 1) Ego exp { . PROBABILITY OF RUIN IN FINITE TIME Recall that IG(x.Varo S1 = f30Eo U2 = S1. the solution of r. () where h (A. C.7(u)/u2} eh(A. IGu+u2. (. means up to o(u1) terms): .. Theorem 5. 9otc0" (0) = 0061 = ul.e.2' where as ususal ry > 0 is the adjustment coefficient for the given risk process.C. The corrected diffusion approximation to be derived is (u.1) IG(x..() The idea of the proof is to improve upon this by an O (u1) term (in the following. i. for brevity.S.1) . (01. u) = euh(a . Vargo S. 0o to.2) . (U. 1) • Since L eatIG (dt. and Si = QoEoU2 = Q B"'['Yo Eo U3 ]... 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.u. _ ^(u) = ST .T) 1+u2 (6.3 applies and yields 1061 U61 Stdlu2/CZdi {W_1(t)}t>0 t>0 which easily leads to 1 StU2 {W( J(t)1t>0 { u S1 t>o Y'(u. u) = IG(x/u2. 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 . The first step in the derivation is to note that µ = k (0) = r0 (00) . . . C) = 2A + (2 .
. . and the dotted line the corrected diffusion approximation (6. the r.h. however. 1.z .2) is indeed o(u1). In ( 1) and (2). of a (defective) r. that the saddlepoint approximation of BarndorffNielsen & Schmidli [59] is a serious competitor and is in fact preferable if 77 is large] .3).52/u where Z has distribution IG (•. (6. A numerical illustration is given in Fig.ry2 . just replace t by Tb1/u2.h.1 As u + oo.d.4. it holds for any fixed A > 0 that Ego exp { Ab1rr(u)/u2} . p = 0. that whereas the proof of Proposition 6. distributed as Z .5) according to (6.3.2). the formal Laplace transform inversion is heuristic: an additional argument would be required to infer that the remainder term in (6. is the c. The solid line represents the exact value . .1 + 629. 6 .7. 9o T 0 in such as way that C = Sou is fixed. The justification for the procedure is the wonderful numerical fit which has been found in numerical examples and which for a small or moderate safety loading 77 is by far the best amoung the various available approximations [note.s.v.1 + u2 I Indeed.5) Once this is established . CORRECTED DIFFUSION APPROXIMATIONS 123 Proposition 6. we get by formal Laplace transform inversion that C 2 u.f. The initial reserve u has been selected such that the infinite horizon ruin probability b(u) is 10% in (1) and (3).exp { h(A.2 ).s. we have p =. But the Laplace transform of such a r. Note.6.'yu /2)(1 + b2/u)} + Aug 1I J . calculated using numerical integration and Proposition 1. To arrive at (6. which is based upon exponential claims with mean µB = 1. in (3) and (4). however .3 = 0. bl I IG I t +2 . u is Eeazead2/++ Eeaz[1 + ab2/u] where the last expression coincides with the r. .v.1 below is exact. 1% in (2) and (4). of (6.
124 0.111 W(U.1 It is seen that the numerical fit is extraordinary for p = 0.T) 0. BarndorffNielsen & Schmidli [59] and Asmussen & Hojgaard [34].u2 2u3 (e .EB 0 p ex p ( 7 S h ^)u .05{ 0.W21 0.114 0.0 0.011 L1 60 T IM 11. The proof of Proposition 6. Similarly.7.TI CHAPTER IV.4 may not be outstanding but nevertheless.^) . A51 7(SAT 3 3 h(X.aa1 .2 e.7 or at values of Vi(u) like 1% is unsatisfying.19)2 11 20 20 i0 T 1n0 Figure 6.(061 0. see Asmussen [12].T1 00. For further numerical illustrations.08 a.08 0.00 0.07 0..199 0. .T) 111 0.01 0. PROBABILITY OF RUIN IN FINITE TIME 0. (Inc 0s 0.1 W IU.1 proceeds in several steps.02 I 90 120 160 2W A0 Z WT 40 80 120 160 100 240 280 T 111 WI. it gives the right order of magnitude and the ordinary diffusion approximation hopelessly fails for this value of p.OOIi O. Note that the ordinary diffusion approximation requires p to be close to 1 and '0 (u) to be not too small.() Lemma 6. and all of the numerical studies the author knows of indicate that its fit at p = 0.. OM 0. the fit at p = 0.
co (e) .h.61a2T (B3 . () 62 Eeo exp u u2 J . (6.4 Ea.7) 2 2 . (6. + a1b2 + .s.00)(u +C)  'r (. exp ue } al 1J 3 exP I [2)..C2 = 2). in Lemma 6. 3 lim Eof (u) = EoC(oo) = a2 Ep = 3EoU2 uroo Proof By partial integration .. 1 = PB(T < oo) = Eo0 exp 125 {(B .3 EoU2 + 103OoEoU3 + " 2 6 Using d2 . CORRECTED DIFFUSION APPROXIMATIONS Proof For a>0.C)C/u . () + C and note that 2 KO (0) = 102.. 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.co ((/u)) } Let 8 = (2a + (2)1/2 = h().2u (B3 .6) u U3 Lemma 6 .r0 (00)) } Replacing B by 8/u and Bo by C/u yields e(B() = E eo exp { (e .1) h(A.(3) J t _ aa1T l + eh(A.4) that the r. the formulas Po(C(0) > x) Po(C(co) > x) imply 1 °° Po(ST(o) > x) = EIU fIP0 (U>y)dy .. the result follows.(3)Eea LauT exp i 3J .2 behaves like C l Eeo eXp r _ ^81T 1 Sl u2 1 u 2u3 [1+h(AC) S .+ h (A. 1 / Po(C(0) > y) dy EoC(0) x k EDUk + 1 k Eo[(0)k+1 EoC(0) _ (k + 1)EoU' EoC(^) _ (k + 1) Eo£(0) Lemma 6 .T (co (8/u) .
we get the correct asymptotic exponential decay parameter ^/ in the approximation ( 6.2u [2A+ (2 3 .S) d e 62 .2. Thus a2 y = 290 + O (u2). 0 The last step is to replace h(A.7) and using eh(a.h.s. There are two reasons for this : in this way.() . C) ( 1+ u2 The result follows by combining Lemma 6 .\ + () 1 2 / .2 (^/2 + 3y9o + 390) + O(u3). letting formally T * oo yields 7/)(u) C'e7u where C' = e7a2). l Lemma 6 . 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 6.6  d h(A. [2+ (2 .6) and 7co (Oo) = ico('y + Bo) to get 0 = 21 (^/2 + 2y90) + 1112 (_Y3 + 3_Y200 + 3y9o) + O(u4). yu/2).(2A + ()1/21 exp S h(A.x. PROBABILITY OF RUIN IN FINITE TIME The last term is approximately (e 3 (3) 27.h (A. 5 exp { _h(A) (1 + / y u J)) exp 1. 2 + 00 = . yu/2) h(A.1 (y/2 + Oo)u .4..126 CHAPTER IV. () by h(\. () . and inserting this and 9o 2 = S/u on the r.() I 1 + u2 ) y .\+ (2 (3 e 2u [ (2. yields +90 62 0 + O(u 3) 2u2 +O(u 3). we get h(A.2) for O(u) (indeed .e h(aS)h (^^ 262 exp {_h(. 2 and (6. Thus by Taylor expansion around ( = 90u.
The adaptation to risk theory has not been carried out. i. 'yu/2) 127 ( i+ M pz^ exP { h (A. The answer is similar: the process behaved as if it changed its whole distribution to FL. 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. Hogan [200] considered a variant of the corrected diffusion approximation which does not require exponential moments. that is. and to the Markovmodulated model of Chapter VI in Asmussen [16].T) has not been carried out and seems nontrivial. .e. () I 1 + u2 )I 2u L 2A+C2_(2 exp { _h. Fuh [148] considers the closely related case of discrete time Markov additive processes. HOW DOES RUIN OCCUR? exp { h (x. His ideas were adapted by Asmussen & Binswanger [27] to derive approximations for the infinite horizon ruin probability 'i(u) when claims are heavytailed.1: Just insert Lemma 6. the 'typical' value (say in sense of the conditional mean) was umL. ()} . 7 How does ruin occur? We saw in Section 4 that given that ruin occurs. with the translation to risk processes being carried out by the author [12].(i+ 62 exP{ h(A. () I 1 + u 2 ) } S 1 . 0 1 Proof of Proposition 6.7. The corrected diffusion approximation was extended to the renewal model in Asmussen & Hojgaard [34]. () (i+a ) 2A + (2 .4.1 (y/2 + Oo)u )} 1 (i + U ) [2+ C2 2u 62 S Pt^ exP { J 62(2 exp { h(A. the same as for the unconditional Lundberg process. u Notes and references Corrected diffusion approximations were introduced by Siegmund [345] in a discrete random walk setting.5 in Lemma 6. We shall now generalize this question by asking what a sample path of the risk process looks like given it leads to ruin. the analogous analysis of finite horizon ruin probabilities O(u. ()} 3 h (A. the approach to the finite horizon case is in part different and uses local central limit theorems. In Siegmund's book [346].
(u)_ is that i. the Poisson rate changes from . the numerator becomes e'ruELe7^ (u)PL(F( u)t) = e7uCFL (F(u)°) when F(u) E . Recall that 13L = (3B[ry] and BL(dx) = e'rxB(dx)/B[7]. Theorem 7 . P(u) and rate = aL w.r. In the exponential case. In fact. and let M(u) be the index of the claim leading to ruin (thus T(u) = Ti + T2 + .r.(u)_ and ^(u) are independent . ^(u) is exponential with rate 8 w. F(u)c] ti e' ru]PL (F(u)`) > 0. . we give a typical application of Theorem 7. F(u)c] P(r(u) < oo) ?P(U) < EL[e7u.2 If B is exponential. u * oo. 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.TT(u) _measurable. {ST(u)+t .F. .ST(u)}t> o is just an independent copy of {St}t>o). so in the in the proof.vi(u) Ce'Yu Corollary 7. PROBABILITY OF RUIN IN FINITE TIME changed its arrival rate from 0 to /3L and its claim size distribution from B to BL. + TMOO ).t. Note that basically the difference between FT(u) and . r(u) < oo) .3L and the claim size distribution from B to BL.(u)_ and similarly the denominator is exactly equal to Ce7u. stating roughly that under F(u).1 Let {F(u)}u>0 be any family of events with F(u) E F. . Proof P(u) (F(u)c) = F(flu)c.EL[e7S. then P(u) and FL coincide on .F. Recall that .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.3 to .(u) is not . (u) and satifying PL(F(u)) * 1.FT(u) = o' (T(u ).t.T.128 CHAPTER IV. {St}0< t<T(u)) Proof Write e'rsr(u ) = e'rue'r£(u). Then also P(u)(F(u)) + 1. FL As example.1..(u)..
who also treated the heavytailed case. A somewhat similar study was carried out in the queueing setting by Anantharam [6].3 M(u) pcu) 1 . take I(Tk < x) . however. the subject treated in this section leads into the area of large deviations theory.(1 . From a mathematical point of view. Notes and references The results of the present section are part of a more general study carried out by the author [11]. HOW DOES RUIN OCCUR? Corollary 7.3. 129 M(u) >2 I(Tk < x) M(tu) p(u) M(u) >2 I(Uk < x) BL(x). This is currently a very active area of research. Proof For the first assertion. . the queueing results are of a somewhat different type because of the presence of reflection at 0. see further XI.eALx) M(u) k=1 u The proof of the second is similar.eaLx.7.
This page is intentionally left blank .
1 Define p = !µ. are i.Chapter V Renewal arrivals 1 Introduction The basic assumption of this chapter states that the arrival epochs O'1. D'2.d. Var(St) = 11Ba2A + I4AaB lim t goo t PA 131 . Thus the premium rate is 1. and M is the maximum of {St}.i. We use much of the same notation as in Chapter I.Q...7). Then the arrival process is stationary which could be a reasonable assumption in many cases (for these and further basic facts from renewal theory. In the socalled zerodelayed case. with Nt = # {n: Un <t} the number of arrivals before t. and the one corresponding to T1 = s by 1/i8 (u). A different important possibility is Al to be the stationary delay distribution A° with density A(x)/µA. see A.1).1. of the risk process form a renewal process: letting Tn = Qn . r(u) the time to ruin. . the Tn are independent. {St} is the claim surplus process given by I.(1.Then no matter the distribution Al of T1i B. Proposition 1.1 (T1 = a1).. The ruin probability corresponding to the zerodelayed case is denoted by 1/'(u). the one corresponding to the stationary case by 00)(u). the claim sizes U1. the distribution Al of T1 is A as well.. .. with common distribution B.. . with the same distribution A (say) for T2.. T3. AA t*oo lim St = lim ESt t tioo t = p . U2. .
Here are two special cases of the renewal model with a similar direct interpretation: Example 1.0 > 0. by the elementary renewal theorem (cf. Nt = Var(PBNt) + E(4Nt) Q2 2 0`2 A tpB B + o(t). The simplest case is of course the Poisson case where A and Al are both exponential with rate 0.1 of the safety loading appears reasonable here as well. A.St] = a(p . For (1 .3 (SWITCHED POISSON ARRIVALS) Assume that the process has a random environment with two states ON. From this ( 1. after E. t 4oo Proof Obviously. If the environment is Markovian with transition rate A from on to off and u from OFF to ON. The renewal model is often referedd to as the Sparre Andersen process.a is really the accumulated claims over a period u of length a. Thus.2 (DETERMINISTIC ARRIVALS) If A is degenerate. Sparre Andersen whose 1959 paper [7] was the first to treat renewal assumptions in risk theory in more depth. This has a direct physical interpretation (a large portfolio with claims arising with small rates and independently). the definition 77 = 1/p .1) follows . 3) follows similarly by Blackwell 's renewal theorem. s + t µA PA 0 Of course.t.Nt] * a/PA. Nt ESt = E E UI Nt t = ENt•pB . RENEWAL ARRIVALS lim E [St+a . say at a. Proposition 1. Nt + EVar U. CHAPTER V. the . we get similarly by using known facts about ENt and Var Nt that Nt Var(St) = Var E Nt U. OFF. such that no arrivals occur in the off state.1 gives the desired interpretation of the constant p as the expected claims per unit time.1). one could imagine that the claims are recorded only at discrete epochs (say each week or month) and thus each U. Example 1 . and (1 . However .132 Furthermore for any a > 0. but the arrival rate in the ON state is .1) ENt/t + 1/µA. stating that E[Nt+a . 2).
r. integrate (1. u For later use..d. if for nothing else then for the mathematical elegance of the subject. Ao. INTRODUCTION 133 interarrival times become i.T between a claim U and an interarrival time T. and the present author agrees to a large extent to this criticism. initial vector (1 0) and phase generator 11 However. Proof The essence of the argument is that ruin can only occur at claim times.4) fo Indeed. Therefore.. More precisely. and for historical reasons. (an arrival occurs necessarily in the ON state..t.1. 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 . we have From this the result immediately follows.1. the relevance of the model has been questioned repeatedly. the first term represents the probability F(U1 .} with {S(d)} a discrete time random walk with increments distributed as the independent difference U .2. arrival times. For the stationary case.y)B(dy).s.. the fundamental connections to the theory of queues and random walks. we feel it reasonable to present at least some basic features of the model. . The values of the claim surplus process just after claims has the same distri bution as {Snd^ }• Since the claim surplus process {St} decreases in between max St = max ^d^.4) w. U1 . The following representation of the ruin probability will be a basic vehicle for studying the ruin probabilities: Proposition 1.i.s < u). A is phasetype (Example 1. 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 .. in general the mechanism generating a renewal arrival process appears much harder to understand.1.oFF}.. (1.4 The ruin probabilities for the zerodelayed case can be represented as 0(u) = P(M(d) > u) where M(d) = Max {Snd) : n = 0. However. S o<t<oo n=0.4) with phase space {oN. and then the whole process repeats itself).s > u) of ruin at the time s of the first claim whereas the second is P(r(u) < oo.
d. t. we shall be able to compute the ruin probabilities i(i* (u) for this model very quickly (. i. each of which receive a payment at constant rate during the lifetime . resp .Ut. St = t . 00).3* (say ) and the U. A typical sample path of {Rt } is illustrated in Fig.a*PB• > 1. (2. The compound Poisson model with negative claims We first consider a variant of the compound Poisson model obtained essentially by signreversion . b=1 !=1 where {Nt } is a Poisson process with rate . the claim surplus process are given by Nt Nt Rt = u+^U. are independent of {Nt} and i.1 If. That is . If .1) +ry.1. then 0 * (u) = 1 for all u > 0. The initial reserve is obtained by prepayments from the policy holders.0* (u) = P (rr* (u) < oo) where rr* (u) = inf It > 0: Rt < 0} ) . RENEWAL ARRIVALS 2 Exponential claims. the remaining part of the prepayment (if any ) is made available to the company. Using Lundberg conjugation . then 0*(u) = e 'r" where ry > 0 is the unique solution of 0 = k*(ry) = *(B*[ry] .1) . Theorem 2 . At the time of death .1 r* (u) One situation where this model is argued to be relevant is life annuities. with common distribution B* (say) concentrated on (0. U Figure 2.134 CHAPTER V. < 1. the claims and the premium rate are negative so that the risk reserve process . A simple sample path comparison will then provide us with the ruin probabilities for the renewal model with exponential claim size distribution.3* pB. 2.
and thus 1 = P(T. 0) and has typically the shape on Fig.. (a) is*(a) (b) . T_ (u) = inf { t > 0 : St = u 'r* (u). . Let B(dx) = ^e7x B*(dx). > 1 . EXPONENTIAL OR NEGATIVE CLAIMS [Note that r.f.2 Assume now .Rt.2 sup St = inf St = 00 t>o t>o and hence 0* (u) = 1 follows. T_ ( u) < 001 e7"P(T_ (u) < oo) = e"V)* (u). B. B* [7] and let {St} be a compound Poisson risk process with parameters . St=Rtu=St.2(a). Then { St } is the claim surplus process of a standard compound Poisson risk process with parameters 0 *.UB. 0 Now return to the renewal model.(u ) < oo) = E {e7sr_ (u). B*. If I3*pB* < 1.2(b).3*. of {St} is c(a) = is*(a7).0. Fig.(a) 7 Figure 2. Hence T_(u) < oo a. Then the function k* is defined on the whole of (oo. Since ic'(0) < 0. Define T_ (u) = inf It > 0 : St = u} . Then the c.s.2. cf.1.g. 2. and the Lundberg conjugate of {St} is { St } and vice versa. 2. Hence y exists and is unique.* (a) = log Ee'st I. the safety loading of { St} is > 0. then by Proposition 111. Proof Define 135 St =u .
Y a I....e.. 3* = 6.Tn} n=0.1 it is seen that ruin is equivalent to one of these values being > u.1 means that M* is exponentially distributed with rate ry..4 goes as follows: define 7r+ = P(M(d) > 0) and consider {St*} only when the process is at a maximum value.g.Ui . u Hence P(M(d) > u) _ 1r+e'r". and (2 . the failure rate of this process is y. However.Un } = max St = t>0 n=0..'s and noting that V)*(u) = P(M* > u) so that Theorem 2.)(u) _ 1r+e7" where ry > 0 is the unique solution of 1 = Ee'Y(uT ) = S 8 A[..136 CHAPTER V..1. respectively. alternatively termination occurs at a jump time (having rate 8). To + max {Ul+•••+UnTI. 2. and hence the failure rate .+Tn U1 Un..Y] (2.u+ and lr+.•. Taking m..a) = 1 .7r+ 7r EeTo b/(Sa) + +.f.4. then . A variant of the last part of the proof.. we get Ee'M(d) = Ee°M* _ Y/(. 1) means that 8(A[ry] . RENEWAL ARRIVALS Theorem 2 . with rate S (say).1. and from Fig . which has the advantage of avoiding transforms and leading up to the basic ideas of the study of the phasetype case in VIII... Hence M* max {To + Ti + • • • + Tn .• • • .Tr+. Then B* = A. with the probability that a particular jump time is not followed by any later maximum values being 1 . the distribution of M(d) is a mixture of an atom at zero and an exponential distribution with rate parameter ry with weights 1 .2).1. According to Theorem 2.Ti = U1. To + M(d) in the notation of Proposition 1. and 5PA > 1.2 If B is exponential.1) + ry = 0 which is easily seen to be the same as (2.2) 7 and7r+=1Proof We can couple the renewal model { St} and the compound Poisson model {St*} with negative claims in such a way the interarrival times of { St*} are To . Now the value of {St*} just before the nth claim is To +T1* +. T2 = U2..
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] .. Hence the failure rate of M(') is 6(1 . resp. which states that for a given a. the relevant exponential change of measure corresponds to changing the distribution F(d) of Y = U .6.3. 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^.. a ladder step is the overshoot of a claim size.5. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 137 is b(1. we see that ry = 6(1.B(dx). The probability that the first ladder step is finite is 7r+. : S(d) > u} . the imbedded discrete time random walk and Markov additive processes. 0 3 Change of measure via exponential families We shall discuss two points of view.y/b.2.3 A[a] B[a] F( d) [a +)3] F(d) [a] = Fad) [^] Letting M(u) = inf in = 1.7r+). Bads (dx) = . consider instead the failure rate of M(d) and decompose M(d) into ladder steps as in II. we have ] A[a )3] E«d'efl' = Bad> [a] A ad> [Q] = B[a +. letting P(d) refer to the renewal risk model with these changed parameters . However...7r+) and hence r+ = 1.2.4.7r+) = ry and hence P(M(d) > u) = P(M(d) > 0)e7u = 7r+e'r". Putting this equal to y. B^d) where Aad> (dt) = ^[ a] A(dt). Thus a ladder step terminates at rate b and is followed by one more with probability 7r+. This follow since. 3a The imbedded random walk The key steps have already been carried out in Corollary 11. 111. Furthermore. hence exponential with rate b.
provided the distribution F of U . and claim (a) follows immediately from this and e (u) > 0. cf. E(d)e 1' (u). just note that F7(d) is nonlattice when F is so .3 For the delayed case Tl = s. in the easiest nonexponential case where B is phasetype. In fact.e.138 CHAPTER V. 00)(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 claim (b). we get: Proposition 3.1) is explicit given 7.C8e7u where Cs = Ce78B[7].1)..t.T is nonlattice. Consider now the Lundberg case. O(u) = eauE (d)ea{ (u)+M(u)K (d)(a) .u the overshoot .2 In the zerodelayed case.4.r. (d) (7) _ 0.1 For any a such that k(d)' (a) > 0.2 p. the evaluation of C is at the same level of difficulty as the evaluation of i/i(u) in matrixexponential form.Ce"u where C = limu. i . VIII. RENEWAL ARRIVALS be the number of claims leading to ruin and ^(u) u = SM(u) .p)/($B'[7] . Proof Proposition 3. (a) '(u) < eryu. This is known to be sufficient for ^(O) ]p (d) ([APQ] Proposition 3.(u) . 7µA . Corollary 3.1 implies Cu) = e«uE ( 7d)e«^(u) . (b) V)(u) . We have the following versions of Lundberg' s inequality and the CramerLundberg approximation: Theorem 3 . to converge in distribution since p(yd) (r(0) < oo) = 1 because of r (d)' (y) > 0.C(°)eryu where C(O) = C0[7] . ik.. let 7 > 0 be the solution of r. For the stationary case. 187) and thereby for ^(u) to be nonlattice w.
another use of dominated convergence combined with Ao[s] = (A[s] 1)/SPA yields 00 u) e7u iP8(u) Ao(ds) + f 0 = CB['Y](A[y] . h(s) = e(a +x( a))8 (3. we invoke the behavior at the 1 = h«(0.3.1) (normalizing by h(0) = 1). For s > 0. B(x) = o(e7x) and dominated convergence.4)./c. (u + s . where G is the infinitesimal generator of {Xt} = {(Jt. we look for a function h(s) and a k (both depending on a) such that Gh. (s.(°) ( Ce8B[7] Ao(ds) similar manner. y) = e°yh(s). 3b Markov additive processes We take the Markov additive point of view of II.St)} can be defined by taking Jt as the residual time until the next arrival.h'(s).5. Here K. According to Remark 11. Equating this to tch(s) and dividing by h(s) yields h'(s)/h(s) _ . St)} and h.1) = C(O).9.y) B(dy) 0 For the stationary case.0 ) = Eo[ha ( Jdt. delayed version of Lundberg's inequality can be obtained in a e7u. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES Proof Using (1.a . Sdt) = h(s .5.Sdt] = Ee'uh(T) means 1 = f ' e°^B(dy) f ' h( s)A(ds). 0) = ah (s) .. 0) = tc(a)h(s). we get r u +8 e"8(u) 139 e7uB(u + s) + 4 0 + L 00 J e7(v8)e7(u+8v). Let P8f E8 refer to the case Jo = s.. IPA 0 Of course. The underlying Markov process {Jt} for the Markov additive process {Xt} = {(Jt. 0 0 . The expressions are slightly more complicated and we omit the details.dt(ah ( s) + h'(s)) so that Gha ( s.(s. E8h0 (Jdt.dt ) eadt = h ( s) . To determine boundary 0.
Note that the changed distributions of A and B are in general not the same for Pa.c(a)] which shows that U1. i..2) means 1 = B[a]/3/(/3+a+rc (a)). Proposition 3. . T2.s(Jo = s) = 1 follows trivially from Lo = 1. (3. . we can now for each a define a new probability measure Pa. = J8 = T2. St)}too by letting the likelihood ratio Lt restricted to Yt = a((J.tK( a)e. however. A[a .rc(a)] = B = Ba[13]Aa[5].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.S„):0<v< )be Lt = eaSt tK(a) h(Jt) = east .2) As in 11..4 The probability measure Pa.5. J n+1 u are independent with distribution Aa for the Tk and Ba for the Uk.13]A[b . rc(a) = 0 (B[a] ..( a+r' (a))(Jt s) h(s) where c(a) is the solution of (3. (3. resp. 5 For the compound Poisson case where A is exponential with rate . Ba where Aa (dt) . [e1U1 + 6T2ea ( U1s)stc ( a)e(a+K(a ))( T2s)I B[a +. Remark 3 .140 CHAPTER V. Ease AU1+6T2 [ AU1+6T2 = Ea a LT. ] = E. Further. Ba(dx) = B(dx).. [a + /3] A[b . Aa as asserted .2). resp. An important exception is. since JT..e.a . T2 are independent with distributions Ba.1)a in agreement u with Chapter III.s and P(d).rc(a)] = 1.a .c(a)] B[a] Proof Pa.s governing {(Jt.e(«+k(a))t esy A(dt). the determination of the adjustment coefficient ry where the defining equations rc(d) (ry) = 0 and rc(ry) = 0 are the same. . RENEWAL ARRIVALS B[a]A[a .. . . . An easy extension of the argument shows that U1. U.8.rc(a)] B[a] A[a .
yu) = F'ay. . Label the customers 1. T(u) < yu h(JT(u)) < eayu+yuk(ay ) ( Eia y Le(a(+k(ay))s v.(u.y yuAa y [ay + K(ay) .6 Let y < let ay > 0 be the solution of ic'(ay) = 1/y. yu). not after time T. that is. Using the Markov additive approach yields for example the following analogue of Theorem IV. Let M(u) be the number of claims leading to ruin .yx(ay). Then "^ e(ay+w(aY))8 Ys(u. which is the same as the asserted inequality for 0.4.)+1 e J j e(ay+w(ay))8 e . Then J(rr(u)) = TM(u)+1 and hence Ws(u.rc( ay)] = e(aa+(r ))sb[a ]e7yu L y1 In particular.g. see e. and U„ the service time of customer n. and define yy = ay . and assume that T„ is the time between the arrivals of customers n . or FCFS = first come first served) queueing discipline and renewal interarrival times. that is. The claim for the zerodelayed case follows by integration w.4. see in particular Dassios & Embrechts [98] and Asmussen & Rubinstein [45]. XII.yu ) e7vu A[ay . 2. The virtual waiting time Vt at time t is the residual amount of work at time t..5 Proposition 3. for the zerodelayed case zp8(u.1 and n. For the approach via Markov additive processes. Notes and references 4 The duality with queueing theory We first review some basic facts about the GI/G/1 queue. [APQ] Ch. the amount of . u The approach via the embedded random walk is standard.s eaysr(")+r(u ) K(ay) h (s) . say finite horizon ruin probabilities where the approach via the imbedded random walk yields results on the probability of ruin after N claims. The actual waiting time Wn of customer n is defined as his time spent in queue (excluding the service time).t. . the time from he arrives to the queue till he starts service. . 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. Proof As in the proof of Theorem IV. defined as the single server queue with first in first out (FIFO.ay+ray))TM(. yu ) < e7yu. it is easily seen that ic(ay ) > 0.r.4. A(ds).5..
.4. the proposition follows.4. Thus. (4.1).2 Let Mnd) = maxk=o.Tn)+. equivalently. and obviously z/'(u) = limN.4): Proposition 4. the waiting time a customer would have if he arrived at time t. since customer n arrives at time on.3) Proof Part (a) is contained in Theorem 11. Then P(r(u) < T) is the probability z/iiNi (u) of ruin after at most N claims.1 Wn+1 = (Wn + U. whereas in [On . Then: (a) as n + oo.• • • Tk ). we get: Corollary 4. It then jumps to VQ„ . an+1) = [on.. and we have P(V > u) = ?/iiol(u).3 Assume rl > 0 or. but interchanging the set .1.n1 (U1 +• • •+Uk Tl . The next result summarizes the fundamental duality relations between the steadystate behaviour of the queue and the ruin probabilities (part (a) was essentially derived already in 11. 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.. and combining with (4.4: Proposition 4.. p < 1. Thus Vos}1 _ = (Wn + Un . we have Wn = Van(left limit). If W1 = 0.3."^ Vi(N) (u). but we shall present a slightly different proof via the duality result given in Theorem II.2) (b) as t * oo.1) The following result shows that {Wn} is a Lindley process in the sense of II. Also {Zt}o<t<T evolves like the leftcontinuous version of the virtual waiting time process up to just before the Nth arrival.1. then Wn v M. (4. Vt converges in distribution to a random variable V.+ Un. .2.1 and Corollary 11. Wn converges i n distribution to a random variable W. in which case {V} remains at zero until time on+1. equivalently. The traffic intensity of the queue is p = EU/ET. 0 Applying Theorem 11. on + Tn) the residual work decreases linearly until possibly zero is hit. and we have P(W > u) = V. (u). Let the T there be the random time UN.4.142 CHAPTER V. (4.Tn)+ Proof The amount of residual work just before customer n arrives is VQ„ .
T* < x) fK(x_y)F(dy) (x > 0 is crucial for the second equality!). Letting n oo in Corollary 4. T] form a stationary renewal process with interarrival distribution A.2). Then the corresponding queue is M/G/1. which implies the convergence in distribution and (4. resp . T1. i.. where U*.1..oo in Proposition 4. (4. x > 0. namely W1 = 0 in (a) and Vo = 0. It follows that P(WN > u) =. we obtain: Corollary 4. and hence in particular ZT is distributed as the virtual waiting time just before the Nth arrival.5) Proof Letting n .. { 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). hence (since the residual lifetime at 0 and the age at T have the same distribution .T* = y yields K(x) = P ((W + U* . Corollary 4. Then K(x) = J x00K(x ..T*)+ < x) = P(W + U* .T*)+. by an obvious reversibility argument this does not affect the distribution .2. However. as WN.. and we get: . Then the arrivals of {Rt} in [0. A. conditioning upon U* . we let T be deterministic . we get W = (W + U* .5 (LINDLEY'S INTEGRAL EQUATION) Let F(x) = P(U1T1 < x).4) Tlim F(s) (VT > u) = limo P(s) (r(u) < T) = '+^io) (u)• 0 It should be noted that this argument only establishes the convergence in distribution subject to certain initial conditions..T* are independent and distributed as U1. For part (b).e.4..4 The steadystate actual waiting time W has the same distribution as M(d).le) the same is true for the timereversed point process which is the interarrival process for { Zt}o<t < T• Thus as before .(N)(u) has the limit tp(u) for all u. Hence for x > 0. T1 .. Ti) and similarly for the U. TN) with (TN. but this requires some additional arguments (involving regeneration at 0 but not difficult) that we omit. u Now return to the Poisson case . THE DUALITY WITH QUEUEING THEORY 143 (T1.. K(x) = P(W < x). cf. convergence in distribution hold for arbitrary initial conditions . In fact . (4..y)F(dy).Ao in (b).
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. the actual and the virtual waiting time have the same distribution in the steady state.5) to hold for all x E R and not just x > 0). W v V.5) looks like the convolution equation K = F * K but is not the same (one would need (4.g. . Asmussen [24] and references there. VIII). implying P(W > u) = P(V > u) for all u.g. Note that (4. The equation (4.144 CHAPTER V. RENEWAL ARRIVALS Corollary 4. 0 Notes and references The GI/G/1 queue is a favourite of almost any queueing book (see e . see e. the zerodelayed and the stationary renewal processes are identical. That is. Some early classical papers are Smith [350] and Lindley [246].5) is in fact a homogeneous WienerHopf equation.6 For the M/G/1 queue with p < 1. Proof For the Poisson case. Hence '(u) = Ali(°)(u).
. 145 Oj( u. . • The premium rate when Jt = i is pi. {St} denotes the claim surplus process. and can be computed as the positive solution of WA = 0.f pi. t St = E Ui .T) = Pi (T(u) < T). {Jt} describes the environmental conditions for the risk process.(3i when Jt = i. Ire = 1. here it exists whenever A is irreducible which is assumed throughout. The ruin probabilities with initial environment i are '+ki(u) = pi(T(u ) < oo) = Pi (M > u). The intensity matrix governing {Jt} is denoted by A = (A.)iJEE and its stationary limiting distribution by lr.Chapter VI Risk theory in a Markovian environment 1 Model and examples We assume that arrivals are not homogeneous in time but determined by a Markov process {Jt}0<t<oo with a finite state space E as follows: • The arrival intensity is . • Claims arriving when Jt = i have distribution Bi. As in Chapter I. i=1 0 and r(u) = inf It > 0: St > u}. Thus. M = supt>o St. dv. N.
we could distinguish between normal and icy road conditions.5 below. = iii when j E E(i). in blockpartitioned form. We let p Pi = /ji/AB. r^ = P (1. meaning that accidents occuring during icy road conditions lead to claim amounts which are different from the normal ones. and assume that weather conditions play a major role for the occurence of accidents. we shall assume that pi = 1.4) with representation (E(i).2. cf. the intensity matrix is A OW) T(i) T(n) t(n)a(i) where t(n) = T(n)e. a(i).2 (ALTERNATING RENEWAL ENVIRONMENT) The model of Example 1.1) iEE Then pi is the average amount of claims received per unit time when the environment is in state i. this is no restriction when studying infinite horizon ruin probabilities.11 below.1 implicitly assumes that the sojourn times of the environment in the normal and the icy states are exponential. cf. say. /3 = Nn when j E E(n). the operational time argument given in Example 1. and p is the overall average amount of claims per unit time. which is clearly unrealistic. f3i and claim size distributions Bn.146 CHAPTER VI. P = E 7riPi. Then the state space for the environment is the disjoint union of E(n) and E(i). leading to E having two states n. we can approximate A(i) with a phasetype distribution (cf. Example 1 . assume that the sojourn time in the icy state has a more general distribution A(i). Example 1. According to Theorem A5. For example. MARKOVIAN ENVIRONMENT where as usual Pi refers to the case Jo = i. respectively. T(=)). An example of how such a mechanism could be relevant in risk theory follows. and we have f3. say. i and corresponding arrival intensities Qn. Thus.a(').. 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. Assume similarly that the sojourn time in the normal state has distribution A(n) which we approximate with a phasetype distribution with representation (E(').T(n)). t(i) = T(')e are the exit rates. one expects that 3i > on and presumably also that Bn # Bi. with rates Ani and Ain. Unless otherwise stated. Bi. u .14.1 Consider car insurance. Proposition 1.
n8}. but assume now that the arrival intensity changes during the icy period. Approximating each A('?) by a phasetype distribution with representation (E('l). the state space E for the environment is { ('q... u Example 1. iq (visited in that order) and letfOil >. The simplest model for the arrival intensity amounts to . Then for example wi. depending only on 77.J017. one could for example have H = {i1.n.3.>. This amounts to a family (A(")) ?CH Of sojourn time distributions. (9) where q = CHI. we assume again pi = 1 so that the claim surplus is Nt St = ?Ui_t. 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.a(n).1. u Example 1 . where W = (w.p. A('^).. say.Q.3i/pi. One way to model this would be to take A(') to be Coxian (cf.2. 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) . dt. Qi = ..T(n)).4) with states i1. u From now on. MODEL AND EXAMPLES 147 Example 1 .. St = SB=(t)..5 (MARKOVMODULATED PREMIUMS) Returning for a short while to the case of general premium rates pi depending on the environment i.. w. 0 Then (by standard operational time arguments) {St } is a risk process in a Markovian environment with unit premium rate. i ) : n E H.. T(9) +wggt(9)0. such that a sojourn time of type rt is followed by one of type c w. such that the icy period is of two types (long and short) each with their sojourn time distribution A('L). In the car insurance example.tEH is a transition matrix. resp. is the probability that a long icy period is followed by a short normal one. and 1/ii(u) = t/ii(u). . Example VIII.. and . i8f n1. say it is larger initially. Indeed. and similarly for the normal period. .. t(n) = T("i)e.3i. 1 . it = Jel(t).. the parameters are ^ij = aid/pi.3 Consider again the alternating renewal model for car insurance in Example 1.1..j = . i E E(n) }. let T 9(T) = f pi.
o = 0. 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. More precisely. t l=1 Note that the last statement of the proposition just means that in the limit. N > 1(Ul < x) a4 B*(x).e(A(Oi)d'sg)xe. the Markov additive structure will be used for exponential change of measure and thereby versions of Lundberg's inequality and the CramerLundberg approximation.7 The Pidistribution of T1 is phasetype with representation (ei. 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. 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.(3iBi(dx). the empirical distribution of the claims is B*. vi(dx) = . )3*. . Nt Nt a .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.8 As t oo. MARKOVIAN ENVIRONMENT We now turn to some more mathematically oriented basic discussion. The key property for much of the analysis presented below is the following immediate observation: Proposition 1. one can associate in a natural way a standard Poisson one by averaging over the environment. Next we note a semiMarkov structure of the arrival process: Proposition 1. iEE iEE )3 These parameters are the ones which the statistician would estimate if he ignored the presence of Markovmodulation: Proposition 1. JT1 = j) = Qj • e. dx. . Pi (Ti E dx. qij = 0 in the notation of Chapter 11. B* = 1 /^* Bi. In particular.148 CHAPTER VI.A . we put )3* = E 7fi/3i.5.
6. oo) as a 4 oo. 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$. Bi. B*. By Proposition A5.2. i. and furthermore in the limit JT. Nt a' t t iEE Also.. N + oo Hence 1 Nt 1 N`+) Nits Nt E I ( Ut <.* (u) for all u. zli( (u) . In particular. In particular.(/3i)aiag). the Fidistribution of T1 in {St(a ) } is phase type with representation (E.7. given {Jt}0<t<0.. A. has distribution (7ri)3i //3*)iEE and is independent of Ti. Proposition 1. iEE 13 A different interpretation of B* is as the Palm distribution of the claim size. aA. and let {St °i} refer to the one with parameters Pi.aA . Then {St)} + {St*} in D[0. MODEL AND EXAMPLES 149 Proof Let ti = f1 I(JJ = i) ds be the time spent in state i up to time t and Nti) the number of claim arrivals in state i .1. Hence Nt'> a . Example 11. Proof According to Proposition 1.. ^j 7riNi.. we may view Nt`i as the number of events in a Poisson process where the accumulated intensity at time t is Niti.. Bi. Then it is standard that ti lt '4' iri as t > oo. e.4. y Ni) i Nti) t a. The next result shows that we can think of the averaged compound Poisson risk model as the limit of the Markovmodulated one obtained by speeding up the Markovmodulation. {St} to the compound Poisson model with parameters 0 *. the limiting distribution of the first claim size U1 is B*. this converges to the exponential distribution with rate 0* as a * oo. cf.x) = Nt E > I (Uk) X) Nt Bi(x) 1=1 iEE k=1 iEE 1: t5 Bi( x) = B*(x). However .9 Consider a Markovmodulated risk process {St} with param eters Ni. Conditioning . Bi(x).
lines in the path of {St}. with T2 being exponential with rate . which also yield O(a) (u. U2) are independent of .10 Let E_{1. 0 Example 1.2.2 +2 2 = 3. thick. we may imagine that we have two types of claims such that the claim size distributions are E3 and E7. we first get that 3 (3* = 2. the company even suffers an average loss. since E3 is a more dangerous claim size distribution than E7 (the mean is larger and the tail is heavier). A= (  a a ) \ a a 5 5 J 9 3 2 a1=2. resp.... Computing the parameters of the averaged compound Poisson model. 9 . The fact that indeed 0(a) (U) 3 0* (u) follows.1 with periods with positive drift alternating with periods with negative drift.=1. That is. s 5 in state 2. shows similarly that in the limit (T2..31µB 2 = 2 5 3 7 70 Thus in state 1 where p.. > 1. marked by thin. from Theorem 3. Continuing in this manner shows that the limiting distribution of (T. there are p = 2 background states of {Jt}. those of type E7 with intensity z s = 5 in state 1 and with intensity z . B1=3E3+2E7. U.l3* and U2 having distribution B*. and (at least when a is small such that state changes of the environment are infrequent). is as in {St }.FT.1 of [145].s = o in state 1 and with intensity 1 .2}.). the paths of the surplus process will exhibit the type of behaviour in Fig. 5 5 where E5 denotes the exponential distribution with intensity parameter 5 and a > 0 is arbitrary. e. the overall drift is negative since it = (2 2) so that p = 71P1 + 112P2 = 7. oo).g. On Fig. From this the convergence in distribution follows by general facts on weak convergence in D[0.. T) + ?P* (u.1. Claims of type E3 arrive with intensity 2 . 132=2. B2=1E3+4E7. 1..150 CHAPTER VI. 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 = . 1. MARKOVIAN ENVIRONMENT upon FT. Thus.s = 1o in state 2. state 1 appears as more dangerous than state 2.. T) for all u and T.2..
iEE .11 (a ) ESt/t * p .1 Thus.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) 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. note first that EN Uk')/N a4' µgi.1 a.1.(3. MODEL AND EXAMPLES 151 Figure 1. = P. That is.8. t + oo. (b) St/t * p . the averaged compound Poisson model is the same as in III. a fraction r.3* = 3/4 of the claims occur in state 1 and the remaining fraction 1/4 in state 2. Hence B* = 415E3+5E7/ +4 ( 51E3 +5 E7) = 1E 3 +2E7..s. 0 The definition (1. t * oo. For (b). 01 /. 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). Proof In the notation of Proposition 1.1.
12 If 77 < 0.. [302]. 38) Eiw1 = 1/ir. with X2. and hence wn /n a4. n n Thus {SWn l is a discrete time random walk with mean zero.1 of St / t is > 0. let some state i be fixed and define w=wl=inf{t >0:Jt_#i.ld.0i(u) < 1 for all i and u. and hence 1/ii(u ) = 1 for all i and u. and hence oscillates between 0o and oo so that also here M = oo. with some important improvements being obtained in Asmussen [17] in the queueing setting and being implemented numerically in Asmussen & Rolski [43]. If 77 > 0.4.. 136 or A. Statistical aspects are not treated here. having the Pidistribution of X.s. see [APQ] p. s. X 1 =Sty. Then by standard Markov process formulas (e.. see the Notes to Section 7. Theorem II. X2 =SW2 So.152 CHAPTER VI. 2 The ladder height distribution Our mathematical treatment of the ruin problem follows the model of Chapter III for the simple compound Poisson model. [APQ]. and a more comprehensive treatment in Asmussen [16].2(a) p. [315]. The proof of Proposition 1.Jt=i}. Proposition 1. 0 Notes and references The Markovmodulated Poisson process has become very popular in queueing theory during the last decade.1 jEE = (p .s.Eiw o'o Eiw • E ^ifjµs.. . EiX = 0.1(b) is essentially the same as the proof of the strong law of large numbers for cumulative processes. and involves a version of the . Now obviously the w.g. and so on. MARKOVIAN ENVIRONMENT Corollary 1. w2=inf {t>w1:Jt_#i. [212]. X3.1 and the Corollary are standard. The case 77 > 0 is similarly easy. then M = 00 a.. In risk theory. PB. . Eiw. Since the X„ are independent . some early studies are in Janssen & Reinhard [211]. limit p . dt . also + . Now let r) = 0.. and hence M = 00.1)Eiw = 0... and .. Proof The case 77 < 0 is trivial since then the a. See Meier [258] and Ryden [314]. then M < oo a.\ i and EiX1 Ei f 13 J. + Xn SWn ](1 a . The mainstream of the present chapter follows [16]..Jt=i}.a form a renewal process . There seems still to be more to be done in this area.
IIG +II)e. (y. dx)/jBj(y . j.Jr+ =j. n=0 (2. THE LADDER HEIGHT DISTRIBUTION 153 PollaczeckKhinchine formula (see Proposition 2.Jt=j)dt.6. For measurevalued matrices. Proposition 2. but is substantially more involved than for the compound Poisson case . Let further R denote the preT+ occupation kernel. j.A) =ZI(St E.(u) = Pi(M < u) = e' E G+ (u)(I . •) II = JG+(i. which represents a nice simplified form of the ladder height distribution G+ when taking certain averages : starting {Jt} stationary. IIG+ II denotes the matrix with ijth element IIG+(i. •)• kEE Also. by specializing results for general stationary risk processes (Theorem II . .j.A) = Pt(ST+ E A. oo)) = f R(i. •).a/i.6*. j.x).j E E. The form of G+ turns out to be explicit (or at least computable).i. T+ < oo) and let G+ be the measurevalued matrix with ijth element G+(i. Thus. let G+(i. B* in Section 1. and S (dx) the measure valued diagonal matrix with /3 Bj(dx) as ith diagonal element. 6. we get the same ladder height distribution as for the averaged compound Poisson model. oo)). j. •) * G +(k. e.1 irG+(dy)e =.EA. However . only with the product of real numbers replaced by convolution of measures. oo) = J ao 0 G+(i. j.2) R(dx)S((y .x.4) we obtain the following result . G+ is the matrix whose ijth element is E G +(i.g. T R(i.1) 0 (b) G+ (y. That is.j.2(a) below ) where the ladder height distribution is evaluated by a time reversion argument. for i. Proposition 2. 00 (2.3*B *(y)dy. Define the ladder epoch T+ by T+ = inf It : St > 0} = r(0).dx).. cf. we define the convolution operation by the same rule as for multiplication of realvalued matrices. the definition of .2 (a) The distribution of M is given by 00 1 . see also Example II.2. k.5.
and let further {my} be the Evalued process obtained by observing {Jt } only when {St*} is at a minimum value.1 for an illustration in the case of p = 2 environmental states of {Jt}. we need to invoke the timereversed version {Jt } of {Jt} . JJ = j. thick. hence uniquely specified by its intensity matrix Q (say). St < S* for u < t. the intensity matrix A* has ijth element * 7r ^i3 7ri and we have Pi(JT = j) = 7rj P2(JT = i)7ri (2.3) We let {St*} be defined as {St}. . see Figure 2. mx = j when for some (necessarily unique) t we have St = x. III to bring R and G+ on a more explicit form .3 When q > 0. To make Proposition 2.2 useful . That is.6. 0  x Figure 2. only with {Jt} replaced by {Jt } (the /3i and Bi are the same ). resp.1 The following observation is immediate: Proposition 2. G+ the probability that there are no further ladder steps starting from environment j is e^ ( I . we need as in Chapters II. marked by thin.1) follows by summing over n and j. {mx} is a non terminating Markov process on E. and that the environment is j at the nth when we start from i is e .IIG+II)e.2) is just the same as the proof of Lemma 11. MARKOVIAN ENVIRONMENT Proof The probability that there are n proper ladder steps not exceeding x and (x)ej. lines in the path of {St}. To this end . The u proof of (2.154 CHAPTER VI. From this (2.3.
we recursively define the depth of an excursion as 1 plus the maximal depth of a subexcursion.*.2.0. Furthermore. 0 mms1   ^O \ T. Figure 2. Note that the integral in the definition of W(Q) is the matrix whose ith row is the ith row of _ 3 f e2Bi(dx). ( Q( n)) converges monotonically to Q. and S(dx) is the diagonal matrix with the f3iBi(dx) on the diagonal. the sequence {Q(n)} A* defined by Q(O) = . and the excursion ends at time s = inf {v > t : S.and a jump (claim arrival) occurs at time t. Proof The argument relies on an interpretation in terms of excursions. we say that the excursion has depth 0. } is a minimum value at v = t. In general. Otherwise each jump at a minimum level during the excursion starts a subexcursion.(/3i) diag. THE LADDER HEIGHT DISTRIBUTION 155 Proposition 2. An excursion of {St*} above level x starts at time t if St = x.2 . The definitions are illustrated on Fig. s]. and the excursion is said to have depth 1 if each of these subexcursions have depth 0.(/3i)diag + T S(dx) eQx.2. If there are no jumps in (t. For example the excursion of depth 2 has one subexcursion which is of depth 1. = x}. Q( n+l) _ ^. corresponding to two subexcursions of depth 0. 2.4 Q satisfies the nonlinear matrix equation Q = W(Q) where 0 co(Q) = n* . {S..2 where there are three excursions of depth 1.
= j.s. mx+dx = j) occurs in two ways . StEA .. Similarly by induction .6) . the subintensity matrix of {min+i ) } is cp (Q(n)) = Q(n +l) which implies that qgj +1) = \!. A). j. Suppose mx = i. Then a jump to j (i. A) = L' U(j.4). Writing out in matrix notation . It is clear that { mini } is a terminating Markov process and that { mio) } has subintensity matrix A* . either due to a jump of {Jt } which occurs with intensity A= j. we first compute qij for i $ j.u< t). p1^) Define a further kernel U by f U(i.5) A (note that we use A = {x : x E Al on the r. Similarly.j +/3ipij. of the definition to make U be concentrated on (co.St <S*. Theorem 2 .4) To show Q = cp(Q). e. (2. Q = W(Q) follows.St EA. h. 7rE Proof We shall show that Fi(Jt=j.j. By considering minimum values within the excursion. or through an arrival starting an excursion terminating with J. The proof of Q = W(Q) then immediately carries over to show that the subintensity matrix of {mil) } is cp (Q(o)) = Q(l). It follows that qij = A. Fi(mh =i ) = 1 + =hflh+Qihpii+o(h) implies qii = 'iii /i +)3ipii. 0)). A) = f Pi(mx = j) dx eie4xej dx A u (2. 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. it becomes clear that pij = r [eQh] 0 ij Bi (dy) • (2.(01)diag = Q.T+>t) _ ^iF 7ri (JJ =i. Now let {m ( n) } be {mx } killed at the first time i7n (say) a subexcursion of depth at least n occurs . i.Qi + )%pij) Now just note that t pij and insert (2.5 R(i..156 CHAPTER VI.
To this end... 0<u<t). where A is the diagonal matrix with 7r on the diagonal: Corollary 2. THE LADDER HEIGHT DISTRIBUTION 157 from which the result immediately follows by integrating from 0 to oo w.7 It is instructive to see how Proposition 2. St E A.r.=StSt. From Qe = 0. oo)) = f o' eIXS((x + z.Qi)diag. the CramerLundberg approximation (Section 3). 0 +1) = cp (K( n)) defined by K(o) = A .S„<0.. (Jo = j. (b) for z > 0. G+((z. e. St < St U. We may then assume Ju=Jtu.1 can be rederived using the more detailed form of G+ in Corollary 2. and this immediately yields (2.g. 0 < u < t) = 7rjPj(Jt =i. u It is convenient at this stage to rewrite the above results in terms of the matrix K = 0'Q'A.(.z+>t) = P.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. K( n (d) the sequence converges monotonically to K.6 is hardly all that explicit in general.2.(Jt=j. 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). S.0<u<t.0<u<t) = P.StEA.St <Su.6 (a) R(dx) = eKxdx. dt. oo))dx.. and to obtain a simple solution in the . Jt = i. x < 0. `` {K(n)} [the W(•) here is of course not the same as in Proposition 2. we shall see that nevertheless we have enough information to derive. and get irPi(Jt =j. consider stationary versions of {Jt}.Jo=i. Remark 2.St EA..6). and we let k be the corresponding right eigenvector normalized by Irk = 1. it is readily checked that 7r is a left eigenvector of K corresponding to the eigenvalue 0 (when p < 1).4]. {Jt }.t.
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
5) will produce the maximal ryy for which the argument works. hj P .(ay)}. av 'i [h.7. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 165 hiav)e _avuE. (3.j.3.00 su e7( ( 3. (u. 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. for a vector <p(u) = (cpi (u))iEE of functions .V)i(u.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).y)G+(z. if y > 1lk'(ry). r(u) yu o)(y)eavuEav. r(u) < yu] hiay)C+ h=av)C+ o) (y)eayu+yuw(av).8 ) Then for all i E E and all u > 0. Chie ryu < Vi(u ) < C+hie 7u.5).i [e*(u)K(av).9) For the proof. yu < r(u) < 00 h 4(u) < h(av)C+o)(y)eavuEav .i I (a) exp {aye(u) + r(u)r. we shall need the matrices G+ and R of Section 2. dy)• o iEE jEE .j) * coj)(u) _ f u ^Pj(u . 1 Similarly. we have ic(ay) < 0 and get 'i(u) . exp {e() + r(u))} .. as in the classical case (3. oo)) and. yu) f h(av)e v avuE«v.i [eT(u)K(av ). we let G+ * W(u) be the vector with ith component E(G+(i. However. We further write G(u) for the vector with ith component Gi(u) = EiEE 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.
166 CHAPTER VI. Then iterating the defining equation ip(n+1) = G + G+ * V(n) we get W(N+1) = UN * G + G+N+1) * ^(o) However. n > oo. j. jEE u 0 j. dy) = aj f Bj(dy . Hence lim cp(n) exists and equals U * G.(0) ] (u) < sup Jt t.& (u). j. j.x) jEE 00 u 0 //^^ C+E. Then cpin)(u) sit (u) as n + oo. 00 Thus C+ > hj f"o e7(Yu)G +(i.ery(&u+x)Bj (dy) Bj(u Bj (u . we have G *(N +1) * ^.3jhj // f 00 R(i.u Iv 2°)(u)I Pi(rr+(N + 1) < oo) + 0. MARKOVIAN ENVIRONMENT Lemma 3 . j. U = U". 00 f C_ hj f e(Y)G+(i.u IMP:°) (u) I < oo. dy) 00 C+ ijhj f R(i.dy). 0 G+(i. dy) : 1(u) < C+ > hj u e(1tL)G+(i.12 Assume sup1. Lemma 3 .7.13 For all i and u.x ) R(i. °O . _ To see that the ith component of U * G(u) equals ?Pi (u).j. and define W(n+1) (u) = G(u) + (G+ * tp(n))(u). dx) f e7( vu)Bj (dy . dx) 100 C . just note that the recursion <p(n+1) = G + G+ * (p(n) holds for the particular case where cpin)(u) is the probability of ruin after at most n ladder steps and that then obviously u cp2n) (u) + t.j.x) x) jEE 0 E Qj f jEE R(i. Proof Write UN = EN G+ . dx ) Bj (u . if r+ (n) is the nth ladder epoch.x ) = Gi(u). dx). = Eo G+ G.
13 Let first cp=°)(u) = C_ hie"u in Lemma 3. Indeed. j.M>u) = Ei [VGJT (u . j. 14 Let yo > 0 be the solution of 'c'(yo ) = 0.3.12) Proof We first note that just as in the proof of Theorem 3.13. 167 u Proof of Theorem 3. dy) jEE u U +C_ hje7( yu)G jEE"" +(i. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES proving the upper inequality. ST < u] < C+(yo)e7ouEi [h^7o)e70ST1 l T J = C h(7o)e7ou8T .tpi(u.T) = Pi(M > u) . and the proof of the lower one is similar.Pi(MT > u) = Pi(MT < u. this is obvious if n = 0. and let 8 = e'(70).MT<u.13) Hence. (3. j. 9 for the last equality in (3. let C+(yo) be as in (3. taking cps°) (u) = 0. MT < u.11. j.n) ( u . jEE estimating the first term in (3. (3. we have Vii (u) .M > u) = Pi(ST<u. +i . u The proof of the upper inequality is similar . Then 0< Vi (u )  0i(u. and using Lemma 3 . Here is an estimate of the rate of convergence of the finite horizon ruin probabilities 'i (u. dy) jEE o (3. We claim by induction that then cpin) (u) > C_ hie7u for all n. y]hj = C_ e7uhi.10: Theorem 3 .(u) < T ) to 0i (u) which is different from Theorem 3.10) C_ 1 f hje7(y. we get Wo n +1) (u) = ? 7 i ( U ) + E J u gyp. from which the lower inequality follows by letting n * oo. and assuming it shown for n. letting MT = maxo<t<T St.u)G+(i. T) = Pi (7.y)G+(i.8) with y replaced by yo and hi by h=7o ). 13 and the second by the induction hypothesis .11). dy) (3.ST).10 ) by Lemma 3 .11) C_e7u 57 O+[i. T) < C+(')' o)hi7u)e7ou8T . it follows that Vi(u) < C_(yo) h=70)e7ou.
.o. but that in general the picture is more diverse...4) To avoid trivialities. < . where it has been observed repeatedly that Markovmodulation increases waiting times and in fact some partial results had been obtained. Further related discussion is given in Grigelionis [176]. (4. MARKOVIAN ENVIRONMENT Notes and references The results and proofs are from Asmussen and Rolski [44]. where o*(u) is the ruin probability for the averaged compound Poisson model defined in Section 1 and . M" of the corresponding two claim surplus proceses (note that 0'(u) _ P(M' > u). Occasionally we strengthen (4.3) to B = Bi does not depend on i. (4. 0"(u) = P(M" > u)) Now consider the risk process in a Markovian environment and define i' (u) _ >iEE irioi(u). (4. 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. this correponds to the usual stochastic ordering of the maxima M'. 4 Comparisons with the compound Poisson model 4a Ordering of the ruin functions For two risk functions 0'. The Markov process {Jt} is stochastically monotone (4.31:5)32 . The conditions which play a role in the following are: . we refer to .1) Obviously.168 CHAPTER VI.3p.33 or Bi 0 Bj. ".2) alone just amounts to an ordering of the states.0..3). It was long conjectured that 0* Vi. For the notion of monotone Markov processes. u > 0. this is not the case for (4.o. we also assume that there exist i # j such that either /3i <. Bp. B2 <_s.. The results to be presented show that quite often this is so. [177].o.o..2) (4. <s. The motivation that such a result should be true came in part from numerical studies. in part from the folklore principle that any added stochastic variation increases the risk. and finally in part from queueing theory. V)" if z/i'(u) <'c"" (u)..3) Bl <_s..5) Note that whereas (4.
Section 4.3* f uB(x) z/^.10) Q*B*(u)+. p). it follows by a standard . Comparing (4. ^i 7ri = 1. Proof of Theorem 4.. Then V.13* J0 u 0*(u . .. = b. Conditioning upon the first ladder epoch. we need two lemmas.4) say basically that if i < j .x)B*(x) dx. The first is a standard result going back to Chebycheff and appearing in a more general form in Esary.4.4) hold. Lemma 4 . Proposition 2. Theorem 4 .6). Proschan & Walkup [140]..3 for the second) *(u) _ /3 *B* (u) +... Conditions (4..1 for the first term in (4. E 7r i Wi(u . 7(0) < oo) = pirf+).x)dx _ /3*B*(u) + f u / ^ t=1 > 3 * B* ( ) + f (4.7) and Lemma 4. T(0) < oo) = Bi(x) dx/tcai . 1:7riaibi > E 7riai i=1 i=1 j=1 The equality holds if and only if a1 = . Lemma 4 .r (Sr(o) E dx Jr(o) = i. < bp and 7ri > 0 (i = 1.. we obtain (cf.* For the proof.. = aP or b1 = .10) and (4.8) ^j Tri/iBd(x) .5 (cf.2..x) of i and using Lemma 4. b1 < .9 ) below)..r(u x)dx. . then P P P 7rjbj.r (JT(o) = i.2)(4.2. where 7r2+) = QiµBilri/p.9) (4. 0 Here (4.7) 7ri..4) is automatic in some simple examples like birthdeath processes or p = 2 . also Proposition 2.6.1) which with basically the same proof can be found in Asmussen & Schmidt [49]. then j is the more risky state .. note that (4.2)(4.1. < a...3iBi(x)YPi(u . dx (4. the second follows from an extension of Theorem I1. (b) P.6) 7r= fl*B*(u) + p> s=1 +) fu 0 b (u  x)Bt (x) /pB.1 Assume that conditions (4. COMPARISONS WITH THE COMPOUND POISSON MODEL 169 Stoyan [352]. 2 If al < .9) follows by considering the increasing functions 3iBi (x) and Oi (u . (4.x) dx u o i =1 i=1 (4. and it is in fact easy to show that Vii(u ) < t/j(u) (this is used in the derivation of (4. 3 (a) P.
4 Assume that . Frey.h.r (u ) fails for all sufficiently small e > 0.0*• i=1 But it is intuitively clear (see Theorem 3. i=1 i=1 7'r(0) _ EFioiwi(0) . they are at present not quite complete.0. that P P /^ 1r1NiµBi /^2 /^ ^i/ji pBi < 1il3i i=1 i=1 (4.3µi < 1 for all i. What is missing in relation to Theorem 4. u To see that Proposition 4.11) i=1` and that A has the form eAo for some fixed intensity matrix A0.8) we get P P '*' (0) = 3* + /3*1* (0) _ > lri'3qqi • E 7i/ipBi . dominates the solution 0* to the renewal equation (4.3i. MARKOVIAN ENVIRONMENT argument from renewal theory that tk.4 is the understanding of whether the stochastic monotonicity condition (4. of order 101. Proof Since 0..* (0). Using (4.6). Q2 = 1. (4.1 of [145] for a formal proof) that z/ii(u) converges to the ruin probability for the compound Poisson model with parameters . For u = 0. Then the l. Recall that .s. of (4.s. it is sufficient to show that 0'.4) is essential (the present author conjectures it is). = 102. Notes and references The results are from Asmussen. this ruin probability is /3iPBi.6). As is seen.h.4 is not vacuous.11) is of order 104 and the r.(0) = V. µB2 = 104.2./3*..(0) < b *'(0) for e small enough. Then i/i*(u) < . (u) may fail for some u...170 CHAPTER VI. except possibly for a very special situation . 0. Rolski & Schmidt [32]. 01 = 103. (u) is not in general true: Proposition 4. it will hold for all sufficiently large u.1 and Proposition 4. let = ( 1/2 1/2 ) . 4b Ordering of adjustment coefficients Despite the fact that V)* (u) < *. µB. and from this the claim follows.. u Here is a counterexample showing that the inequality tp* (u) < V). Bi as e J.
with strict inequality unless rci (y*) does not depend on iEE.a = E irirci(a). Hence if 5i 54 0 for some i E E. and by Proposition II.)a.g. cf. it follows by Proposition A1.13) implies A(a) > 0 for all a.g.13) (4. COMPARISONS WITH THE COMPOUND POISSON MODEL 171 the adjustment coefficient for the Markovmodulated model is defined as the solution y > 0 of ic(y) = 0 where c(a) is the eigenvalue with maximal real part of the matrix A + (rci(a))diag where rci(a) = ai(Bi[a] . e. Jt = i])' EE = vA+n(6. The adjustment coefficient y* for the averaged compound Poisson model is the solution > 0 of rc*(ry*) = 0 where rc*(a) _ 13*(B*[a] . This implies that A is strictly convex.4.4(b) that the limit in (4. in particular . (4.12) iEE Theorem 4. It is clear that the distribution of X..5.a. which in view of EiEE 1ibi = 0 is only possible if Si = 0 for all i E E.5 y < ry*.14) A„(O) iioo varXt t t By convexity. Lemma 4. Proof Define X= f &ids.Jt=kI A (the return time of k) where k E E is some arbitrary but fixed state. 0 . Asmussen [20]) as discussed in 11.ld) with generic cycle w = inf{t>0: Jt_54 k. Xt)} is a Markov additive process (a socalled Markovian fluid model.. Now we can view {Xt} as a cumulative process (see A.7) )i is convex with A'(0) = lim EXt tioo t = iEE 70i = 0. (4.(a) > 0 for all a 0 0.6 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. Further (see Corollary 11.5. Then {(Jt.5.1) . is nondegenerate unless bi does not depend on i E E.1) .14) is nonzero so that A"(0) > 0. with strict inequality unless a = 0 or bi = 0 for all i E E. (4.2 we have (Ei[e"X'.
p yi and compute 8y 8a a=0 In both cases. Let bi = rci(y*).16) Differentiating (4. y. Further a(1) = rc(y*) by definition of A(.) and rc (•).e7r)1 (Ici(Y*))diage. Notes and references Theorem 4. Since ic is convex with rc'(0) < 0 . The corresponding adjustment coefficient is denoted by ry(e). whereas the . In the case of e.15) yields 0 = (Ii(y*)) diage + Aoh'(0) = (rci('Y*)) diage + (Ao .6. a = 1 in Lemma 4. improving upon more incomplete results from Asmussen. and our aim is to compute the sensitivity ay e a E=O A dual result deals with the limit a 4 oo.15) Normalizing h by 7rh = 0.. 0 = ((ri(Y))diag + ery (4{('Y))diag)h + (A0 + e(?i'Y))diag)h'. we have 7rh' = 0. note that y(a) + mins=1. where A. (4. Frey. Here we put a = 1/e.5.5 is from Asmussen & O'Cinneide [40]. MARKOVIAN ENVIRONMENT Proof of Theorem 4. Hence rc (y*) > 0. this implies that the solution y > 0 of K(y) = 0 must satisfy y < y*. Thus y(e) * y* as e 10.172 CHAPTER VI.12) and rc*(y*) = 0. 4c Sensitivity estimates for the adjustment coefficient Now assume that the intensity matrix for the environment is Ae = Ao/ e. h(0) = e.. Then > risi = 0 because of (4. Hence letting e = 0 in (4. the basic equation is (A + (rci(y))diag)h = 0. If rci(y* ) is not a constant function of i E E. we get rc (y*) > 0 which in a similar manner implies that u y < y*.Qi and Bi are fixed . h'(0) = (Ao . multiply the basic equation by a to obtain 0 = (A0 + e(r£i(y))diag)h. (4. h depend on the parameter (e or a).. Rolski & Schmidt [32].15) once more and letting e = 0 we get ..eir)h'(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.16) yields Proposition 4.20) and multiplying by el to the left we get 0 = All + 7'(0)rci (0) + 0 (here we used icl (ry(0)) = 0 to infer that the first component of K[7(0)]h'( 0) is 0).17) (4. multiplying (4. then 8a a=o All rci (0) Notes and references The results are from Asmussen.7 8ry aE = 1 7r(ci ('Y*))diag ( Ao e7r)1(Xi(Y*))diage *=0 P Now turn to the case of a. which has recently received much attention in the queueing literature. (4..18).i(7' *))diagh'(0). and may have some relevance in risk theory as well (though this still remains to be implemented). the intensity for such a transition (referred to as marked in the following) is denoted by Aii l and the remaining intensity .5. i = 2.19) Then 'y ^ ryl as a ^ 0 and we may take h(0) = el (the first unit vector). We get 0 = (aAo + ( lc&Y))diag)h. p. Inserting (4. THE MARKOVIAN ARRIVAL PROCESS 173 0 = 27'(0)(ri(`Y *)) diage + 2(ci('Y* )) diag h' (0) + Aoh" (0) . ..20) Letting a = 0 in (4. Rolski & Schmidt [32].19) holds. and we have proved: Proposition 4. Frey.18) 0 = 27'(0)p+27r(rs. . 0 = (Ao + ry'(ii(Y)) diag )h + (aAo + (Ki(7'))diag)h'.8 If (4.17) by 7r to the left to get (4. (4.8 when ryi < 0 for some i is open. 5 The Markovian arrival process We shall here briefly survey an extension of the model. We assume that 0 < y < 7i. (4. The analogue of Proposition 4.
where qij is the probability that a transition i * j is accompanied by a claim with distribution. Jt2)) (2. Thus . Here are some main examples: Example 5 . A(l) = tv. Bii = Bi . the definition of Bi is redundant because of f3i = 0. This is the only way in which arrivals can occur.d. For i = j. let { Jt 1) }. A ( 2) = A (2`1 ) ® A. T). Note that the case that 0 < qij < 1.2 (SUPERPOSITIONS) A nice feature of the setup is that it is closed under superposition of independent arrival streams . A(1) = A . u Example 5 . . Indeed.2). B. with common distribution B. Bij = B. and the marked transitions are then the ones corresponding to arrivals. the Markovmodulated compound Poisson model considered sofar corresponds to A(l) = (. and that are determined by A = A(l ) +A(2) where A is the intensity matrix the governing {Jt}. but the point process of arrivals is not Poisson but renewal with interclaim times having common distribution A of phasetype with representation (v. MARKOVIAN ENVIRONMENT f o r a transition i + j by A . j(2) } be two independent environmental processes and let E(k). We then let (see the Appendix for the Kronecker E = E(1) x E(2).2) A(1) = A(' 1) ® A(1.174 CHAPTER VI.i. refer to notation) { Jt k) }.6i ) diag. the claim surplus is a Markov additive process (cf. Again . In the above setting. then {Nt} is a Markov additive process if and only if it corresponds to an arrival mechanism of the type just considered. II. Jt = (Jtl). that Bii = Bi . we may let {Jt} represent the phase processes of the individual interarrival times glued together (see further VIII. is neither 0 or 1 is covered by letting Bij have an atom of size qij at 0. A(1'k) A(2 k1). and thus 1i = 0. the definition of Bij is redundant for i i4 j. The extension of the model can also be motivated via Markov additive processes: if {Nt} is the counting process of a point process. A(l) = T.1 (PHASETYPE RENEWAL ARRIVALS) Consider a risk process where the claim sizes are i.2 for details).(13i )diag.4).^) etc. we use the convention that a1i = f3i where 3i is the Poisson rate in state i.
. all Al i2. 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 .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). 11. WIDOWED. say. Bilo.. and that the policy then expires.. the idea of arrivals at transition epochs can be found in Hermann [193] and Rudemo [313].kj = Bik) B13 4k = Bak) 175  (the definition of the remaining Bij. possibly having a general phasetype sojourn time...5. RETIRED. with rate ai.1i2. Hermann [193 ] and Asmussen & Koole [37] showed that in some appropriate ..iil... iN. or.. iN. The versatility of the setup is even greater than for the Markovmodulated model.. assume that there is a finite number N of policies. where ik = 0 means that the kth policy has not yet expired and ik = 1 that it has expired. after which it starts afresh. u Example 5 . Easy modifications apply to allow for • the time until expiration of the kth policy is general phasetype rather than exponential.iN C17 AilO... Example 5 .iN = C27 All other offdiagonal elements of A are zero so that all other Bii are redundant. However .. E = { WORKING. Assume further that the ith policy leads to a claim having distribution Ci after a time which is exponential. DIVORCED. DEAD etc. MARRIED.}. 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. more recently.4 (A SINGLE LIFE INSURANCE POLICY ) Consider the life insurance of a single policy holder which can be in one of several states.... Thus. the kth policy enters a recovering state. In fact .kl is redundant).iN are zero and all Bi are redundant..iN.. In this way we can model. E 10. This means that the environmental states are of the form i1i2 • • • iN with il. claims occur only at state transitions for the environment so that AN2... iN = all BOi2. i2i .g. e.iN = a2. INVALIDIZED. • upon a claim. superpositions of renewal processes. Similarly. THE MARKOVIAN ARRIVAL PROCESS Bij.1i2 .iil.. as the Markovian arrival process ( MAP).iN.
Without loss of generality. Neuts [271] and Asmussen & Perry [42]. a claim arrives with rate /3(s + t) and is distributed according to B(8+0 . 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. Sengupta [336]. • Claims arriving at time t of the year have distribution B(t). from an application point of view.3*µs • p = f /3(v) dv 0 0 (6. • The premium rate at time t of the year is p(t). but now exhibiting (deterministic) periodic fluctuations rather than (random ) Markovian ones. 1). For the Markovmodulated model. one needs to assume also (as a minimum) that they are measurable in t. )3 t 1 J (6. 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 . p(t) and B(t) are defined also for t t [0. Lucantoni [248]. we talk of s as the 'time of the year'. Lucantoni et at. B* = J f B(t) ((*) dt. Obviously. let the period be 1. for s E E = [0.1) Then the average arrival rate is /3* and the safety loading rt is 77 = (p* . By periodic extension. We denote throughout the initial season by s and by P(8) the corresponding governing probability measure for the risk process. continuity would hold in presumably all reasonable examples. [248]. Let 1 1 /3* _ f /3(t) dt.p)/p.2) Note that p is the average net claim amount per unit time and µ* = p//3* the average mean claim size. one limitation for approximation purposes is the inequality Var Nt > ENt which needs not hold for all arrival streams. we may assume that the functions /3(t). Some main queueing references using the MAP are Ramaswami [298]. The basic assumptions are as follows: • The arrival intensity at time t of the year is 3(t) for a certain function /3(t). 0 < t < 1. where i f00 xB(°) (dx) _ . p * = 0 p(t) dt. Thus at time t the premium rate is p(s + t).176 CHAPTER VI. 1). .
B*.9). it turns out that they have the same adjustment coefficient. one may think of the standard compound Poisson model with parameters 3*. It is easily seen that . Thus . The claim surplus process {St } two is defined in the obvious way as St = ^N° Ui . and thus the averaged standard compound Poisson models have the same risk for all A.w(t). and we recall from there that the ruin probability is 24 1 *(u) _ 3 5eu + 35e6u. Section 4b). We u assume in the rest of this section that p(t) . the discussion in 111. or. p* = A whereas B* is a mixture of exponential distributions with intensities 3 and 7 and weights 1/2 for each (1/2 = ff w(t)dt = f o (1. the average compound Poisson model is the same as in III.t.3* = 3A. the conditional distribution . since the added variation is deterministic. Thus. 0 Then (by standard operational time arguments ) {St} is a periodic risk process with unit premium rate and the same infinite horizon ruin probabilities. (6.10. In contrast.6. St = SeI(t). respectively. let . 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 . RISK THEORY IN A PERIODIC ENVIRONMENT 177 In a similar manner as in Proposition 1.(3. p* as an averaged version of the periodic model. In contrast. Many of the results given below indicate that the averaged and the periodic model share a number of main features. not random. The behaviour of the periodic model needs not to be seen as a violation of this principle. The arrival process {Nt}t>0 is a timeinhomogeneous Poisson process with intensity function {/3(s + t)}t>0 . we shall see that for the periodic model increasing A increases the effect of the periodic fluctuations.2 Define T 6(T) = p(t ) dt.w(t)) dt).1 As an example to be used for numerical illustration throughout this section. u Remark 6 .1) and Example 1.8.3) Note that A enters just as a scaling factor of the time axis.3(t) = 3A(1 + sin 27rt). for Markovmodulated model typically the adjustment coefficient is larger than for the averaged model (cf. in agreement with the general principle of added variation increasing the risk (cf. In particular.1. of the periodic model as arising from the compound Poisson model by adding some extra variability. Example 6 . equivalently.
[44] (the literature in the mathematical equivalent setting of queueing theory is somewhat more extensive. a) etw*(a) h(s+t. with some variants in the proofs.Q(v) (B(„) [a] . and define h(s.a) Proof Conditioning upon whether a claim occurs in [t.1]) .3(v)(B(vl [a] . let f 8+1 tc *(a) _ (B* [a] .adt +.1) . [101] . 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. 3 E(8)eaSt = h(s. 6b Lundberg conjugation Motivated by the discussion in Chapter II. J Theorem 6 . Jt = (s + t) mod 1 P(8) .g.a .(3(s + t)dt)e«St adt + /3(s + t)dt .. Notes and references The model has been studied in risk theory by.^8 [.f.178 CHAPTER VL MARKOVIAN ENVIRONMENT of U.8). r(u) _ inf It > 0 : St > u} is the time to ruin . (6. given that the ith claim occurs at time t is B(8+t).a.e.3(s + t)dt[B(8+t)[a] .tc* (a)] dv then h (.5. we obtain E.1) dv .. 0 (5)(u. e.(1 .s .(8) [eaSt+dt I7t] = = (1 .. The claim surplus process {St} may be seen as a Markov additive process.T) = P(8)(r(u) <T). . The exposition of the present chapter is basically an extract from [44]. To this end.g.1) a = J8 .g. a) is periodic on R.5 (see in particular Remark 11. we start by deriving formulas giving the m. see the Notes to Section 7). Dassios & Embrechts [98] and Asmussen & Rolski [43]. of the claim surplus process.east B(8+t) [a] east .al.a be the c. of the averaged compound Poisson model (the last expression is independent of s by periodicity). Daykin et.4) At a first sight this point of view may appear quite artificial.a) = exp { . 1). t + dt] or not. with the underlying Markov process {Jt} being deterministic period motion on E = [0. i. As usual.f. and the ruin probabilities are 0(8) (U) = P(s )(r(u) < 00).
a) = h(s.s.3. so that obviously {Lo. a) et.t} is a multiplicative functional for the Markov process { (Jt.2.9) eastt.adt +.(e) Let = h( h(Jo. at + f log h(s + t. a) = exp I f t3(v)(kv)[a) .log h(s.1].0(s + t)dt[B(8+t)[a] .* (a) h(s. RISK THEORY IN A PERIODIC ENVIRONMENT E(8)east+ dt d Et.1)dv l og E(8) et where atetk•(a) h(t. 0) exist and are finite.4 For each 0 such that the integrals in the definition of h(t .6 . it then suffices to note that E(8)Le.5.3(s + t)[D(8 +t)[a] . St)} .1]) .. E (8)east (a +. a) Thus E(8)east = h(s + t.c* (e) {Le. B) eoSt t. a) .9 as follows.6. h(s + t. + v)(B([a] . 9) is a P ( 8)martingale with mean one. dt log E(8)east a + f3(s + t) [B(8+t) [a] .(8)east 179 = = = = = E(8)east (1 . u Remark 6.5 The formula for h(s) = h(s.1]) . Proof In the Markov additive sense of (6. St)} and . a) h(s + t.t = 1 by Theorem 6.1)dv  o h(t. With g the infinitesimal generator of {Xt} = {(Jt.t. According to Remark 11. a) as well as the fact that rc = k` (a) is the correct exponential growth rate of Eeast can be derived via Remark 11. we can write Lo Jt.4). a) Corollary 6. 0) P(8)a.t}t>o = h(s. a).
Proposition 6. (iv) finally.6 The P(s). St)} with governing probability measures Fes).3(s)ks)[a]h(s)} ah(s) 13(s)h(s) + h'(s) +.5 that we can define a new Markov process {(Jt. However. 0 < s < 1. Equating this to rch (s) and dividing by h(s) yields h(s ) = h(s) = a + . say.a . the requirement is cha(i. Bet)(dx) = ^ B(t ) (dx).3(s)h(s) + h'(s) +. Lemma 6 . correspond to a new periodic risk model with parameters ex . the restrictions of Plsi and Pest to Ft are equivalent with likelihood ratio Le. Proposition 6. A further important constant is the value yo (located in (0. 0) = h(s) + dt {ah(s) .T.(3(s)dt) +. yo is determined by 0 = k* (70) = QB*. ry)) at which n* (a) attains its minimum.f.2. .7 When a > yo.'y). MARKOVIAN ENVIRONMENT ha(s.4. (ii) use Markovmodulated approximations (Section 6c).g.tc] dv} (normalizing by h(0) = 1). it follows by Theorem II. Now define 'y as the positive solution of the Lundberg equation for the averaged model. such that for any s and T < oo.6 ( s ) exp { 0( s )&s) [a] + tc . P(s) (T(u) < oo) = 1 for all u > 0.3(s)B(s) [a]h(s). Proof (i) Check that m. Sdt) = h(s + dt) eadt (1 .y) = eayh(s). That rc = is*(a) then follows by noting that h(1) _ u h(0) by periodicity. That is.1) . [70] .0) = Kh(s). see [44] for 11 a formal proof.60(t) = a(t)B(t)[0].3(s)dt • B(s)[a]h(s) = gha(s. ( iii) use approximations with piecewiese constant /3(s). y solves n* (y) = 0. we put for short h(s) = h(s.180 CHAPTER VI. When a = y. That is. For each 0 satisfying the conditions of Corollary 6.3(v)( Bi"i [a] . of St is as for the asserted periodic risk model. B(s).1.3. as above E (s) ha(Jdt. cf. J s [.
0(u)) * (b(oo). considered with governing probability measures { E(8) }E[ . a) e«uE(8 ) e «^ .9 Assume that there exist open intervals I C [0. s E I. J C R+ such that the B(8). have components with densities b(8)(x) satisfying inf sEI. 9(u)) for any bounded continuous function (e.9) and noting that weak convergence entails convergence of E f (^(u). 1). ^(u) = ST(u) . Corollary 6.g. Lemma 6 . a) a > ry0 (6.9(u))} u>0. Wu).10) Then for each a. f (x. The relevant likelihood ratio representation of the ruin probabilities now follows immediately from Corollary 11.9) 0(')(u) = h(s. Here and in the following.6(v) dv Jo ' xe«xB (°) (dx) r^ xe«xB'(dx) = Q'B' [ a] = ^' J 0 = ^c"'(a) + 1. a)e«uE (a iP(s) (u) = h( s)e7uE(` ) h(O(u)) To obtain the CramerLundberg approximation from Corollary 3. which is not used elsewhere in the book. and we refer to [44].1) the distribution of (l: (oo). RISK THEORY IN A PERIODIC ENVIRONMENT Proof According to (6. T(u) < (6.8) (6. e(cc)) Letting u > oo in (6.u is the overshoot and 9(u) = (T(u) + s) mod 1 the season at the time of ruin.2. The proof involves machinery from the ergodic theory of Markov chains on a general state space. and no matter what is the initial season s. has a unique stationary distribution. say s0.2).4. a) TI h(9(u). (6. we need the following auxiliary result .8 The ruin probabilities can be computed as (u)+T(u)k'(a) ^/i(8) (u. we get: . q) = eryx/h(q)). T) = h(s. xEJ 0 (s)b(8)(x) > 0.7) h(B(u). the mean number of claims per unit time is p« 181 = Jo 1. the Markov process {(^(u).1.6. u which is > 1 by convexity. B(oo)).
11) Note that ( 6.10) of Lemma 3. (6. Among other things.W. Noting that ^(u) > 0 in ( 6. 6. we obtain immediately the following version of Lundberg ' s inequality which is a direct parallel of the result given in Corollary 3. MARKOVIAN ENVIRONMENT Theorem 6.1. 1. For our basic Example 6 .ir) } Plots of h for different values of A are given in Fig. it does not seem within the range of our methods to compute C explicitly. Theorem 6 . Vi(8) (u) . 10 shows that certainly ry is the correct Lundberg exponent.10 Under the condition (6.) C = E1 h(B(oo)) u + oo. this provides an algorithm for computing C as a limit.1 In contrast to h. At this stage .Ch(s)ery". illustrating that the effect of seasonality increases with A.9). elementary calculus yields h(s) = exp { A C 2^ cos 2irs  4^ sin 21rs + 11 cos 41rs .16. A=1/4 A=1 A=4 0 Figure 6.11) gives an interpretation of h(s ) as a measure of how the risks of different initial seasons s vary. which may provide one among many motivations for the Markovmodulated approximation procedure to be considered in Section 6c. 11 7/'O (u) < C+°)h(s) ery".182 CHAPTER VI.1. where C(o) = 1 + info < t<i h(t) .6 for the Markovmodulated model: Theorem 6 . where e.
the proofs are basically the same as in Section 3 and we refer to [44] for details. .15) The next result improves upon the constant C+) in front of eryu in Theorem 6.(8) (u.17) (6. RISK THEORY IN A PERIODIC ENVIRONMENT Thus. we obtain Co) = 1. e7 ( yx)B(t)(dy) > Then for all $ E [0.4. #c( ay) < 0 when y > 1/tc'('y). whereas ay < y. e.16 In order to apply Theorem 6.(s)(u) < C+h(s)e7".w ) • 7e u{w • 3e3x + ( 1 .11 as well as it supplements with a lower bound. We state the results below. Lundberg's inequality can be con siderably sharpened and extended.0(8) (u+ yu) (6. yu) 000 (u) .w)e4u dx 9w + 7(1 .12) As for the Markovian environment model. ay) • (6.167r I Cu.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. in our basic example with A = 1.42 • exp {J_ cos 27rs .47r sin 27rs + 167r cos 47rs . r.w)e4u . (ay). 1 (6.7x j dx _7x } _ 6w + 6(1 .w) .g. Just as in IV.3x + (1 .6.13 to our basic example. T) and replace the Lundberg exponent ry by ryy = ay . 1 ) and all u > 0.42 so that 183 tp(8) (u) < 1. Theorem 6 . (6.yr.12 Let 00)(y) 1 Then info < t<i h(t. we substitute T = yu in 0(u.. we first note that the function fu° exu {w • 3e . where ay is the unique solution of W(ay) =y• (6. Theorem 6.13) Elementary convexity arguments show that we always have ryy > Y and ay > ry.(ay) > 0 when y < 1/ic' (7). Consider first the timedependent version of Lundberg's inequality.7e .14) < C+)(y)h(s) e7yu. C_h(s)e7u < V.
1/i18 1 s (u) > 0.9 3 0<8<1 p 27r 47r 167r 161r 2 _ _e. and in fact.T) < C+('Yo)h( s. Thus.L sin 27rs + 1 I cos 47rs .013. C+ = 1. Finally.. . Thus C_ = 2 inf ex cos 2irs .20).cos 27rs .1 sin 27rs + 1 cos 47rs . 1).19 } 0 <8<1 8 + cos 21rs Thus e.(8)(u. and let 8 = er' (Y0). completing a cycle .1 sin 2irs + 16_ cos 47rs . the nth Markovian environmental process {Jt} moves cyclically on {1.g.16.16) with 'y replaced by yo and h(t) by h(t. 6c Markovmodulated approximations A periodic risk model may be seen as a varying environment model.20 •exp { 2n cos 27rs .4^ sin 2irs + 16^ cos 41rs .184 CHAPTER VI.18) Notes and references The material is from Asmussen & Rolski [44]. 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. with s the initial season.\ 3 C+ = sup 6 exp { A (. but thereby also slightly longer. where the environment at time t is (s + t) mod 1 E [0.0.\ = 0 .19 I eu.66. MARKOVIAN ENVIRONMENT attains its minimum 2 /3 for u = oo and its maximum 6 /(7 + 2w) for u = 0.I eu. Of course. yo). This observation motivates to look for a more formal connection between the periodic model and the one evolving in a finite Markovian environment.013. 14 Let C+('yo) be as in (6.'Yo)e (6. such a deterministic periodic environment may be seen as a special case of a Markovian one (allowing a continuous state space E = [0. 1) for the environment). exp 2^ cos 21rs . Then 7oudT . Some of the present proofs are more elementary by avoiding the general point process machinery of [44]. .. we have the following result: Theorem 6 .66. for A = 1 (where 3 e0. n}. 0 <'p(8)(u ) .181 s(u) < 1. . The idea is basically to approximate the (deterministic) continuous clock by a discrete (random) Markovian one with n 'months'.
Let 0j. it is desirable to have formulas permitting freely to translate from one setting into the other. since the settings are equivalent from a mathematical point of view. We let {Stn)} (6.1 ((i 1)/n) ) and Bni = B . so that the intensity matrix is A(n) given by n n 0 ••• 0 0 n n ••• 0 A(n) _ (6. To this end. M(n) = Supt>o Stn). 7 Dual queueing models The essence of the results of the present section is that the ruin probabilities i/ (u). (6.21) which serves as an approximation to 0(1)(u) whenever n is large and i/n s. . This queue is commonly denoted as the Markovmodulated M/G/1 queue and has received considerable attention in the last decade. one simple choice is Oni = 0( i . 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 . 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. z/'i (u.jEE of the risk process.7. Bi. AE= Aii'r?/7ri• The arrival intensity is /3i when Jt = i.20) be the claim surplus process of t>o the nth approximating Markovmodulated model. but others are also possible. and the ruin probability corresponding to the initial state i of the environment is then Y'yn)(t) = F (M(n) > t).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]. 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. Thus.
=1 . and the virtual waiting time (workload) process {Vt}too are defined exactly as for the renewal model in Chapter V.1) 7ri In particular. For (7. JT = i} coincide.n(VT > u. JT = Z). JT = j) = 7rjPj(VT > u. (7.1 Assume V0 = 0. and (7.P(V > u. . (7.. . JT = i) = 'P.oo in u (7. JT = j) = LjPj (VT > u. just sum (7. I* = i).2). {Jt }o<t<T• Then we may assume that Jt = JTt. Then Pi(T(u) < T. The first conclusion of that result then states that the events {T(u) < T. Now let In denote the environment when customer n arrives and I* the steadystate limit. • The queueing discipline is FIFO. JJ = i). ii (u) = it /3 P(W > u.2) Oi(u) = 1. Taking probabilities and using the stationarity yields 7riPi(T(u) < T. (VT > u I JT = 2).. and for (7.2) and use that limF (VT > u. Jo = j. (7. T) = 7ri 1 P. 2 .4) where 0* = >jEE 7rj/3j. (7.T(V > u I J* = i).1). J* = i) for all j. let T . In particular. I* )3i P(V > u. Proof Consider stationary versions of {Jt}o<t<T.3). JT = j} and {VT > u.2 The relation between the steadystate distributions of the actual and the virtual waiting time distribution is given by F(W > u. MARKOVIAN ENVIRONMENT • Customers arriving when Jt = i have service time distribution Bi. The actual waiting time process 1W1. Proposition 7.1) over j. J*) is the steadystate limit of (Vt.3) 7ri where (V. Jt )..3. Jo = i.1) follows. J* = i). J* = i) = P.0i (u .186 CHAPTER VI. JT = i) = P(V > u. Proposition 7. 0 < t < T and that the risk process {Rt}o<t<T is coupled to the virtual waiting process {Vt}o<t<T as in the basic dualitylemma (Theorem 11.
.=i) a4. that /3(t).7) of that paper. Proposition 7.7. J* = i) see W > u. I*) with the timeaverage . see Regterschot & van Doorn [123]. P(. With {Vt} denoting the workload process of the periodic queue.6) (7. u Notes and references One of the earliest papers drawing attention to the Markovmodulated M/G/1 queue is Burman & Smith [84]. a general formalism allowing this type of conclusion is 'conditional PASTA'.I *=i). n=1 N However. see in particular Harrison & Lemoine [186]. [243]. DUAL QUEUEING MODELS 187 Proof Identifying the distribution of (W. The first comprehensive solution of the waiting time problem is Regterschot & de Smit [301].4). >u. N * oo. . with (7. P(W >u. and (7. In the setting of the periodic model of Section 6.1 is from Asmussen [16].4) and (7. on average 0*T customers arrive in [0.T)(T(u) <T) = P(8)(VT > u). 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. P(1')(r(u) < oo) = P(')(00) > u). Lemoine [242]. and one has PI'>(rr(u) < T) = P('_T)(VT > u). on average /32TP(V > u. B(t) have been periodically extended to negative t). The relation (7. we have 1: I(W.3) improving somewhat upon (2. a paper relying heavily on classical complex plane methods.l. Taking the ratio yields (7. (7. p < 1 then ensures that V(*) = limNloo VN+9 exists in distribution.5) follows from (7.o. and Rolski [306].I. A more probabilistic treatment was given by Asmussen [17]. I* = i.3). T].g.8) For treatments of periodic M/G/1 queue.7) (7. if T is large. and of these.4) can be found in Regterschot & de Smit [301]. and further references (to which we add Prabhu & Zhu [296]) can be found there.
This page is intentionally left blank .
finite horizon. t] are Nt At = Ui (1. with common distribution B and independent of {Nt}. 189 . and the evolution of the reserve may be described by the equation Rt = u . the premium charged is assumed to depend upon the current reserve Rt so that the premium rate is p(r) when Rt = r. . and T(u) = inf {t > 0 : Rt < u} is the time to ruin starting from Ro = u so that '(u) = F(T(u) < oo). U2.T) = F(T(u) < T). the aggregate claims in [0..2) tk(u. are i. Thus.6. i&(u.i..Chapter VII Premiums depending on the current reserve 1 Introduction We assume as in Chapter III that the claim arrival process {Nt} is Poisson with rate .At + p(R8) ds. Thus in between jumps. {Rt} moves according to the differential equation R = p(R). z/i(u) = F IinffRt< 0IRo=u 1 (1. and that the claim sizes U1.1) (other terms are accumulated claims or total claims). Zt As earlier. resp .d. However .T) = FloinfTRt< OIRo=u1 denote the ruin probabilities with/initial reserve u and infinite.
the payout rate of interest is Sx and absolute ruin occurs when this exceeds the premium inflow p. A basic question is thus which premium rules p(r) ensure that 'O(u) < 1. i. rather than when the reserve itself becomes negative.'(u)) > 0 so that V'(v) < 1. say e. or o(u) < 1 for all u. If Ro = v < u. Hence in terms of survival probabilities. it seems reasonable to assume monotonicity (p(r) is u . we can put Rt = Rt + p/S.190 CHAPTER VII. However. and the probability of absolute ruin with initial reserve u E [p/S. No tractable necessary and sufficient condition is known in complete generality of the model. say at interest rate b. Example 1.p2. but assume now that the company borrows the deficit in the bank when the reserve goes negative. 1 . Proposition 1. that {Rt} will reach level u before the first claim arrives. That is.i(u) = 1 for all u.Vi(v) u > e(1 .p/S) r > p/S p5(p/5r) 0<r<p/5 Then the ruin problem for {Rt } is of the type defined above.2. but when the reserve comes above v. dividends are paid out at rate pi .3 (ABSOLUTE RUIN) Consider the same situation as in Example 1. Proof Obviously '(u) < ilb(v) when u > v. In this situation. where one would try to attract new customers as soon as the business has become reasonably safe. Assume 0(u) < 1 for some u. Thus at deficit x > 0 (meaning Rt = x). we get p(r) = p + er. pi > p2 and p(r) = One reason could be competition.1 Assume that the company reduces the premium rate from pi to p2 when the reserve comes above some critical value v. RESERVEDEPENDENT PREMIUMS The following examples provide some main motivation for studying the model: Example 1 .e. there is positive probability. P(r) _ p + e(r . Now return to the general model. oo) is given by i (u + p/S). when x > p/S.2 (INTEREST) If the company charges a constant premium rate p u but invests its money at interest rate e. Another could be the payout of dividends: here the premium paid by the policy holders is the same for all r. Example 1.4 Either i.
INTRODUCTION 191 decreasing in Example 1. { Vt} remains at 0 until the next arrival). We next recall the following results. However.4.2(d)).. instead of (1. obviously infu<uo z/'(u) > 0. Hence ik(u) < 1 for all u by Proposition III.I3IB requires a more detailed analysis and that µB < oo is not always necessary for O(u) < 1 when p(r) 4 oo. [APQ] pp. In particular. then l/i(u) < 1 for all u.o(uo) = 1 so that t/'(u) = 1 for all u by Proposition 1.2) for r sufficiently large so that p(oo) = limr. (1. . (1. we have z/i(u) <p(u .1.1.1 and increasing in Example 1. let uo be chosen such that p(r) > p = 0ILB + e for r > uo.2 once more. say V. appealing to Proposition 111. the probability that Rt < uo for some t is at least tp(0) = 1 (cf.1. {Vt} decreases at rate p(v) when Vt = v (i.uo) and. then ?(u) = 1 for all u.6 For any T < oo. Proof This follows by a simple comparison with the compound Poisson model.uo) < 1. one can couple the risk process and the storage process on [0. Here {Vt}two is a storage process which has reflection at zero and initial condition Vo = 0. (b) If p(r) > /3µB + e for all sufficiently large r and some e > 0.4) 0 and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0. 296297): Theorem 1. Then if u > no. In case (a).3. that u zPp(u . In case (b). and hence by a geometric trials argument.1. cf. hence Rt < uo also for a whole sequence of is converging to oo.5) and the process {Vt} has a proper limit in distribution . T] i n such a way that the events {r(u) <T} and {VT > u} coincide.f p(Vs) ds. and P(Rt + oo) > 0. Let Op(u) refer to the compound Poisson model with the same 0.T) = P(VT > u). Starting from Ro = uo.b(u.2(d). V = p(V)).6) .2) we have t Vt = At . if and only if V)(u) < 1 for all u. In between jumps.e. Then 0(u) = P(V > u). B and (constant) premium rate p. This is basically covered by the following result (but note that the case p(r) .+ p(r) exists.5 (a) If p(r) < /. which was proved in 11. (1.3µB for all sufficiently large r. let uo be chosen such that p(r) < p = /3µB for r > uo. That is. Theorem 1. Proposition I1I.
8) is the rate of downcrossings (the event of an arrival in [t.y.h.8) as the rate of upcrossings.y)g(y) dy.Sx} dx.s. we arrive at the desired interpretation of the r. Considering the cases y = 0 and 0 < y < x separately. Oeax f x e'Yg (y) dy } = p) eaxa(x) . In view of the path structure of {V t }. t + dt] can be neglected so that a path of {Vt} corresponds to a downcrossing in [t. of (1. we thus need to look more into the stationary distribution G. say when {Vt} is in state y. yo ^ 1 + oo Q exp {. It is intuitively obvious and not too hard to prove that G is a mixture of two components. one having an atom at 0 of size 'yo.6w(x) . say.192 CHAPTER VII.7 p(x)g(x) = tofB (x) + a f (x .9) Proof We may rewrite (1. oo).7) Proposition 1. It follows in particular that 0(u) = fg(Y)dy.8) as g(x) = p 1 {yo13e_6x +. Then the ruin probability is tp (u) = f' g(y)dy. An attempt of an upcrossing occurs as result of an arrival. B(x) = e. the l. RESERVEDEPENDENT PREMIUMS In order to make Theorem 1.Sx}. Corollary 1.h. (1.8) Proof In stationarity. where g(x) = p( ^ exp {.6x and that w(x) < oo for all x > 0. the flow of mass from [0.s.Qw(x) . say. Now obviously. Jo AX) (1. of (1. x] to (x. x + p(x)dt]). for the storage process {Vt}. say if p(r) goes to 0 at rate 1 /r or faster as r j 0. and the other being given by a density g(x) on (0. (1. t + dt] if and only if Vt E [x. Note that it may happen that w (x) = oo for all x > 0. and is succesful if the jump size is larger than x .6 applicable. oo) must be the same as the flow the other way. this means that the rate of upcrossings of level x must be the same as the rate of downcrossings. u Define ^x 1 w(x) Jo p(t) dt. Then w(x) is the time it takes for the reserve to reach level x provided it starts with Ro = 0 and no claims arrive.8 Assume that B is exponential with rate b.
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
2y +µ ) dy .5 The analysis leading to Theorem 2.pa. and the c. i. of Z is IogH[a] = f ytc(y)dy = e fa (0. As an example.a) . assume that {Wt} is Brownian motion with drift µ and variance v2. 13/E).3./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. . 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 .g.200 CHAPTER VII.3/E) By the memoryless property of the exponential distribution.pa. RESERVEDEPENDENT PREMIUMS u from which (2.b P/E dx /' P/ ' (p/  x)p/e 150/f I' (/3/E) (6P1'E. with density x(3/e1aQ/e fV (x) _ e 6X ' x > 0.V < x)]0 + f P(V > p/E ) + eby fv (p/E .3 is also valid if {Rt} is obtained by adding interest to a more general process {Wt} with stationary independent increments. H(u) = P(Z r < u) = P(V > u + p/E) = (8(p + Eu)/E.e.01'E) + (p/E)al aO l febP/E } IF (0 /0 jF From this (2. Proof 2 We use Theorem 2.V. RT(u) has an exponential distribution with rate (S) and hence E [H(RT(u))I r(u) < oo] L Pe6'r (P/C .3a/ (5 . where V is Gamma(b.13 /E) r (. From ic(a) = .x) dx e.2) follows by elementary algebra.f. r (j3/E) In particular.2) follows by elementary algebra. /^ u Example 2 . The process {St} corresponds to {Wt} so that c(a) or2a2/2 .
[282].3) was derived by Emanuel et at. Gerber [155]. [283].g.1) . write y* for the solution of the Lundberg equation f3(B[ry *] . write Vi* (u) for the ruin probability etc. THE LOCAL ADJUSTMENT COEFFICIENT _ Q2a2 pa 4e E 201 I. Further studies of the model with interest can be found in Boogaert & Crijns [71]. [357]. it follows that the ruin probability is Cu) H(u) H(0) 11 Notes and references Theorem 2. as in the proof of Proposition 2. [129] and Harrison [185]. that the analysis does not seem to carry over to general phasetype distributions. 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).i. Paulsen & Gjessing [286] found some remarkable explicit formulas for 0(u) beyond the exponential case in Corollary 1. and recall Lundberg 's inequality .v. A r.e. se e. Q2/2E). Paulsen & Gjessing [286] and Sundt & Teugels [356]. The formula (2.. however.Y*p* W*(u) < ery*u = 0. Delbaen & Haezendonck [104]. Paulsen [281]. Corollary 2. Goldie & Griibel [167]. Some of these references also go into a stochastic interest rate.p*.d. it is also used as basis for a diffusion approximation by these authors.g. Gerber [157] p. or to nonlinear premium rules p(•). see e.8. 134 (the time scale there is discrete but the argument is easily adapted to the continuous case)..3 is from Harrison [185]. not even Erlang(3) or H3.2 is a special case of a perpetuity. and since RT = 0 by the continuity of Brownian motion. 3 The local adjustment coefficient. of the form Ei° p"X" with the X„ i. for a martingale proof.4 is classical. Z is normal (p/E. Logarithmic asymptotics For the classical risk model with constant premium rule p(x) . Emanuel et at. It must be noted. [129].3.
. and (for simplicity) that inf p(x) > (3µs . 1) and for a given E > 0. Let y* < So.1. we will use the local adjustment coefficient 'y(x). choose c(.*(u) .5) which implies inf.4) we assume existence of y(x) for all x.w (u) J dt > c(3)eeu v 1 p(u+ t) .1 ).ap(x).C*ef*". then log u (u) In the proof as well as in the remaining part of the section . (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) . (3. Then lim sup u>oo u and e E''p(r) + 0. For the last asssertion . 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). 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 . i.202 CHAPTER VII.e. B(x) > C(2)e(ao+f)x for all x.>o 7(x) > 0. which in turn by Lundberg's inequality can be bounded by ery*(1E)" Hence limsup„. let p* be a in (3. oo for all E > 0. choose uo such that p( x) > p* when x > u0E. c(.. as solution of the equation n(x.log '(u)/u < ry*(1 . e(1o+e)2 (x ) u > 00. as will hold under the steepness assumption of Theorem 3.e. x * oo. a first step is the following: Theorem 3 . If 60 s f 6o.'y ( x)) = 0 where r. Letting first E * 0 and next ry * T 5o yields the first statement of the theorem. Proof of Theorem 3. (3. When u > uo.1 Assume that for some 0 < 5o < oo. it holds that f3[s] T oo.3) When trying to extend these results to the model of this chapter where p(x) depends on x. and that p(x) * oo.1) .i)eex. The steepness assumption and p(x) + oo ensure 'y(x) * So. i.1. x>0 (3.E). log ?i(u) < < 00 JO .2) such that p(x) < c(. a) = f3(B[a] . 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. RESERVEDEPENDENT PREMIUMS and the CramerLundberg approximation V.
The slow Markov walk limit is appropriate if p(x) does not vary too much compared to the given mean interarrival time 1/0 and the size U of the claims.3 Assume that either (a) p(r) is a non decreasing function of r. UJU > x cannot have a much heavier tail than the claim U itself. Theorem 3 . or (b) Condition 3. Bucklew [81]).6) The second main result to be derived states that the bound in Theorem 3..' (u) < eI(").13 is a technical condition on the claim size distribution B. (u) = O(u/e).0 are the same. let 0e (u) be evaluated for the process only with 3 replaced by /0/e and U. and hence '(u) > c(4)eeuc( 2)e(do+e)u The truth of this for all e > 0 implies lim inf log V. {Rte)} defined as in (1. and in particular. Theorem 3.(u) > so. Therefore the probability that a claim arrives ( before the reserve has reached level u + v is at least c(. 3. one can then assume that e = 1 is small enough for Theorem 3. 2 Assume that p(x) is a nondecreasing function of x and let I(u) = fo ry(x)dx. the result is not very informative if bo = oo. the limit is not u + oo but the slow Markov walk limit in large deviations theory (see e. which essentially says that an overshoot r.2). The first main result in this direction is the following version of Lundberg's inequality: Theorem 3 . The form of the result is superficially similar to the CramerLundberg approximation. (3. If p(x) = pis constant .13 below holds. the asymptotics u * oo and c .7) CIO Remarks: 1. Condition 3. I.1 only presents a first step. 2.2 is also an approximation under appropriate conditions.ea°/(ecf1)).3. Then limelog l/ie (u) = I(u). u Obviously.4)eE" Given such an arrival.v. noting that in many cases the constant C is close to 1. THE LOCAL ADJUSTMENT COEFFICIENT 203 where c. then Rte) = CRtie for all t so that V). . 3) = (1 .e. However. Then . ruin will occur if the claim is at least u + v. (3. For e > 0. The rest of this section deals with tail estimates involving the local adjustment coefficient.3 to be reasonably precise and use e` (u) as approximation to 0 (u).g. by cU2.
4 Consider again the exponential case B(x) = eax as in Corollary 1.bu}.bx}]o + b /' oo exp low (x) .2.bx} dx 1+0. and r j 1 'Yo v(x)dx = bu  a J0 p(x)ldx = Integrating by parts.1 + b f e. we consider some simple examples.(x) dx. the logaritmic form of (3. RESERVEDEPENDENT PREMIUMS 4.3.3.bx} dx = 1 + J0 dodx(x) exp {.bx} dx oo exp low(x) bx dx 70 Ju r oo = b J exp low (x) . As typical in large deviations theory. 3a Examples Before giving the proofs of Theorems 3.(3/p(x). we show how to rewrite the explicit solution for ti(u) in Corollary 1. we get = 1+ J" AX) exp {(3w (x) . One would expect the behaviour in 2) to be important for the quantitative performance of the Lundberg inequality (3.exp {/33w(u) . First.8 in terms of I(u) when the claims are exponential: Example 3 . u .204 CHAPTER VII.6). 3.bx} dx fo 00 1 + [exp {/(3w(x) . 5.8. say. rather than eI(u)).(iw(x) .bx} dx . However. J0 ^oo g(x ) dx f AX) lexp IOW (X ) bx + b u 1 exp low(x) . it is formally needed only for Theorem 3.7) 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). Then y(x) = b .
ry(x /b I u o e f0 °° e  e. note first that the appropriate slow Markov walk assumption amounts to u. in the definition of AE converges to 0.fo 7(x)dx/Edy f . In particular.10 or Karlin & Taylor [222] pp.f y(x)dxd y If 7(x) is increasing .fo 7(x) dx /E dy > av 'yo /Edy = E (1 . > lime log e = 0 and AE * 0.7) follows. oo) with drift µ(x) and variance a2 (x) > 0 at x.I ( v )dy fo +u) dxdy .9) yields e log .3.2(X) = ev2(x) so that 7e(x) = 7(x)/e.9 ) 11000 eI(v)dy f000 e.5) is infx>o 7(x) > 0 which implies that f °O .e.1. (X) = µ(x).BE.8) 7(x)dxdy 11 We next give a direct derivations of Theorems 3. It is well known that (see Theorem XI.10) where AE = e log 000 e.. u .3 in the particularly simple case of diffusions: Example 3. and (3.5 Assume that {Rt} is a diffusion on [0. Similarly.2.fa 7(x+u)dx/Edy o The analogue of (3.. IE(u) = I(u)/e. and (3. The appropriate definition of the local adjustment coefficient 7(x) is then as the one 2p(x)la2(x) for the locally approximating Brownian motion. Choosing yo. 3.I ( u) fool. BE * 0. (u) = I(u) + AE . 70 > 0 such that 7(x) < 7o for y < yo.3. For Theorem 3.1/8 . (3.ev 0 O /E) J0 70 70 Yo This implies lim inf A. 0. 191195) that 1P (U) = fu0 eI(v)dy = eI(u) follo e.0. THE LOCAL ADJUSTMENT COEFFICIENT and hence 205 f°° eI(v )dy . (3.fory(x+u)dxdy ( 3. we get r 00 e.2. applying the inequality 7(x + u) > 7(x) yields immediately the conclusion of Theorem 3. Be = e log U000 e. the integral is bounded by 1 eventually and hence lim sup AE < lim sup a log 1 = 0.
7o C 15 I I.5 for risk processes with exponential claims is as follows: Example 3 . ..6/p* so that u 1 I (U) = bu . G. we have 5 > 7o and get lim inf AE > lime log e .5. the slow Markov walk assumption means 5E = b/c. the slow Markov limit a * 0 and the limit u walk approximation deteriorates as x becomes large. 0 Now (3.7) follows just as in Example We next investigate what the upper bound / approximation aI (°) looks like in the case p(x) = a + bx (interest) subject to various forms of the tail B(x) of B. this leads to (3. Thus 7e(x) _7(x)/e and (3.Q/p*. E+o e*O By (3. (u) representing the first few terms in the asymptotic expansion of I(u) as u + oo. the results are suggestive in their form and much more explicit than anything else in the literature. G.1 3.4.10) holds if we redefine AE as AE = flog (j °° efo 7(x)dx/edy _ E/5 I and similarly for B.+1 (u) = o( 1). _ . > ./3 1 AX dx.5. + Gq(u) + o(G9(u))• Gi (u) It should be noted . 0. I. . Nevertheless .0.(u) oo. As in Example 3. Then the solution of the Lundberg equation is y* = b .. however . that the interchange of the slow Markov walk oo is not justified and in fact.5) and 7* = 5 .0) = 0.0/e. . . Ignoring 1/5 in the formula there. Of course. ) Note that this expression shows up also in the explicit formula for lk(u) in the form given in Example 3.5. 7(x) is typically not explicit. RESERVEDEPENDENT PREMIUMS The analogue of Example 3. so our approach is to determine standard functions Gl (u).6) exactly as in Example 3.e.. Further.206 CHAPTER VII. I(u ) = G1(u) + . lim sup Af < lim sup c log(1 .. . G.6 Assume that B is exponential with rate S.
x)n1. fu I(u) Su .1 =$ f cse8 Sn f e"B(x)dx = e8 Jo s eIB ( 1 . .4) or gamma distributions.8).7 Assume that B(x) . ry* loge*+ g7loglogp*. It follows from (3. THE LOCAL ADJUSTMENT COEFFICIENT Example 3 . x T 1.11) with a > 0. if the phase generator is irreducible ( Proposition VIII. say 1 is the upper limit and B(x) . 2. Here B[s] is defined for all s and B[s] . u(logu + r7loglogu).s)C' f "o o as s T S. B[s] = 1 + s exB(x)dx = 1 +c1SF(a) ('+o(')) (S .1) leads to (S7T N Ocp a.C2p* C2 = (3clr( a))11'.g.:. I(u) Pt. in the phasetype case . e. the typical case is a = 1 which holds . 1.. Hence (3. u Example 3 . . c4 = c2b 1/'/(1 .c2 Su a dx ) Su a<1 Su .c3 logu a= 1 J 0 a + bx 1/ ( c4ul 1/° a > 1 where c3 = c2 /b. more generally.1/a).cs(1 .Y . This covers mixtures or convolutions of exponentials or. 77 = 1 if B is degenerate at 1. phasetype distributions (Example 1.1) leads to .clxale5x 207 (3.3cse7*I7(77) . For example. More precisely.3.11) that b[s] * co as s f S and hence 7* T S as p* + oo.12) with y > 1.ry*°p*. and hence (3. y = 2 if B is uniform on (0. (3.y/s)dy sn 1 1 f ' evy'7ldy = cse8r(T7) as s T oc.8 Assume next that B has bounded support. 1) and 17 = k + 1 if B is the convolution of k uniforms on (0.1/k).
r„(Ti).f. of the increment in a small time interval [0. h 10.c8 log .4) is the formula h logEues ( Rhu) .u . h]. (3.12). (3.f.(T1)) > Ee7o(u)(ul+vr»(Ti)).ru(TI)) .208 CHAPTER VII.•.u is a nondecreasing function of u.10 Assume that p(x) is a nondecreasing function of x.B[7o (u)] . one could also have considered the increment ru (T1) .9 As a case intermediate between (3. ec78)2/2c7 dx C7 .14) for the m .Ul up to the first claim (here ru (•) denotes the solution of i = p (r) starting from ru(0) = u).Ote7o( u)(u.e7o ( u)(ul+u r. 1 0 3e.sp(u).css 2%rc7eC782/2. (3. this is only possible if 7o(v) 2 7o(u)• . The assumption implies that ru(t) .log p*.2 We first remark that the definition (3. assume that B(x) CO x2/2c7.(t))dt.. (b) 'y(x) <'Yo(x)• Proof That 7(x) is nondecreasing follows easily by inspection of (3.r^. 3b Proof of Theorem 3.1) . I (u) c8u log u 0 where c8 = 2/c7. 1 = E.13) We get b[s] .4) of the local adjustment coefficient is not the only possible one: whereas the motivation for (3.1 Cgs o"O 0 esxex2/2c7 dx = cgsec782/2 f . RESERVEDEPENDENT PREMIUMS Example 3 . This leads to an alternative local adjustment coefficient 7o(u) defined as solution of 1 = Ee''o(u)(vi+u . 7 * .g.11) and (3. g.3 (B[s] . Hence for u<V. . of U1 + v .15) Proposition 3. x f oo .4). By convexity of the m . Then: (a) y(x) and 7o(x) are also nondecreasing functions of x.
1) . We shall show by induction that (' Y'(n) (u) < e fo 'Yo(x)dx (3. Then (u) < efo Yo(x)dx. Assume (3. we obtain „I. note that the assumption implies that ru(t) . Also.11 Assume that p(x) is a nondecreasing function of x. THE LOCAL ADJUSTMENT COEFFICIENT For (b). u We prove Theorem 3. (3.es'Yo(u)Fu(dx)} o0 e fo yo( x)dx j.2 in terms of 7o.3. The case n = 0 is clear since here To = 0 so that ik(°)(u) = 0. Hence 1 = EeYo(u)(U1+uru(T1)) < E.(n+l) (u) 1 .(n+1) (u) efo Yo(x)dxI^"Q exyo( I u u)Fu(dx )+ J .17) from which the theorem follows by letting n + oo.u > tp(u).16) Proof Define 411(n)(u) = P('r(u) < on) as the ruin probability after at most n claims (on = TI + • • • + Tn).4) considered as function of 7 is convex and 0 for y = 0.u[70(u)] fo eyo(x)dx . the case of 7 then follows immediately by Proposition 3.e70(u)(U1P(u)T1) 209 0 + 7o(u)p(u)' 0 <_ 00['Yo( u)] .17) shown for n and let Fu(x) = P(U1 + u .7o (u)p(u)• Since (3. Separating after whether ruin occurs at the first claim or not. it is easily seen that fu x7o(y)dy < xyo (u).10(b): Theorem 3.ru(T1) < x). fa 7o(y)dy < u7o(u) < xyo (u) for x > u. Hence „/.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. this is only possible if yo(u) > 7(u).Fu(u ) + J ^(n)(u .
(u). {RtE)} (starting from u = un. and.12 lim sup4^o f log O.n = ku. For these reasons. in . 0. Let Ck... the value of {R(E)} at the time of downcrossing is < unl.n) pn niE (u /n) n n_1 n. W O . (un2. 3c Proof of Theorem 3.. Further. resp.E (u/ n) Y'E (un .11 be reasonably tight something like the slow Markov walk conditions in Theorem 3.E (u/n) Now as e . the probability that ruin occurs in the CramerLundberg model with p* = pn.n AX).n. given downcrossing occurs. (u) < I(u).10(a) for some of the inequalities. To this end. pk n = uk_l.11 is sharper than the one given by Theorem 3.E (u) denote the ruin probability for the classical model with 0 replaced by .n..n. x + x/n] by two classical risk processes with a constant p and appeal to the classical results (3. However.n = sup p(x).n) > k =1 II v ^k n. ryk. 0 It follows from Proposition 3.n. op*.3 The idea of the proof is to bound { R( f) } above and below in a small interval [x . for either of Theorems 3.: y(u).n so that n. The probability of this is at least n n. P k. RESERVEDEPENDENT PREMIUMS where the last identity immediately follows from (3.I.2.2.n u k}1. define uk. 3.n) must first downcross unl. we have chosen to work with y(u) as the fundamental local adjustment coefficient.15).3 is required.n <Z auk}l.3). Proof For ruin to occur.nbe C*.3).210 CHAPTER VII. Also.3/e and U. yo(u) appears more difficult to evaluate than y(u). we used also Proposition 3.10(b ) that the bound provided by Theorem 3.e (u) = v'. C*e where the first equality follows by an easy scaling argument and the approximation by (3.n inf n uk1. (3. let Op*.E ( u/n) ^•e. Lemma 3..E (u/n). by €U=.2). in accordance with the notation i/iE (u).x/n. y* evaluated for p* = Pk. Y*u /E.n (starting from u/n) without that 2u/n is upcrossed before ruin.. and here it is easily seen that yo(u) . 0.
for all x .e..nk=1 limsupelogv). /' (u/n) 'T nk.. i.nu /En) o(1)). so that Theorem 3.log Ck. k=1 k=1 n u _ nE7 k.a( u)z. we need the following condition: Condition 3. B(x) (3.nu /fn( 1 Ck  e. 0 with n and u fixed. uk_1.n + 0(1).12 completes the proof.F (2u/n). *p•.13 There exists a r. 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. also ryk. THE LOCAL ADJUSTMENT COEFFICIENT particular. y > 0 it holds that F(U>x +yIU>x) B(x + y) < F (V > y).18) (ii) the family of claim overshoot distributions is stochastically dominated by V.n . (3.i. It follows that n log V'C (u) k =1 log Ypk.n = sup ?'(x).. 11 Theorem 3. (u) CIO < Letting n 4 oo and using a Riemann sum approximation completes the proof. Indeed . since ry' is an increasing function of p'.nu/en(1 + where o(1) refers to the limit e . v..2 gives 7PE (u) < eIi"i/f = lim inf Clog 0E (u) > I (u). 40 Combining with the upper bound of Lemma 3.19) .n. ne7k.n cE (2u/n) Ck ne7k. in obvious notation one has tC (x) = y(x)/e. 211 Clearly. In case (b). 3 now follows easily in case (a).3.n <X<Uk.E (u/n) OP +^p•.7k. .E (urn) < \ *I.E (u/n) Op•.! (u/n) n n m 7k.
u/n)) I T(E) (u.( . Write EO. Ei + E2 < e71.n V. V < u/En] + P(V > u/En) (u/En . where El is the contribution from the event that the process does not reach level 2u/n before ruin and E2 is the rest.^(E) (u.QEU 1 . we first note that the number of downcrossings of 2u/n starting from RoE) = 2u/n is bounded by a geometric r. infx>2u /n P(x) .212 CHAPTER VII. For E2. T() (u.E(E) (u. T(E) (u. u/n) < oo] E [OE (u/n . (R. N with EN < 1 = infx>2u/nA(x) = 0(1).V) = e71 nu/Eno(l) (using (3. ) (u u /n)) .E (2u/n .E (u/n . . The probability of ruin in between two downcrossings is bounded by Epp . let v < u and define T(E) (u.nu /EnE [e71. u/n) < oo) EV).of:>2 in n(x).n < ery1.EV) = e.eV) • P (T(E) (u.R<) (u v). u /en 0(i) _n so that E2 < e2ryl nu/En0(1).2y 1 ' . (u/n ..EV) = EiI 1 .EV) = El + E2. Then the standard Lundberg inequality yields El < E?.^'' = E [ .. v ) = inf { t > 0 : R(c ) < v R) = u } .nu/En0(1) . u /n) < oo] l = = < E [OE (u/n .1 n. (u/n . u/n)) .18) for the last equality). Then Y'E (u) ^(E) (u. u/n) < oo) .5) and the standard formula for b(0). RESERVEDEPENDENT PREMIUMS To complete the proof. P (T(E) (u. u/n) < oo] . (3.v. v ) = v .E (0) cf...
J y(Rs)dR. Djehiche [122] gives an approximation for tp(u.g. [89]. Similarly.1. Bucklew [81]).T) = P „(info<t <T Rt < 0) via related large deviations techniques. (u) 40 213 lim inf e log(Ei +E2) + logP (r(`) (u. l o JJJ o . u Notes and references With the exception of Theorem 3.7(x)) (3. Whereas the result of [122] is given in terms of an action integral which does not look very explicit. 0 and b T 00 are interchangeable in the setting of [89]. T) is maximized over T by taking T as the time for (3. THE LOCAL ADJUSTMENT COEFFICIENT Hence lim inf e log Ali.7) for ruin probabilities in the presence of an upper barrier b appears in Cottrell et al.=1 J An approximation similar to (3.3EU) (3. Typically.)ds + Y(R2. the results are from Asmussen & Nielsen [39].r. one can in fact arrive at the optimal path by showing that the approximation for 0(u. it might be possible to show that the limits e .) = exp .u/n) < oo) CI  > u n n ryi n' i=1 Another Riemann sum approximation completes the proof. the approximation (3. where the key mathematical tool is the deep WentzellFreidlin theory of slow Markov walks (see e . . the risk process itself is close to the solution of the differential equation r(x) _ r (x.4) and the prime meaning differentiation w. s) as in (3. Comparing these references with the present work shows that in the slow Markov walk setup.J r(Rs)p(R.13.t.)Ui } . s). 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. they also discuss simulation based upon 'local exponential change of measure' for which the likelihood ratio is ( /'t /'t Ns Lt = exp S .7) then comes out (at least heuristically) by analytical manipulations with the action integral.21) (the initial condition is r(0) = u in both cases). u/n) < oo { 40 )I U nryl n+liminfelogP (T(')(u. whereas the most probable path leading to ruin is the solution of r(x) _ k (x.3.21) to pass from u to 0. 0 ) (= p(x) .20) (with ic(x.
. however. e. RESERVEDEPENDENT PREMIUMS the simplest being to require b[s] to be defined for all s > 0 (thus excluding .g.214 CHAPTER VII. . For different types of applications of large deviations to ruin probabilities . see XI. We should like.3. the exponential distribution ). to point out as a maybe much more important fact that the present approach is far more elementary and selfcontained than that using large deviations theory.
Chapter VIII Matrixanalytic methods 1 Definition and basic properties of phasetype distributions Phasetype distributions are the computational vehicle of much of modern applied probability. oo) is said to be of phasetype if B is the distribution of the lifetime of a terminating Markov process {Jt}t>o with finitely many states and time homogeneous transition rates. Typically. We often write p for the number of elements of E. a terminating Markov process {Jt} with state space E and intensity matrix T is defined as the restriction to E of a Markov process {Jt}o<t<. we write Pv for the case where Jo has distribution v so that Pv = KER viPi• 215 . P. and not in other cases. A proper knowledge of phasetype distributions seems therefore a must for anyone working in an applied probability area like risk theory. This implies in particular that the intensity matrix for { it } can be written in blockpartitioned form as T 0 0 . that is.1) is 'Here as usual . refers to the case Jo = i. More precisely. on Eo = E U {A} where A is some extra state which is absorbing. if a problem can be solved explicitly when the relevant distributions are exponentials. if v = (vi)iEE is a probability distribution. A distribution B on (0. then the problem may admit an algorithmic solution involving a reasonable degree of computational effort if one allows for the more general assumption of phasetype structure. F (Jt = A eventually) = 1 for all i E E 1 and where all states i E E are transient. Note that since (1.
j. Then a = a1 = 1. tij > 0 for i 54 j and EjEE tij < 0 .e. the rows sum to one which in matrix notation can be rewritten as t + Te = 0 where e is the column Evector with all components equal to one.3.T)) if B is the Padistribution of the absorption time C = inf{t > 0 : it = A}. E = {i. We now say that B is of phasetype with representation (E. i.e. 0 2this means that tii < 0. k}. and the phasetype distribution is the lifetime of a particle with constant failure rate /3.216 CHAPTER VIII. B(t) = Fa(^ < t ). T is a subintensity matrix2.2) The interpretation of the column vector t is as the exit rate vector. that is. i.1 Suppose that p = 1 and write . T) (or sometimes just (a. the exit rates ti and the transition rates (intensities) tij: tj 3 aj ai i ti tk tjk FkJ ak Figure 1. and we have t = Te.1 The phase diagram of a phasetype distribution with 3 phases. Thus the phasetype distributions with p = 1 is exactly the class of exponential distributions. an exponential distribution with rate parameter . The initial vector a is written as a row vector. t1 = /3. Equivalently. MATRIXANALYTIC METHODS the intensity matrix of a nonterminating Markov process. In particular. A convenient graphical representation is the phase diagram in terms of the entrance probabilities ai. the ith component ti gives the intensity in state i for leaving E and going to the absorbing state A. a.0 = t11. C is the lifetime sup It > 0 : Jt E E} of {Jt}. Here are some important special cases: Example 1 . (1.
. so that the density is P E ai6ie6.. 0 SP 0 and the phase diagram is (p = 2) . a = (1 0 0 . 0 S 6 Example 1..1)!e Since this corresponds to a convolution of p exponential densities with the same rate S. 0 0 0 0 S 6 . . the EP distribution may be represented by the phase diagram (p = 3) Figure 1..2 corresponding to E = {1. . PHASETYPE DISTRIBUTIONS 217 Example 1.. 6.. 00)) S s o . p}.1..2 The Erlang distribution EP with p phases is defined Gamma distribution with integer parameter p and density bp XP1 6x (p... .. 0 •. . . 0 0 0 T= t= 0 ••• S S 0 0 0 0 0 0 ... .x i=1 Thus E _ Si 0 T 0 S2 0 0 . 0 ••• 0 0 Sp1 0 0 t= 0 0 00 •.3 The hyperexponential distribution HP with p parallel channels is defined as a mixture of p exponential distributions with rates 51...
Proof Let P8 = (p ^) be the sstep EA x EA transition matrix for {Jt } and P8 the sstep E x Etransition matrix for {Jt} .1)"n! aT"e. j E E. ds^ = ds' = ttlaj + tikpkj. B[s] = f0°O esxB (dx) is a(sI T)lt (d) the nth moment f0°O xnB(dx) is (. [APQ ] p. see A. 5 Let B be phasetype with representation (E.t2 yt bP.3 . dp. the restriction of P8 to E. 36) yields s d. a. E t ikp kj = kEE kEE 3For a number of additional important properties of matrixexponentials and discussion of computational aspects . p:.d. and is defined as the class of phasetype distributions with a phase diagram of the following form: 1 617 ti t2 2 b2. the backwards equation for {Jt} (e.g.1 tP1 1 Figure 1. The basic analytical properties of phase type distributions are given by the following result . . Then for i . Then: (a) the c. (c) the m.218 CHAPTER VIII. Recall that the matrixexponential eK is defined by the standard series expansion Eo K"/n! 3. the Erlang distribution is a special case of a Coxian distribution.4 For example.f.e.3 0 Example 1 . i. (b) the density is b(x ) = B'(x) = aeTxt.g.f is B (x) = 1 . Theorem 1 .4 (COXIAN DISTRIBUTIONS) This class of distributions is popular in much of the applied literature.aeTxe. MATRIXANALYTIC METHODS Figure 1. T).
5) as hi(tii + s) = ti  t ij hj.1 ) n +l n ! a (s I + T ) . Since 1 . the solution is P8 = eT8. d" dsn a (.f. or w.. i. and since obviously P° = I.tii we go to A. B(n)[0] = _ Alternatively.p. Then h tit ti + ti3 h j .g. tij / .B(x) = 1'a (( > x) = P.s j# tii i (1. for n = 1 we may put ki = Ei( and get as in (1. .5) Indeed .f. (Jx E E) = this proves (a). the rule (A. Alternatively. d8 P8 = TP8.1. and since b[s] = ah.g. we i w..tii is the rate of the exponential holding time of state i and hence (tii)/(tii . hj . and (b) then follows from 1: aipF.s I . in which case the time to absorption is 0 with m . of the initial sojourn in state i.n1t = (1)nn!aTn1Te (1)nn! aTne. PHASETYPE DISTRIBUTIONS 219 That is.g.5) ki = 1 + tii L jj:Ai tii (1. For (c).s) is the m . ti/ .tii tii . i. h = (T + sI)1t.tii and have an additional time to absorption either go to state j which has m . this means in vector notation that (T + sI)h = t. Part (d) follows by differentiating the m. we arrive once more at the stated expression for B[s].6) .jEE B'(x) _ cx Pxe = aeTxTe = aeTxt (since T and eTx commute).T) 't = (.f. After that. j#i jEE tijhj + his = ti. 1. Rewriting ( 1.f.g.n lt .p.e. = aPxe.12) for integrating matrixexponentials yields B[s] = J esxaeTxt dx = a ( f°°e(81+T)dx ) t a(sI . define hi = Eie8S.T) 1t. (1) n+1n!aT .
T= 2 111 so that 2 2 Then (cf.h.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. are idempotent. Consider for example 3 9 a= (2 2).s.220 CHAPTER VIII. we get the density as 9 9 6 (1 1) 10 7 1 0 10 2 aeTyt = e x . One obvious instance is the hyperexponential distribution. Example A3."n! ( ( l 2 2 ) 17 9 0 \ 1 / 10 10 32 n! 35 6" +n!353 Similarly. MATRIXANALYTIC METHODS which is solved as above to get k = aTle. see the Appendix. another the case p = 2 where explicit diagonalization formulas are always available. making the problem trivial. there are some examples where it is appealing to write T on diagonal form. 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. 0 Example 1.6 Though typically the evaluation of matrixexponentials is most conveniently carried out on a computer.
29) and Proposition A4.7) Proof According to (A. This is the traditional choice in the literature.e a mixture of a phasetype distribution with representation (a/llall. B[Q] of B is f3[Q] = J e'1zB(dx) _ (v (9 I)(T ® Q)1(t ® I). i.4.4b for definitions and basic rules): Proposition 1.g. • The phasetype distribution B is zeromodified.e 11BIJ = 1laDD < 1. or one just lets U be undefined on this additional set. i. PHASETYPE DISTRIBUTIONS 1 10 7 10 221 9 6 70 7 9 10 2 +e 6x (1 11 2 2 35ex + 18e6x 35 The following result becomes basic in Sections 4. .hall.f. 5 and serves at this stage to introduce Kronecker notation and calculus (see A. < 1.11aDD. T) is then defined to be oo on a set of probability 1. and in fact one also most often there allows a to have a component ao at A. 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). a random variable U having a defective phasetype distribution with representation (a. There are two ways to interpret this: • The phasetype distribution B is defective. then the matrix m.T) with weight hall and an atom at zero with weight 1 .1.T). 0 Sometimes it is relevant also to consider phasetype distributions. (1. hail = E=EE a.7 If B is phasetype with representation (v. where the initial vector a is substochastic.
one has k = 0. Rolski. (1. Example A5. In Proposition A5.1 of the Appendix. 1. v. All material of the present section is standard. Using B(x) = aeTxe .hve7x. 0 Of course. Neuts. . No satisfying . it is easily seen that the asymptotic form of the tail of a general phasetype distribution has the form B(x) _ Cxkenx. O'Cinneide [276] gave a necessary and sufficient for a distribution B with a rational m. B[s] = p(s)/q(s) to be phasetype: the density b(x) should be strictly positive for x > 0 and the root of q(s) with the smallest real part should be unique (not necessarily simple.F.222 CHAPTER VIII.. In older literature. MATRIXANALYTIC METHODS la Asymptotic exponentiality Writing T on the Jordan canonical form. 77 > 0 and k = 0.g. the conditions of Proposition 1. see his book [269] (a historical important intermediate step is Jensen [214]).8 Let B be phasetype with representation (a.q be the eigenvalue of largest real part of T. Other expositions of the basic theory of phasetype distributions can be found in [APQ]. Schmidli. not only in the tail but in the whole distribution. The Erlang distribution gives an example where k > 0 (in fact. Notes and references The idea behind using phasetype distributions goes back to Erlang.f.4c). cf. but in many practical cases.g. where C. B(x) . we give a criterion for asymptotical exponentiality of a phasetype distribution B. the result follows (with C = (ah)(ve)). let . the Erlang case). but todays interest in the topic was largely initiated by M. cf. Here is a sufficient condition: Proposition 1.8) Proof By PerronFrobenius theory (A. here k = p1). distributions with a rational m. and we have eTx . Then the tail B(x) is asymptotically exponential. Schmidt & Teugels [307] and Wolff [384]. the text is essentially identical to Section 2 of Asmussen [26]. T). x * oo. (or Laplace transform) are often used where one would now work instead with phasetype distributions. 2. See in particular the notes to Section 6. i is real and positive. let v. but the relevant T is not irreducible.Ce7'. Lipsky [247]..8 are far from necessary ( a mixture of phasetype distributions with the respective T(') irreducible has obviously an asymptotically exponential tail. assume that T is irreducible .8). h can be chosen with strictly positive component.f. h be the corresponding left and right eigenvectors normalized by vh = 1 and define C = ah • ve .
but is in part repeated below. Jt={Jt?ul}. .t. Let U1. if U is absolutely continuous on (0. If B is exponential with rate 0.. .. the renewals form a Poisson process and we have u(x) = 0. .i. The explicit calculation of the renewal density (or the renewal measure) is often thought of as infeasible for other distributions.. and U(A) is then the expected number of replacements (renewals) in A. 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 +. 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. oo) w.1.2. +UnEA} 00 = EEI(U1 +.. we refer to U as the renewal measure.. known.T). A related important unsolved problem deals with minimal representations: given a phasetype distribution . with common distribution B and define4 U(A) = E# {n = 0.1) Proof Let {Jtk)} be the governing phase process for Uk and define {Jt} by piecing the { J(k) } together.. or the density is available ) is.. JtJt1) Then { 0<t<U1 . be i.+UnEA). (2... U1<t < U1+U2. Lebesgue measure... as the lifetimes of items (say electrical bulbs) which are replaced upon failure. however.g..r. U2. + U0 is 0 .: U1 + .1 Consider a renewal process with interarrivals which are phasetype with representation (cr. is Markov and has two types of jumps . we denote the density by u(x) and refer to u as the renewal density. Then the renewal density exists and is given by u(x) = ae(T+ta)xt. n=O We may think of the U..f. RENEWAL THEORY 223 algorithm for finding a phase representation of a distribution B (which is known to be phasetype and for which the m.d. but nevertheless.1 of the Appendix. the problem has an algorithmically tractable solution if B is phasetype: Theorem 2. For this reason.
IIafl < 1. i. Hence the intensity matrix is T + ta. This is defined as U1 + . T).T) where v = aT1 /µB.T).1) remains valid for that case. However. . (b) £(t) has a limiting distribution as t * oo. the lifetime of the renewal process. fi(t) U2 U1 .2 Consider a terminating renewal process with interarrivals which are defective phasetype with representation (a. define the excess life e(t) at time t as the time until the next renewal following t. Then: (a) the excess life t(t) at time t is phasetype with representation ( vt. the phasetype assumptions also yield the distribution of a further quantity of fundamental importance in later parts of this chapter .1.3 Consider a renewal process with interarrivals which are phasetype with representation (a. this is welldefined. 2. since Uk = oo with probability 1 . and the distribution of Jx is ae ( T+t«)x.. i. B is defective .U1 U3 U2 U3 U4 Figure 2. and the jumps of the first type are governed by T.224 CHAPTER VIII.e. The renewal density at x is now just the rate of jumps of the second type. . which is ti in state i. Proof Just note that { it } is a governing phase process for the lifetime. and hence ( 2. that is. + Uit_1 where s. which is phase type with representation (v. and let µB = aTle be the mean of B. u The argument goes through without change if the renewal process is terminating. Then the lifetime is zeromodified phase type with representation (a.IIBII which is > 0 in the defective case. Equivalently.1) follows by the law of total probability.. as the time of the last renewal. the density is veTxt = B(x)/µB.T + ta).e. MATRIXANALYTIC METHODS of the last type from i to j occurs at rate tiaj . u Returning to nonterminating renewal processes . is the first k with Uk = 00. Hence ( 2.1 Corollary 2. see Fig. Corollary 2.T) where vt = ae (T+ta)t .
6. Al. The formulas involve the matrixexponential of the intensity matrix Q = T + to = ( tll + tlal t12 + t2al tlz + tlaz _ q1 ql t22 + t2a2 q2 q2 (say). (ii) First check the asserted identity for the density: since T.e. hence e(t) is phasetype with representation (vt.2): aT1 e = AB = 1 µB µB a + aT'Tea aT1(T + ta) µB PB a + aea a + a µB µB =0.T) where vt is the distribution of it which is obviously given by the expression in (a).2) v(T + ta) = 0.e. The renewal density is then aeQtt = (al a2) ( 7i 7"2.2. According to Example A3.) ( t2 ) . cf. T1 and eTx commute. Hence in (b) it is immediate that v exists and is the stationary limiting distribution of it. we first compute the stationary distribution of Q. i.4 Consider a nonterminating renewal process with two phases. Next appeal to the standard fact from renewal theory that the limiting distribution of e(x) has density B(x)/µB. u Example 2 .1. RENEWAL THEORY 225 Proof Consider again the process { Jt } in the proof of Theorem 2. we get B(x) aeTxe aT1eTxTe µB µB PB = veTxt. The time of the next renewal after t is the time of the next jump of the second type. Here are two different arguments that this yields the asserted expression: (i) Just check that aT1/µB satisfies (2. (2. = qz ql (x1 xz) = ql + qz ql + q ' and the nonzero eigenvalue A = ql . the unique positive solution of ve = 1.q2.
and Example 2.a27r1) (t1 .52) 25152 51x2+5251 51a2+5251 Notes and references Renewal theory for phasetype distributions is treated in Neuts [268] and Kao [221]. t1B 0 Example 2 .t2) 1 + eat (a17r2 .t2) . )t (51 .e2bt) 13 Example 2 . MATRIXANALYTIC METHODS e. Then Q= 0 55 )+(1o)=( j ad ).`t (al a2) + C 11 172 ir12 / \ t 2 ) r1 (7r1 7r2) ( t2 7rltl + J + eAt (al a2) ( 71(t2 .4 yields the renewal density as u(t) = 5152 e. A = 25.(biaz + aza.6 Let B be hyperexponential.tl) 7r2t2 + eat (a17r2 .5 Let B be Erlang(2). .a27rl) (tl . and Example 2. Hence 7r = (1/2 1/2).226 CHAPTER VIII.4 yields the renewal density as u(t) = 2 (1 . The present treatment is somewhat more probabilistic.52a1. Then _ Q Hence 51 0 0 52 + 51 52 _ 5152 51a2 ) (al a2) 52a1 62a1 Slat + 52a1 51a2 51a2+52a1 A = 51a2 .
Considering the first. T(0) < oo) the ladder height distribution and M = supt>o St. Here we have taken the terminating Markov process underlying B with two states. Then: (a) G+ is defective phasetype with representation (a+. (b) V. Thus the total rate is tip + tia+. add a more selfcontained explanation of why of the phasetype structure is preserved. The stars represent the ladder points ST+(k). B the claim size distribution. Then each claim (jump) corresponds to one (finite) sample path of the Markov process. Now just observe that the initial vector of {mx} is a+ and that the lifelength is M.) = F(ST(o) E •. {St} the claim surplus process. Proof The result follows immediately by combining the PollaczeckKhinchine formula by general results on phasetype distributions: for (a).f3aT1. 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.e. itself phasetype with the same phase generator T and the initial vector a+ being the distribution of the upcrossing Markov process at time ST+_. the Markov processes representing ladder steps can be pieced together to one {my}.2. 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. i.p. Corollary 3. and M is zeromodified phasetype with representation (a+. which occurs at rate ti. marked by thin and thick lines on the figure. T). represent the maximum M as the lifetime of a terminating renewal process and use Corollary 2. T + to+). we shall. The essence is contained in Fig. Next. Corollary 2. . and if there is a subsequent ladder step starting in j whic occurs w.1 Assume that the claim size distribution B is phasetype with representation (a. a+j. with 0 denoting the Poisson intensity. use the phasetype representation of Bo. G+(. T). We asssume that B is phasetype with representation (a.(u) = a+e(T+tQ+)u Note in particular that p = IIG+II = a+e. 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. cf. T) where a+ is given by a+ = .3. Since the results is so basic. For (b).i. and rewriting in matrix form yields the phase generator of {my} as T + ta+. 3. Within ladder steps.1 on the next page. however.3. r(u) the time of ruin with initial reserve u.
228 CHAPTER VIII..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 .QaT1. see Corollary 2.. 7e7x 2 2 Thus b is hyperexponential (a mixture of exponential distributions) with a (2 2 ). MATRIXANALYTIC METHODS t .t t d kkt S.M {mx} ST+(2)  S . 0 Example 3.1 . Figure 3.Q = 3 and b(x) = .2 Assume that ..1 This derivation is a complete proof except for the identification of a+ with .3. T = (3 . This is in fact a simple consequence of the form of the excess distribution B0. 3e3x + .
^(u) = a+e( T+ta+)ue = 24eu + 1 e6u 35 35 0 Notes and references Corollary 3. In the next sections. see Section 6.4. cf. 0(8) (u) (recall that z/i(u) refers to the zerodelayed case and iY(8) (u) to the stationary case).1): Proposition 4.and Markovmodulated models.4. 3. the duality result given in Corollary 11. We shall derive phasetype representations of the ruin probabilities V) (u). T). 3.j). 4 The renewal model We consider the renewal model in the notation of Chapter V. the discussion around Fig. this was obtained in Section 3. The result carries over to B being matrixexponential. Fig.6). The parameters of Example 3. For an attempt. T) for some vector a+ = (a+. his derivation of +'(u) is different. For the compound Poisson model. (a) G+ is of phasetype with representation (a+. 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 . see Shin [340]. but there the vector a+ is not explicit but needs to be calculated (typically by an iteration). For further more or less explicit computations of ruin probabilities.6. see Stanford & Stroinski [351] .1 In the zerodelayed case.T). We assume p = PB/µA < 1 and that B is phasetype with representation (a. with A denoting the interarrival distribution and B the service time distribution. we encounter similar expressions for the ruin probabilities in the renewal.1 which does not use that A is exponential) by noting that the distribution G+ of the ascending ladder height ST+ is necessarily (defective) phasetype with representation (a+. THE RENEWAL MODEL This is the same matrix as is Example 1.2 are taken from Gerber [157]. but that such a simple and general solution exists does not appear to have been well known to the risk theoretic community. if we define {mz} just as for the Poisson case (cf.1 can be found in Neuts [269] (in the setting of M/G/1 queues. It is notable that the phasetype assumption does not seem to simplify the computation of finite horizon ruin probabilities substantially. and the argument for the renewal case starts in just the same way (cf. That is. where a+ is the (defective) .
In fact.1) Proof We condition upon T1 = y and define {m. where u w(a +) = aA[T + to+) = a J0 e(T+t+)1A(dy). Hence by Theorem 11. (c) {mx } is a (terminating) Markov process on E. but with initial distribution a rather than a+. obviously mo = m.*'} is Markov with the same transition intensities as {mx}.230 distribution of mo. the form in which we derive a+ for the renewal model is as the unique solution of a fixpoint problem a+ = cp(a+). Then {m.6.T). Proof Obviously. cf.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).Sy} in the same way as {mx} is defined from {St}. with intensity matrix Q given by Q = T + to+.4 Consider the renewal model with interarrival distribution A and the claim size distribution B being of phasetype with representation (a. (4.1). But by Corollary 2.3 a+ satisfies a+ = V(a+). G(') = pBo.*} from {St+y . the Palm distribution of the claim size is just B. where B0 is the stationary excess life distribution corresponding to B.5. Fig. MATRIXANALYTIC METHODS (b) The maximum claim surplus M is the lifetime of {mx}. Nevertheless.T). Also.3. where a(8) = aT1/PA. it follows by integrating y out that the distribution a+ u of mo is given by the final expression in (4. We have now almost collected all pieces of the main result of this section: Theorem 4 . 4. which for numerical purposes can be solved by iteration. B0 is phasetype with representation (aT1/µa.T)• Proposition 4. The key difference from the Poisson case is that it is more difficult to evaluate a+. the calculation of the first ladder height is simple in the stationary case: Proposition 4. Then .1. Since the conditional distribution of my given T1 = y is ae4y. CHAPTER VIII.
e.3) (defined on the domain of subprobability vectors . (4.1).1(b). It remains to prove convergence of the iteration scheme (4.1 . . THE RENEWAL MODEL 231 .•. The second follows in a similar way by noting that only the first ladder step has a different distribution in the stationary case. only with initial distribution a(*) for mo.0.^(u) = a+e ( T+ta+)xe. and that this is given by Proposition 4. I {mx} .. a+2) = ^p (a+l)) .1/pA. (4.^(8)(u) = a ( 8)e(T+ta +) xe. The term tf3 in cp(i3) represents feedback with rate vector t and feedback probability vector (3..0) is an increasing function of /3.2. Furthermore . a+l ) = cp (a+°)) . i. i y ^ T1= y `•r Figure 4. the maximum claim surplus for the stationary case has a similar representation as in Proposition 4.1) and a(8) _ aT.3).3) Proof The first expression in (4. In particular . a+ can be computed by iteration of (4.2 ) follows from Proposition 4. . by a+ = lim a +n) where a+°) .M. Hence ^p(.. a+) > 0 = a+o) implies a+) _ (a+) > W (a+)) = a+) .4.1 by noting that the distribution of mo is a+.. thus .2) where a+ satisfies (4.
Then (4.T1. the normalization is equivalent to F(s) = 1. and hence we may assume that h has been normalized such that ahA[s] = 1. Theorem 4. Obviously.f. MATRIXANALYTIC METHODS and (by induction ) that { a+ n) } is an increasing sequence such that limn. a+ ) exists .. so to complete the proof it suffices to show that &+ < a+) for all n. Then s is an eigenvalue of Q = T + ta+ if and only if 1 =. s ¢ sp(T).4). both quantities are just 0 .1 arrivals (n arrivals are excluded because of the initial arrival at time T1 ).T)1t. Then each subexcursion of {St+Tl .T)It. To prove the converse inequality. Thus by (4. Proof Suppose first Qh = sh.5) yields h = (sI . However. 0 = a+) < a+ yields a+) _ (a+0)) (a+) = a+ (n and by induction that a(n) < a+ for all n . which links together the phasetype setting and the classical complex plane approach to the renewal model (see further the notes). we use an argument similar to the proof of Proposition VI. Then e4'h = e82h and hence sh = Qh = (T + taA[Q])h = Th + A[s]tah. this implies that ahA[s] # 0.) = P(mTl = i.ST. Similarly.} can contain at most n .232 CHAPTER VIII. Let Fn = {T1 + • • • + Tn+1 > r+}be the event that {my} has at most n arrivals in [T1. and let &+".T)'t • A[s] (4. Assume the assertion shown for n . In that case.4.P[s] = A[s]B[s]. 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).g.4) makes sense and provides an analytic continuation of F[•] as long as s ¢ sp(T). let F be the distribution of U1 . (4. . the corresponding right eigenvector may be taken as (sI . n) &+n) T a+.1. F[s] being interpreted in the sense of the analytical continuation of the m. 7+ ].2. Then F[s] = a(sI . with B[s]. Fn ). For n = 0. 0 0 We next give an alternative algorithm. Thus . (4.5 Let s be some complex number with k(s) > 0.5) Since s $ sp(T).4) whenever EeR(S)U < oo. limn4oo a ) < a+. To this end.
.. Pd with corresponding eigenvectors hl.5. Given T has been computed. we get at a(Q ..5(c) means that a+(sI T)1t = 1. . . and the solution is . Hence with h = (sI T).T)lt = sh. yd satisfying R(ryi) > 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 ).9) we have G+[s] = 1 which according to Theorem 1.. 0). hd.6) i=1 i=1 Proof Appealing to Theorem 4. and define hi = (piI . Then G+ is phase.. pdhd. This immediately implies that Q has the form CD1 and the last assertion on the diagonal form ..T)lt + t = s(sI . Corollary 4.. ..p1i .T) = 1 ata+ = a+. 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. 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.. (4.6 Suppose u < 0..T)It.6. Further. . In older literature . THE RENEWAL MODEL 233 Suppose next F(s) = 1. This gives d roots 'y.. 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. we have IG_ [s] I < 1 . the matrix Q in Theorem 2. . and hence by the WienerHopf factorization identity (A. T) with a+ = a(QT)/at. in turn. As in Corollary 4..4. Pd in the domain ER(s) > 0 . letting vi be the left eigenvector of Q corresponding to pi and normalised by vihi = 1 . hd. t(ry) > 0. . Since R(s) > 0 and G _ is concentrated on (oo.lt we get Qh = (T + to+)h = T(sI . W v M(d) in the notation of Chapter V). Notes and references Results like those of the present section have a long history.' that the equation F(s) = 1 has d distinct roots p1. D that with columns p1 hl. Q = CD1 where C is the matrix with columns hl... Let d denote the number of phases.type with representation (a+.1 has the d distinct eigenvalues .. ...
5 Markovmodulated input We consider a risk process {St } in a Markovian environment in the notation of Chapter VI. MATRIXANALYTIC METHODS d F 1 + a J e°" ip(u) du = Ee°w = 11(t. That is . In risk theory. which contains somewhat stronger results concerning the fixpoint problem and the iteration scheme. For surveys . The number of elements of El=> is denoted by q. T(').F. a pioneering paper in this direction is Tacklind [373].exponential form of the distribution was found by Sengupta [335] and the phasetype form by the author [18]. The distribution of W comes out from the approach but in a rather complicated form . Here phase.type assumption . The matrix.g. [119]. but the models solved are basically Markov chains and processes with countably many states ( for example queue length processes ). e.... whereas the approach was introduced in queueing theory by Smith [350]. an alternative approach (the matrixgeometric method ) has been developed largely by M. We assume that each B. the ruin probability can be found in matrixexponential form just as for the renewal model. and the distribution of an arrival claim is B. starting around in 1975. For further explicit computations of ruin probabilities in the phasetype renewal case . with representation say (a(' ). the fixpoint problems look like R=Ao+RAI+R2A2+ . similar discussion appears in Kemperman [227] and much of the queueing literature like Cohen [88].234 then in transform terms CHAPTER VIII. and appears already in some early work by Wallace [377]. Numerical examples appear in Asmussen & Rolski [43].4. It turns out that subject to the phase. The solutions are based upon iterations schemes like in Theorem 4. E(t)). the intensity matrix is A and the stationary row vector is ir .) d (see. The exposition here is based upon [18]. The arrival rate in background state i is a. is phasetype. the background Markov process with p states is {Jt}.type assumptions are basic. In queueing theory. involving . 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.contained derivation). where R is an unknown matrix. [270] and Latouche & Ramaswami [241]. see Neuts [269]. Asmussen & O'Cinneide [ 41] for a short self. see Dickson & Hipp [118]. Neuts and his students.
We start in Section 5a with an algorithm involving roots in a similar manner as Corollary 4.1 In Fig. Section 5b then gives a representation along the lines of Theorem 4.Vt)} obtained by time reversing the I component. Diagonalization Consider a process {(It. for which the relevant fixpoint problem and iteration scheme has already been studied in VI. 5. 5. The stationary distribution is obtained by finding the maximum of the Vcomponent of the version of {(It. 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). O. The version of the process obtained by imposing reflection on the V component is denoted a Markovian fluid and is of considerable interest in telecommunications engineering as model for an ATM (Asynchronuous Transfer Mode) switch. states . The connection between the two models is a fluid representation of the Markovmodulated risk process given in Fig.1. has states o.5. The two environmental states are denoted o.1. The key unknown is the matrix K. the phase space E(°) for B. 5a Calculations via fluid models. •.4. MARKOVMODULATED INPUT 235 some parameters like the ones T or a+ for the renewal model which need to be determined by similar algorithms. p = ql = Q2 = 2. (a) 0 0 ♦ o ° tl ♦ • 0 0 o } o o (b) 0 } ♦ • 0 o f o Figure 5.6.2. Vt)}t>o such that {It} is a Markov process with a finite state space F and {Vt} has piecewiese linear paths. This calculation in a special case gives also the ruin probabilities for the Markovmodulated risk process with phasetype claims. However. say with slope r(i) on intervals where It = i. and the one E(•) for B.
4}.1))diag + sII = 0 (5.1 A complex number s satisfies 'A+ (f3i(Bi[s] .1))diag ) a = sa and the eigenvector b = . j = 1. In the general formulation .236 CHAPTER VIII.31a(l) (/3i)diag . a E E(i) } .Vt)} is then obtained by changing the vertical jumps to segments with slope 1. a) of {It}.1) if and only if s is an eigenvalue of E. This implies that in the fluid context. '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. Recall that in the phasetype case. First. 4.1(b) {(It . Second. o. F is the disjoint union of E and the Eli). V. resp. of E into components indexed by E. 5. F = E U { (i. A claim in state i can then be represented by an E()valued Markov process as on Fig. MATRIXANALYTIC METHODS 4.A 0 Or 1A/ _ t(i) 0 t(2) 0 0 0 0 0 t(3) 0 T1 0 0 0 . The fluid model on Fig . If s is such a number.1(a). < oo for all s. a) = 1. Thus F = {o. •.(Ni)diag r(i. whereas Ee8s' = oo for all t and all s > so where so < oo.92a(2) 0 0 T(2) 0 0 0 f33a(3) 0 0 T(3) with the four blocks denoted by Ei„ i. in the fluid model Eel'. Let E denote the matrix . 4. Bi[s] = a(i)(T(i) + sI)it('). 5. i E E. t. The intensity matrix for { It} is (taking p = 3 for simplicity) I A . r(i) _ 1. 2. corresponding to the partitioning + Epp). consider the vector a satisfying (A + (13i(Bi[ s] . 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. we have more martingales at our disposal. a) : i E E. Eli) + Proposition 5.
E22)1 E21a E21a .(sI .sI 0 0 0 T(2) . E(1) + + E(P).sI+ ((3ia(i)(T(i) . c = a.sI + E12 (sI .sI)1t)) iag I = 0 which is the same as (5.Nla(1) 0 0 T 1.5.sI) (sI .sI ()3i)diag .T('))1t(i) .1).E21a + sd = sd. MARKOVMODULATED INPUT 237 indexed by E. Then E21c+E22d = E21a . 0 . d = (sI E22)1E21a = E ai(sI . For the assertions on the eigenvectors. it follows that Ell E12 ( E 21 E22) (d) = s 1 d I .A . Then (up to a constant) c = a.sI 0 0 0 T(3) .32a(2) (/3i)diag . with Eii replaced by Eii .E22 . and let d = (sI .sI.E12E22 E21 I . d correspond to the partitioning of b into components Proof Using the wellknown determinant identity Ell E12 E21 E22 E22 I ' I Ell . it follows that if Qla(1) 0 0 . where c. assume that a is chosen as asserted which means (Ell . resp .sI 0 0 t(3) 0 0 = 0. Noting that E11c + E12d = se by definition.A . t(1) 0 0 then also 0 t(2) 0 .E22)1 E21) a = 0.E22)1 E21a. iEE (a> of 0* 1 AI.
. j. a )d(a + e8 °vpi (u . v) yields C{V) = e8 . v.2 Assume that E = Or 'Al has q = ql + + qp distinct eigenvalues si. I' i( V P2 (w (u) < oo. c j.4 that {e"1b(v) is a martingale . it follows by Proposition II. d("))1 e. . u) Iw(u.. Example 5 . v > 0. Then . j) pi( u . pi(u.sv)b(v) = 0.. a) = Pi (Vw(u. MATRIXANALYTIC METHODS Theorem 5. For u. j.a)d^ ).v) = = p i( u . a) = (j..v) = Optional stopping at time w (u. the result u follows.v) = j).upi(u.. Iw(u. e89uc(e)) (d(1) ....v) = v) I. a)).v}.. j. Thus 0(u) = esu/d = pe7 ° as u should be. a)).4 Assume that E has two states and that B1. a) and noting that i1 (u) = >I j.3 Consider the Poisson model with exponential claims with rate 5. . . . q. sq with $2s.238 CHAPTER VIII.( u.v. Letting v ^ oo and using Rsv < 0 yields e8'u = Epi(u.. we first look for the negative eigenvalue s of E = I 0 I which is s = ry with yy = b . We can take a = c = 1 and get d = (s + b)16 = 5/(3 = 1/p.j. j.v) = (j.pi(u. To determine 0 (u).v)=inf{t >0:Vtu orVt=. B2 are both exponential with rates 51 i b2. Then we get V)i (u) as sum of two exponential terms where the rates s1.. v. w(u)=inf{t >O:Vtu}. j. < 0 and let b(v) = I d(„)) be the right eigenvector corresponding to s.O.. Example 5 . w(u. s2 are the negative eigenvalues of Al +01 A1 E _ A 2 b1 0 52 A2 +32 0 ..j)c v ./' u = e' (esiuc ( 1) .Q. a). .a Solving for the pi(u. v. Proof Writing Or'Alb( v) = svb( v) as (AI .. Here E has one state only.5. . define w(u. v = 1.
(y. j. j.33(e = 0 a(j))(K ®T ( j))(ej (9 I). In terms of K. i. the Pidistribution of M is phasetype with representation (E(1) + + E(P). U) where t(j) + t(j)O(j j = k uja. oo)) j)ye. 9('). MARKOVMODULATED INPUT 239 5b Computations via K Recall the definition of the matrix K from VI. (5.k. (') a T( However .h. •) is phasetype with representation (E(i).2) the l. dx)Bj(y .xxej • a 00 oo el . 0 Theorem 5 .b (u) = Pi(M > u) = 9(i)euue.3) .6 For i E E. 8^')IT(j)) where e 3^') =.x) 00 f ° (') (j) eT (yy)edx . Proof We must show that G+ (i. we get the following phasetype representation for the ladder heights (see the Appendix for the definition of the Kronecker product 0 and the Kronecker sum ®): Proposition 5. according to VI.5 G+(i. is 0 /3 f R(i . j.Qj eie 0 f e (j) T(') x T(j)y ej a e dx e e 00 00 eKx ® e T(')' dx (ej (& I)e T(')ye eKa®T(')x dx (ej (9 I)eT(') Ye e(i)eT(')ye.5.3j eye.y = to B k7 j # k In particular.2.s.( 2.
Numerical illustrations are given in Asmussen & Rolski [43].2) . T. t = (0 0 . Piecing together these phase processes yields a terminating Markov process with state space EiEE E('). equivalently. we have sofar concentrated on a claim size distribution B of phasetype.5). An alternative characterization is that such a distribution is matrixexponential..g. Furthermore... bn1 bn). with phase space EU> whenever the corresponding arrival occurs in environmental state j (the ladder step is of type j). Associated with each ladder step is a phase process. This yields the asserted form of uja. a m. and it just remains to check that U has the asserted form. i.1 Let b(x) be an integrable function on [0. we have the additional possibility of a phase change from a to ry within the ladder step. 0 1)'. oo) and b* [0] = f °O eBxb(x) dx the Laplace transform. 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.) which is rational.e. say. intensity matrix U. which occurs at rate t(i). and a new ladder step of type k must start in phase y. 6 Matrixexponential distributions When deriving explicit or algorithmically tractable expressions for the ruin probability. a) to (k. that the density b(x) can be written as aeTxt for some row vector a. For j = k.. +aii10+anI then a matrixexponential representation is given by b(x) = aeTxt where a = (b1 b2 . . i. the current ladder step of type j must terminate. u Notes and references Section 5a is based upon Asmussen [21] and Section 5b upon Asmussen [17]. the ratio between two polynomials (for the form of the density. a) is obviously chosen according to e(`). (6.240 CHAPTER VIII. some square matrix T and some column vector t (the triple (a. Bk7 .. if b* [0] = b1 +b20+b302 +...e.2.p. Then b*[0] is rational if and only b(x) is matrixexponential. which occurs w.k y. which occurs at rate t^^7. For a transition from (j. and lifelength M.. MATRIXANALYTIC METHODS Proof We decompose M in the familiar way as sum of ladder steps . see Example 1.y) to occur when j # k. +bn0i1 0n +a10n1 +.f. Starting from Jo = i. t) is the representation of the matrixexponential distribution/density): Proposition 6. the initial value of (i. However.
(6.47r2 3 .. 0 1 an an1 an _2 . namely to asssume the roots 6l. t= 0 . S. For a proof. One of his elementary criteria. s) is given by 27r i ..1 0 .. (6. Example 6 .. t). 0 0 0 0 1 0 0 . S = f c/2 0 21ri ... of (6. but as follows from Proposition 6. personal communication). .4) 0 0 1 c This representation is complex. . b(x) > 0 for x > 0. then b*[0] = a(0I T)1t which is rational since each element of (01 . since 1 + 4ir2 03 + 302 + (3 + 47x2)0 + 1 + 47r2 it follows by (6.2 A remarkable feature of Proposition 6..T)1 is so. Thus. T. see Asmussen & Bladt [29] (the representation (6.3) that we can take 0 1 0 0 a= (1 + 47r2 0 0)..e(tai1)x/2 + e'T) it follows that a matrixexponential representation ()3.2). 1 .3 A set of necessary and sufficient conditions for a distribution to be phasetype are given in O'Cinneide [276].3) 0 0 0 0 0 . Namely. . cannot be phasetype.1 0 0 )3 = (111)./(0 + bi).6. .(6.s.1. shows that the distribution B with density b(x) = c(1 cos(21r x))ex. The converse follows from the last statement of the theorem. Writing b(x) = c(e( 2ni1 ) y/2 .3) was suggested by Colm O'Cinneide. where c = 1 + 1/47r 2.47x2 3 1 0 . . u Remark 6. s = c/ 2 . (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. u giving b(x) = E 1 ciebiz/bY. a2 a1 Proof If b(x) = aeTxt.h. matrixexponentiality implies a rational transform.. we can always obtain a real one (a.an_3 an _ 4 . MATRIXEXPONENTIAL DISTRIBUTIONS 241 T = 0 1 0 0 0 .1) as E 1 c.2). 0 0 . bn of the denominator to be distinct and expand the r. T= 0 0 1 .
and present two algorithms for calculating '(u) in that setting.242 CHAPTER VIII. We recall (see Section 3. that despite that the proof of (6. recall that t = Te) that if B is phasetype and (a.6) The remarkable fact is. leading to matrix calculus in high dimensions when b is small.6) in Section 3 seems to use the probabilistic interpretation of phasetype distribution in an essential way.5) Thus.1 to get i (u) = f3esus. q are polynomials without common roots. For the second algorithm. MATRIXANALYTIC METHODS Example 6 . Corollary 111. 0. As for the role of matrixexponential distributions in ruin probability calculations. T. 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. then 5(u) = a+e(T+t+)uTle where a+ = /3aT1.5 (6. we have represented ti* [0] as ratio between polynomials (note that 0 must necessarily be a root of the numerator and cancels). we take as starting point a representation of b* [0] as p( O)/q(9) where p.4) the Laplace transform of the ruin probability is /g(e)PO 0*[e] _ /' eeu^G(u)dU = 0 9(/3a0p(9)ap (9)/q(9)) . (6.1)2 + 6).6) holds true also in the matrixexponential case. and can use this to invert by the method of Proposition 6. T.4 This example shows why it is sometimes useful to work with matrixexponential distributions instead of phasetype distributions: for dimension reasons . . we use a representation (a. Then (cf.1 shows that a matrixexponential representation can always be u obtained in dimension only 3 independently of J. t) a phasetype representation with a the initial vector. T the phase generator and t = Te. (6. we shall only consider the compound Poisson model with arrival rate 0 and a matrixexponential claim size distribution B. Consider the distribution with density = 15 ((2e2x . 7 + 155ex b(x) Then it is known from O'Cinneide [276] that b is phasetype when 6 > 0. then: Proposition 6. For the first. and that the minimal number of phases in a phasetype representation increases to 0o as 5 .3. t) of b(x).
T)1T 2 = and 1 = AB IT2 + 82T .1BVA1.T .T)1T1t.T)1t ( l .T). b+ = a +(BI .5 ).1t = b* .T)1 J0 00 b(x) dx = f aT1t.T)1t)1a +(9I .T)1 + 1 ib* (91. since (91T)1T .1UB(B + BVA1UB). with A = 91T. we get (91.T)1 + (6I . (6. xb(x) dx = aT2t.b* (6.to+)1T . the assertion is equivalent to a+(BI . (91.T)1ta+(OI . MATRIXEXPONENTIAL DISTRIBUTIONS 243 Proof Write b* = a(9I .T)1 (91.T)1t.1t du = .T)1 so that b* b** b** a+(9I .t. this can be verified by analytic continuation from the phasetype domain to the matrixexponential domain .T .a+(9I . we get b+ = 0aT1(9I T). but we shall give an algebraic proof.1t = f3a (0I T)1T1t . .B=land V=a+.to+)1 = (BI .6b* .1 + b+ = b++ 1 . Presumably.6. b+ = a+(9I . From the general matrix identity ([331] p.1 = ^(T1 + ( 91T)1).to+)1T .7) 9( cf.T .1 = A1 .6).A .'t. (6. U =. Now. Then in Laplace transform formulation . 519) (A + UBV ).1 + 82 (9I .
3 is taken). MATRIXANALYTIC METHODS .1. 7a Computing O(u) via differential equations The representation we use is essentially the same as the ones used in Sections 3 and 4. which is selfexplanatory given Fig. From this it is straightforward to check that b**/(b+ .3a (1 0 T 2 + 1 T 102 (9I + 02 1 T)1) t P + 7. 0 Notes and references As noted in the references to section 4. 7 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate 0.h. a. see Asmussen & Bladt [29].la.1) is the same as the r. T).8 a(T1 + (01. but the argument of [286] does not apply in any reasonable generality).s. A key tool is identifying poles and zeroes of transforms via WienerHopf factorization. In Corollary VII. 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. 7. some key early references using distributions with a rational transform for applied probability calculations are Tacklind [373] (ruin probabilities) and Smith [350] (queueing theory). For expositions on the general theory of matrixexponential distributions. .1. the ruin probability(u) was found in explicit form for the case of B being exponential. (for some remarkable explicit formulas due to Paulsen & Gjessing [286].1t = /3a (9I .1. Much of the flavor of this classical approach and many examples are in Cohen [88].T)1T. cf.T)1)t = 8 (1 . /3aT1(0I . See Fig.82b*. see the Notes to VII.7). The proof of Proposition 6. Lipsky [247] and Asmussen & O'Cinneide [41].b*). a key early paper is Cox [90] (from where the distribution in Example 6. We present here first a computational approach for the general phasetype case (Section 7a) and next (Section 7b) a set of formulas covering the case of a twostep premium rule.8.T)1T2t . 3. to piece together the phases at downcrossing times of {Rt} (upcrossing times of {St}) to a Markov process {mx} with state space E.5 is similar to arguments used in [29] for formulas in renewal theory. VII.1. of (6.244 CHAPTER VIII.
7. is no longer timehomogeneous. t + s) . Figure 7. RESERVEDEPENDENT PREMIUMS 245 Rt l0 u .1z I. Let P(tl.tl < t2 < u. Proof The first statement is clear by definition. in contrast to Section 3.1) where A(t) = v(u)P(0.1 A(0) = v(u) and A'(t) = A(t)(T + tv(u .u)e = A(u)e (7.t)). >iEE Vi (U) is the ruin probability for a risk process with initial reserve 0 and premium function p(u + •).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.1 The difference from the case p(r) = p is that {m2}. i. Note that in general >iEE Vi (U) < 1. P(tl. the A(t) and hence Vi(u) is available by solving differential equations: Proposition 7. Since v(u) = (vi(u))iEE is the (defective) initial probability vector for {m8}. Ai(t) = P(mt = i). 0 < t < u. 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. though still Markov. In fact. we obtain V)(u) = P(m„ E E) = v(u)P(0. By general results on timeinhomogeneous Markov processes.e. the definition of {m8} depends on the initial reserve u = Ro. t) is the vector of state probabilities for mt. Also.I] I 80 { tq f Q(v) dvl t1 1 . O<.
Qdt) vi(u) + vi'(u)p(u)dt + p(u) dt E{tji+tjvi(u)}.t and being followed by a downcrossing. 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.2 For i E E. jEE .Sj i)p(u)dt • tji = Sji + p(u)tji dt. Given A. dt].3dt. A'(t) = A(t)Q(t) = A(t)(T + tv(u . Given this occurs. (7. 0 < t < u.(u) p ( u) = . the probability that level u is downcrossed for the first time in phase i is ai. vi.t)). 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. In the first case. whereas in the second case the probability is p(u)dt • tjvi(u). 0 Thus. from a computational point of view the remaining problem is to evaluate the v(t). MATRIXANALYTIC METHODS However. given A.4) jEE jEE Proof Consider the event A that there are no arrivals in the interval [0.(tai + vi(u) E vj(u)tjp (u)  Q + vj (u)tjip ( u). we get vi(u) = aidt + (1 . Proposition 7. the probability that level u + p(u)dt is downcrossed for the first time in phase j is vj (u + p(u)dt).t). {mx} has jumps of two types.t) for the second. The intensity of a jump from i to j is tij for jumps of the first type and tivj(u . those corresponding to state changes in the underlying phase process and those corresponding to the present jump of {Rt} being terminated at level u . Given A'. the probability of which is 1 . two things can happen: either the current jump continues from u + p(u)dt to u. or it stops between level u + p(u)dt and u.246 CHAPTER VIII. Thus. Hence Q(t) _ T + tv(u . the probability of downcrossing level u in phase i is 8ji(1 + p(u)dt • tii) + (1 .
<oo..)P"(AnB.h. Let p" (t).) + 0 as v + oo. we can first for a given v solve (7. F" etc.3 For any fixed u > 0. Now since both P(A n Bv) 3 0 and P"(A n Bv) . consider a modification of the original process {Rt} by linearizing the process with some rate p. When solving the differential equation in Proposition 7. vi (U) = lim v= (u).0 as v + 00 we have P(A) P"(A) = P(AnBv)+P(AnBv) P"(AnB.5). Since the processes Rt and Rt coincide under level B. RESERVEDEPENDENT PREMIUMS 247 Subtracting v. and similarly P"(Bv) . From Section 3. V .00 Proof Let A be the event that the process downcrosses level u in phase i given that it starts at u and let B" be the event By={o. then P(A n Bv) _ P"(A n BV').7. (u) on both side and dividing by dt yields the asserted differential u equation. after a certain level v.. supRt>v l t<7 I where o. Rt .. say.2.s. .. starting from v"(v) = .4) backwards for {va (t)}v>t>o.^ 0. of (7.) P"(AnBv) = P(AnB. P u which implies that v. Then pv(r) p(r) r < v p r>v ' and (no matter how p is chosen) we have: Lemma 7.) is the tail of a (defective) random variable so that P(Bv) + 0 as v 4 oo. we have p(r) = p = vi (u) 0aTe. Then P(B. This yields v. (v) is given by the r. refer to the modified process. we face the difficulty that no boundary conditions is immediately available. Thus.i7rT1/p. To deal with this. denotes the time of downcrossing level u . (u) for any values of u and v such that u < v. say.
Then v(u) is the initial distribution of the undershoot when downcrossing level v given that the process starts at u.1.zp1(u)/(1 . r<v r > v. where v = inf It > 0 : Rt < v}.. 0 < u < v. i.. p(r) P. which is available since the z/i'(. Recall that q(w) is the probability of upcrossing level v before ruin given the process starts at w < v. Let ii'( u) = a+'ie(T+ta +^)"e denote the ruin probability for R't where a+ = a+i) = laT1/pi.248 CHAPTER VIII. The f iin in (7.7) equals 01 (v . < 0}). 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. numerically implemented in Schock Petersen [288]) and the present one based upon differential equations require both discretization along a discrete grid 0. cf. assuming u > v for the moment. The precision depends on the particular quadrature rule being employed. T). the probability of ruin between a and the next upcrossing of v. v = u..V" M 0 . We recall from Propositon VII. the evaluation of Vi(u) requires q(u) = 1 . say. as well p1(u). 3u etc.e. 1/n. (u)}. However..1. where. such that Rt coincide with RI under level v and with Rt above level v.7) f o (the integral is the contribution from {R. 7b Twostep premium rules We now assume the premium function to be constant in two levels as in VII.RQ (defined for or < oo only) is defective phasetype with representation (v(u). let v(u) = a+2ieiT +ta+>)(uv).1. p2.) are so.1a.z51(v)). for u > v the distribution of v . Therefore u pl(vvueTa t 1.10 that in addition to the O'(•). The trapezoidal rule used in [288] gives a precision of 0(n 3).q(v dx +( ) ) = ( ) ( q( )) vueTva (7. (7. MATRIXANALYTIC METHODS Next consider a sequence of solutions obtained from a sequence of initial values {v. while the fourthorder RungeKutta method implemented in [30] gives 0(n5). 2/n.. Corollary 3. > 0} and the last term the contribu tion from {R. 2u. The algorithm based upon numerical solution of a Volterra integral equation (Remark VII.v v(u)eTat 1 1 . To evaluate p1(u). typically the complexity in n is 0(n2) for integral equations but 0(n) for integral equations.6) We may think of process Rt as pieced together of two standard risk processes RI and Rte with constant premiums p1.9.x) dx f v(u)eT xt dx .
RESERVEDEPENDENT PREMIUMS 1 .2.u .e6u 35 .e.and A2 = 3 .e6v Let Al = 3 + 2V'2. p2 < 3.24ev .v) 1eai(u v) + 7 7 1 e\2(u v) 1 3 ^') eA2 (u.4 Let {Rt } be as in Example 3.(7 The arrival rate is (i = 3.21 = ? yields 0(u) = 1. Example 7.01 (v . so we consider the nontrivial case example p2 = 4 and p1 = 1. Since µB = 5/21. From Example 3. 01(u) _ 24 u + 35 e6u 1 35 e 4(u) _ 35 .2V"2. B is hyperexponential corresponding to 3 0 3 a(2 2)' T= ( 0 7 t. Then one gets X20 20 21 f 1ea1(u v) + 1 3 3 ^ A 2(u e .24e.^1(v) 1 ..7.1 from which we see that pl (u) = 1 + 1 249  1 .v(u)eTVe .v(u ) eTV e J v(u)eTxtz/)l (v .4) can be written as (Y(u) ®a+)e(T+t°+>)°1 (T ® (T . I.be the eigenvalues of T + to( 2 ).v(u)eTve).v) + (2^ + 3v2 ea'(u " .x) dx} V 1 1(v) f V v(u) eTxt.x) dx 1 ^(v) ( 1 .8) The integral in (7.2. (7.8) equals v v(u)eTxta+2) e(T+ta +))( vx)edx which using Kronecker calculus (see A. all quantities involved in the computation of b(u) have been found in matrix form.jl (t ®e) Thus.to+))11 {e{T®(Ttoy+ ))}„ .
24e5v . and one gets 12e5" . ) e sv + ( 2v/2.1 Thus.1 V2 = 4e5"+6 35e6v ./2 ea1(u") .24es" .b(v) = 192esv +8 35e6v + 168esv + 7* Thus all terms involved in the formulae for the ruin probability have been exu plicitly derived.2 35e6v .21(35e6v . 21 3 In particular. pi (u) = p12(u)/p1 l(u) where p1i(u) p12(u) 35e6v . Notes and references [30].+ it (3 4'I 1 ea2(uv e1\2(u") 7 + ( 32 +4../2) ea 1(u .1)' ?.250 CHAPTER VIII.v)esv + 7 4_ 2.7) we see that we can write pi (u) = v(u)V2 where V2 depends only on v. 192esv + 8 P1 .24es" .24e5v . MATRIXANALYTIC METHODS From (7.1. The analysis and the example are from Asmussen & Bladt .
For further examples.4. For the definition .B(x). (b) the lognormal distribution (the distribution of eu where U .Chapter IX Ruin probabilities in the presence of heavy tails 1 Subexponential distributions We are concerned with distributions B with a heavy right tail B(x) = 1. For example.46 and at numerous later occasions require a light tail.N(µ.2b. The definition b[s] = oo for all s > 0 of heavy tails is too general to allow for a general nontrivial results on ruin probabilities. x 2iror2 (c) the Weibull distribution with decreasing failure rate . III. a2)) with density 1 e(logyFh) 2/2az . the exponential change of measure techniques discussed in II.g. oo ) and say then that B is subexponential (B E S) if 251 . 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. for all t > 0. Some main cases where this lighttail criterion are violated are (a) distributions with a regularly varying tail. B(x) = L(x)/x" where a > 0 and L(x) is slowly varying. x 4 oo. we require that B is concentrated on (0. and instead we shall work within the class S of subexponential distributions . see I. A rough distinction between light and heavy tails is that the m. B(x) = ex0 with 0<0<1. L(tx)/L(x) 4 1.f.
X2) > x}. Then X1 +X2 has an Erlang(2) distribution with density yeY so that B*2(x) xex. X2) > x) is P(X1 > x) + P(X2 > x) . Thus the liminf in Proposition 1. we have {max(Xi. (1.p.'s X1. X1 is w. As contrast to Proposition 1. proving (a).B(x)2 . X2) > x) ^' 2B(x).p. one can check that x x where U is uniform on (0.'s. To capture the intuition behind this definition. the distribution of independent r. then (with high probau bility) so are both of X1.2B(x). B(x) Here B*2 is the convolution square. X2) > x} C {X1 + X2 > x}. Proof By the inclusionexclusion formula. oo). (b) liminf BB(() ) > 2. 1/2 it has the distribution of X1I X1 > x. We later show: Proposition 1. x 3 00.1 Let B be any distribution on (0. Since B is concentrated on (0. note first the following fact: Proposition 1. Then: (a) P(max(Xi.2 If B E S. .v. In contrast. given X1 + X2 > x. oo). X2) > u x)/B(x) = 2. That is. HEAVY TAILS B*2\ 2. 1). The proof shows that the condition for B E S is that the probability of the set {X1 + X2 > x} is asymptotically the same as the probability of its subset {max(Xi. that is.1(b) is oo. the r.252 CHAPTER IX. X2 with distribution B. the behaviour in the lighttailed case is illustrated in the following example: Example 1. P(max(Xi. X2 but none of them exceeds x. P(Xi <yI Xi+X2>x) 1B(y). and thus the lim inf in (b) is at least lim inf P(max(Xi. X2 > x) = 2B(x) . if X1 + X2 is large . Thus .2. 1/2 'typical' (with distribution B) and w. then P(X1>xI X1+X2>x)* 2.v.F(X1 > x.1) then means P(X1 +X2 > x) 2P(Xi > x).3 Consider the standard exponential distribution. in the subexponential case the only way X1 + X2 can get large is by one of the Xi becoming large. In terms of r. That is. B(x) ax.v.
Let 0 < 5 < 1/2.B*n(x .4 Any B with a regularly varying tail is subexponential. Hence lim sup a+oo B*2(x) 2B((1 . Finally lim inf B(x .1. If X1 + X2 > x.B(y)) . B( 0 . 253 Proof Assume B(x) = L(x)/xa with L slowly varying and a > 0. or they both exceed Sx.1(b) we get B*2(x)/B(x) * 2.y)/B(x) > 1 since y > 0. We now turn to the mathematical theory of subexponential distributions.5)x)' + 0 _ 2 L(x)l xa (16) Letting S 10.xIX > x converges in distribution tooo. SUBEXPONENTIAL DISTRIBUTIONS Here is the simplest example of subexponentiality: Proposition 1. Using the identity B*(n+1)(x) = 1+ + 1)(x) 1+ 2 1 .B E S.6)x)/((1 . Proposition 1. 1 < B(x ) B( x) Y) < B( 0). This follows since the probability of the overshoot to exceed y is B (x + y)/B(x ) which has limit 1. yo] as X + 00.y)/B(x) > 1. we therefore get lim sup B*2(x)/B(x) > 1+B(y)+ 1 .S)x.B(y) = 2. and combining with Proposition u 1. then B(B(x)y) * 1 uniformly in y E [0.] Proof Consider first a fixed y. we get limsupB*2(x)/B(x) < 2.S)x + B(Sx)2 < lim sup B(x) xaoo B(x) lim sup 2L((1 x^oo . a contradiction.v.5 If B E S.B*(n ) B(dz) (1. x]. The uniformity now follows from what has been shown for y = yo and the obvious inequality y E [0.'s: if X . [In terms of r.2) B(x) B(x ) B(x) Jo with n = 1 and splitting the integral into two corresponding to the intervals [0. then either one of the Xi exceeds (1 . we get BZ(x)) > 1 + B(y) + B(B()y) (B(x) .yo].z B(x) . y] and (y. If lim sup B(x . then the overshoot X .
5 and the induction hypothesis.B(x .7 If B E S. Proof We use induction. The case n = 2 is just the definition. 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 .254 CHAPTER IX. and this immediately yields the desired conclusions.z) B(dz) _y B(x) 111 Lx B .B*2 (x) B(x) (x . O The following result is extremely important and is often taken as definition of the class S. x oo. HEAVY TAILS Corollary 1. Given e > 0. so assume the proposition has been shown for n.nI < e for x > y.1) for all large n so that B(n) > cle6n for all n.. B*(n+1) (x I xy + Jxx y) W.z) B(dz) (n + O(e)) ^x JO B(x) (n + 0(0) I B (x) .y) sup v>o B(v) B(x) which converges to 0 by Proposition 1. Proposition 1.6 If B E 8. 0 Proof of Proposition 1. choose y such that IB*n(x)/B(x) .z) B(x) Here the second integral can be bounded by B*n(y) B(x) . Then by (1.z) B(dz). then for any n B*n(x)/B(x) * n. The first integral is y B(x . b[c] = oo for all e > 0. This implies B(x) > c2e5x for all x. we have by Proposition 1. then e"R(x) * oo.z) B(dz) 2B(x) o rv 2 0 2 using Proposition 1.5 and dominated convergence.2).5 that B(n) > e6B(n . its intuitive content is the same as discussed in the case n = 2 above.2. Proof For 0 < 5 < e. B(x) \Jo _ B(x .z ) (x ) = 1 + (^ B(x .
X1 > xv. For any fixed v.1.4) .B(x .y)B(dy) = B(x)ov (1)• v (1. it follows that it is necessary and sufficient for the assertion to be true that JX_VA (x .v. oo) such that Ai (x) _ aiB(x) for some B E S and some constants al. Xi <= v Ai (x .ala2B(x)2 which can be neglected.8 If B E S.(al + a2)B(x).ajB(x)Ai(v) = ajB( x)(1+o„(1)) (j = 3 .2). choose T such that (B(x)B*2(x))/B(x) < 1 + b for x > T and let A = 1/B(T). Then Al * A2 (x) . a2 with a1 + a2 > 0. Then Al * A2(x) = P(X1 + X2 > x). SUBEXPONENTIAL DISTRIBUTIONS 255 Here the first term in {•} converges to 1 (by the definition of B E S) and the second to 0 since it is bounded by (B(x) .z) B(x . X2 be independent r. Proposition 1. Proof Define 5 > 0 by (1+5)2 = 1+e.z) B(dz) < 1 + A + an(1 + d) .y))/B(x).'s such that Xi has distribution Ai.y)Ai(dy) v) f o . an = supx>o B*n(x)/B(x). A2 be distributions on (0.5 easily yields P(X1 + X2 > x. Then by (1.i). 0 Lemma 1.9 Let A1. Combining these estimates and letting a 4. Since P(X1+X2 > x.X2 > xv) < A1(xv)A2(x v) .z) B(x) < 1 + A + an sup f x B(x .0 completes the proof.z) B(dz ) + sup < 1 + sup f x<T B ( x) x>T 0 B(x .3) Using the necessity part in the case Al = A2 = B yields f xv B(x .z) B(dz) x .y)Ai(dy) = (x)o(1) (1. Proof Let X1. e > 0. 0 Proposition 1. then there exists a constant K = KE such that B*n(x) < K(1 + e)nB(x) for all n and x. an+1 fX B*n( *n(x . x>T o B(x) The truth of this for all n together with al = 1 implies an < K(1 + 5)2n where K = (1 + A)/e.
A(x) = o(B(x)).y)B(dy). the l. L2 are slowly varying.Bl (x) + B2 (x) follows precisely as in the proof of Proposition 1. However. B1 * B2 (x) . HEAVY TAILS 'VV B(x . In the regularly varying case.2. B2 E S. Hence Corollary 1.3) follows if CHAPTER LX.y)Ai(dy) = B(x)o„(1).2aB(x) .(x) is decreasing for x > x0 with limit 0 at oo. a2 = 1. V (1. B1 * B2 E S does not hold in full generality (but once B1 * B2 E S has been shown. a1 = a2 = a yields A*2(x) . then so is L = L1 + L2.4).v)Ai(v) . Proof Taking Al = A2 = A. A2 = B so that a1 = 0.12 Assume that Bi(x) = Li(x)lxa.5) Here approximately the last term is B(x)o„(1) by ( 1. with a > 0 and L1. Then L = L1 + L2 is slowly varying and B1 * B2(x) sim L(x)/x«. if q(x) aB(x) for some B E S and some constant a > 0. f " By a change of variables. of (1.v)B(v) + _'U Aq(x . u Corollary 1.10 The class S is closed under tailequivalence. We next give a classical sufficient (and close to necessary) condition for subexponentiality due to Pitman [290]. Then A * B E S and A * B(x) . That is. Recall that the failure rate A(x) of a distribution B with density b is A(x) = b(x)/B(x) Proposition 1.s. i = 1.h. it is easy to see that if L1. That is.11 Let B E S and let A be any distribution with a ligther tail. .9).5) becomes x B(x . u Corollary 1.Bl (x) + B2 (x) when B1.B(x) Proof Take Al = A.2A(x).aiB(v)) = B(x)o„(1).13 Let B have density b and failure rate A(x) such that . u It is tempting to conjecture that S is closed under convolution. it should hold that B1 * B2 E S and B1 * B2 (x) . Then B E S provided fo "O exA(x) b(x) dx < oo. whereas the two first yield B(x)(Ai(v) . L2 slowly varying.256 Now (1.Ai(x . then A E S.
A(xy)A ( y). f ' L(y) dy .e009xv)2/2a2/(x 2irv2) logx ( ) 't ((logx . Then b(x) = Ox0lexp.16 For L(x) slowly varying and a > 1.y ) b(y)dy = B (x) o ox _ J = ox/2 eA( x)A(xy ). 0 A(x) . SUBEXPONENTIAL DISTRIBUTIONS 257 Proof We may assume that A(x) is everywhere decreasing (otherwise.y) * 0. The middle bound shows that it converges to b(y) for any fixed y since \ (x . subexponentiality has alrady been proved in Corollary 1. Thus A(x) is everywhere decreasing. Then B(x) = eA(x).U) /or) v 2x This yields easily that ex. the first integral has limit 1 .A(x .3 < 1. we first quote Karamata's theorem (Bingham. elementary but tedious calculations (which we omit) show that A(x) is ultimately decreasing.y) < yA(x .13 works in this setting. an integrable function by assumption. Thus B*2(x )/ B(x) . Goldie & Teugels [66]): Proposition 1. proving B E S. To illustrate how Proposition 1.`(x)b(x) is integrable.y) dy.(y) dy. and exa(x)b(x) = (3x01e(10)x9 is integrable. By (1. Example 1.(x .y) y\(y)• The rightmost bound shows that the integrand in the first integral is bounded by ey"(v). Further.15 In the lognormal distribution. the DFR Weibull distriu bution is subexponential. Jo For y < x/2. x .1)xcl1 .y) < A (y) for y < x/2. Define A(x) = fo .14 Consider the DFR Weibull case B(x) = ex0 with 0 <.1. replace B by a tail equivalent distribution with a failure rate which is everywhere decreasing). Thus.. Thus by dominated convergence . In the regularly varying case. B*2(x) .12.1 has limit 1 + 0. the u lognormal distribution is subexponential. Since ) (x . Example 1. a(x) = ax01..A(y)\(y) dy + fox/ 2 eA(x ). we can use the same domination for the second integral but now the integrand has limit 0 .1 B(x) eA( x)A(xv )A(y)A(y) dy f B(x . L(x) y° (a .A(y)a(y ) = ev'(y) b(y). Thus.2).
1/A(x) and P(X ixil'Y (x) > y) * e'.ea b(x) is integrable. However. Thus exa(x)b(x) .1))a .258 From this we get CHAPTER IX.1))^ ' (b) Assume that for any yo )t(x + y/A(x)) 1 A(x) uniformly for y E (0.17 If B has a density of the form b(x) = aL(x)/x°+1 with L(x) slowly varying and a > 1.xjX > x.)/Y(x) > y) (1 + y/(a .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 .L(x)/x" and )t(x) .1)X(x)/x > y) = P(X > x[1 + y/(a . we get (1.4 is necessary in full generality.1)] I X > x) L(x[1 + y/(a . then B(x) . We conclude with a property of subexponential distributions which is often extremely important: under some mild smoothness assumptions.1)]) xa L(x) (x[1 + y/(a . Proof ( a): Using Karamata's theorem. then 7(x) x/(a . the monotonicity condition in Proposition 1. Then: Proposition 1. 'y(x) = EXix>.13 may present a problem in some cases so that the direct proof in Proposition 1. Then 7(x) .1) and P(X (. (c) Under the assumptions of either ( a) or (b). f O B(y) dy . (1 + y/(a . .6) EX(x) .18 (a) If B has a density of the form b(x) = aL(x)/xa with L(x) slowly varying and a > 1. the overshoot properly normalized has a limit which is Pareto if B is regularly varying and exponential for distributions like the lognormal or Weibull.1)])a 1 1 . let X W = X . More precisely.y(x)B(x). yo] .a/x. HEAVY TAILS Proposition 1.E(X .
Y2. Kliippelberg & Mikosch [134].8) in (b) then follows from P (A(x)X (x) > y) = F(X > x + y/. The remaining statement (1. 2 The compound Poisson model Consider the compound Poisson model with arrival intensity /3 and claim size distribution B. d. cf.A(x) I X > x) = exp {A(x) . St > u}.f yl 0 0 = exp {y (1 + 0(1))} 0 fY A( x + u /A( x)) a(x) du } The property (1. Recall that B0 denotes the stationary excess distribution.1 If Bo E S. Notes and references A good general reference for subexponential distribution is Embrechts.t be the claim surplus at time t and M = sups>0 St..A(x + y/A(x))} =a(x) a(x + x) dx = ex p ex P . then Vi(u) P Bo(u).2.(x). nG(u). Theorem 2 .2 Let Y1.7) is referred to as 1/A(x) being selfneglecting. . with EzK < oo for some z > 1. P The proof is based upon the following lemma (stated slightly more generally than needed at present).1/. Let St = Ei ` Ui .EK G(u).+YK> u) = ^•P(K = n)G* n(u ) .n0 1•P(K= n)•n = EK.. u a oo. Examples 1. . i. Lemma 2.and lognormal distributions . We get p(yl+. be i.14. Then P(Y1 + • • • + YK > u) .nn.15. with common distribution G E S and let K be an independent integervalued r. We assume p = /3µB < 1 and are interested in the ruin probability V)(u) = P(M > u) = P(r(u) < oo). Bo(x) = f0 B(y) dy / µB. 0 G(u) L G(u) . r(u) = inf it > 0. THE COMPOUND POISSON MODEL 259 We omit the proof of (c) and that EX (x) .v.. 1. It is trivially verified to hold for the Weibull.. 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.
Proof of Theorem 2. The tail of Bo is easily expressed in terms of the tail of B and the function y(x) in Proposition 1. x 4 00. P(K = k) = (1. Weibull) one has Bo(x ( B(x) .p)p'.2.?(xµ 8 (x).13).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 ) . a]. The approximation in Theorem 2. Borovkov [73] and Pakes [280].1 is essentially due to von Bahr [56]. .x400 PBB(x) PB Leta+oo. For some numerical studies.1 is notoriously not very accurate. (2. Since EK = p/(1.µ J ) .µB(01 .18. u The condition Bo E S is for all practical purposes equivalent to B E S. HEAVY TAILS u using dominated convergence with >2 P(K = n) Dz" as majorant.. In general: Proposition 2. lognormal .x^ ) B(x) _ f or ( lox . Bo E S. Proof Since B(x + y)/B(x) * 1 uniformly in y E [0. r(1/Q) xlQexp B(x) = ex' From this .1) In particular .1. Note that in these examples . However.p) and EzK < oo whenever pz < 1.260 CHAPTER IX. and for the lognormal and Weibull cases it can be verified using Pitman 's criterion (Proposition 1.. Bo is more heavytailed than B . the result follows immediately from Lemma 2. we have fx B(y)dy = a B0 (x) > lim inf lim inf x+oo B(x) . mathematically one must note that there exist (quite intricate) examples where B E S. see Abate.2) M = Yl + • • • +YK where the Yt have distribution Bo and K is geometric with parameter p.3 If B E S. See also Embrechts & Veraverbeeke [136]. Bo E S is immediate in the regularly varying case. in our three main examples (regular variation . then Bo(x)/B(x) + 00. u x+a Notes and references Theorem 2. as well as examples where B ¢ S. The PollaczeckKhinchine formula states that (in the setup of Lemma 2. Bo ¢ S. The problem is a very slow rate of convergence as u ^ oo. _ B(x^sx Bo(x) µ8 I aoB(y )dy = (^) .
Let U= be the ith claim .. t9(u) = inf {n : Snd> > u} .} Then ik(u) = F ( M > u) = P(i9 (u) < oo)..3. Somewhat related work is in Omey & Willekens [278].+ E A. Based upon ideas of Hogan [200]. This shows that even the approximation is asymptotically correct in the tail. + Xn.9+ < oo) = P(S.1 when u is small or moderately large. Then K M=EY.1 gives 1010.Ti. THE RENEWAL MODEL 261 Choudhury & Whitt [1]. i.. 3 The renewal model We consider the renewal model with claim size distribution B and interarrival distribution A as in Chapter V. Define further 0 = IIG+II = P(r9+ < oo). G+ (A) = P(Sq+ E A. T+ < oo) where r+ = T1 + • • • + T.] The proof is based upon the observation that also in the renewal setting. 1 Assume that (a) the stationary excess distribution Bo of B is subexponential and that (b) B itself satisfies B(x .y + as usual denotes the first ascending ladder epoch of the continuous time claim surplus process {St}..1.e.. In [1].y)/B (x) > 1 uniformly on compact y internals. [279]. also a second order term is introduced but unfortunately it does not present a great improvement. i=1 . in [219] p. M = sup s$ .g. there is a representation of M similar to the PollaczeckKhinchine formula. Then l/i(u) 1 P P [Note that (b) in particular holds if B E S. E. let t9+ = i9(0) be the first ascending ladder epoch of {Snd> }. (3.1) this end . The main result is: Theorem 3 . . Thus G+ is the ascending ladder height distribution (which is defective because of PB < PA). one may have to go out to values of 1/'(u) which are unrealistically small before the fit is reasonable. We assume positive safety loading.. Asmussen & Binswanger [27] suggested an approximation which is substantially better than Theorem 2. Snd) = Xl +.. {n= 0. 195 there are numerical examples where tp(u) is of order 105 but Theorem 2. To Bo(u) u + 00. T1 the ith interarrival time and Xi = U. Kalashnikov [219] and Asmussen & Binswanger [27]. p = iB /µA < 1.
x > 0. the contribution from the interval (.. A(dy) = 1.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_. Then 0 0 F( x . with distribution G+/9 (the distribution of S. whereas for large y . we will use the fact that the proof of (3.Y2.1 IPG_ I / F(x . P(K = k) = (1 .2). Write G+( x) = G+ ( x.1) holds for a general random walk satisfying the analogues of (a). d+ < oo). u a 00. and hence FI(x) . Proof By dominated convergence and (b). Lemma 3 .N.FI(u). Proof Let R+(A) = E E'+ ' I(S. FI (x) _ fz ° F(y) dy.. cf. 0] to the integral is O(F(x)) = o(FI(x)). (3.g+ > x. (b) and does not rely on the structure Xi = Ui .y_ E A) the descending ladder height distribution (IIG II = 1 because of PB < P A) and let PG_ be the mean of G_. Lemma 3 . A.9)9'' and Y1. 0 The lemma implies that (3. B(x) _ J O° B(B(x)y) A(dy) f 1 . 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. oo) = F(S. 0] normalized by IPG_ I so that we should have to G+(x) . U_ (dy) is close to Lebesgue measure on (.d.y+ given r+ < oo). As for the compound Poisson model. The heuristics is now that because of (b). x + oo.i..Ti).FI(x) /IPG_I.262 CHAPTER IX.2 F(x) . are independent of K and i. G_(A) = P(S.y) dy = 1 Pi (X) oo IPG_ I . Let further 19_ _ inf {n > 0: S^d^ < 0} be the first descending ladder epoch.oo.1) is equivalent to P(M > u) " . Let F denote the distribution of the Xi and F1 the integrated tail.3) and we will prove it in this form (in the next Section.d)) E A) denote the pre19+ occupation measure and let and U_ = Eo G'_" be the renewal measure corresponding to G_.3 G+ (x) .(.PBBo(x).B(x). HEAVY TAILS where K is geometric with parameter 9. x * oo.
2). n] < (1 + e)/1µc_ I for n > N.3.(dy) fN FI ( x) + lim sup ZY00 N F(x . > (1 . THE RENEWAL MODEL 263 We now make this precise.9) 1 . n] F1 ( n=N _1 1+e E F(x+n) 0 + limsup xr00 FI(x) FAG. (3.3. n] is just the probability of a renewal at n. the proof is complete.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 0.9)IpG_ I Differentiating the WienerHopf factorization identity (A. F(Y= > x) FI(x)/(OIp _ 1). If G_ is nonlattice. choose N such that F(n .1.y) U.O[s])(1 . Given e. we can assume that the span is 1 and then the same conclusion holds since then U(n .=1 BIp G_ I (1. 0] x+00 FI(x) 00 + lim up 1 x) E F(x + n) U_ (n .1 I .0)0k k I(u) A.1.2) yields 00 F F I (u) P(M > u) _ E(1 .F[s] = (1 . By Lemma 3.1. n] + 1/I µG_ I.e) z lim inf G+(x)  FI (x) Ip G_ I Letting a 10.I n=N (1 E)2 r00 F(x + y) dy + e) lim sup . We then get lim sup G+(x) xro0 Fj(x) < lim sup X)00 o F(x .1.y) U_ (dy) 00 FI (x) < lim sup F(x) U(N. then by Blackwell 's renewal theorem U_ (n .G+[s]) . In the lattice case.UG_ I x. and that U_(n . Similarly.1)/F(n) < 1 + e for n > N (this is possible by (b) and Lemma 3.1. u Proof of Theorem 3. Hence using dominated convergence precisely as for the compound Poisson model. and in the last that FI is asymptotically proportional to Bo E S.
Mn < u}.264 and letting s = 0 yields CHAPTER IX. S+q(u) .Se(u)_1 < a) = o(Fj(u)). .a.u)) = o(P (M > u)) = o(FI(u)).. u)) > P(w(u) < oo)(i lp (0))• On the other hand. FJ(u) UBBO(U) PBo(u) N = (10)Ipc_I JUA . Notes and references Theorem 3. 10(0) But since P(M > u . Then P(M E (u . u)). allowing also for possible dependence between the arrival process and the claim sizes.4 For any a < oo.l. 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..So( u)_1 < a) < P (w(u) < oo)j/i(0) < 0(0) P(M E (u . 4 Models with dependent input We now generalize one step further and consider risk processes with dependent interclaim times.yiui_1.0)ua_ .SS(u)}n=o. see Asmussen & Kliippelberg [36]. on the set {M > u.IIG+II)µc_ = (1 . Note that substantially sharper statements than Lemma 3. u).2. and {Su.a.a. with roots in von Bahr [56] and Pakes [280]. we have P(M E (u .1)6+[0] .(u)+n . HEAVY TAILS µF = (1 .1 is due to Embrechts & Veraverbeke [136]. Proof Let w(u) = inf {n : Sid) E (u .So(u)) are available. Therefore by Lemma 3.a) N P(M > u).. S+9(u) .(1 .4 on the joint distribution of (S.a. Sty(u) .AB iP We conclude by a lemma needed in the next section: Lemma 3 . must attain a maximum > 0 so that P(M > u. P(M > u.Sty(u)_I < a} we have w(u) < oo.
We return to this point in Example 4. 2. We let F* denote the Podistribution of Si.1 except for the first one) is a random walk.1 = max k=0..Sxk}o<t<xk+1xk is the same for all k = 1.1 where the filled circles symbolize a regeneration in the path.. Figure 4. t>0 S. M = sup St. Schmidli & Schmidt [47].. (viewed as random elements of the space of Dfunctions with finite lifelengths) are i. such that {SXo+t  SXo}0<t< X 1Xo .... = Sx. < 0 and EoX < oo where X = X2 . For further approaches. See Fig. G(x) (4. {Sn}n=o..F*(X) = P0(Si > x) ..i.. . Thus the assumption . and apply it to the Markovmodulated model of Chapter VI. MODELS WITH DEPENDENT INPUT 265 Various criteria for this to be true were recently given by Asmussen.. 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 <.. The zerodelayed case corresponds to Xo = Xl = 0 and we write then F0. M. 4. 4.... E0.1) .4 below. (corresponding to the filled circles on Fig..1 based upon a regenerative assumption. 0o(u) etc.d. 4... We give here one of them. +1.1) is assumed.X1 is the generic cycle.X2 < .n n=0. and the distribution of {Sxk+t .1.4.1.Sxi}0<t<x2Xl . {SX1+t . Theorem 4.. Define S. M* = max S. see [47].. assume pp.. The idea is now to observe that in the zerodelayed case.1 Note that no specific sample path structure of {St} (like in Fig.. examples and counterexamples.
3) where Mnx) = sup o<t<xn +1 X..3) applicable so that F(M* > u) 141 F*(u). u p 00. Then '00 (u) = Fo(M > u) .1 Assume that (4.2. Sxn +t . Since clearly M(x) > Sl .2) to show F(M* > u) > 1. Proof Since M > M*. See Fig.1) and (4.266 CHAPTER IX.* i o<t<xn+1x. The one we focus on is Fo (Mix) > x) . it suffices by (4. jF11 F* (U). (4. the assumption means that Mix) and Sl are not too far away.2 Theorem 4.2) Imposing suitable conditions on the behaviour of {St} within a cycle will then ensure that M and M* are sufficiently close to be tail equivalent.S.. 4. (4. (4. HEAVY TAILS for some G such that both G E S and Go E S makes (3.4) liminf u>oo F(M > u) .3) hold.Sxn = sup Sxn+t . Fo(Si > X)... N N Xi=0 N Figure 4.
/3(u) = inf{n=1.4.: S. Theorem 4. Given e > 0.6) 1 p pBo(u) u where B is the Palm distribution of claims and p . 0 yields (4. Po(M* > u) . assume the path structure Nt St = EUit+Zt i=1 .. Letting first u + oo and next e . )) > (1 . .: Sn > u} ..a. Then by Lemma 3.E) Po ( n max St u.2.Sn+1Sn>aV(uSn*)) n=1 00 > (1E)EPo(Mn<u.1 = limti00 St/t.e)Po (MMX> > x). S... MODELS WITH DEPENDENT INPUT Define 79* (u) = inf {n = 1 . MW O(u)+1 < a) IN ( U n=1 A1.( u)1 > a) 00 1: Po(Mn<u. Under suitable conditions .Po (M* > u.S. u)} < P(M* E (u . E (u .5) which follows since Po (M > u..(1 . Let a > 0 be fixed.e)Po (M > U). x > a. u))/P(M* = 0) = o(Po(M* > u)).Mn +1 >aV(u n=1 00 S.(u) . 2.e)Po (M > u. Mn+l > a V (u . We shall use the estimate Po(M > u) Miu^+ 1 < a) = o(Po (M > u)) (4. M^xu)+l > a) .4..1 can be rewritten as 00 (U) (4.+Mn+1>u} 267 (note that {M> u} = {3(u) < oo}). choose a such that Po(Si > x ) > (1 . To this end.a.Sn 0<t<x„+j ( 1 ..4).
cf. The same is true for Sl.'s order EoNx • B(x).4). HEAVY TAILS N` U.3PB.v.268 with {Zt} continuous. i=1 (4.6) holds with p = . independent of {> CHAPTER IX. Proof It is easily seen that the r. (ii) EozNX < oo for some z > 1.} and satisfying Zt/t N. since the tail of Zx is lighter than B(x) by (iv). Corollary 4.1 is in force. and also for Mix) since Nx FNX U.p) Ju P Bo(u) 1p 0 .7). Assume further that (i) both B and Bo are subexponential. and the rest is just rewriting of constants: since p = 1+tlim St = 1+ . and ENX Ui . a4' 0.8) x Write . Mix) < > UE + i=1 o<t<x Thus Theorem 4. we get 00 (u) 1 IPF. Then the Palm distribution of claims is B(x) = E N Eo 0 I( U1 < x) .Q = EoNx/EoX.2 Assume that {St} is regenerative and satisfies (4.X both have tails of sup Zt. are Fmeasurable and NX Po J:U=>x i=1 (iv) Po sup Zt > x / (0:5t<x o(B(x)) Then (4. oX (see Proposition A1. (iii) For some o field Y. the proof of Lemma 4. X and N.I u J Po(Sl > x) dx 1 EoNxB(x) dx EoX(1 .6 below.
i. X3 = 2. Zt .4. The arrival rate is /3i and the claim size distribution Bi when Jt = i. we assume that B E S.. and taking F = o. consider the periodic model of VI. We now return to the Markovmodulated risk model of Chapter VI with background Markov process {Jt} with p < oo states and stationary distribution 7r. .. . of claims arriving in [0. 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. Again . 3 The average arrival rate / and the Palm distribution B of the claim sizes are given by P P Q = ir i/i. . (iii) is obvious.t} is standard compound Poisson and {Zt} an independent Brownian motion with mean zero and variance constant a2.3 As a first quick application. The regenerative assumption is satisfied if we take Xo = Xi = 0. We consider the case where one or more of the claim size distributions Bi are heavytailed. . Taking again Xo = Xi = 0.t + EN'I Ui where {>N`1 Ui . Bo E S.0 (thus (iv) is trivial).. MODELS WITH DEPENDENT INPUT 269 Example 4 ..4 Assume that St = Zt .6) holds. Bo E S. and for some constants ci < oo such that cl + • • • + c.e. Thus we conclude that (4. 1) is Poisson with rate /3 = fo /3(s) ds so that (ii) holds. in particular lighttailed. > 0..(NX). Assume that B E S.6) holds.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).3 that (4. Example 4 . then (iv) holds since the distribution of supo<t<i Z(t) is the same as that of I Zl 1. More precisely. Theorem 4.6) u holds.9). 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 . X2 = 1. (i) holds. < 1. X2 = 1. we conclude just as in Example 4.5 Consider the Markovmodulated risk model with claim size distributions satisfying (4. i=1 B = >2 7riaiBi i=1 and we assume p = 014 B = Ep ri/3ipB. The key step of the proof is the following lemma. Then (4. In particular. X3 = 2. The number N.
It follows by a slight extension of results from Section 1 that P P(Yo > x I Y) G( x) ci Ni. oo) and define p Ni Yx = EEX'i .. u Proof of Theorem 4.. i=1 Proof Consider first the case X = 0.. Let {Fi}t=1 P be a family of distributions on [0. .2. oo) such that G E S and some c1.. and that for some + cp distribution G on [0. as x a oo. 2 . 1. . we can define the regenerations points as the times of returns to i.ciG(x). Then P P(Yx > x) ... . X > 0 a r."+Np . Assume EzN1+"'+Np < oo for some z > 1 and all i.270 CHAPTER IX. P P P(YX and > x I.F) < CG(x)zn'1+. cp with cl + > 0 it holds that Fi(x) . and the rest of the argument is then just as the proof of Corollary 4.. and F a aalgebra such that (N1.v. are independent with distribution Fi for Xij. i =1 P(Yo > x I ^ ) < CG(x)zN1+ +Np for some C = C(z) < oo. For lighttailed distributions.Fmeasurable. The same dominated convergence argument completes the proof. i=1 P(Yx > x ^) < P(Y0 > x I. . NP ) be a random vector in {0.5.^•) G(x) P ^ E ciNi = C.}P. .G( x ) > ciNi . 6 Let (N1.F) = P(Yo > X+x I •^) G (x +x)>2ciNi i=1 . An easy conditioning argument then yields the result when Jo is u random.X i=1 j=1 where conditionally upon F the Xi. . If Jo = i. Thus dominated convergence yields ( P(Yo>x P(Yo>x .c'(x) where c = ciENi . . Markovmodulation typically decreases the adjustment coefficient y and thereby changes the order of magnitude of the ruin . HEAVY TAILS Lemma 4 . NP ) and X are . i1 = E\ G(x) In the general case..
5 Finitehorizon ruin probabilities We consider the compound Poisson model with p = /3pB < 1 and the stationary excess distribution Bo subexponential.. Within the class of risk processes in a Markovian environment.2 and Example 4. i.4. 5 was first proved by Asmussen. The main result of this section. states that under mild additional conditions. VI. I T(u) < oo). Notes and references Theorem 4. we let PN"N = P(. see Schlegel [316].. ) form a general stationary sequence and the U. FINITEHORIZON RUIN PROBABILITIES 271 probabilities for large u. Combined with the approximation for O(u). cf. Theorem 5.5. IV.1.4. cf. Theorem 4. in particular Proposition 2. and independent of (T1. cf. > 0) matter for determining the order of magnitude of the ruin probabilities in the heavytailed case. In contrast. and the final reduction by Jelenkovic & Lazar [213].e.3. Floe Henriksen & Kliippelberg [31] by a lengthy argument which did not provide the constant in front of Bo(u) in final form. this is applied for example to risk processes with Poisson cluster arrivals.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.. We start by reviewing some general facts which are fundamental for the analysis.p)Bo(u). For further studies of perturbations like in Corollary 4. ). The present approach via Theorem 4. That paper also contains further criteria for regenerative input (in particular also a treatment of the delayed case which we have omitted here).d.6) to hold in a situation where the interclaim times (T1.4.T2. An improvement was given in Asmussen & Hojgaard [33].1 is from Asmussen.. Essentially.i. i. 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). It follows from Theorem 4. 5a Excursion theory for Markov processes Let until further notice {St} be an arbitrary Markov process with state space E (we write Px when So = x) and m a stationary measure. Theorem 2. Then O(u) .pl(1 . this then easily yields approximations for the finite horizon ruin probabilities (Corollary 5. Schmidli & Schmidt [47]. m is a (orfinite) . the discussion provides an alternative point of view to some results in Chapter IV.7). this should be compared with the normal limit for the lighttailed case. As usual..4. as well as a condition for (4. for lighttailed distributions the value of the adjustment coefficient y is given by a delicate interaction between all B.T2.. r(u) is the time of ruin and as in IV.7.5 shows that basically only the tail dominant claim size distributions (those with c.
St is distributed as y . m.h. an excursion in F starting from x E F is the (typically finite) piece of sample path' {St}o<t<w(F°) I So = x where w(Fc) = inf It > 0: St 0 F} .t.2) with t = 1 means m.r. Lebesgue measure. but the example of relevance for us is the following: Proposition 5. u For F C E. a main difficulty is to make sense to such excursions also when Px(w(F°) = 0) = 1.z) dx G(dz) = ffh(y + z) k(y)dy G(dz). w(Fc) < oo ) 'In general Markov process theory.s. We let QS be the corresponding distribution and Qx. y = 0). to the r.>N` Ui. . 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.s. follows by the substitution y = x . and starting from So = y. Then (5.2) for all bounded measurable functions h. in the terminology of general Markov process theory.)k(x . Rt is distributed as x + t .rij = mjsji where r13.1 A compound Poisson risk process {Rt} and its associated claim surplus process {St} are in classical duality w .t. y to vary in. t. r. and (5. say. HEAVY TAILS measure on E such that L for all measurable A C E and all t > 0.272 CHAPTER IX. . The simplest example is a discrete time discrete state space chain. however . {St} and {Rt} are in classical duality w. a familiar case is time reversion (here m is the stationary distribution). resp. The equality of the l. {Rt}.h.y = Qx (. oo). to consider only the case Px(w(F`) = 0) 0. Sw(F. Thus. x = 0+ and F = (0. Proof Starting from Ro = x. Say {St} is reflected Brownian motion on [0.00).t + EI U.2) means ffh(a. for states i. k on E. (note that we allow x. For the present purposes it suffices . where we can take h. Let G denote the distribution of ENt U. the whole of R and not as usual impose the restrictions x > 0.s=j are the transition probabilities for {St}. k as indicator functions.= y.). j.z.
in = y. Qx. . io = x. when p = .2.SS(F.. Thus. .y(2p21 . Sw(F)1 = y) . i1.y as a measure on all strings of the form i0i1 . w(0. Qx y is the distribution of an excursion of {St} conditioned to start in x E F and terminate in y E F. 0]. The theorem states that the path in (b) has the same distribution as an excursion of {Rt} conditioned to start in y < 0 and to end in x = 0. this simply means the distribution of the path of {Rt} starting from y and stopped when 0 is hit. In particular: Corollary 5. The theorem is illustrated in Fig ..3 The distribution of r(0) given r(0) < oo.y() = P ({SW(F`)t} 0<t<w(F °) E So = x. But in the risk theory example (corresponding to which the sample paths are drawn).2 Qy. We can then view Qy. . 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). /^s x (S1 = Z1. That is. in with i0.. the one in (b) is the time reversed path. Sn+1 E Fc) nx...(0)_ = y < 0 is the same as the distribution of w(y) where w(z) = inf It > 0 : Rt = z}. Sn = in = y.. FINITEHORIZON RUIN PROBABILITIES 273 y E F (in discrete time. x = 0. z > 0. oo) = r(0) x= St y (a) Figure 5. .y = Qy Q.. 5. S.13AB < 1] Proof of Theorem 5.5.1 for the case F = (oo. QR and QRy are defined similarly. Sw(Fo)_ should be interpreted as Sw(F^)_1). in E F.= y) Theorem 5 .).s. [note that w(z) < oo a.1 The sample path in (a) is the excursion of {St} conditioned to start in x = 0 and to end in y > 0..itt) = P Px(w(Fc) < 00. and we let Qy y refer to the time reversed excursion .
Silt' E Sxik_1 .. note first that Pt' (R l = il... in) = Qx. in) = oo jEF^ Sxin1 .. Rn = in = x.ik_1EF . . . Sn+1 E Fc) n=1 i1. (Fc)1 = y) 00 CHAPTER IX.. Si1y k=1 i1 . R . 20 = y.. Si l io E mjSjx... in = x.... Si11 S 1 ... Si1y 00 jEF° E E 5xik_ 1 . 2n) = Qx.. 2p).(F<)1 = Y) S S and Qx y( ipil ..ik1EF Similarly but easier Sxin_1 ... i0) Q x. Rn+1 E FC) TioilTili2... t' y and Qy x are measures on all strings of the form ipi l . . Rn = in = x.. = in = y.....ii ..y(inin _ E SYj jEF` 00 Sxik _1 . in)  Pt' (R1 = ii. HEAVY TAILS E E Px (Si = 21i .......i„_iEF Similarly.TI( 2n2n _1 . 2p) when 20. R Qy x(2p21 ...ik_1EF Sxin_1 .... i0 = y.gilt' k=1 ii . Rn+1 E F`) F (w(Fc) < 00.in1 ... To show Q y x (i0 i 1 . ...... Silt' E SO k=1 i1.. . . S... . 21 .. MY Thus Qx(ioii ...J (i..rin_1in E Txj jEFC m21 s2120 m2252221 m in Ssn n1 mjSjx Mx m2p mil min1 jEF` 1 Sinin _ 1 ... S. in E F.in E F. ....274 note that Fx(w(Fc) < 00. in with 20. ... in = x...
(o) > y} = {T(0) < oo. That is. Y > y} . Z) is described in Theorem 111.2 The distribution of (Y. Z follows the excess distribution B(Y) given by B(Y) (x) _ B(y + x)/B(y). y > u. the case r (O) < oo.t.')density of Y is B(y)/[.2. P(o) ). 7(0) < oo. the distribution w.r. Let Y = Yl = Sr+( 1) be the value of the claim surplus process just after the first ladder epoch . the P(u. We are interested in the conditional distribution of T(u) = T(0) given {T(0) < oo.5. The formulation relevant for the present purposes states that Y has distribution Bo and that conditionally upon Y = y.UBBo(u)].t. P(") = P(.v. Y > u).p. Bo") is also the P(u. ST(o) > y.2. that is. see Fig.r.B(a) +a PBBo(u) .')distribution of Z since P(Z>aIY>u) = 1 °° B(y) B(y + a) dy FLBBo(u) B (y) J°° (z) dy . FINITEHORIZON RUIN PROBABILITIES 275 5b The time to ruin Our approach to the study of the asymptotic distribution of the ruin time is to decompose the path of { St} in ladder segments .2. 1 w .')distribution of Yu is Bo"). that is. S. U T(O) = T (u) Y Figure 5. To clarify the ideas we first consider the case where ruin occurs already in the first ladder segment . 5. Now the P(u.'s are defined w. Z = Zl = ST+( 1)_ the value just before the first ladder epoch (these r.
Fig. Zn). . Then.. Then Corollary 5. and since its dominates the first n . Z/'y(u) * W in Pi "' ')distribution .e.3) where the distribution of W is Pareto with mean one in case ( a) and exponential with mean one in case (b). . are i. and distributed as (Y.18(c) Bo")(yY (u)) + P(W > y) ( 5... 1/(1 .e. it therefore follows that T(u)/Z converges in Pi"'')probability to 1/(1 . Z1).3.. However.: r+ (n) < oo. We now turn to the general case and will see that this conclusion also is true in P(")distribution: Theorem 5 .1. this in principle determines the asymptotic behaviour of r(u). It is straightforward that under the conditions of Proposition 1.. in particular of Z.p). The idea is now to observe that if K(u) = n. (Y. z ^ oo. i. Recall the definition of the auxiliary function y(x) in Section 1. That is . Zk be defined similarly as Y = Y1. must be large and Z1. Now Bo E S implies that the Bo ")(a) + 0 for any fixed a.p). Zn_1 'typical' which implies that the first n1 ladder segment must be short and the last long. we get the same asymptotics as when n = 1. In the proof.d. the random vectors (YI.. Since the conditional distribution of Z is known (viz. . Z = ZI but relative to the kth ladder segment. Bo") ). Since w(z)/z a$..p) in Pi"'')distribution. Z). HEAVY TAILS Let {w(z)}Z^.. . cf. a slight rewriting may be more appealing.. Then 7(u)/y(u) ^ W/(1 .. . Yn_1 'typical'.3 implies that the P("'1)distribution of T(u) = r(0) is that of w(Z). P(Z < a I Y > u) 3 0.o be defined by w(z) = inf It > 0: Rt = z} where {Rt} is is independent of {St}. let r+(1) = T(0).. . . then by the subexponential property Yn must be large..p) in F(u) distribution. Hence Z. and YI. K(u) = n).r+ (n .p) then yields the final result T(u)/y(u) + W/(1 . Y1 + • • • + Yn > u} denote the number of ladder steps leading to ruin and P("'n) = P(• I r(u) < oo. 5..i.T+(2). the duration T+ (n) .3) holds. We let K(u) = inf In = 1. 4 Assume that Bo E S and that (5. i. denote the ladder epochs and let Yk. > u with high probability.. conditionally upon r+ (n) < oo.276 CHAPTER IX. more precisely. r(u)/Z 4 1/(1 . 2.1) of the last ladder segment can be estimated by the same approach as we used above when n = 1..
I A'(u)) = P(u. Yn .u) E •) . > u}.. I A"(u ))II + 0..Yn1iYn . then IIP( I A'(u)) Taking A'(u) = {Y. Further. .n) (y1. Lemma 5.Yl+ +Yf1>u}...u) E • I A'(u)) = Bo (n1) ®Bou) . P (Yj. Y„1.5. +Yn1<u. ..2. FINITEHORIZON RUIN PROBABILITIES 277 16 Z3 Z1 r+(1) T+(1) T+(1) Figure 5. ..5 Ilp(u. Proof We shall use the easily proved fact that if A'(u). .Bo (ri1) ®B( ..3 In the following. A"(u) are events such that P(A'(u) AA"(u)) = o(F (A'(u)) (A = symmetrical difference of events).u) II 0. P(. II ' II denotes the total variation norm between probability measures and ® product measure. A"(u) _ {K(u)=n} = {Y1+ P(.n). suitably adapted). the condition on A'(u) A A"(u) follows from Bo being subexponential (Proposition 1.
The same calculation as given above when n = 1 shows then that the marginal distribution of Zn is Bou). .. be independent random vectors such that the conditional distribution of Zk given Y.. Y1 +.278 Lemma 5 . copies of {w(z)}. . ...P(Z' E •)II > 0 (here Y. the F'distribution of r(u) is the same as the P'distribution of w1(Zl) + • • • + wn(Zn). then 11P(Z E •) . n_1 < u.2. + Y" > u) Flul (K (u ) = n) _ Cu) P"F(1'i +.P(Y' E •)II * 0. +wn(Z n))l7( u ) > 1y) ^' P(u'n)(wn (Zn)/7(u) > y) 4 NW/(1 . Notes and references Excursion theory for general Markov processes is a fairly abstract and advanced topic.P) Bo(u) for n = 1.. the density of Yn is B(y)/[IBBO(u)]..i.+y 1 p"F(Yn > u) P)Pn1 P/(1 . ..4. By Lemma 5.. Then according to Section 5a. Z' are arbitrary random vectors.1.6 IIPIu'n ) CHAPTER IX. Similarly (replace u by 0).6. and clearly Zi. see Fitzsimmons [144]).. HEAVY TAILS ((Z1'. Y") u etc. . n.. . . whereas wn(Zn) has the same limit behaviour as when n = 1 (cf..... For Theorem 5. It therefore suffices to show that the P(u'")distribution of T(u) has the asserted limit. {wn(z)} be i.1). Now use that if the conditional distribution of Z' given Y' is the same as the conditional distribution of Z given Y and JIF(Y E •) . (Y.u has distribution Bout That is.). Zn are independent.1 P PBo(u) • P(W/(1 . in our example Y = (Y1. wk(Zk) has a proper limit distribution as u + oo for k < n.d.. Thus F(u'n)(T(u) /7(u) > y) = F(u'n)((wl (Z1) + .p) < y).7 O (u.4).. Z. ..P) > y) Corollary 5.y(u)T) .r. y > u... .Bo (n1) ®Bo' 0.. the discussion just before the statement of Theorem 5. k = 1. 2.t. and that Yk has marginal distribution B0 for k = 1....1 and Y„ . Let {wl(z)}. Proof Let (Y11. . P(u) since by Theorem 2. The first step is to observe that K(u) has a proper limit distribution w. in particular his Proposition (2.' = y is BM.. Proof of Theorem 5. n .. . Z11). .. Zn) E •) .. Y'.. the marginal distribution of Zk is Bo for k < n. Zn)..
6 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate /3. Proof of Theorem 6. u (6. Assume for simplicity that {Vt} regenerates in state 0 .1. T) when T + oo with u fixed. More precisely. Corollary II. 3. nontrivial and we refer to Asmussen [22]./3Ea B(u). one expects the level y form which the big jump occurs to be 0(1).1 Assume that B is subexponential and that p(x) > 00.B(u). max VB>0I Vo=0^ o<s<t J11JJJ Lemma 6 . x > oo.y) . cf.2.2 Define M. p(Y) and the result follows. We will show that the stationary density f (x) of {Vt} satisfies f (x) /B(x) r(x) We then get V.(u) = P(V > u) = f f (y) dy . Theorem 6 . V. that fo p(x)1 dx < oo. and premium rate p(x) at level x of the reserve. 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).e.1) The key step in the proof is the following lemma on the cycle maximum of the associated storage process {Vt}. . claim size distribution B. the results only cover the regularly varying case. and define the cycle as a = inf{t>0: Vt=0.6. however.. the probability that is exceeds u is then B(u . 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 . Extensions to the Markovmodulated model of Chapter VI are in Asmussen & Hojgaard [33].(3 u u J B(y) dy . Then P(MT > u) . = supo<t<0. i. The rigorous proof is. Asmussen & Teugels [53] studied approximations of i (u. The heuristic motivation is the usual in the heavytailed area. that MQ becomes large as consequence of one big jump. Then 0 (u) Qf "O ^) dy.
q ( u)) 1 . 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. . Then D(u) = f(u)p(u) and. D(u) = DQ(u)/µ. there exist constants c(u) 4 0 such that the limiting distribution of r(u)/c(u) given r(u) < oo is exponential. where also the (easier) case of p(x) having a finite limit is treated .P(MT > u) $B(u) Ft µ(1 . Hence f (u)r(u) = D(u) = Do(u) . Further the conditional distribution of the number of downcrossings of u during a cycle given Mo > u is geometric with parameter q(u) = P(Mo > u I Vo = u). u Notes and references The results are from Asmussen [22]. It is also shown in that paper that typically.q(u) Now just use that p(x) * oo implies q (x) + 0.280 CHAPTER IX. by regenerative process theory.
i.96s z f (1. The crude Monte Carlo ( CMC) method then amounts to simulating i. Hence 1..z) 4 N(0.d. . Fox & Schrage [77]. z) 2 = Zit NE ii ii According to standard central limit theory .Chapter X Simulation methodology 1 Generalities This section gives a summary of some basic issues in simulation and Monte Carlo methods .2) is an asymptotic 95% confidence interval . replicates Zl.. topics of direct relevance for the study of ruin probabilities are treated in more depth.. 4Z). and this is the form in which the result of the simulation experiment is commonly reported. Rubinstein [310] or Rubinstein & Melamed [311] for more detail . vrN(z . estimating z by the empirical mean (Z1 + • • + ZN)/N and the variance of Z by the empirical variance N s2 = E(Z{  N 2. We shall be brief concerning general aspects and refer to standard textbooks like Bratley. 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. where a2 = Var(Z ). ZN. 281 . Ripley [304].
there are others which are widely used in other areas and potentially useful also for ruin probabilities. v. writing Var(Z) = Var(E [Z I Y]) + E(Var[Z I Y]) . and a longer CPU time to produce one replication. SIMULATION METHODOLOGY In the setting of ruin probabilities. conditional Monte Carlo and importance sampling. We survey two methods which are used below to study ruin probabilities. typically by modifying Z to an alternative estimator Z' with EZ' = EZ = z and (hopefully) Var(Z') < Var(Z). an added programming effort. We mention in particular ( regression adjusted) control variates and common random numbers. Say that Var(Z') = Var(Z)/2. 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. 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'. generated at the same time as Z. Sections 24 deal with alternative representations of Vi(u) allowing to overcome this difficulty.282 CHAPTER X. However. variance reduction is hardly worthwhile. Typically variance reduction involves both some theoretical idea (in some cases also a mathematical calculation). Therefore. 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. lb Variance reduction techniques The purpose of the techniques we study is to reduce the variance on a CMC estimator Z of z. one can argue that unless Var(Z') is considerable smaller than Var(Z). we then have EZ = EZ = z. it is straightforward to use the CMC method to simulate the finite horizon ruin probability z = i.b(u. Further. and many sophisticated ideas have been developed. Conditional Monte Carlo Let Z be a CMC estimator and Y some other r . The situation is more intricate for the infinite horizon ruin probability 0(u). so that Z' is a candidate for a Monte Carlo estimator of z. Z = I inf Rt < 0 (0<t<T = I('r(u) < T). This is a classical area of the simulation literature. and in most cases this modest increase of N is totally unproblematic. Letting Z' = E[Z I Y].
and the problem is to make an efficient choice. i. a crucial observation is that there is an optimal choice of P: define P by dP/dP = Z/EZ = Z/z.zrs. In order to achieve (1. Thus we cannot compute L = Z/z (further..z2 = 0. it gives a guidance: choose P such that dP/dP is as proportional to Z as possible.3) Thus. using the CMC method one generates (Z1. it may often be impossible to describe P in such a way that it is straightforward to simulate from P).[E(LZ)] = E Z2 Zz . Thus. Then z Var(LZ) = E(LZ)2 .3).zrs) = 2 1 N 2 2 2 i=1 i=1 N > Lt Zi . it appears that we have produced an estimator with variance zero. However. (ZN.96 sis v^ N 2 1 where srs = N j(LiZi . (1. L1). Variance reduction may or may not be obtained: it depends on the choice of the alternative measure P. even if the optimal change of measure is not practical.1. the obvious possibility is to take F and P mutually equivalent and L = dP/dP as the likelihood ratio. L = z/Z (the event {Z = 0} is not a concern because P(Z = 0) = 0). ..E [Z Z]2 = z2 . one would try to choose P to make large values of Z more likely. This may also be difficult to assess . GENERALITIES 283 and ignoring the last term shows that Var(Z') < Var(Z) so that conditional Monte Carlo always leads to variance reduction. Importance sampling The idea is to compute z = EZ by simulating from a probability measure P different from the given probability measure F and having the property that there exists a r.v.e. . LN) from P and uses the estimator N zrs = N > L:Zj i=1 and the confidence interval zrs f 1. the argument cheats because we are simulating since z is not avaliable analytically. To this end. but tentatively. L such that z = EZ = E[LZ]. . Nevertheless.
assume that the rare event A = A(u) depends on a parameter u (say A = {r(u) < oo}). as is the case of typical interest. This leads to the equation 1. In ruin probability theory. I.0. it does not help telling whether z is of the magnitude 104. if z is small. let z(u) = P(A(u)). The optimal change of measure ( as discussed above) is given by P(B) = E [ Z] i. We shall focuse on importance sampling as a potential (though not the only) way to overcome this problem. Again.. and further it is usually not practicable to simulate from P(•IA). 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.284 CHAPTER X.1.96oz /(zV) = 0.100 . An example where this works out nicely is given in Section 3. However. u + oo.1. but if the point estimate z is of the order 105. the issue is not so much that the precision is good as that relative precision is bad: oZ z(1 . A = {T(u) < T} or A = {r(u) < oo} and the rare events assumption amount to u being large. and let Z(u) be a Monte Carlo estimator of z(u).96 2Z ( 1 .5 or even much smaller . large sample sizes are required. a confidence interval of width 10 4 may look small. The CMC method leads to a variance of oZ = z(1 .z) 1001. we may try to make P look as much like P(•IA) as possible. Z z V5 In other words .B = iP(AB) = P(BIA).. We then . However. For each u.96 2 z2 z increases like z1 as z . the optimal P is the conditional distribution given A. i. say 10%.e. N . z I. assume that the A(u) are rare in the sense that z(u) * 0.e.z) which tends to zero as z ^ 0. say of the order 103 or less. Z = I(A) and A is a rare event. To introduce these. Two established efficiency criteria in rare events simulation are bounded relative error and logarithmic efficiency.e.z) 1 > 00. in terms of the halfwidth of the confidence interval. 10 . SIMULATION METHODOLOGY 1c Rare events simulation The problem is to estimate z = P(A) when z is small . Thus. Another way to illustrate the problem is in terms of the sample size N needed to acquire a given relative precision .
where X1. P(K = k) = (1 . The term logarithmic comes from the equivalent form . Therefore . 3.e. but as a CMC method .4) for any e > 0. 2. which gives a logarithmically efficient estimator . so that NE (u) may go to infinity.1) may be written as V) (u) = P(M > u). XK from the density bo(x).. logarithmic efficiency is almost as good as bounded relative error.X1 + + XK. . with common density bo(x) = B(x)/µB and K is geometric with parameter p. . If M > u. 2 Simulation via the PollaczeckKhinchine formula For the compound Poisson model.p)pk. this means that the sample size N = NE(u) required to obtain a given fixed relative precision (say a =10%) remains bounded.log z(u) of (1. the mathematical definition puts certain restrictions on this growth rate.d. Notes and references For surveys on rare events simulation. Var(Z(u)) hm sup U+00 z (u) 2E < oo (1.log Var(Z(u)) lim inf > 2 u+oo . Generate K as geometric. Otherwise. where M = X1 + • • • + XK. X2. The algorithm gives a solution to the infinite horizon problem .i. are i. the PollaczeckKhinchine formula III. Generate X1. Thus.4).(2. and in practice. We shall here present an algorithm developed by Asmussen & Binswanger [ 271. i. According to the above discussion.. However. it is appealing to combine with some variance reduction method .1. see Asmussen & Rubinstein [45] and Heidelberger [190]. let Z +. where Z = I(M > u) may be generated as follows: 1. it is not efficient for large u . Let M .p)pk.2. F(K = k) = (1 ... . let Z +. 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.0. O (u) = z = EZ. 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. This allows Var(Z(u)) to decrease slightly slower than z(u)2.
L(x)/x`' with a > 0 and L(x) slowly varying.SK1)2.. compute Y = u . assume in the following that Bo(x) ..p/(l ...1) V)(u) . X1 + + XK_ 1 > x when X1 > x... asymptotically it presents no improvement : the variance is of the same order of magnitude F(x).. However. and considering only the remaining ones. Then (cf. So.. and let Z(2)(u) = _ P (SK B0((u > u I X(l). For the simulation...+XK > uIXl. form the order statistics X(1) < X(2) < .XK_1). Z(1) (u) has a smaller variance than Zl (x).X(n_1)) Bo(X(„_l) V X) Bo(X(n1)) . 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.XK_1 and let Z( 1)(u) = Bo (Y) (if K = 0. As a conditional Monte Carlo estimator .. XK..p)Bo(x).. . To see this.S( K_1)) V X(K1)) / Bo(X(K 1)) where S(K_l) = X(1) + X(2) + • • • + X(K_1). The idea of [27] is to avoid this problem by discarding the largest X. . Xl > x... Thus. just note that EZ(1)(u ) 2 > E[Bo (x ....X(2)....286 CHAPTER X. .. note first that To check the formula for the P(X(n) > x I X(1). K > 2] = P2p(Xl > x) = P2Bo(x) (here we used that by positivity of the X.X1 ..b(u) = P (Xl +•••+XK>u) = EF[Xl + . < X(K) throw away the largest one X(K).. . This calculation shows that the reason that this algorithm does not work well is that the probability of one single Xi to become large is too big. SIMULATION METHODOLOGY when the claim size distribution B (and hence Bo) has a regularly varying tail.2..X(2). A first obvious idea is to use conditional Monte Carlo: write i.. Theorem IX.XK1] = EBo(uX1 . conditional probability.X(K1)) . XK1. we generate only X1.Xl .. we thus generate K and X1i .. and that Bo(y) = 1. y < 0). Z(1)(u) is defined as 0).
. Notes and references The proof of Theorem 2. Also in other respects the findings of [28] are quite negative: the large deviations ideas which are the main approach to rare events simulation in the lighttailed case do not seem to work for heavy tails. For practical purposes.y. that is. . l)) BO(X(n1)) Theorem 2 . Compute y > 0 as solution of the Lundberg equation 0 = K(y) = )3(B[y] . IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 287 We then get P(S" > x I X( 1). However . X(n1)) P(X(TZ) + S(.modulated model P(r+ < oo) and G+ are not explicit ). X(2). and we refer to [27]. using 13L. the algorithm for generating Z = Z(u) is: 1. BL(dx) = e7sB(dx)/B[y].u is the representation 0(u) = e7sr(u) overshoot (cf. Then the algorithm given by { Z (2) (u) } is logarithmically efficient. Asmussen . and that paper contains one more logarithmically efficient algorithm for the compound Poisson model using the Pollaczeck. .Khinchine formula and importance sampling . . 3 Importance sampling via Lundberg conjugation We consider again the compound Poisson model and assume the conditions of Ce7". l)) .. Binswanger and HOjgaard of [28] give a general survey of rare events simulation for heavy tailed distributions . BL by I3L = /3B[y]. X . X(2).1 is elementary but lengty. 111. 1 Assume that Bo (x) = L(x)/x° with L(x) slowly varying.S(n_1) I X(1).. B.. the continuoustime process {St} is simulated by considering it at the discrete epochs {Qk} corresponding to claim arrivals. and simulate from FL. 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. use the the CramerLundberg approximation so that z(u) = '(u) = e7"ELe7E(") where ^(u) = ST(") . The algorithm is sofar the only efficient one which has been developed for the heavytailed case.. BL instead of 0. X (.1) ._1) > P(X(n) > _ X X(1). Thus. X(2). and define )3L.. for the purpose of recording Z(u) = erysr(u). . in the renewal or Markov.3.5). . X(n1)) Bo((x .S (n1)) V X (..
. let S. In detail . More precisely. BL could improve the variance of the estimator . BL). the discussion at the end of Section 1b.2ryu _ z (u)2/C2. namely ELery£(").7 tell that P(.1 The estimator Z(u) = e'rs* "u) (simulated from FL) has bounded relative error. Let FL (dx) = 'For the renewal model.r(u) < oo) = 1.l3 and U from B. return to 3.(u)) are asymptotically coincide on {r(u)} < oo. so that changing the measure to FL is close to the optimal scheme for importance sampling . cf. Let S .288 2. b) # (/3L. the results of IV. b different from . 4.2 The estimator (3. be i. There are various intuitive reasons that this should be a good algorithm.. SIMULATION METHODOLOGY 3. u It is tempting to ask whether choosing importance sampling parameters . r(u) < oo) and FL (both measures restricted to. If S > u. let Z F e_'s. and avoid simulating the known part e7". B) is not logarithmically efficient when (/3. and we have: Theorem 3. one must restrict attention to the case 4µB > 1. . Ti. It resolves the infinite horizon problem since FL(.d.. P'[y] < oo for some ry > 0.QL. to deal with the infinite horizon problem . The algorithm generalizes easily to the renewal model . The answer is no. Generate T as being exponential with parameter . M(u) = inf {n : S„ > u}.. = X1 + . Let X1.Q. Proof Just note that EZ(u)2 < e . and the change of measure F r FL corresponds to B > BL. X2.. A > AL as in Chapter V.T.1) (simulated with parameters ^3. + X. and assume that µF < 0 and that F[y] = 1. . Otherwise. The proof is given below as a corollary to Theorem 3.. We formulate this in a slightly more general random walk setting '. Xi = U.3.i..F. 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.S+U . Let Sf0 CHAPTER X. with distribution F. We may expect a small variance since we have used our knowledge of the form of 0(u) to isolate what is really unknown. In fact: Theorem 3.
1.2 > 0.. where Kl og (X) (j) 2 ) = log dFL (Xi) . The importance sampling estimator is then Z( u) = e'rSM( ).P = FL. .. Jensen's inequality and Wald's identity yield EpZ(u)2 > exp {EL(K1 + .yu+elu u +oo etry' 1 > lim up C2e2. For the second. = c'.3 The estimator (3.. (3. More generally.2) dF Theorem 3. write W(F IF) _ F(XI). where e' = EL Iog dFL (Xi) > 0 by the information inequality.i. + KM(u))} = exp {ELM(u)(E . Here ELK. EFZ(u)2 = EeW2(FIF) = Ep [W2(FIFL)W2(FLIF)] = EL [W2 ( FIFL)w(FLIF)] = ELexp {Kl+. When F # FL.2ryELXi. Proof The first statement is proved exactly as Theorem 3 . let F be an importance sampling distribution equivalent to F and M(u) dF Z(u) _ I (Xi) . EFZ(u)2 EFZ(u)2 lim sup z(u)2eeU = lim cop C2e2.. it thus follows that for 0 < e < e'/ELXi. F(XM(u)). .2'X1 . By the chain rule for RadonNikodym derivatives. IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 289 e7yF(dx).. K2...d..2) (simulated with distribution F of the X3 has bounded relative error when . Since K1.+KM(u)}.. is not logarithmically efficient. Since ELM(u)/u + 1/ELXi.3.2ryELXi)} . are i.yu = G.
U". then the estimator Z(u) = e7Sr(°)I(r(u) < yu) (simulated with parameters /3L. The queueing literature on related algorithms is extensive .T" > x) J /3"e0 yB (x + y) dy = . u Notes and references The importance sampling method was suggested by Siegmund [343] for discrete time random walks and further studied by Asmussen [ 13] in the setting of compound Poisson risk models .T".'(y) or y > 1/r. see e. claim size distributions B'. Further discussion is in Lehtonen & Nyrhinen [245]. optimality is discussed in a heavy traffic limit y 10 rather than when u + oo.4 indicate that we can expect a major difference according to whether y < 1/r. The easy case is y > 1/k'(y) where O(u.T' D U" . In fact: Proposition 4. BL) has bounded relative error. U' .i. Next. CHAPTER X. As in IV.'(y).T' has a left exponential tail with rate /3' and U" .1 is from Lehtonen & Nyrhinen [244]. T) with T < oo.1 If y > 1/ic'('y). Consider compound Poisson risk process with intensities /3'. we conclude by differentiation that Bo(x)=B' (x)forallx > 0. with the present (shorter and more elementary) proof taken from Asmussen & Rubinstein [45]. T".T" has a left exponential tail with rate /3'. The extension to the Markovian environment model is straightforward and was suggested in Asmussen [ 16]. all that needs to be shown is that if U' . yu) is close to zk(u).290 which completes the proof.4.3'eO'x f f P (U" .2. Then according to Theorem 3. In [13].3. This immediately yields / = 3". 4 Importance sampling for the finite horizon case The problem is to produce efficient simulation estimators for '0 (u. from 3' P(U'T'>x) ^ = ^ eQ'zB (z) dz.B'=B".T". the references in Asmussen & Rubinstein [45] and Heidelberger [190].e. B" and generic claim sizes U'. /3". First by the memoryless distribution of the exponential distribution . U' . The optimality result Theorem 3. SIMULATION METHODOLOGY u Proof of Theorem 3. we write T = yu. then /3' B' = B".T' = U" .3"eQ x 0 J eQ zB (z) dz x (x > 0) and /3' = /3". generic interarrival times T' . so that one would expect the change of measure F 4 FL to produce close to optimal results.g. . The results of IV. /3'eQ'YR'( x + y) dy = .
1.4.(ay). .5) follows.to g x ( u ) u u so that (1. Bounding u ELZ(u)2 above by a7u. T( u) < yu] e2ryyuEay le 2ay^(u). (4.3) (simulated with parameters /gay.2) Since the definition of ay is equivalent to Eay r(u) .4.8). Bay) is logarithmically efficient. T(u) < yu] . 7y (4. and in fact. We recall that ay is defined as the solution of a'(a) = 1/y.log 4')u) 4 u (Theorem IV.2 The estimator (4.3) and we have: Theorem 4. that ryy = ay .log Var(Z(u)) _ . yu) = eayu Eay Leay^(u)+r(u)K(ay). T(u) < yu] e Hence by (4.yu.8 has a stronger conclusion than (4. and that ryy > ry. IMPORTANCE SAMPLING FOR THE FINITE HORIZON CASE 291 Proof The assumption y > 1/n'(y) ensures that 1fi(u.1) so that z(u) = zP(u.4.yk(ay) determines the order of magnitude of z'(u. one would expect that the change of measure F Pay is in some sense optimal. we have ic(ay ) > 0 and get Eay Z(u)2 = Eay [e  2aySr( u)+2r(u )r.yy> 2 . We next consider the case y < 1/r. The corresponding estimator is Z(u) = eavS' ( u)+T(u)K (ay)I(T( u) < yu).4.O(u. Further .1).(u) * 1 (Theorem IV. Let Qy2 = . (4.1).1) which is all that is needed here can be showed much easier . yu) is of order of magnitude a71. the result follows as in the proof of Theorem 3. Remark 4 . Proof Since ryy > y.'(7). yu) in the sense that . 3 Theorem IV.log Var(Z(u)) l im of . lim inf uoo 27yu . yu)/z.
Z (u)2 above.yu1/2 <1 T(u) < yu l r > e7vu +avul/ 2r. SIMULATION METHODOLOGY Vara„ (r(u))/u so that (T(u) . In Asmussen [13].(av)Eav l e. yu . In most of the simulation literature (say in queueing applications). yu . Hence lira inf log ryyu + vyu 1/2 tc(ay) . the object of interest is {Vt} rather than {Rt}. the algorithm in Section 3 produces simulation estimates for the tail P(W > u) of the GI/G/1 waiting time W).2) . we believe that there are examples also in risk theory where (5. zi(u) = INV.3.4.1) where the identity for Vi(u) requires that Vt has a limit in distribution V. N(0.Qyu1/2 < T(u) C yu e.1) is used to study Voo by simulating {Rt} (for example.a yu +l/ur' (av)Ei`av reav^(u)+(T(++)(U) yu .3: for many risk processes {Rt }.292 CHAPTER X. Then z(u) = Eay Z(u) > Eay avS'(u)+T( u)k(av 1 ).4. 0 rather than when u 3 oo. 5 Regenerative simulation Our starting point is the duality representations in 11.ryyu +oy u1/2K'(av)Eo l v 1/2) where the last step follows by Stam's lemma (Proposition IV. '%(u) = P I info Rt < 0) = P(VV > u).o.uaoo U That lim sup < follows similarly but easier as when estimating En.yu)/(uyu1/2) .4). there exists a dual process { V t} such that i. related discussion is given in a heavy traffic limit q J. > u) = E f I(VV > u) dt 0 (5.T) = P O<t<T inf Rt < 0 = P(VT > u). One main example is {Vt} being regenerative (see A.2).1) may be useful.u1/2 < r(u) < yu Le ] l = e.1) (see Proposition IV.o ..1): then by Proposition A1.7y x(u) > hm inf u+Oo U . 0 Notes and references The results of the present section are new.a vt(u).b(u. and (5. (5. However.
+Z(N) z 1. Then (Z1z1i Z2z2 ) 4 N2(0.5. oh) for h : R2 ^ R and Ch = VhEVh.. Then Z(1)... and Z2'>) where Zi'i = w. z2)) > N(O. Thus the method provides one answer on to how to avoid to simulate { Rt} for an infinitely long time period. let E denote the 2 x 2 covariance matrix of Z(').. i. i (^(u) .. a standard transformation technique (sometimes called the delta method) yields 1 V 2 (h (Zi. EZ1'i = z1 = Ew. . which we survey below .E). Z2 .t(u)) 4 N(0. Vh = (8h/8z1 8h/ 8z2).. + Z1N>) . the regenerative estimator z%(u) is consistent. 2. provides estimates for F ( V.. record Zi'i = (Z1'). . )). z2) z2/z1 yields Vh = (z2/z2 1/zl).3) . Z2 = N (X21' + . Simulate a zerodelayed version of {V t } until a large number N of cycles have been completed. consider first the case of independent cycles ..h (zl. + Z2N)) . The method of regenerative simulation. (u) ?2 = E fo I(Vt > u) dt = 0( u ) zl Ew as N > oo. Z2 a4* z2. is the cycle length. j = 1. Thus. Taking h(zl. Therefore . Z1 = (Zl1i +.. Zl the LLN yields Z1 a$' Z(1) +. letting J0 'o I (Vt > u) dt . For details .. For the ith cycle. EZ2'i = z2 = E Thus. REGENERATIVE SIMULATION 293 where w is the generic cycle for {Vt}. Z(N) are i . > u) (and more general expectations Eg(V.. 02) (5.d. To derive confidence intervals . Z2'> the time during the cycle where { Vt} exceeds u and zj = EZJ').
g.2 E1 2 z1 z1 Z2 The natural estimator for E is the empirical covariance matrix N S = N 1 12 (ZW . say risk processes with a complicated structure of the point process of claim arrivals and heavy tailed claims . There is potential also for combining with some variance reduction method. Then z(() = f cp(x) f (x. Let X have a density f (x. The regenerative method is not likely to be efficient for large u but rather a brute force one. v. In 111. However . see e. in some situations it may be the only one resolving the infinite horizon problem .5) Z1 Z1 Z1 and the 95% confidence interval is z1 (u) ± 1.C)dx = f w(x) d( f ( x. 9. Notes and references The literature on regenerative simulation is extensive. Here are the ideas of the two main appfoaches in today 's simulation literature: The score function ( SF) method . Rubinstein [310] and Rubinstein & Melamed [311]. asymptotic estimates were derived using the renewal equation for z /i(u). the expectation z = EZ of a single r. () dx f Ax) (dl d()f (x' () f ( z. () dx so that differentiation yields zS d( fco(x)f(x.z^ i=1 so a2 can be estimated by 2 2 = 72 S11+ 12 S22 .g S12 (5. We here consider simulation algorithms which have the potential of applying to substantially more complex situations. to evaluate the sensitivity z/i( (u ) = (d/d() 0(u) where ( is some parameter governing the risk process .9. 6 Sensitivity analysis We return to the problem of 111 . () depending on C. Before going into the complications of ruin probabilities . () dx = E[SZ] f(X. with distribution depending on a parameter (. SIMULATION METHODOLOGY 2 Eli = Z2 z1 + 2 E22 .96s/v"N.v. consider an extremely simple example .0 . Z of the form Z = ^p(X) where X is a r .294 where 01 2 CHAPTER X.Z) ^Z(=) .2.
nonpathological examples where sample path derivatives fail to produce estimators with the correct expectation. SENSITIVITY ANALYSIS where 295 S = (d/d()f (X.1).log U/(. IPA will estimate zS by 0 which is obviously not correct. the Poisson rate /3 in the compound Poisson model. () = d log f (X. Thus. with density f (x. I(r(u) . one. ()) is 0 w . For the SF method. ()) h((U. C)). giving h( (U. p. say W(x) = I(x > xo) and assume that h(U. () Thus.t. Then z(() = Ecp(h(U. () can be generated as h(U. The derivations of these two estimators is heuristic in that both use an interchange of expectation and differentiation that needs to be justified. if f (x. A related difficulty occurs in situations involving the Poisson number Nt of claims: also here the sample path derivative w. cp(h(U. () is increasing in C. ( where h( (u. To see this. = E [`d (h(U. ()) d( hc(U. For IPA there are. ()) is 0 for C < Co and 1 for C > Co so that the sample path derivative cp'(h(U. SZ is an unbiased Monte Carlo estimator of z(. this is usually unproblematic and involves some application of dominated convergence . cp' (h(U.r. Example 6 .1 Consider the sensitivity tka(u) w. The likelihood ratio up to r(u) for two Poisson processes with rates /3. Thus . however .t. /3o is M(u) Oe (3T: < oo) .r. for some Co = (o(U). this phenomenon is particularly unpleasant since indicators occur widely in the CMC estimators . () where U is uniform(0. just take cp as an indicator function . () is an unbiased Monte Carlo estimator of zS. ()). Let M(u) be the number of claims up to the time r(u) of ruin (thus. zc = E [d co(h(U. Then . C) f(X. r(u) = Tl + • • • +TM(u)). So assume that a r.v. For example .() d( is the score function familiar from statistics . C). In the setting of ruin probabilities . () = . () = (8/8()h (u. () _ (eSx.6. one can take h (U. () = log U/(2. /3 is 0. The following example demonstrates how the SF method handles this situation. 11 /3oeOoT. Infinitesimal perturbation analysis (IPA) uses sample path derivatives.
change the measure to FL as when simulating tp(u). There have been much work on resolving the difficulties associated with IPA pointed out above. We then arrive at the estimator ZZ(u) = (M(u) .3 (u) (to generate Zp (u). we get 1 M(u) 00(u) = E (_Ti)I(T(U)<) E [(M(u) . Example 6. BL). whereas for the SF method we refer to Rubinstein & Shapiro [312]. In the setting of ruin probabilities. SIMULATION METHODOLOGY Taking expectation. since ELZp(u)2 < (M(U) _T(u) \ 1 2 a2ryu = O(u2)e27u. ) we have VarL(ZQ(u)) ZO(u)2 O(u2)e2 u2e2ryu yu . for different models and for the sensitivities w.r. the risk process should be simulated with parameters . However. a relevant reference is VazquezAbad [374]. different parameters. 0 Notes and references A survey of IPA and references is given by Glasserman [161] (see also Suri [358] for a tutorial).t. in part for different measures of risk than ruin probabilities. differentiating w.3L.296 CHAPTER X.1 is from Asmussen & Rubinstein [46] who also work out a number of similar sensitivity estimators.T(u)) e7uerVu) for ?P.9 .t. Thus. j3 and letting flo = 0. 4) that V5.3 (u) is of the order of magnitude ue7u. .T(u)) I(T(u) < co) ] . We recall (Proposition 111. the estimation of z(ip(u) is subject to the same problem concerning relative precision as in rare events simulation . To resolve the infinite horizon problem .r.0(1) so that in fact the estimator Zf(u) has bounded relative error.
.d. ... }).. Besides its intrinsic interest . R„ = u+X. are i..1}valued .. and {1. '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 ). X2.Chapter XI Miscellaneous topics 1 The ruin problem for Bernoulli random walk and Brownian motion. Consider first a Bernoulli random walk. a) = r(u))..g. Y'a(U) = P(T (u) = r+(a)) = 1 .(u) = 0 ) = 0) or it is trivial to translate from one setup to the other. defined as Ro = u (with u E {0. Oa(U ) can also be a useful vehicle for computing t/i(u) by letting a * oo. in most cases . with P(Xk = 1) = 9.. a) = r(u) A T+(a). T+(a) = inf It > 0 : Rt > al. That is.P(•r(u. as e.. 297 .1. where X1.(u) is defined as the probability of being ruined (starting from u) before the reserve reaches level a > u. T(u. in the Bernoulli random walk example below.i.. either this makes no difference (P(R. wherel T(u) = inf {t > 0 : Rt < 0} . The twobarrier ruin problem The twobarrier ruin probability 0.+• • •+X.
then 'Oa(u) _ au a We give two proofs . C1_0\a.e.. the solution of F[.o)'t/1a(a . i.a) = 0) + zap ( R.. u Proof 2. By optional stopping. = z°Va(u) + za(1  ...1 For a Bernoulli random walk with 0 0 1/2.2).2) Oa(a . Proof 1.. tba(2) _ (1 .1.1) is solution. where a is any number such that Ee°X = F[a] <oo.o)T/la (1) + 8z/'u(3).o» = z°P (RT ( u.. u + 1.298 CHAPTER XI. and in view of the discrete nature of a Bernoulli random walk we write z = e7. z and the solution is z = (1 .y] = 1.(4.1) o If 0 = 1/ 2.0)/0. and insertion shows that ( 1. The martingale is then {zuzXl+•••+X„ } = {zR° }.1) = (19)4/'0(a3)+9ba(a1). The Lundberg equation becomes 1=F[ry]=(19)+9z. and the other more advanced but applicable also in some other settings.(1B)u oJ 0.+Xn) F[ a]n n=0.. zu = EzRO = EzRT(u. (1.a(u))... one elementary but difficult to generalize to other models. 7/la(a . Conditioning upon X1 yields immediately the recursion 'a(1) = 19+00a(2). In a general random walk setting . We choose a = ry where ry is the Lundberg exponent.a) Y.(u) I\ e = 1 oa ' ()i a = u. MISCELLANEOUS TOPICS Proposition 1. = (1 .r(u.4) by ea(u+Xl+. Wald's exponential martingale is defined as in 11.
{Rt} is itself a martingale and just the same calculation as in the u proof of Proposition 1. BROWNIAN MOTION. } yields e7u = Ee7R° = e°Wa(u) + e7a(1 . Then for p 0 0.u)/u. If p<0. {R. and solving for 9/la(u) yields Z/)a(u) = (e 76 .e7u)/(e7° .1). RANDOM WALK.1) for p # 0.2) is trivial (z = 1). then Vi(u) = 1..1.3 Let {Rt} be Brownian motion starting from u and with drift p and unit variance . thenz1 (u)=1. u Proposition 1. 1h (u) = a el u \1 If 9 < 1/ 2.1 yields 't/la(u) = (a . Corollary 1. Applying optional stopping to the exponential martingale {e7R. (1.ba(u) = e2µa .• a2µa e2µu .0a(u)). TWO BARRIERS 299 and solving for 4/la(u) yields t/ia(u) = (za . i1(u) = e211 .2 For a Bernoulli random walk with 9 > 1/2. . then Proof Since 'Oa (U)  au a Eea(R°.1). If 9 = 1/2.4 For a Brownian motion with drift u > 0. Proof Let a+ oo in (1.} is then itself a martingale and we get in a similar manner u = ER° = ER ra( u) = 0 • Y'a (u) + all  au Y'a( u)).1 If p = 0.u) = et(a2 /2 +aµ) the Lundberg equation is rye/2'yp = 0 with solution y = 2p. (1. pa( u) _ u Corollary 1. If p = 0.5) . However.zu)/(za .
0 (u) (where u p =. It may then be easier to first compute the onebarrier ruin probability O(u): Proposition 1.a) < 0) + e7°P (R(u.a) = a ) + e ' ° ( 1 .0(a) 0 < u < a.300 Proof Let a * oo in (1. passing to even more general cases the method quickly becomes unfeasible (see. letting a * oo yields the standard expression pe7u for .3.a) = r+ (a)} and similarly for the boundary 0. (1.a ) < 0) + e 7aF ( R (u. VIII. 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) = 5 y = P (R (u.+^a(u))^(a) If 7k(a) < 1. say. 5).5a). 7O(u) = 7/la(u) + (1 . the paths are upwards skipfree but not downwards.vi(a) Proof By the upwards skipfree property.7/la(u)). we obtain 'Oa a7u . Ic 5ry 'pa(u) Using y = 6 . and thus one encounters the problem of controlling the undershoot under level 0. 0. Here the undershoot under 0 is exponential with rate 5. this immediately yields (1.e7a Again . However. Here is one more case where this is feasible: Example 1..7) . For most standard risk processes . a) I R(u a ) < 0] P (R(u .616).e7a (u) = 6 /0 . (u) _ O(u) . and hence e7u = Ee7Ro E [e7R(.5 Consider the compound Poisson model with exponential claims (with rate.7). implying R(u. CHAPTER XI. . 7/'(u) = 1).6 If the paths of {Rt} are upwards skipfree and 7//(a) < 1.a) = a on {r (u.4). however. valid if p < 1 (otherwise .
T ) = P(T(u) < T ) = 241. TWO BARRIERS 301 Note thas this argument has already been used in VII.1. (1. and (1 . Then the density and c. ( 1.10) Pµ (T(u) < T) !. T(u) E dT.ST<u) = P(MT>u.d. For the symmetric (drift 0) case these are easily computable by means of the reflection principle: Proposition 1. Hence P(MT>u. and hence Pµ('r(u) E dT) = Eo [e µsr(. of r(u) are ( U2 Pµ (T(u ) E dT) = 2^T 3/2 exp µu . = eµuTµ2/2Po (T( u) E dT) 2 eµuTµ2/2 u T3/2 ex p u 27r p 12 T . MT > U) = P(ST > u) + P(ST > u) (1.11 ) is the same as (1. BROWNIAN MOTION.. For µ # 0. We now return to Bernoulli random walk and Brownian motion to consider finite horizon ruin probabilities.4) I = . 0(u. P(MT > u) = P(ST > u) + P(ST < u. Here {St } is Brownian motion with drift 0 (starting from 0).2 .8 ). = 1 . (1.9) = 2P(ST > u). we have ili(u.1a for computing ruin probabilities for a twostep premium function. MT > u) = P (ST > u) + P (ST > u. the density dPµ / dP0 of St is eµsttµ2/2. 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 ..11) VIT ) Proof For p = 0.µ%T (1. + µ2T) } .8) Proof In terms of the claim surplus process { St} = {u . RANDOM WALK.ST>U).f.µ T I + e2µ"4) ( .Rt}.7 For Brownian motion with drift 0. 10) follows then by straightforward differentiation.r(u)..)_ _( u)µ2 /2.8 Let {Rt} be Brownian motion with drift .µ so that {St} is Brownian motion with drift µ ..T) P(MT > u) where MT = maxo<t<T St.. Corollary 1. (i).
Here {2T( (v}TT)/2) v=T.10) and that the value at 0 is 0. 0 0 (1.g.13) with 0 as lower limit of integration. The same argument as used for Corollary 1. S(oo) = f c s(y)dy.10). is finite for all x > 0.12) is the same as ( 1.12) P(ST = v) = 0 otherwise.9 For Bernoulli random walk with 9 = 1/2.T)dx.3 we can define the local adjustment coefficient y(x) as the one 2µ(x)/a2(x) for the locally approximating Brownian motion.302 CHAPTER XI.. oo). The expression for F ( ST = v) is just a standard formula for the u binomial distribution. as defined in (1. Vi(u.h.g. We assume that u(x) and a2 (x) are continuous with a2 (x) > 0 for x > 0. Theorem 1.s. and (1. We finally consider a general diffusion {Rt} on [0. If this assumption fails. is (1. Proof The argument leading to ( 1. as defined above as the probability of actually hitting 0. close to x {Rt} behaves as Brownian motion with drift µ = u(x) and variance a2 = a2(x).13) The following results gives a complete solution of the ruin problem for the diffusion subject to the assumption that S(x). (1.. the behaviour at the boundary 0 is more complicated and it may happen. that 0(u). is zero for all u > 0 but that nevertheless Rt ^4 0 (the problem leads into the complicated area of boundary classification of diffusions. whenever u.T+2. and in a similar spirit as in VII.8 also applies to the case 9 54 1/2. Breiman [78] or Karlin & Taylor [222] p. e.. MISCELLANEOUS TOPICS which is the same as (1.T (1. but we omit the details. such that the drift µ(x) and the variance a2(x) are continuous functions of x and that a2(x) > 0 . Thus..T2.11) then follows by checking that the derivative of the r. u Small modifications also apply to Bernoulli random walks: Proposition 1. Let s(y) = ef0 ry(.10 Consider a diffusion process {Rt} on [0.T) = P(ST = u) + 2P (ST > u). oo) with drift µ(x) and variance a2 (x) at x.9) goes through unchanged. see e. S(x) = f x s(y)dy. 226). T are integervalued and nonnegative.9).
b(u)dt. 0 in (1. If (1.1.14) fails. where Lq(u) = 0'22u) q "(u) + p(u)q(u) is the differential operator associated with the diffusion. see in particular pp. A classical reference for further aspects of Bernoulli random walks is Feller [142].b(Rdt).b = 0 implies that VQ b/s is constant. E„ q(Rdt) = q(u)+Lq(u)dt. In view of (1.(u) < 1 for all u > 0 and ^ S^ Conversely. TWO BARRIERS 303 for x > 0. so that Y)n.b(b) = 1. Letting b J.e. b = 0.S(u)/S(a). i.S(b) Proof Recall that under mild conditions on q.10.b('u) = Eu &0. 1'. (1. Then YIa. Further references on twobarrier ruin problems include Dickson & Gray [116]. then 0 < 2l.16) yields 4b (u) = 1 . 191195 for material related to Theorem 1. O.0(u) = 1 for all u > 0. The obvious boundary conditions '0a. For generalizations of Proposition 1. Using s'/ s = 2p/a2.b (Rdt) = Oa. BROWNIAN MOTION.S(u) (1.16) S(a) . (1 . If b < u < a.14) S(oo) < 00. 0 Proof of Theorem 1.11 Let 0 < b < u < a and let t&0.10. if (1. see Asmussen & Perry [42].6 to Markovmodulated models . and we get Wo.b(u) = S(a) . we can ignore the possibility of ruin or hitting the upper barrier a before dt.e LVa. .17) Hence L.13) is finite for all x > 0. Lemma 1.b(u) + L. [117]. Wa. Notes and references All material of the present section is standard. then.ba.16).b(u) be the probability that {Rt} hits b before a starting from u. Letting a T oo and considering the cases S(oo) = oo.ba. 15) i. the function S(x) is .b('u) = Eu . Assume further that S (x) as defined in (1.b = a+/3S. RANDOM WALK. S(oo) < oo separately u completes the proof. elementary calculus shows that we can rewrite L as Lq(u) d 1a2 (u)s(u)d [ s (u) ? ] . A good introduction to diffusions is in Karlin & Taylor [222].b(a) = 0 then yield the result.
yu) where W (ay) = y. See Asmussen [20] and Rogers [305] for some recent treatments and references to the vast literature. MISCELLANEOUS TOPICS referred to as the natural scale in the general theory of diffusions (in case of integrability problems at 0. (2.(7) .4. y > . and here are alternative martingale proofs of the rest . one works instead with a lower limit 5 > 0 of integration in (1. 7y = ay . IV.1. yielding eau = Ee. (2. Another basic quantity is the speed measure M . Lo is a martingale (cf. much of the literature dels with the pure drift case. information on ruin probabilities can be obtained .aRo .)AT .5) A martingale proof of (2.o•K(a) = Ee . 1 y < k (y). variance 0.9 ) and optional stopping applied to the stopping time r(u) A T..13)).4. which is motivated from the study of modern ATM (asynchronous transfer mode ) technology in telecommunications. is currently an extremely active area of research.5.1 ) was given already in II.aR.(a) (2.3) < e 7yu. equivalently. correponding to piecewise linear paths or .(T(u)AT) r. (2. 2 Further applications of martingales Consider the compound Poisson model with adjustment coefficient ry and the following versions of Lundberg 's inequality (see Theorems 111.304 CHAPTER XI. where C_ = B(x) _ B(x) sup 2no fy° e7(Y )B(dy)' f2e7(Y2)B(dy)' C+ i/i(u.5): _ z/'(u) < e 7u.4) I.3.6) . defined by the density 1/va(u)s(u) showing up in (1. Lo I. with the drift and the variance depending on an underlying Markov process . yu) '+/1(u) .1) (2.ytc (ay). (2.17).2. Markovmodulated Brownian models .(.6. 111 . They all use the fact that ( tx(a) l ( eaRt = eau + aSttx(a) < e7yu.t&(u. Remark 11. but by duality.2) C_e7u < t(u) < C+e _7u. The emphasis is often on stationary distributions .
Rt has distribution B(r + dy)/B(r).)r(u)r.6). Proof of ( 2. For (2. it follows easily from (2..T(u)K(ay) I yu < r(u) < T] F(yu < r(u) < T) > e.d.1. A claim leading to ruin at time t has c. dr) denote the conditional distribution of (T(u).1. .7R. dr JO Zoo ) f e7'B(r + dy) B(r) Jo ^00 ^00 H(dt. (2.1 . y > r. Equivalently.3).(ay)I T(u) < yu] P(r(u) < yu) (using RT(u) < 0).4): We take a = ay in (2. dr) e 7( yr)B(dy) B(r) f oo o 0 r > H(dt. and the proof of the lower inequality is similar.B(r))/B(r). Let H(dt.(u.yuk (ay)(u&(u.( u ) I T(U) < 00] . dr) 1 = 1 I0 /o C+ C+ From this the upper inequality follows. RT(u)_) given r(u) < oo. we have tc(ay) > 0 and we can bound (2.6) with = 'y that eyu . so that i/1(uL yu) < eayu . yu))• Letting T + oo yield e_ayu > eyur4ay)(0(u)  Notes and references See II. (B(y) . eyuk (ay) = e7yu e > eyu"(ay ) ij(u.2. when Rt_ = r.3). Hence E [e7Rr (u) Jr(u) < ool ^00 H( dt.yu))• b(u.6) below by 1 E Le7Rr(. u Proof of (2.E [e. we have ic(ay ) < 0 and use the lower bound E [e7Rr („).yu) Y Similarly for (2.4).2): As noted in Proposition II.f.T)  V. 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).
if x > EX1.(B) = log EeOX 1 is defined for sufficiently many 0. however .1) where we return to the values of 0. the parameter will be u rather than n). Example 3.3na with a < 1. Cramer considered a random walk Sn = X1 + .1) is an example of sharp asymptotics : . not quite so much in insurance risk...1 We will go into some more detail concerning (3. 1) but only the dominant exponential term . (3. we will write fn 1. logarithmic asymptotics . MISCELLANEOUS TOPICS 3 Large deviations The area of large deviations is a set of asymptotic results on rare event probabilities and a set of methods to derive such results. ri. which in the setting of (3. large deviations results been. gn 4 0. Thus.1) amounts to the weaker statement lim 1 log P I Sn > x I = 17. nroo n n /// Note in particular that (3. its generality.?n typically only give the dominant term in an asymptotic expression . + X.306 CHAPTER XI. Accordingly.gn if nioo lim 109 fn = 1 log gn (later in this section. og For sequences fn.^ e nn 1 > x n 0o 2xn (3.2). then P C S.. and that a considerable body of theory has been developed.1) does not capture the \ in (3. However . . The classical result in the area is Cramer's theorem. v2 later. gn with fn + 0 . For example.means (as at other places in the book) that the ratio is one in the limit (here n * oo). large deviations results have usually a weaker form./n E I) for intervals I C R. e.1). 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 .the correct sharp asymptotics might as well have +. The advantage of the large deviations approach is. and gave sharp asymptotics for probabilities of the form P (S..g. cle . . such that the cumulant generating function r.2) can be rewritten as F (Sn/n > x) 1g a'fin. Thus . The limit result (3. logarithmic asymptotics is usually much easier to derive than sharp asymptotics but also less informative . (3.nn or C2e.
rc*(x) = sup(Ox .1). Since P nn > x) = E {e_8 ' ( 9). the LegendreFenchel transform or just the Legendre transform or the large deviations rate function).960/) * 0.r.r.rc(0) where 0 = 0(x) is the solution of x = rc'(0). More precisely.q = rc* (x).2). and hence for large n P(Sn/n > x) > E [e.the mean rc'(0) of the distribution of X1 exponentially tilted with 0.96o /] > 0.e.9S„+n' ( 9). P with mean nx and variance no.t. 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 . (3.3) is put equal to x. i. nx < Sn < nx + 1.4) immediately yields (3.(e)i XI E dx]. if we replace Sn by nx + o / V where V is N(0.sseo f which in conjunction with (3.4 enn +1. S rtn > x 1. Define .425. which is a saddlepoint equation . of P(X1 E dx) = E[e9X1K.3. In fact. we have P(nx < Sn < nx + 1. V > 0 e.(0)) e 307 (other names are the entropy. the sup in the definition of rc* can be evaluated by differentiation: rc*(x) = Ox . exponential change of measure is a key tool in large deviations methods. since Sn is asymptotically normal w.4) n Next. Most often. 2 where o2 = o2(x) = rc"(0). replacing Sn in the exponent and ignoring the indicator yields the Chernoff bound P Sn > x 1 < e°n (3. we get P(Sn/n > x) E [e9nx +nK(9)9" '.
We further write µ = tc'(ry). In the application of large deviations to ruin probabilities. Xn given by Fn(dxl.. be a sequence of r... n Icn(0) exists and is finite for ry . to be made rigorous... Assume that there exists 'y. e > 0 such that (i) Kn (0) = log Ee°Sn is welldefined and finite for 'y .. see Jensen u [215] or [APQ] p. however. The substitution by V needs. which is a version of Cramer's theorem where independence is weakened to the existence of c(O) = limn.h.dxn) where Fn is the distribution of (X1i . that is.. 260 for details. Further main results in large deviations theory are the GartnerEllis theorem. and the WentzellFreidlin theory of slow Markov walks.dxn) = 05nKn(7)Fn(dx1.. 1) and no such that Sn . integrates to 1 by the definition of Icn). is differentiable at ry with 0 < K'(y) < 00.. 1]. which is of similar spirit as the dicussion in VII. Xn) and sn = x1 + • • • + xn (note that the r. .308 CHAPTER XI. ..3 For each i > 0. we introduce a change of measure for X1. and write Sn = X1 + • • • + Xn.'s..1). We shall need: Lemma 3 . Then i/..e < 8 < y + e. Ee9X n < oo for e < 0 < e. we shall concentrate on a result which give asymptotics under conditions similar to the GartnerEllis theorem: Theorem 3 ..2 (GLYNN & WHITT [163]) Let X1..p > 7 < zn. commonly denoted as is the saddlepoint approximation.s. For the proof.e < 8 < y + e. . Mogulskii's theorem which gives path asymptotics. r(u) = inf {n : Sn > u} and o(u) = P('r(u) < oo).. (iii) #c (8) = limn. (ii) lim supn.. . (iv) tc(ry) = 0 and r. Pn Sn1 .3.o log Ee9Sn /n./^ >7 < zn n for n n0. X2. Sanov's theorem which give rare events asymptotics for empirical distributions. there exists z E (0.'(u) )Ng a"u.v. asymptotics for probabilities of the form P ({S[nti/n}o<t<l E r) for a suitable set r of functions on [0. .. MISCELLANEOUS TOPICS which is the same as (3.
the r . it is easy to see that the r. Since I EeqOX „ ] 1/q is bounded for large n by (ii)...µ?7 .s. h.n > u ) = [ Em [em Em 1e.+r. P n(Sn/n > {c+77) < e no(µ 309 +n)Enees n +n)elcn(B +7). This proves the existence of z < 1 and no such that Pn (Sn/n > µ.Bµ .y) .s. The corresponding claim for Pn(Sn/n < µ . S.. 0.ne(µ limsup 1 log Pn (Sn/n > µ + 17) < ic(9 + ry) . u Proof of Theorem 3. LARGE DEVIATIONS Proof Let 0 < 9 < e where a is as in Theorem 3. > 1 +17] m(7).]1/q = e.077 n^oo n and by Taylor expansion and (iv ).2.77) follows by symmetry (note that the argument did not use µ > 0). in particular the r.W. Then V.ne(p+ 17). Let r7 > 0 be given and let m = m(77) = [u(1 + 77)/µ] + 1. log zl'(u)/u > 'y.r (7) n = e. ( U) P(S. we get lim sup 1 log Pn (Sn1 /n > µ + r7) < 0(1i + r7) + i(p(0 +'Y))/p n+oo n and by Taylor expansion. The rest of the argument is as before.3.2.71 < e and jq9j < e.m(7).+r7) < zn for n > no. 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. S. We first show that lim inf„_. This establishes the first claim of the lemma .91) + o(O ) as 0 J. mµ Sm > u] km e7Sm+n. for Sn. is of order .YS.s..> .h. Clearly.Kn(7)e'n (p(O +7))/p I Ee geXn]1/q where we used Holder's inequality with 1/p+ 1/q = 1 and p chosen so close to 1 and 0 so close to 0 that j p(0 +.n e(µ +o)w"(7) [Eep(B +7)Sn]1 /p [EegoX. h.. can be chosen strictly negative by taking p close enough to 1 and 0 close enough to 0. can be chosen strictly negative by taking 9 small enough.n m µ 1 + rl .
n=1 n=n(b)+1 00 Lu(1 +6) /µJ 13 F( T (u) = P(T(u) n). MISCELLANEOUS TOPICS (7).n Yµ 1 + m + r ('Y) } U n \ 77 m µ µ7 1 < 1+ 77 ) Here E. For lim supu.7) so that n(b) I1 < e'Yu E en. n=1 .(Y). we get lum inf z/i(u) 1 +12r7 >_ ry + 77 Letting r7 J.6) for some z < 1 and all n > n(E).log z) /2 and Sn Fn\ n >lb+S) <Zn. this is possible by (iii). logO(u)/u > ry...(•) goes to 1 by Lemma 3.3. Pn \ > la+ 8 I < zn (3.. (iv) and Lemma 3.I < µl1 1+77 I M 1_ 1+277 S. we write P(T(u) = n) = Il + I2 + I3 + I4 'i/I(u) _ E00 n=1 where n(b) Lu(10/µJ Ii = 1: F(T(u) = n). I2 = F(T(u) = n).n(ry)/u 4 0andm/u* (1 + r7)/µ..310 ]Em I e. and since Ic.+wn(7). Sn > u] < eYu+Kn(7)pn(Sn > u) (3.. I > IL exp `S...YS +^c CHAPTER XI. 14 = = E Lu(16)/aJ+1 Lu(1+6)/µJ+l = n) and n(S) is chosen such that icn('y )/n < 6 A (. 0 yields liminfu __. 3. P(T(u) = n) < P(Sn > u) = En [e7S. Obviously.0 log i'(u )/u < 'y.
3. LARGE DEVIATIONS Lu(16)/µJ 311 I2 < e"u n=n(6)+1 e'n(Y)P(Sn > u) < Lu(16)/µJ ^. we get lim sup log u/00 O (U) < y + b(1 + b) U Letbl0.10) 00 I4 < E F(Sn_1 < u. 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 . Sn1 C U. S. 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 . u .11) [u(1+6)/µJ+1 1  Thus an upper bound for z/'(u) is n(6) e'Yu n=1 eKn (7) + 2 + (28U + 1) e6u(1+6)/µ Fi 1 zl /2 and using (i).zl/z en6 [u(1 +6)/µJ 1u (1 +6) /µJ ekn(7) < e' 13 < C" E Yu l u(16)/lij+1 Lu(16)/µJ+l1 < e7U Finally. eryu en logz/2p n nt n. C 26u `p / +1 I e6u(1+6)/µ (3.' 1 + b) n e7u x 1 /2 1 n x n / 2x (3. > u) Lu(1+6) /µJ +l 00 )^n 'YSn+kn (7) .
312 CHAPTER XI.b)/i(7).3ui where .(u) = I1+I2+I3+I4'^ ery( u).7' a"ju. Corollary 3. we need to redefine n(b) as L. it holds for each b > 0 that 0(u) 1' g F(T(u) E (u(1 . MISCELLANEOUS TOPICS The following corollary shows that given that ruin occurs.2.z 1/z For I1. Letting c11 = maxn<n.4. 4 there is an aj > 0 and a cj < oo such that Ij < c3e. this is straightforward since the last inequality in (3.4 Under the assumptions of Theorem 3. the typical time is u/rc'(7) just as for the compound Poisson model. the last steps of (3. 2. (7 + a) < 2arc'(7). > u) < e"' E eIsn = ectueKn (a+'Y)Kn(7) where 0 < a < e and a is so small that r. For 12.('+'Y). say n n1.. 13 = P(T (u) E (u(1 b)l^ (7).4/3rc'(y) > 0.9) can then be sharpened to x LQuJ /2 I2 < e7u 1 . Then for n large. IV.u(1+b)/rc'(7)).8) by P(S. we have rcn (a + 7) < 2n^c(7 + a) < 4narc' (7)..xl/2 to give the desired conclusion. I2. e'. For I. For 14. u(1 + b)/i(7)) Proof Since V. 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.11 ) can be sharpened to x 4 [u(1+6)/µJ /2 1 . we replace the bound P(Sn > u ) < 1 used in (3. cf. it suffices to show that for j = 1. u . ryue«iu .Q is so small that w = 1 .
Assuming that the further regularity conditions can be verified.2 then immediately yields the estimate log F( sup Skh > u) a7u (3.. Theorem 3.12) k=0. the key condition similar to (iii). 11 Inspection of the proof of Theorem 3. The reader not satisfied by this gap in the argument can easily construct a discrete time version of the models! The following formula (3..LARGE DEVIATIONS 313 Example 3 ..3.f. but nevertheless. 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.1.13) One would expect this to hold in considerable generality. whether P ( sup St > u ltg a ^" 0<t<oo // (3. Thus the total reward in the interval [0. An event occuring at time s is rewarded by a r. 09(9).g..5 Assume the Xn form a stationary Gaussian sequence with mean p < 0. i. The problem is whether this is also the correct logarithmic asymptotics for the (larger) ruin probability O(u) of the whole process. and we conclude that Theorem 3 . for the ruin probability z/'h(u) of any discrete skeleton {Skh}k=0. Obviously many of the most interesting examples have a continuous time scale.2 shows that the discrete time structure is used in an essential way. t] is Rt = E V (Un) n: o„ <t . criteria are given in Duffield & O'Connell [124]. If {St}t> 0 is the claims surplus process. r.14) is needed in both examples . Hence z z\ 2 z nrn(9) _ n Cn0p+BZn/ * . Xk+l) k=1 00 naoo n provided the sum converges absolutely. we shall give two continuous time examples and tacitly assume that this can be done. 2 is in force with y = 2p/wz.(O) = 9µ+02 for all 9 E R..e.3(s) at time s. (iv) becomes existence of a limit tc(9) of tct(9) _ log Ee8S° It and a y > 0 with a(y) = 0.v...'(y) > 0. and in fact. To verify these in concrete examples may well present considerable difficulties. Let {Nt}t>0 be a possibly inhomogeneous Poisson process with arrival rate . V(s) with m..1.
then the payments from the company in [on.. we have S.g. Of course. MISCELLANEOUS TOPICS are the event times. the Cramer. derive .6 We assume that claims arrive according to a homogeneous Poisson process with intensity 0 . where Ft = a(A8 : 0 < s < t).1) ds .14) (to see this . is At . a differential equation in t). the best estimator of /3µB based upon Ft. one would take p(t) = (1 + rt)At/ t. Then logEeOR° = J0 /3(s)(^8(9) .noise model is the same as the one for the Cramer Lundberg model where a claim is immediately settled by the amount Un.1) ds rt (3. leading to St = At(1+77) Joo t S8 ds. More precisely. this is not realistic . e > 0 such that ic('y) = 0 and that r. 0 and since EeOUn(8) + Ee°U^ as s * oo. assuming a continuous premium inflow at unit rate.9t = /3 J t (Ee8U° i8l .t.14). Un(s). O'n +S] is a r . We further assume that the processes {U1(s)}8>0 are i. we have rct (9)/t 4 ic (9). if the nth claim arrives at time a. we conclude that Cu) log e7 u (cf.1) ds ..Lundberg model has the larger ruin probability. = U„ ( t .'`1 U. n: o. 0 Example 3 .9t. . Un represents the total payment for the nth claim). i. (3. Thus by (3. the CramerLundberg model implicitly assumes that the Poisson intensity /3 and the claim size distribution B (or at least its mean µB) are known. but that a claim is not settled immediately. <t which is a shotnoise process. (9) < oo for 9 < 'y + C. Example 3. At = . nondecreasing and with finite limits Un as s T oo ( thus. Kt (0) t (Ee9U"it8i J0 . Thus.v. it contributes to St by the amount Un(t .It.d.Q„) . An apparent solution to this problem is to calculate the premium rate p = p(t) at time t based upon claims statistics . If the nth claim arrives at time Qn = s.15) . e.s). Thus. 7 Given the safety loading 77. Most obviously.2 are trivial to verify. It is interesting and intuitively reasonable to note that the adjustment coefficient ry for the shot . Of course .1) ..0 and assume there are y. Since the remaining conditions of Theorem 3. We let ic (9) = 3(EeWU° . the above discussion of discrete skeletons)..314 where the an CHAPTER XI.
the solution of /3(Eelu . standard exponential .19) with equality if and only if U is degenerate.b(u) IN a'Yu (cf./3.1) . It then follows from (3.(1 + r7) log t (3. Indeed.d. equivalently. 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 . To see this . and since the remaining conditions are trivial to verify.(1 + 17)0µB = 0. (3. Thus. Ui Nt / t 01i 315 St = Ui .17) K(a) f o 1 O (a[I + (1 + 77) log u]) du )3.20) (3.(1 +i) f > i= 1 s ds = E Ui 1 .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 . typically the adaptive premium rule leads to a ruin probability which is asymptotically smaller than for the CramerLundberg model . LARGE DEVIATIONS With the Qi the arrival times.d.i. the Vi = . 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. (3.2 hold.3. we have Nt t N.1) or . one has y > y' (3. uniform (0.log Oi are i.16) i=1 o i=1 Let ict (a) = log Eeast .e. rewrite first rc as te(a) _ /3E 1 1 +(1+77)aUJ eau 1 .i.14) that rt _ 13 Jo _ (a [1_( i+77)log]) ds_flt = t (a) (3. we conclude that t. i.18) Thus (iii) of Theorem 3.
a* (s) are convex with tc'(0) < 0 . k'(0) < 0. For Example 3. we are interested in estimating P(A > x) for large x. say one year. though we do not always spell this out.19).1 .20) is due to Tatyana Turova. assuming that the U. and since tc(s). the function k(x) = e7*x .2. see also Nyrhinen [275] for Theorem 3.2 expressing the finite horizon ruin probabilities in terms of the distribution of A. rc*' (0 ) < 0. In addition to Glynn & Whitt [163]. Therefore e7'U _ k(U) E [1+(1+77)y*U] .1 E [1+(1+77)y*U] 0 k (+ *y B(+ 1 + (1(+71)y*y B(dy) L xa 1 + f + (1 + rl) Y* xo jJxo k(y) B(dy ) + f' k(y) B(dy) } = 0. MartinL6f [256]. [245]. 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. and k(x) < 0. the proof of (3. so there exists a unique zero xo = xo(r7) > 0 such that k(x) > 0. this in turn yields y > y*. y = y* can only occur if U .xo. 11 Notes and references Some standard textbooks on large deviations are Bucklew [81]. . 0 < x < x0. 4 The distribution of the aggregate claims We study the distribution of the aggregate claims A = ^N' U. Dembo & Zeitouni [105] and Shwartz & Weiss [339]. much of the analysis carries over to more general cases.7. x > x0. we then take t = 1 so that p. In particular. For notational simplicity. the study is motivated from the formulas in IV. The main example is Nt being Poisson with rate fit. with common distribution B and independent of Nt. are i. Further applications of large deviations idea in risk theory occur in Djehiche [122]. Lehtonen & Nyrhinen [244]. at time t. This implies n(y*) < 0.d. using that Ek(U) = 0 because of (3. = P(N = n) = e(3an However. [257] and Nyrhinen [275]. MISCELLANEOUS TOPICS Next. Further..316 CHAPTER XI.i. k(0) = 0.(1 + ri)y*x is convex with k(oo) = 00. see Nyrhinen [275] and Asmussen [25]. Further.
The exponential family generated by A is given by Pe(A E dx) = E [eeA K(9). For a given x. The analysis largely follows Example 3. A E dx] . A > x) eex+K(e ) ee AB°[ely 1 ev2/2 dy 0 2^ 00 9x+p(e) e ezez2/(2BZpB „[9)) dz 9 27r/3B" [9] fo eex+w ( e) oo z x)] ] 0 27r /3B" [9] o e 9 2 /3B" [9] J eex+w(B) dz .9(Ax). Then as x * oo.4.1). e9x+K(°) P(A > x) B 2ir /3 B" [9] Proof Since EBA = x."(0) = . no(a) = logE9e'A = rc(a + 9) . K'(0) _ ic'(9) = x.[s])3/2 = 0. This shows that the Pedistribution of A has a similar compound Poisson form as the Fdistribution. In particular. (4.3e(bo[a] .ic(9) = . 818' where s' = sup{s : B[s] < oo}.x)//3B"[9] is standard normal. Then Ee"A = e'(") where x(a) _ 0(B[a] .1 Assume that lim8T8.3B[9] and Be is the distribution given by eox B9(dx) = B [9] B(dx).1). Proposition 4.2) implies that the limiting Pedistribution of (A . Hence P(A > x) = E e [e9A+ ic(9). i. B"' [s] lim (B". we define the saddlepoint 9 = 9(x) by EBA = x.1) where )30 = .1. B"[s] = oo. A > x)] = eex+K( e)E9 [e .e.3B"[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. Vare(A) = s.
For a rigorous proof.318 CHAPTER XI. under the Poisson assumption (4.x') where x' = sup {x : b(x) > 0}.1). For example. b is logconcave.v. 2).1 goes all the way back to Esscher [141]. For details. or.2i and that (A . Var(A) _ ^3p.3) The result to be surveyed below improve upon this and related approximations by taking into account second order terms from the Edgeworth expansion. i. A covers the exponential distribution and phasetype distributions.(3µB)/(0µB^))1/2 has a limiting standard normal distribution as Q ^ oo. Remark 4 .l3pB.(D X . Y satisfies 9(u) ti eu2/2(1 + ibu3) (4. The present proof is somewhat heuristical in the CLT steps. large x. [138]. For example. 4b The NP approximation In many cases . more generally. Thus . B covers distributions with finite support or with a density not too far from ax° with a > 1. Jensen [215] and references therein. where q(x) is bounded away from 0 and oo and h (x) is convex on an interval of the form [xo. In fact. see Embrechts et al. some regularity of the density b(x) of B is required. Notes and references Proposition 4.ycix °ie6x B.e.3) and related results u for the case of main interest . either of the following is sufficient: A.2) is often referred to as the Esscher approximation. it is quite questionable to use (4. MISCELLANEOUS TOPICS It should be noted that the heavytailed asymptotics is much more straightforward. it holds that EA = . 1 . b(x) = q(x)eh(z). The (first order) Edgeworth expansion states that if the characteristic function g(u) = Ee"`}' of a r. In particular. 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 yields: Proposition 4. Furthermore 00 b(x)Sdx < oo for some ( E (1. b is gammalike.EN B(x). bounded with b(x) . leading to P(A > x) :.Q{AB (4.2 If B is subexponential and EzN < oo for some z > 1. then P(A > x) . and (4. the distribution of A is approximately normal . just the same dominated convergence argument as in the proof of Theorem 2.4) .
as u2 u3 u4 9(u) = Ee'uY = exp {iuci . one needs to show that 163. K2 = Var (Y).2K3 + 4i 64 + . Let Y = (A .5 (y3 . which is often denoted VaR (the Value at Risk).y2)^P(y)• 319 Note as a further warning that the r..f. A particular case is a.6(1 . resp. the standard normal distribution. and so as a first approximation we obtain a1_E = EA + yle Var(A) . the NP (normal power) approximation deals with the quantile al_E.. are small. of (4. one expects the u3 term to dominate the terms of order u4.c2i.99. the CLT for Y = Y6 is usually derived via expanding the ch. . and from this (4. THE DISTRIBUTION OF THE AGGREGATE CLAIMS where b is a small parameter.l = EY. s.3!).. K3 = E(Y . are the cumulants . Heuristically.. ylE should be close to zl_E (cf.2 2 . If this holds . Rather than with the tail probabilities F(A > x). u5.i 3 K3 } Pt^ exp . defined as the the solution of P(A < yle) = 1 .5) follows by integration.3& (y). . f °o 9(y) = 1 e'uye u2/2(1 + iSu3) du 27r _ cc(y) .s. the density of Y is 1 °° _ eiuy f(u) du 2x _.equantile in the distribution of Y. Thus if EY = 0. (4. If the distribution of Y is close to N(0.6) .1).: EA + zl_E Var(A) .4..EA)/ Var(A) and let yl_E. Var(Y) = 1 as above . .2 ^ \1 . Remark 4. in particular. so that 1(u) 3 exp { .h...5). In concrete examples .e. zl_e be the 1 .5) is obtained by noting that by Fourier inversion.EY)3. where Kl . then P(Y < y) 4(y) . K4 .i 6 r 1 3 so that we should take b = ic3/6 in (4.2X2 .. (4. however.5) may be negative and is not necessarily an increasing function of y for jyj large..
E(/3PB^1 )1^2 + s(z1E .1). For example.1) E (A .1)! n ^eQ .6 (1 .k = /3µB^1 / (.zl E)V(zl_E) . however. Note. that [101] distinguishes between the NP and Edgeworth approximations. K5 . k3 is small for large /3 but dominates 1c4.zlE )w(zl _E) = which combined with S = EY3/6 leads to q^ 1 Y1 .....5(1 . this holds with a = 0.(y) terms dominate the S(1 .. Using Y = (A .y2)cp( y) term.6pBki) d/2.S(1 .E)A1 l E) 1 E 4)(yl E) ^' .. as required .E + (yl.E = z1E + S(zi_E . MISCELLANEOUS TOPICS A correction term may be computed from (4.1)^ 2) µ'E Notes and references We have followed largely Sundt [354]. this yields the NP approximation 6(Z1 _E . 21 . We can rewrite (4..3ni /3 . n = 1. let pn Pn = (a+ = P(N = n). Another main reference is Daykin et at.7) as 1 (3) a1E = Qµa +z1 . This leads to t( yl E) .zl E)^o(zl E) .zlE)W(zlE) 1 .yi.320 CHAPTER XI. [101].5) by noting that the 4. b = /3 for the Poisson distribution with rate /3 since Pn = Pn1 n! n (n . In particular . 4c Panjer 's recursion Consider A = constants a. b such that EN 1 U%.zi.1)EY3.E .EA ) / Var(A).EA)3 a1_E = EA + z1_E(Var (A))1/2 + 1 Var(A) Under the Poisson assumption (4.E )Azl E) 4(z1E) + ( ylE . . and assume that there exist n ) Pn_i .S(1 . the kth cumulant of A is /3PBk' and so s.
n = k=n1 9k(n1 )9j k • (4.10) f o = po. . (4. The expression for fo is obvious. 2. the value of (4.. n. u Proof of Proposition 4.k . (4. 1. E[a +bU=I >Ui =j l i=1 J (4. then j (a + b!) 1ag k_1 3 gkfj.12). and calculating the gj*n recursively by 9*1 = 9j.. . . j = 1.4 is that the algorithm is much faster than the naive method.4 Assume that B is concentrated on {0.4.5 The crux of Proposition 4. .. . (4.14) is therefore a + b/n.4. . 2... the complexity (number of arithmetic operations required) is O(j3) for (4. Then fo = >20 9onpn and fi = 1 E In particular. fj = P(A = j).13) Namely.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 .} and write gj = 2 ..1.13) but only O(j2) for Proposition 4. Since the sum over i is na + b.14) is independent of i = 1.4.. j = 0.. .. . Hence by (4. .11) Remark 4. j = 1.. By symmetry..9). j1 g. (4. fj = E (a+ b k =1 )9kfi_k . if go = 0.. THE DISTRIBUTION OF THE AGGREGATE CLAIMS 321 Proposition 4.12) where g*n is the nth convolution power of g. which would consist in noting that (in the case go = 0) fj = pn9jn n=1 (4. 2.
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)+ with b determined by E(U .P > EexP{7[X . and so on. u But since ry > 0.P } < 1 = Eexp E[Ui. Local reinsurance with h(u) = (u .4. Eexp 7 [E [Ui .h(U) (as in the proof of Proposition 6. then (6. we get EX = EN • EU. Proof As in the proof of Proposition 6.b)+ = Eh(U) (and the same P) satisfies 71 > ry. Then for any local retention function u .P I = EC [7]N.4) and u g(x) = e7x in Ohlin's lemma.h * (X) . it suffices to show that Eexp {ry ii 'UiAb.h(U)]. y = Ei [Ui . Remark 6. that 01[ry] < 0[y] where 0[y] = Ee'r(U^') .i. i.h(Ui)P JJJ l:='l {ry ] or. ' ii (6. (6.6).P. however.5 Because of the independence assumptions .6.h(Ui)] .4). as often local as global.h( UU) = EN • E[U .h(u) and any P satisfying (6. this implies 7* > 7. the excess ofloss rule hl (u) = (u .5) reduce quite a lot.4). Applying the inequality Ecp(Y ) > EW(E (YIX )) (with W convex ) to W(y ) = eryy.. This follows by taking Xl = U A b.6) u where C[ry] = Ee'r(u4(u)). we get N 1 = Eexp ry E[Ui ii . (6.6 Assume the Ui are i. REINSURANCE Proof Define N 329 h* (x) = E > h(Ui) X = x .3). X2 = U .P]}. N E X .h(Ui)] .4). Assuming for simplicity that the Ui are i. ry* > 0 because of (6.d. appealing to (6. The arrangement used in practice is. expectations like those in (6.b)+ is referred to as excessofloss reinsurance and plays a particular role: Proposition 6.5) holds trivially.d.h(Ui)] .
g. see also Sundt [354]. Heilman [191] and Sundt [354]. [76]. The present proof is from van Dawen [99]. Bowers et at. 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.330 CHAPTER XI. See further Hesselager [194] and Dickson & Waters [120]. e.many texts on insurance mathematics. .
.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. of interarrival times and the time Yo = To of the first arrival (that is. stating that U(t+a)U (t) ^ a. t +a]).. 2h. The number max k : Tk_j < t of renewals in [0. of epochs or the set Y1. Y2. The point process is called a renewal process if Yo. . denoted by F in the following and referred to as the interarrival distribution.T„_1).. not concentrated on {h. Then Blackwell 's renewal theorem holds. all have the same distribution.r. . Y. t]) so that U(t + a) .e. note in particular that U({0}) = 1. are independent and Y1. .U(t) is the expected number of renewals in (t. U(A) is the expected number of renewals in A C R in a zerodelayed renewal process.t. Y1. when t is large. some condition is needed: that F is nonlattice. Lebesgue measure dt normalized by the mean to of F. That is. The mathematical representation is either the ordered set 0 < To < T1 < . The associated renewal measure U is defined by U = u F*" where F*" is the nth convolution power of F.. = T„ .. . If Yo = 0. Y2. If F satisfies the stronger condition of being spreadout (F*' is nonsingular w . then Stone 's decomposition holds : U = U. the renewal process is called zerodelayed.. Technically. Lebesgue measure for some n > 1).} for any h > 0... i. t] is denoted by Nt. + U2 where U1 is a finite measure and U2(dt) = u(t)dt where 331 . t 00 (A.. The renewal theorem asserts that U(dt) is close to dt/µ.. the distribution of Yo is called the delay distribution.1) (here U(t) = U([0..
IV).3) Further.332 APPENDIX u(t) has limit 1/µ as t 4 oo. that z(u) has a limit z(oo) (say) as u 4 oo.4) that z is Lebesgue integrable with limZ.a. wee shall need the following less standard parallel to the key renewal theorem: Proposition A1. Equivalently.2) has the unique solution Z = U * z. in convolution notation Z = z + F * Z.e.out.2) Z(u) = J0 u z(x)U(dx). oo). then Z(u) i f0 z(x)dx .R.. and F(dx) a known probability measure . (A. Note in particular that F is spreadout if F has a density f. i. (A. IV).i.4) If F is spread. resp. 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. and that F has a bounded density2. Then Z(u) 4 z(oo). stating that U(t)/t > 1/p.5) 2This condition can be weakened considerably .EN(t) . (A. but suffices for the present purposes . U Z(u . µF (A.i".9.x)F(dx). ENt 4 1 lb Renewal equations and the key renewal theorem The renewal equation is the convolution equation Z(u) = z(u) + f where Z(u) is an unknown function of u E [0 . u u PF 4 00. see [APQ] Ch. Under weak regularity conditions (see [APQJ Ch. the asymptotic behavior of Z(u) is given by the key renewal theorem: Proposition A1. z(x) = 0. z(u) a known function.2 Assume that Z solves the renewal equation (A. In 111. the statements being EN(t + a) . then it suffices for (A. Both result are valid for delayed renewal processes. (A.2).
The distribution F of Y1. . . that the existence of y may fail for heavytailed F. A regenerative process converges in distribution under very mild conditions: . this expression is to be interpreted as a random element of the space of all Evalued sequences with finite lifelengths. Here the relevant F does not have mass one (F is defective). Yk ). asymptotic properties can easily be obtained from the key renewal equation by an exponential transformation also when F(dx) does not integrate to one.t))u(ut) dt 0 0 J f z(oo) • 1 dt = z(OO). or many queueing processes.2) by e7x to obtain Z = z +P * Z where Z(x) = e'Y'Z(x). However. .t..e.APPENDIX 333 Proof The condition on F implies that U(dx) has a bounded density u(x) with limit 1/µF as x * oo. . F(dx) = e7xF(dx). The kth cycle is defined as {XTk+t}o<t<Yk . results from the case fo F(dx) = 1 can then be used to study Z and thereby Z. T1. Hence by dominated convergence. the postTk process {XT. a basic reason that renewal theory is relevant is the renewal equation II. however. this covers discrete Markov chains where we can take the Tn as the instants with Xt = i for some arbitrary but fixed state i.(3. we let µ denote its mean. Y2. of Yo.i. The property of independent cycles is equivalent to the postTk process {XTk+t}t>0 being independent of To.k+t }t>o is independent of To. where the Tn are the instants where a customer enters an empty system (then cycles = busy cycles). multiply (A. cycles. that F is a probability measure. z(x) = e7xz(x). 1c Regenerative processes Let {T.. 0 PF µF 11 In risk theory.} be a renewal process. the present more general definition is needed to deal with say Harris recurrent Markov chains. A stochastic process {Xt}t>0 with a general state space E is called regenerative w. Tk (or. . Note. Z(u) U = 1 u 1 u f z(u .3) satisfied by the ruin probability for the compound Poisson model. T1. .. Y1 . refer to the zerodelayed case. We let FO.5a. and its distribution does not depend on k.. equivalently. • . However. To this end. is called the cycle length distribution and as before.d. i... Assuming that y can be chosen such that f °° Ox F(dx) = 1. Eo etc.r. The simplest case is when {Xt} has i. This program has been carried out in III. {Tn} if for any k. Tk and {Xt }o<t<Tk • For example...x)u(x) dx = z(u( 1 ..
This is the case considered in [APQ] V.t. {Tn}.334 APPENDIX Proposition A1. are i. r.3.r. [0. If p = oo. for n = 1. (b) If in addition Var(Ul ) < oo..v. e(t )) . just the same proof as there carries over to show: Proposition A1. {Tn} if the processes {ZT +t . where the distribution of X. in total variation. but in fact. and q(t) = sup It .+ X. fi (t) = inf {Tk .. cycles (we allow a different distribution of the first cycle).e. is given by Eg(Xoo) = 1 E0 f Ylg (Xt)dt.Tk : t < Tk} as the age. then Xt .tEU1/µ)/f has a limiting normal distribution with mean 0 and variance Var(Ui) + (!)2Var (Yi)_ 2EU1 Cov(U1. Then {e(t)}.r. P(C ( t) < a) 4 0 for any a < oo) and ij (t) * oo.. assume that p < 00 and define Un = ZT}1 . C(t) and ij (t) both have a limiting stationary distribution F0 given by the density F (x)/p.i.6) id Cumulative processes Let {Tn} be a renewal process with i. 0<t<Yi then Zt /t a$• EU1/µ.0 is called cumulative w.t.ZTOI < 00. and we have: holds more generally that (rl(t).e. Then it (ii.d. Then {Zt}t^.0 be cumulative w.oo (i.i.ZT Then: (a) If E sup I ZTo+t . under the condition of Blackwell's renewal theorem.. i. Then Xt Di X.r.t...ZT }0<t<Y„+. then (Zt . We denote the limiting r... An example is Zt = fo f (X8) ds where {Xt} is regenerative w.. {Tn}.4 Let {Zt}t^. oo).. 2. Y1) le Residual and past lifetime Consider a renewal process and define e ( t) as the residual lifetime of the renewal interval straddling t. oo). then e (t) .'s by e.d. (A..3 Consider a regenerative process such that the cycle length distribution is nonlattice with p < oo. C). {i7(t)} are Markov with state spaces (0. µ 0 If F is spreadout.t : t < Tk}. Otherwise . resp ..
we can bound e(t) by M(t) = max {Yk : k < 2t/p}. and the equivalence of (a) with (b)(d) is an easy exercise..6 Consider a renewal process with µ < oo.v. are not i.i. Y1 > t] 4 0. 0 If Markov renewal theory By a Markov renewal process we understand a point process where the interarrival times Yo . Hence for t large enough. Yo > 0] + f Eo^ (t . Since z ( k) < E[Yi . if in addition EYo < oo. For the second.t. the sum is o(t) so that Eo£(t)/t + 0 . Then fi(t)/t a4' 0 and.5 Under the condition of Blackwell's renewal theorem.U(x) (c < oo because it is easily seen that U(x + 1) . V is uniform on (0. EC(t)/t + 0. and the conditional distribution of ri given l. Y1i Y2. .U(x) < U( 1)).'s with finite mean satisfies Mn/n a$• 0 (BorelCantelli).^(t))} as a regenerative process.dy )z(y) < c ^ l z(k) Eoe(t 0 0 k=o where c = sup. r. and the conditional distribution of given 17 = y is the overshoot distribution R0(Y) given by FO(Y) (z) = Fo (y+z)/Fo(y). = z is Foz) The proof of (a) is straightforward by viewing {(r. In the general case.d.(t). W are independent. the joint distribution of (rl. Since the maximum Mn of n i.. (d) the marginal distribution of ^ is FO. In IV. (1 V)W) where V.t.y) = f U(t . but governed by a Markov chain {Jn} (we .. use t E^(t)/t = E[Yo . Yl > t]. 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.4. ^) is given by the following four equivalent statements: (a) P (77 > x.d. assume first the renewal process is zerodelayed. the first statement follows. Hence t t lt ) = f U(dy)z(t .i. (b) the joint distribution of (ri. ^ > y) = 1 f +Y (z)dz. l:) is the same as the distribution of (VW.APPENDIX 335 Theorem A1. U(x + 1) .y)P(Yo E dy) . we used: Proposition A1. (c) the marginal distribution of q is FO. 1) and W has distribution Fw given by dFw/dF(x) = x/pF.
Jo.. Y1.. Further: Proposition A1. Assume that uj = EjYo < oo for all j and that {J„} is irreducible with stationary distribution (v3)jEE.. . Let X1. Notes and references Renewal theory and regenerative processes are treated. Yn. 0] or (0 .. J1 i . be i. Jn = i is the same as the P. Jn +1=j} where J = a(JO. IT.t. . r+ < oo). T_=inf{n>0: Sn<0}. in [APQ].) and (Fij )i.d. Sn = X1 + • • • + Xn the associated random walk.jEE is a family of distributions on (0. namely {Twk } where {Wk } is the sequence of instants w where Jo. oo). .} is nonlattice (it is easily seen that this definition does not depend on i).t... where the distribution of X.+ < x.}. .. Jn_1. G+(x) = P(S.i .. . X2.. .. and define r+=inf{n>0: Sn>0}.. is given by Eg(X00) = 1 YO vjEj f g(Xt) dt µ jEE o where p = ujEEViAj. A Markov renewal process {Tn} contains an imbedded renewal process... = io for some arbitrary but fixed reference state io E E.r. .. We call r+ (T_) the strict ascending (weak descending) ladder epoch and G+ (G_) the corresponding ladder height distributions. Alsmeyer [5] and Thorisson [372].336 APPENDIX assume here that /the state space E is// finite) in the sense that P(Y.r. with common distribution F.g. G_(x) = P(ST_ < x. A2 WienerHopf factorization Let F be a distribution which is not concentrated on (oo. Then Xt 4 Xo. e.and regenerative processes.. < yIJ) = Fij( y) on {Jn= i. A stochastic process {Xt}t>o is called semiregenerative w. The semiregenerative process is then regenerative w. . the Markov renewal process if for any n. . the semiregenerative process is called nonlattice if {T. . distribution ofjXt}t>o itself where Pi refers to the case Jo = i. For example. the conditional distribution of {XT„+t}t>o given Yo.T_ < oo). oo). These facts allow many definitions and results to be reduced to ordinary renewal.7 Consider a nonlattice semiregenerative process.
oo) (A. . the renewal measures U+=>G+. Sr_ _1 is at its minimum . G_. S. 0]).=EGn.7)..x)R+(dx). A C (oo. U.8) (e. m<j<n}. (A. we consider the last such time (to make w unique) so that {w=m. 0].G+ * G_: (b) G_ (A) = f °° F(A . F(A) is the contribution from the event {T_ = 1} = {X1 < 0}.1 . A C (0. define w as the time where the preT_ path S1.T_=n} = {S. n=0 n=0 00 00 and the T+. F(A .7) follows since G+(A) = 0 when A C (oo.g.>0. In (A. u . (e) R_ = U+. F(A) + (G+ * G_)(A)..1 (a) F = G+ + G_ . On {T_ > 2}. n 0 R_(A) = E I(Sn E A). we may rewrite (a) as G_ (A) = G+(A) = F(A) + (G+ * G_)(A).S.and r_ preoccupation measures T+1 r_1 R+(A) = E E I(Sn E A).7) (A.. A C (oo. 0] and (0. G+. (c) G+(A) = f °. 0<j<m.APPENDIX 337 Probabilistic WienerHopf theory deals with the relation between F. More rigorously. 0).x)R_ (dx). oo). (d) R+ = U_. A C (0. .r. oo). Proof Considering the restrictions of measures to (oc. n=0 The basic identities are the following: Theorem A2.=n w=m i Figure A.S. >0.
E du) = P(T_=nm.1) that P(Sj Sn. A.3... (A.du) (G+ * G)(A)• C llecting terms. SmEdu) = P(T+=m. and reversing the order of summation yields P(T_ > 2..F(r_n_mSrEA_u). Sn1 E dx) n=1  F(A ..= n. (b) follows from 00 G+ (A) _ E F(Sn E A._ E A) n1 f P(r_=nw=m Sm EduSrEA) m=1 n1 F(r+=mSr+Edu). m=1 f S mming over n = 2. S.XnEAx) 00 f 0 f 0 00 00 1: F(A . Aso. ST_ E A .+ E du) E P(S. Sr_ E Adu) (s ee again Fig .7) follows.x)R+(dx). r+ = n) n=1 n=1 0  C0 E fF(Sk< 0. . and the proof of (A.u) f0m m=1 n=m+1 00 J0 OO P(S.x)P(Sk < 0.>0.+ E du)P(S. m it follows (see Fig. 0<j<m._ E A .. . A. S.0<k<ri . ST_ E A) P(T+ = m.. ST+Edu).1. m < j <n._ = n . .3 8 APPENDIX Reversing the time points 0.. 0 < k < n.1).8) is similar. It follows that for n > 2 F (7. SnEAIS.Sn_1Edx.. clearly (Sj Sm>0.m.
6.. the survey [15] by the author and the extensive list of references there.1. In this generality of.g. 6+ [s]. For example.2 In terms of m.O<k<n.1(a) is from Kennedy [228]. see for example Bingham [65]. WienerHopf theory is only used at a few places in this book. and G+. u Notes and references In its above discrete time version.T+> n) = P(Sk < O.P as a product H+H_ of functions with such properties. G_ are trivial. In continuous time. 0]. if {St} is Brownian motion. u Remark A2. a number of related identities can be derived. The classical analytical form of the WienerHopf problem is to write 1 . However. P(SnEA .'s. consider a fixed n and let Xk = Xn_k+l.g.Sn_k. this holds always on the line its = 0.0<k<n. Another main extension of the theory deals with Markov dependence.s. it serves as model and motivation for a number of results and arguments in continuous time. Sk = X1 + • • • + Xk = Sn . Summing over n yields R+ (A) = U_ (A).F[s] = (1 .1).g. being concentrated at 0.. and similarly H_ (s) = 1 . Again.G_[s]) (A. G_ [s] are defined at the same time. there is no direct analogue of Theorem A2.SnEA) = P(SnSn_ k.SnEA) is the probability that n is a weak descending ladder point with Sn E A. Then for A C (oo. we can rewrite (a) as 1 . For (d).1. which is basic for the PollaczeckKhinchine formula. and the proof of (e) is similar. cf. the analogue of a random walk is a process with stationary independent increments (a Levy process.G_ [s] is defined and bounded in the halfplane is : ERs > 01 and nonzero in Is : ERs > 0}. there are direct analogues of Theorem A2.SnEA) = P(Sn<Sk. H+ (s) = 1G+[s] is defined and bounded in the halfplane Is : ERs < 0} and nonzero in Is: Rs < 01 (because IIG+lI _< 1). 11. The present proof of Theorem A2. In discrete time. oo). .0<k<n. E. see e.4).SnEA) = P(Sn<Sk. and using timereversion as in (d) to obtain the explicit form of R+ (Lebesgue measure).O<k<n. the derivation of the form of G+ for the compound Poisson model (Theorem 11.0+[s])(1 .APPENDIX 339 and the proof of (c) is similar. Nevertheless. then T+ = inf It > 0 : St = 0} is 0 a. and sometimes in a larger strip. is based upon representing G+ as in (b). such developments motivate the approach in Chapter VI on the Markovian environment model. Since G+ is concentrated on (0.9) whenever F[s].f.
12) eA'AO = Ale AA (A. Thus. three of the c rrently most widely used ones: xample A3. It is seen from Theorem VIII. write eQ = (eK)m where = Q/m for some suitable integer m (this is the scaling step).10) d dteAt = AeAt = eAtA (A. if m is s fficiently large.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. _I 0 (A. Some fundamental properties are the following: sp(eA) = {e' : A E sp(A)} (A. 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).13) henever A is a diagonal matrix with all diagonal elements nonzero.5 that when handling phase type distributi ons. JAI = max {Jjt : µ E sp(A)} and sp(A) is the set of all eigenvalues of A (the spectrum). To circumvent this. Eo Kn/n! converges rapidly and can be evaluated without p oblems. Here it is standard to compute matrixinverses by GaussJordan el imination with full pivoting . 1. whereas there is no similar single established a proach in the case of matrix exponentials.11) A f eAtdt = eA. Here are.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. 0 . ere A is the eigenvalue of largest absolute value. and eQ can then be computed as the mth power (by squaring if = 2). however . one needs to compute matrix inverses Q1 and matrix exponentials eQt ( r just eQ ).
3 i (A. .e. Zo = h)..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}.e. In practice. the procedure consists in choosing some suitable i > 0. 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.. i. some jumps are dummy in the sense that no state transition occurs ). One then can reduce to p linear differential equations by noting that k = ZQ. p different eigenvalues Aj i . Here is a further method which appears quite appealing at a first sight: Example A3 . .14) holds is therefore that the tstep transition matrix for {fft} is eQt = E ent (.3 (DIFFERENTIAL EQUATIONS) Letting Kt = eQt. construction of {Xt} by realizing the jump times as a thinning of a Poisson process {Nt } with constant intensity 77. To this end. Zo = a (Z = QZ.. vp be the corresponding left . 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 .]t)n (A.2 (UNIFORMIZATION) Formally. assume that Q is the intensity matrix for {Xt} and choose q with rt > max J%J = max qii• 1. 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. i.15) Then it is easily checked that P is a transition matrix . letting P = I + Q/i and truncating the series in the identity = e17t 00 Pn(..4 (DIAGONALIZATION) Assume that Q has diagonal form.7t) n=0 n! u °O n Pn (to see this. Ap. The approach is in particular convenient if one wants eQt for many different u values of t. what is needed is quite often only Zt = TreQt (or eQth) with it (h) a given row (column) vector.. The probabilistic reason that (A. condition upon the number n of Poisson events in [Olt])  Example A3. Let vi. However .APPENDIX 341 Example A3.
.. and we need to have access to software permitting calculations with complex numbers or to perform the cumbersome translation into real and imaginary parts.18) Namely..16) (A. Complex calculus : Typically. say A = (Ai)diag. Then P P Q = > Aihivi = E Aihi (9 vi. i=1 i=1 Thus. (A.. and vihi ¢ 0. however. two serious drawbacks of this approach: u Numerical instability : If the A5 are too close.342 APPENDIX (row) eigenvectors and hl. and we may adapt some normalization convention ensuring vihi = 1. i= 1 i=1 P P (A. hi have been computed. Example A3. some cases remain where diagonalization may still be appealing.g H1.. Qhi = vihi. of largest real part is often real (say. The phenomenon occurs not least when the dimension p is large. hp. we can take H as the matrix with columns hl.17) eQt = E e\`thivi = E ea:thi ® vi. not all ai are real. under the conditions of the PerronFrobenius theorem). this last step is equivalent to finding a matrix H such that H1QH is a diagonal matrix. i # j.. say Al. D = ) 2 2 . Nevertheless. we have an explicit formula for eQt once the A j. the eigenvalue.18) contains terms which almost cancel and the loss of digits may be disasterous. and writing eQt as eQt = He°tH1 = H (e\it)di. There are. Then vihj = 0. (A.. hp the corresponding right (column) eigenvectors. and hence A2 is so because of A2 = tr(Q). In view of this phenomenon alone care should be taken when using diagonalization as a general tool for computing matrixexponentials. v5Q = Aivi. vi.. Everything is nice and explicit here: 411+q2+D' )12_g11+q2^^ where (411422z + 4412421.5 If Q= ( 411 ( q21 q12 q22 is 2 x 2.
e. replacing ai by A2. Of course. Then 7r = (ir1 7r2 ) = a (q21 Al . b are any constants ensuring//Irk = 1. i.20) ir = q2 ql qi +q 2 9l +q2 (A.APPENDIX 343 Write 7r (= v1) for the left eigenvector corresponding to a1 and k (= hl) for the right eigenvector.q.Q2i and after some trivial calculus one gets eQt = 7r 1 112 + eat 7r1 7r2 / (7fl 7r2) = ( 7r2 1r2 7r1 IF. However.19) Example A3 .21) Here the first term is the stationary limit and the second term thus describes the rate of convergence to stationarity. The other eigenvalue is A = A2 = q1 . h2 = Thus. Then Al = 0 and the corresponding left and right eigenvectors are the stationary probability distribution 7r and e. l ab (g12g21 + (A1  411) 2) = 1. u Example A3. 1) . k  C k2 ) =b ( A1 q 1 Q11 / where a . v2 and h2 can be computed in just the same way. eqt = eNlt ( ir1ki i2k1 \ ir1 k2 72 k2 + e azt 7r2k2 i2k1 7ri k2 7r1 k1 (A.k1). where (A.7 Let 3 9 2 14 7 11 2 2 .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 .
ir =a(2 9 9 14 2 1 3 2 2)' k=b 14 =b 1+ 2 ir1 k1 ir2 k1 _ 9 2 10 5 7 9 70 1 ' 7r1 k2 7r2 k2 10 9 9 10 10 + 7 1 10 10 10 1 10 7 10 9 70 9 10 0 e4" = e_.11/2 .5 . (AA+)' = AA+.6. They are most often constructed by imposing some additional properties .satisfying AAA = A. Generalized inverses play an important role in statistics. A+AA+ = A+. A2 = 3/2 .22) Note that in this generality it is not assumed that A is necessarily square. 2 2 1=ab(142+(1+2)2 ) = tab.11/2 + 5 1.344 Then D= 2+ 11)' 7 T4 2 =52..23) . and a generalized inverse may not unique. (A.. (A+A)' = A+A. (A. APPENDIX x1 3/2 . for example AA+A = A. e_6u A4 Some linear algebra 4a Generalized inverses A generalized inverse of a matrix A is defined as any matrix A. but only that dimensions match .
.24) = te7r . lt o eAx dx = te7r + D(eAt . Am+1 = . 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.P + e7r ).... and define D = (A .D + O(ebt).I) (A. Assume that a unique stationary distribution w exists .g.23) is called the MoorePenrose inverse of A. and exists and is unique (see for example Rao [300]).1 goes under the name fundamental matrix of the Markov chain). ( Q .. most often either an intensity matrix Q or a matrix of the form IP where P is a transition matrix.ew..= (I . . are ordered such that Al > 0. Then for some b > 0. one then works with Q = (Q .1Q = Q(Q . = 0 where m < p is the rank of A.eir)1 = I . and can define /ail 0 0 0 0 0 0 A+ = C A' 0 0 0 C' .25) .g.e ® 7r)1. (A. then there exists an orthogonal matrix C such that A = CDC' where 0 0 D = AP Here we can assume that the A .e. _ A. These matrices are not generalized inverses but act roughly as inverses except that 7r and e play a particular role . if A is a possibly singular covariance matrix (nonnegative definite).P + e7r)1 (here ( I . 0 01 In applied probability. Rather than with generalized inverses . (I . E. Am > 0.eir )1.P).eir ).APPENDIX 345 A matrix A+ satisfying (A. one is also faced with singular matrices . .
in block notation i2h A®B= ( a11B a21 B a12B a22 B Example A4.I) (A. respectively.I)}. see below. the formulas involving O(e6t) follow by PerronFrobenius theory. 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.h. B'(t) = e7r + DAeAt = eir + (I . resp. Interpreting 7r. B(t) denote the l.. ()®(6 f 6/ 7f 8^ 7 8 )=! ^)( 6 7 8 )=(6^ 7^ 8^) \ u Example A4. Then A(O) _ B(O) = 0.2e7r .27) Proof Let A(t).J {xe^r + D(e . the r.26) follows by integration by parts: t f t /' xeAx dx = [x {xe7r + D(eAx .eir)eAt = eAt = A'(t).3 Let 2 A= 4 3 Vf' N7 5 )' B= ( 8 ).24).91a(2) .s.I) .2 Let it be a row vector with m components and h a column vector with k components. it follows that h ® it is the k x m matrix with ijth element hi7rj .e. For example.DZ(ent . I. o Finally. the rows are proportional to it.h. h ® it reduces to hit in standard matrix notation.I)} dx.26) 2 = 2 e7r + tD . 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.D + D2 + O(ebt). . (A.346 t APPENDIX 2 xe Ax dx = eir + t(D + e7r) + D(eAt . Equivalently. and the columns to h. h as 1 x m and k x 1 matrices. .s. (A. and in fact any rank 1 matrix can be written on this form. Note that h ® it has rank 1.
31) Indeed. and the number of such factors is precisely given by the relevant binomial coefficient.29) If A and B are both square (k1 = ml and k2 = m2).4 eA® B = eA ®eB.APPENDIX 347 Then A®B = 2 f 20.3V8. such a factor is Ak (&B 1k according to (A.3v'6.31). if A ® I occurs k times.k)! ( n0 n=0 t=0 k=0 J _ ® Ak ®Blk r ^.A9. 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.(A. Using (A.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 A4.5v/. and v1B1h1 • v2B2h2 = v1B1h1 ® v2B2h2 = ( v1(&v2 )( B1(&B2 )( h1(&h2 ) .3f 4v/. C2 = h2 are column vectors. then v1B1h1 and v2B2h2 are real numbers. (A B)' = eA®B e! L 1=0 0 .50 6 7 6 4f 4. each of which is A ® I or I ® B.29).4vf. (A.5v'8 5vf9 11 A fundamental formula is (A1B1C1) ®(A2B2C2) = (A1 (9 A2)(B1 (9 B2)(C1®C2). if Al = vi. (A. then the Kronecker sum is defined by A(1) ®A(2) = A(1) ®Ik2 + k ®A(2). it follows that e® ® e B An _ 0o oo oo Bn 7 I F n! = ` k! (I . (AED B)1 = (A®I+I(9 B)l is the sum of all products of t factors. A2 = v2 are row vectors and C1 = h1.3vV/72f 20.
5 Many of the concepts and results in Kronecker calculus have p(2) is the intuitive illustrations in probabilistic terms. P(2). we have P8 = Pal) ® p(2). Yt(2) where independent Markov processes with intensity matri{y(2) } are {Y(1) }. Then 2 0 ire At h • ve Bt kdt = (^®v)(A®B)1(e A®Ba .s.4 can easily be obtained by probabilistic be the sstep transition reasoning along the same lines .33) . p = P(1) ® {X }. { 1't(1) }. P8 = exp {sQ} = exp {s (Q(1) ®Q(2)) } .348 APPENDIX Remark A4. and the form of the bivariate intensity matrix reflects the fact that Yt(2) } cannot change state in both components at due to independence . in the definition (A. (A. Let further it. Ps 1) = exp {sQ ( 1) } > p(2 ) = exp {sQ(2) } can therefore be rewritten as Taking s = 1 for simplicity .I)(h ® k). independent Markov chains. and Q = Q(1) ® Q (2) = Q(1) ® I + I ® Q(2) (A. where transition matrix of the bivariate Markov chain {X n1). k any column vectors. represents ces Q( 1). A special case of Proposition A4. n2 n1 ) {X(2) } are independent Markov chains with transition matrices P(1). P8 = Pal ) ® P82) exp {Q ( 1) ® Q(2)1 = eXp {Q( 1) } ® exp {Q(2) } Also the following formula is basic: B are both square such that a +. { On the other hand.32) is the intensity matrix of the bivariate continuous Markov process {Yt(1). {Yt(1). P(t) Yt(2) }. X ) }. the {Yt(2) } transitions in the {Yt(1) } component and the second transitions in the component . h.32). first term on the r . v whenever a is an eigenvalue of A and 0 is an eigenvalue be any row vectors and h. From what has been said about matrices of {Yt( 1). Let P8f P(Sl). Thus . the same time.6 Suppose that A and of B. Yt(2 ) }.3 < 0 Lemma A4 . resp . Q(2).
8 Let B be an irreducible3 p x pmatrix with nonnegative offdiagonal elements. . We call A irreducible if the pattern of zero and nonzero elements is the same as for an irreducible transition matrix.1 and references there (to which we add Berman & Plemmons [63]): Theorem A4. the integrand can be written as ( 7r (9 v)( eAt ® eBt )(h ®k ) = ( 7r ®v)(eA (DBt)(h (& k).. . il. = j and atk_li. h can be chosen with 3By this. and appeal to (A. E (0. . we have AO = 1. . f o r each i. . j = 1. ao). That is. [APQ] X. n..7 Let A be a p x pmatrix with nonnegative elements. > 0 for k = 1. then An = Aohv+O(µ") = Aoh®v+O(µ") for some u.g. we mean that the pattern of nonzero offdiagonal elements is the same as for an irreducible intensity matrix. . A is called aperiodic if the pattern of zero and nonzero elements is the same as for an aperiodic transition matrix. i. (b) if in addition A is aperiodic.. (A. in such that io = i. h such that vh = 1. and if we normalize v. h = e and v = 7r (the stationary row vector). Similarly. 4c The PerronFrobenius theorem Let A be a p x pmatrix with nonnegative elements. Then the eigenvalue Ao with largest real part is simple and real. . h can be chosen with strictly positive elements. Here is the PerronFrobenius theorem.34) Note that for a transition matrix. Then: (a) The spectral radius Ao = max{JAI : A E sp(A)} is itself a strictly positive and simple eigenvalue of A.12).. .29).3 whenever a is an eigenvalue of A and 3 is an eigenvalue of B. so that by asssumption A ® B is u invertible. then IN < Ao for all A E sp(A). which can be found in a great number of books. and the corresponding left and right eigenvectors v. p there should exist io.The PerronFrobenius theorem has an analogue for matrices B with properties similar to intensity matrices: Corollary A4.APPENDIX 349 Proof According to (A. .. and the corresponding left and right eigenvectors v.. Now note that the eigenvalues of A ® B are of the form a +. see e..
To this end. Proposition A5. Furthermore. then eBt = ea0thv + O(eµt) = eA0th ® v + O(et t) (A. For example.2). if we normalize v. note that we can write the phase generator T as Q . Note that for an intensity matrix.(ti)ding. not only in the tail but in the whole distribution.350 APPENDIX strictly positive elements. I. Then for any (3. the phasetype distribution B(a) with representation (.(3. h such that vh = 1. let t = (ti)iEE # 0 have nonnegative entries and define T(°) = aQ .e. but is an easy consequence of the PerronFrobenius theorem. we have A0 = 0. A5 Complements on phasetype distributions 5a Asymptotic exponentiality In Proposition VIII. the analogy of this procedure with unformization. h = e and v = 7r (the stationary row vector).35) for some p E (oo. Corollary A4. 10) and use the formula me at e Bt = e 00 Antn = e .8. The next result gives a condition for asymptotical exponentiality. relate the eigenvalues of B to those of B via (A. let {Yti°i } be a Markov process with initial distribution a and intensity .1 Let Q be a proper irreducible intensity matrix with stationary distribution a.(ti)diag where Q = T + (ti)diag is a proper intensity matrix (Qe = 0).8 is most often not stated explicitly in textbooks. Bi° (x) + at*x Proof Let { 4 } be the phase process associated with B(a) and (°) its lifelength. the condition is that t is small compared to Q. 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.. it was shown that under mild conditions the tail of a phasetype distribution B is asymptotical exponential. Ao). T(°)) is asymptotically exponential with parameter t* _ r EiEE aiti as a 4 oo. one can consider A = 77I + B where rl > 0 is so large that all diagonal elements of A are strictly positive (then A is irreducible and aperiodic).n t AL n=0 n! (cf.1.
1. in fact .bij) Hence the intensity matrix of { Ix} is (qij/ti)i.(a) > x . from which it is easily checked that the limiting stationary distribution is (aiti/t*)iEE• Now let a' 4 oo with a in such a way that a' < a. v/ t. it states that the state. t < (a).2 Pi (c(a) > x. we get dx F (Idx = j) = (1 + qij t )Sij + qij dt. Then {Ix} is a Markov process with to = Yo. has a limit distribution: Proposition A5.YQ(av) = j) Pi ( ci(a'V) > x. By the law of large numbers for Markov processes . and this easily yields a(x)/x a' 1/t*.a' + oo (e.x (1 . J(()) _ = i) + at•x t tt' . a . fo tY dv/t a$' t*. J^O)_ = j) Pi (v(aaV) > x. {t Y( a) } v>0 . Hence O ((a) aa. Conditioning upon whether { Yt} changes state in [0.9. 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}. and write Yt = Yt(1).APPENDIX 351 ((1) etc. = YQ(x). a'/a + 1. We can think of ( ( a) as the first event in an inhomogeneous Poisson process ( Cox process ) with intensity process matrix aQ .Yj(av) = j f . Then a(a'V)/a (aV) a' 1. Let further V be exponential with intensity V and independent of everything else. We shall . Since JJ(.)_ = Y(a) = 1'aS(a) = Ya(av)^ it follows that Pi ((.jEE. and that Yt(a) = Yat for all t. prove a somewhat more general result which was used in the proof of Proposition VI.g. We can assume that Jta) = Yt(°). a' = a .aE where 0 < e < 1). dx/ti] or not. from which the phase process is terminated . Proof Assume first ti > 0 for all i and let I. In addition to the asymptotic exponentiality.
Et II I a(a^V) > x) at' . and thus the parameter p of the geometric distribution u can be identified with the exit probability vector p.j) and initial distribution a. Then: (a) The point probabilities are bk = aPklp.+ at*x • a't' L ` at t* t* J Reducing the state space of {Ix } to {i E E : t... so we shall be brief.. (c) the nth moment k 1 k"bkis 1)"n!aP"p. these results are in the spirit of rare events theory for regenerative processes (e. Example A5..p)k1 p..zP)'p. 2. K}.Pe. A distribution B on {1. Keilson [223].5 Let B be discrete phasetype with representation (P.3 As the exponential distribution is the simplest continuous phasetype distribution. . is discrete phasetype. . k = 1.. (b) the generating function b[z] _ E' .4 Any discrete distribution B with finite support. u Notes and references Propositions A5.1 and A5.. ' pk 0 k>1 11 Theorem A5.. with point probabilities bk = (1 . P.. k>1.2 do not appear to be in the literature. However. See also Korolyuk. let E and Pkj j=k1. a = b = (bk)k=1.g. . a). say bk = 0. so is the geometric distribution..} is said to be discrete phasetype with representation (E. 2. 5b Discrete phasetype distributions The theory of discrete phasetype distributions is a close parallel of the continuous case. zkbk is za(I . 1 k=1 1 0 otherwise. Penev & Turbin [238]. Example A5. Indeed. > 0}. = 0 for one or more i. Gnedenko & Kovalenko [164] and Glasserman & Kou [162]).352 rr Ia(a'V) Ei I ( > x) P APPENDIX L at (Yo (aV) . a) if B is the lifelength of a terminating Markov chain (in discrete time) on E which has transition matrix P = (p. the simplest discrete phasetype distribution: here E has only one element..x k > K. an easy modification of the argument yields finally the result for the case where t. Then P is substochastic and the vector of exit probabilities is p = e .
T(1)).APPENDIX 353 5c Closure properties Example A5. Jt t > U1 + U2. resp. U2.2 The form of these results is easily recognized if one considers two independent phase processes { Jt 1) }.1 This corresponds to a convolution of r geometric distributions with the same parameter p. Then {Jt} has lifetime U1 + U2 . The discrete counterpart is the negative binomial distribution with point probabilities bk k1) (1 k = r.a(1).6 (CONVOLUTIONS) Let B1. _ i E E(1) T(1) t(1)a(2) i E E(2) .T(2)). A. and piece the processes together by it = 41) 0<t<U1 U1 < t < U1 + U2 2U.6.. (E(2).a(2). as is seen by minor modifications of Example A5. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2)...36) in blockpartitioned notation (where we could also write a as (a (1) 0)).7 (THE NEGATIVE BINOMIAL DISTRIBUTION) The most trivial special case of Example A5. and hence the negative binomial distribution is discrete phaseu type. r . { Jt 2) } with lifetimes U1 . A reduced phase diagram (omitting transitions within the two blocks) is am E(1) t(1) a(2) (2) t(2) Figure A. a' . B2 be phasetype with representations (E(1).{ 0. initial distribution a and phase generator T. . T= ( 0 T(2) ) (A.r + 1. 11 Example A5.6 is the Erlang distribution Er which is the convolution of r exponential distributions. and a=1).. Then the convolution B = B1 * B2 is phasetype with representation (E. resp. a.
and o'i Oa.p)pn1B*n.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. resp. To obtain a phase process for C.p)pn1.O)B2 (0 < 0 < 1) is phasetype with representation (E. Let B(") be the corresponding phasetype distribution. Then the mixture B = 9B1 + (1 .T. Thus.0)a(2))).E) where a(°) = fAa(a)v(da). a. B2 be phasetype with representations (E(1).a(2). then C is the distribution of Ul + • • • + UN. Equivalently. this means that a = (Oa(1) (1 . (E(2)..'). a reduced phase diagram is f a E t Figure A. Then it is trivial to see that B(") is u phasetype with representation (a("). one obvious interpretation of the claim u size distribution B to be a mixture is several types of claims.. p at each termination. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2).a(1).T(1)).4 .d. and consider B(") = fA B(a) v(da) where v is a probability measure on A. i E E(2) 0 T(2) =IT (in blockpartitioned notation.3 In exactly the same way. a mixture of more than two phasetype distributions is seen to be phasetype. a. A reduced phase diagram is 0a(1) E(1) A . if U1. are i. T) and C = EO°_1(1 .37) (1) (1 . with common distribution and N is independent of the Uk and geometrically distributed with parameter p. Example A5. Example A5.0)ai2). U2. P(N = n) = (1 . In risk theory.8 (FINITE MIXTURES) Let B1.p.10 (GEOMETRIC COMPOUNDS) Let B be phasetype with representation (E.0)a(2) E(2) Figure A.. i E E(1) T 0 I (A. we need to restart the phase process for B w.354 APPENDIX Example A5.T(2)).i.
U2.T) if U is phasetype with representation (E. T + pta). U2. if U1. B2 of phasetype with representations (E(').TWWW). X independent of U. then U1 +• is phasetype with representation (E. T) and C = F. of F. U2 be random variables with distributions B1. we then let the governing phase process be {Jt} _ {(411 Jt2))} 2) interpreting exit of either of {4 M }.. f2.x)+. To see this.°_1 f„ B*?l. If we replace x by a r. resp. P). v. (E(2). Equivalently. E).v. resp. T(2) ). let {Jtl)}. but the same T. if B is defective and N + 1 is the first n with U„ = oo. if {Jt} is a phase process for U. T + ta.°. j E F}.2.11 (OVERSHOOTS) The overshoot of U over x is defined as the distribution of (U . let B be a continuous phasetype distribution with representation (F. Then the minimum U1 A U2 and the maximum U1 V U2 are again phasetype. Example A5 . it follows by mixing (Example A5. a. Indeed. a.d.APPENDIX 355 and C is phasetype with representation (E. +UN 2. Example A5 .aeTx.a(1).7.1.. .. let the phase space be E x F = {i j : i E E. Example A5. are i. be the point probabilities of a discrete phasetype distribution with representation (E. Note that this was exactly the structure of the lifetime of a terminating renewal u process. { 4 } as exit of {Jt}. then C is the distribution of U1 + • • • + UN. Thus the representation is (E(1) x E(2).g. then U1 + • • + UN is zeromodified phasetype with representation (a. cf. Minor modifications of the argument show that 1. v. with common distribution B and N is independent of the Uk with P(N = n) = f. say with distribution F. Corollary VIII. If U1 has a different initial vector.T + pta). say v. a(1) ® a(2 ).. a. For U1 A U2.9) that (U .aF[T]. T). 13 (MINIMA AND MAXIMA ) Let U1... T(1) ® T(2)). . It is zeromodified phasetype with representation (E.. cf. i. To obtain a phase representation for C .T) where F[T] = J0 "o eTx F(dx) u is the matrix m.X)+ is zeromodified phasetype with representation (E. 12 (PHASETYPE COMPOUNDS ) Let fl. { Jt2) } be independent with lifetimes U1.f.2. let the initial vector be a ® v and u let the phase generator be I ® T + P ® (ta). a(2).. then Jy has distribution aeTx. Proposition VIII.
B(bk) I < 1/n for n > k.. any distribution B on (0. Then we must find phasetype distributions Bn with B. however.. Then from above. Let the support of Dn be {xl(n).. q(n) q(n) pi(n)a ..(n) = D.14 To a given distribution B on (0. The general case now follows easily from this. and the closedness of the class of phasetype distributions under the formation of finite mixtures. Thus the state space is E(1 ) x E(2) U E(1) U E( 2). i= 1 C. Proof Assume first that B is a onepoint distribution.(bk)'..8.xq(n)(n)}. the initial vector is (a(1) (& a (2) 0 0). Here are the details at two somewhat different levels of abstraction: (diagonal argument .. r # oo.n = I:pi(n)Er v ( __ ) n) ) a= 1 . oo) can be approximated 'arbitrarily close' by a phasetype distribution B: Theorem A5.(bk) + B(bk) for all k. and let Bn be the Erlang distribution E.356 APPENDIX For U1 V U2. the fact that any distribution B can be approximated arbitrarily close by a distribution with finite support. we can assume that ID. cf. Now we can find first a sequence {Dm} of distributions with finite support such that D. with weight pi(n) for xi(n).(Sn) with Sn = n/b. Hence it is immediate that Bn 4 B.(bk) + B(bk) for all k as n * oo. oo).. relies more on matrix algebra than the probabilistic interpretation exploited here). see Neuts [269] (where the proof. 5d Phasetype approximation A fundamental property of phasetype distributions is denseness . say degenerate at b. By the diagonal argument (subsequent thinnings). That is. 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 .2) } to go on (on E(2)) when { i 1) } exits. there is a sequence {B. we need to allow { Jt. Example A5. elementary) Let {bk} be any dense sequence of continuity points for B(x). and vice versa. The mean of B„ is n/Sn = b and the variance is n/Sn = b2/n..} of phasetype distributions such that Bn 3 B as n + oo.
B(bk )I < . oo) such that f (x) = O(e«x). u Theorem A5.. PIT contains all finite mixtures of onepoint distributions. oo).d. say on the claim size distribution B in risk theory.. the class CO of all discrete distributions. we can then approximate Bo by a phasetype B. and we can take Bn = Cr(n).14 is fundamental and can motivate phasetype assumptions. Let E be the class of functions f : [0. Hence G C PET and L = PIT. i = 1. and that cp is known to be continuous.n.r. Then ICr( n ). the topology for weak convergence) PET of the class PET of phasetype distributions contains all onepoint distributions. Corollary A5. oo) and any fl. i. Since PET is closed under the continuous operation of formation of finite mixtures. k < n. replications).i. x 4 oo... u 2 (abstract topological ) The essence of the argument above is that the closure (w.( dx) * f r f{(x)B(dx)..e.APPENDIX 357 Hence we can choose r(n) in such a way that ICr( n). It should be noted. E E. compute W(B) and use this quantity as an approximation to cp(B0). one would use the B given by some statistical fitting procedure (see below).D(bk)I < n. for some a < oo..n( b k ) .. oo) * [0.n (bk) .15 To a given distribution B on (0 . however. f2..(x)Bf. k < n.t. there is a sequence {Bn} of phase type distributions such that Bn Di B as n 4 oo and f ' f. that this procedure should be used with care if ^p(B) is the ruin probability O(u) and u is large. in at least two ways: insensitivity Suppose we are able to verify a specific result when B is of phasetype say that two functionals Cpl (B) and W2 (B) coincide. oo) approximation Assume that we can compute a functional W(B) when B is phasetype. if information on Bo is given in terms of observations (i. In particular. For a general Bo. . 2. But To is the class G of all distributions on [0. If Cpl (B) and ^02(B) are weakly continuous. then it is immediate that WI(B) = p2(B) for all distributions B on [0.
. n.39) Indeed. 2. for each i.f (z) = f = 1 1 1 1n/ o . . and hence it is sufficient to show that we can obtain limsup n4oo fi(x)Bn(dx) < Jo 0 f fi( x)B(dx ).. TO (A.. i=1... we may assume that in the proof of Theorem A5.n(dx) < 1+. if f (x ) = e°x. By (A. .. Bn=En z f f (x)Bn(dx) fof (x)B(dx) = ° (A. then cc f (x)Bn ( dx) = (?!c ) e'= . i = 1.38 ). \\ 0 Corollary A5. and hence we may choose r(n) such that L 9l) f (x)Cr(n)..f ' f (x)B(dx).f (x)B(dx). oo).n(dx) + f 0 fi(x)Dn(dx)..(dx) > J fi(x)B(dx).38) We first show that for each f E E.. f00 fi(x)Cr. f° xtBn(dx ) * f °° x`B( dx). liminf B.  APPENDIX B implies that 00 o o 00 n.2 .358 Proof By Fatou' s lemma. n B=az...16 To a given distribution B on (0 . and the case of a general f then follows from the definition of the class E and a uniform integrability argument. i = 1. i = 1.. . n. .14 Dn has been chosen such that 00 1 °° f fi(x)D n(dx ) < 1++ '  o \ n o f fi(x)B(dx)..39). Now returning to the proof of (A.. there is a sequence {Bn} of phase type distributions such that Bn Di B as n + oo and all moments converge.oo J fi(x)B.
/3) is defined as the unique solution > 0 of B[y] = l+y/j3. /3) = ry for all n. the loggamma or the Weibull have been argued to provide adequate descriptions of claim size distributions.14 is classical. lim sup ryn < 7. but are certainly not unexpected.e. the adjustment coefficient 'y = 7(B.APPENDIX 359 In compound Poisson risk processes with arrival intensity /3 and claim size distribution B satisfying . If ei > 0.} 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. one can obtain 7(Bn. . the remaining results may be slightly stronger than those given in the literature./3).l3µb < 1.. I. oo) with B[y +e] < oo for some e > y = 7(B. 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. Proof Let fi(x) = el'r+E. For practical purposes. and therefore the following result is highly relevant as support for phasetype assumptions in risk theory: Corollary A5. the problem thus arises of how to fit a phasetype distribution B to a given set of data (1. then Bn['Y + ei] * B[y + ei] > 1 + 7 Q implies that 'yn < ry + ei for all sufficiently large n .3)..16. The adjustment coefficient is a fundamental quantity.> y for some sequence {ei} with ei E (0. The present section is a survey of some of the available approaches and software for inplementing this. . (N. from a more conceptual . . (N or a given distribution Bo. 0 as i * oo. . lim inf > is proved similarly. 5e Phasetype fitting As has been mentioned a number of times already. e ) and ei J. . . However.17 To a given /3 > 0 and a given distribution B on (0. O We state without proof the following result: Corollary A5. We shall formulate the problem in the slightly broader setting of fitting a phasetype distribution B to a given set of data (1i . there is substantial advantage in assuming the claim sizes to be phasetype when one wants to compute ruin probabilities. there is a sequence {B. This is motivated in part from the fact that a number of nonphasetype distributions like the lognormal.
A number of approaches restrict the phase type distribution to a suitable class of mixtures of Erlang distributions . Asmussen & Nerman [38] implemented maximum likelihood in the full class of phasetype distributions via the EM algorithm . Of course. and we next describe two such approaches which also have the feature of being based upon the traditional statistical tool of like maximum likelihood. one could argue that the results of the preceding section concerning phasetype approximation contains a solution to our problem : given Bo (or Be). we have constructed a sequence { B. e . and as fitted distribution we may take B.} of phasetype distribution such that Bo. [317] ) has considered an extension of this setup. The earliest such reference is Bux & Herzog [85] who assumed that the Erlang distributions have the same rate parameter. and this is what matters when using phasetype distributions as computational vehicle in say renewal theory.g. Johnson & Taaffe considered a mixture of two Erlangs (with different rates ) and matched (when possible ) the first three moments . The observation is that the statistical problem would be straightforward if the whole ( EAvalued) phase process { Jtk)} o<t<( k associated with each observa . . we do not not want to perform matrix calculus in hundreds or thousands dimensions). risk theory. a program package written in C for the SUN workstation or the PC is available as shareware. [202]. giving mass 1 /N to each S=. The characteristics of all of these methods is that even the number of parameters may be low (e. defined by the absence of loops in the phase diagram .360 APPENDIX point of view the two sets of problems are hardly different : an equivalent representation of a set of data (1 . (N is the empirical distribution Be. for some suitable large n.d. g. It seems therefore a key issue to develop methods allowing for a more general phase diagram. the L1 distance between the c . 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. at a a number of selected points .f. cf.. In a series of papers (e. reliability or queueing theory.'s). [70]) restrict attention to acyclic phase type distributions . and used a nonlinear programming approach . The likelihood function is maximized by a local linearization method allowing to use linear programming techniques. Schmickler (the MEDA package. A method developed by Bobbio and coworkers (see e.g . and in practice this sets a limitation to the usefulness (the curse of dimensionality .f. the number of phases required for a good fit will typically be much larger. d.g. where more than two Erlangs are allowed and in addition to the exact matching of the first three moments a more general deviation measure is minimized (e.. [216] ). three for a mixture of two Erlangs ). B„ The problem is that the constructions of {B„} are not economical : the number of phases grows rapidly.. .g..
the methods of [70] and [38] appear to produce almost identical results. .(N) = E Ea(n). since this is parameterdependent.APPENDIX 361 tion Sk was available. EN where ai = N 1 I ((k) = i) tii=i iEE. E. The general idea of the EM algorithm ([106]) is to replace such unobserved quantities by the conditional expectation given the observations.g. = j) f k=1 k =1 tE[0. e. it seems open whether the restriction to the acyclic case is a severe loss of generality. Nii = = .(k] (Ti is the total time spent in state i and Nii is the total number of jumps from i to j). one is lead to an iterative scheme.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). . N Ti = I(J= i) dt.. (N) tJk Ea ( n). . In practice. Thus. .. jEEA... .T(n) k=1 I (Jti) dt o \f a(n)eT(n )(kt(n) N f:i a(n)eT(n)xei ...T(n) (Ti ^^ 1.T(n) (Nik IC1.. it is easy to see that N (k Ea(n). eieT(n)((k. (N ) (^ 54 k )+ and similarly for the cn+1) The crux is the computation of the conditional expectations. In fact.g.T (n)(TiI(1.. then the estimators would be of simple occurenceexposure type.
This page is intentionally left blank .
14. [3] J. Appl.A. [13] S. Adv. Preprint. [15] S. 1984 . ed. Abate & W. Skand. ibid. AT&T. [12] S. [8] G. Proc. New York. pp. Borokov. Asmussen (1984) Approximations for the probability of ruin within finite time . John Wiley & Sons. 219229. Abramowitz & I. 20. Arfwedson ( 1954) Research in collective risk theory. Whitt ( 1992) The Fourierseries method for inverting transforms of probability distributions . Abate. 587. Anantharam (1988 ) How large delays build up in a GI/GI11 queue. New York. G.G. 143170. Asmussen (1987) Applied Probability and Queues. Dover. Sparre Andersen (1957) On the collective theory of risk in the case of contagion between the claims . Scand. Optimizations Software. 37. 213229. Stegun ( 1972 ) Handbook of Mathematical Functions (10th ed. Probab. Choudhury & W. 345368. Aktuar Tidskr. Arndt (1984) On the distribution of the supremum of a random walk on a Markov chain . Queueing Systems 5 . [9] G. 38. [5] G. 53100. Abate & W. Asmussen (1989a) Aspects of matrix WienerHopf factorisation in applied probability. Arfwedson ( 1955) Research in collective risk theory. Whitt (1994) Waitingtime tail probabilities in queues with longtail service time distributions . with applications to risk reserve processes and the GI /G/1 queue. The case of equal risk sums. 311338. Act. Queueing Systems 16. Whitt (1998) Explicit M/G/1 waitingtime distributions for a class of longtail service time distributions .L. [4] M. [2] J. The Mathematical Scientist 14. New York. B. Stoch. [14] S. Skand. 191223. Appl.). 253267. Aktuar Tidskr. 1985. Chichester New York. Queueing Systems 10. 363 . Alsmeyer (1991) Erneuerungstheorie. Transactions XVth International Congress of Actuaries. 3157 . In: Limit Theorems and Related Problems (A.). 101116.Bibliography [1] J. [10] K. Asmussen ( 1982 ) Conditioned limit theorems relating a random walk to its associate. [7] E. Teubner. Asmussen (1985) Conjugate processes and the simulation of ruin problems. [11) S. J. [6] V. II. 57. Stuttgart.
J. [23] S. Scand. 772789. New York. Kliippelberg (1994) Large claims approximations for risk processes in a Markovian environment. Florida. 2.K.364 BIBLIOGRAPHY [16] S. Asmussen (1998c) A probabilistic look at the WienerHopf equation. 1989 . [26] S. 137168. [19] S. Rolski & V. Act. Frey. [31] S. 303322. 54. Act.). Appl. [22] S. Ann. SIAM Review 40. Ann. Asmussen (2000) Matrixanalytic models and their analysis. eds. Bladt (1996) Phasetype distributions and risk processes with premiums dependent on the current reserve. Stoch. 2149. K. Asmussen. Proc. 20. Bernoulli 6. Schmidt (1995) Does Markovmodulation increase the risk? ASTIN Bull. Binswanger & B. [24] S. [20] S. Stoch. Alfa & S. Asmussen (1995) Stationary distributions via first passage times. MatrixAnalytic Methods in Stochastic Models (A. [27] S. Asmussen (1992b) Light traffic equivalence in single server queues. 37. Asmussen. [25] S. 106119. 1936. J. Astin Bulletin 27. 2943.). Dshalalow ed. Methods V Problems (J. Hojgaard (1996) Finite horizon ruin probabilities for Markovmodulated risk processes with heavy tails. 1999 . Asmussen (1992c) Stationary distributions for fluid flow models and Markovmodulated reflected Brownian motion. Andronov. Asmussen & B. Advances in Queueing: Models. Boca Raton. Hojgaard (2000) Rare events simulation for heavytailed distributions. Random Processes 2. Scand. Appl. Proc. 25. 27. [17] S. Asmussen & K.M. Asmussen (1998b) Extreme value theory for queues via cycle maxima. 354374. L. 4966. [18] S. [29] S. [34] S. Appl. Probab. Extremes 1 . Appl. Asmussen (1999) On the ruin problem for some adapted premium rules.S. 8. J. Asmussen (1991) Ladder heights and the Markovmodulated M/G/1 queue. Asmussen & M. [30] S. 315. Asmussen (1998a) Subexponential asymptotics for stochastic processes: extremal behaviour. Statist. [33] S. Riga Aviation University. [28] S. Floe Henriksen & C. Probab. Ann. 96107. 555574. 313326. . 297318. Kalashnikov & A. Stochastic Models 11. eds. CRC Press. Asmussen & M. 1996. A. Asmussen (1989b) Risk theory in a Markovian environment. Th. Hojgaard (1999) Approximations for finite horizon ruin prob abilities in the renewal model. Chakravarty. Asmussen. 69100. Binswanger (1997) Simulation of ruin probabilities for subexponential claims.). T. Scand. J. [21] S. Probab. 189201. 193226. 79102. Marcel Dekker. Scand. Act. 313341. Asmussen (1992a) Phasetype representations in random walk and queueing problems. stationary distributions and first passage probabilities. Asmussen & B. Probabilistic Analysis of Rare Events (V. Bladt (1996) Renewal theory and queueing algorithms for matrixexponential distributions. [32] S.
Taksar (2000) Optimal risk control and dividend distribution policies . Asmussen & R. Prob 32 . Nerman & M. Asmussen & S. [42] S. 115. Asmussen. Fitting phasetype distributions via the EM algorithm. Kliippelberg (1996) Large deviations results for subexponential tails. Statist. 31. Asmussen & C. O'Cinneide (2000/01) On the tail of the waiting time in a Markovmodulated M/G/1 queue. Supplementary Volume (Kotz. Read eds.Y. [43] S. Rolski (1991) Computational methods in risk theory: a matrixalgorithmic approach. Perry (1992) On cycle maxima. Asmussen . B.M. Asmussen & V. Stoch. Res. 299324. Asmussen & M. 20. Wiley. [44] S. H. Insurance: Mathematics and Economics 10. Asmussen & G. Appl. [46] S. Olsson (1996). CRC Press. Opns. 11251141. (to appear). Probab. [50] S. [45] S. 410433.Y. Methods 8 Problems (J. Advances in Queueing: Models. Schock Petersen (1989) Ruin probabilities expressed in terms of storage processes. [47] S. 64.BIBLIOGRAPHY 365 [35] S. Boca Raton. 105119. Scand. Inf. Asmussen & D. . O. [51] S. Asmussen. 120. Dshalalow ed. Schmidt (1995) Ladder height distributions with marks. [37] S. Appl. Nielsen (1995) Ruin probabilities via local adjustment coefficients . Hojgaard & M. Th. Management Science 45. 365372. Rubinstein (1999) Sensitivity analysis of insurance risk models. Appl. 30. Example of excessoffloss reinsurance for an insurance corporation. Marked point processes as limits of Markovian arrival streams. Koole (1993). Asmussen & T. Res.). Statistica Neerlandica 47. Finance and Stochastics 4. 419441. with applications to insurance risk Stoch. 913916. Appl. Appl. [38] S. Rubinstein (1995) Steadystate rare events simulation in queueing models and its complexity properties. Schmidli & V. [52] S. [40] S. Sc. Proc. Appl. Probab.A. Stochastic Models 8. Rolski (1994) Risk theory in a periodic environment: Lundberg's inequality and the CramerLundberg approximation. Oper. O'Cinneide (2000/2001) Matrixexponential distributions [Distributions with a rational Laplace transform] Encyclopedia of Statistical Sciences. 103125. J. Probab. Insurance: Mathematics and Economics 20. [48] S. Probab. Asmussen & K. 736755.A. J. Asmussen & H. Adv. Asmussen & T. Asmussen & V. [39] S. 429466. 23. Schmidt (1993) The ascending ladder height distribution for a class of dependent random walks. 19. [49] S. Math. [41] S. Florida. Schmidt (1999) Tail approximations for nonstandard risk and queueing processes with subexponential tails. Eng. [36] S. Adv. Proc. Taksar (1997) Controlled diffusion models for optimal dividend payout. Asmussen & C. Sigman (1996) Monotone stochastic recursions and their duals. Asmussen & R. J. 10. first passage problems and extreme value theory for queues. Asmussen & C. 422447. Johnson. 259274. 421458. 58.).
[69] P.366 BIBLIOGRAPHY [53] S. 9. J.J. 275279. Insurance: Mathematics and Economics 4. Schmidli (1995) Saddlepoint approximations for the probability of ruin in finite time. Scand. J. Act.B . 1988 . New York. 9599. 1995. Bingham. Goldie & J. Probab. Soc.M. Bloomfield & D. Springer . Scand. Beekman (1969) A ruin function approximation. von Bahr (1975) Asymptotic ruin probabilities when exponential moments do not exist. Scand. [73] A.A. SpringerVerlag. Boogaert & V. Berlin. C. [72] P. J. Bobbio & M. 148156. Crijns (1987) Upper bounds on ruin probabilities in case of negative loadings and positive interest rates. [59] O. [58] O. Grandell (1988) Exponential inequalities for ruin probabilities in the Cox case. 832840. Bjork & J. 7. Stochastic Models 10. Act. Trans. Actuaries 21.H. Grandell (1985) An insensitivity property of the ruin probability.H. Baker (1977) The Numerical Solution of Integral Equations. 1974. Oxford. Wiley. [71] P. Cambridge University Press. Act. 610. 661667. Asmussen & J.H. Teugels ( 1996) Convergence rates for M /G/1 queues and ruin problems with heavy tails . Bjork & J. 1985 . J. [62] J.A. [65] N. Halsted Press. [66] N. 169186. Borovkov (1976) Asymptotic Methods in Queueing Theory. [60] J.T. J. [64] P. Cambridge. de Waegenaere (1990) Simulation of ruin probabilities. Berman & R. 4148. Billingsley (1968) Convergence of Probability Measures. [67] T. Clarendon Press. Beekman (1985) A series for infinite time ruin probabilities. Beekman (1974) Two Stochastic Processes.L. Wiley. Act. Ney ( 1972) Branching Processes . [63] A. Teugels (1987) Regular Variation. 705766. Scand. [55] B. [68] T. BarndorffNielsen (1978) Information and Exponential Families in Statistical Theory. Appl. Appl. 190204. von Bahr (1974) Ruin probabilities expressed in terms of ladder height distributions. Plemmons (1994) Nonnegative Matrices in the Mathematical Sciences. New York. 129134. Insurance: Mathematics and Economics 6. Adv. Telek (1994) A bencmark for PH estimation algorithms: results for acyclic PH.L. New York. Probab. Cox (1972) A low traffic approximation for queues. [54] K. 1975. [70] A. [57] C. Probab. J. Appl. [61] J. Insurance: Mathematics and Economics 9. BarndorffNielsen & H. 77111. Scand. 11811190.R. Bingham (1975) Fluctuation theory in continuous time. Act. Athreya & P. [56] B. SIAM. Boogaert & A. . 221232. 33. J.
Computer Performance (K. H.R. Malgouyres (1983) Large deviations and rare events in the study of stochastic algorithms . Cohen (1982) The Single Server Queue (2nd ed. Croux & N.Y . Hickman.L. Berlin. [87] E.M. B. Queueing Systems 33. Chung (1974) A Course in Probability Theory (2nd ed. [86] K. Insurance : Mathematics and Economics 9. Res. New York San Francisco London. Gerber. Smith ( 1983) Asymptotic analysis of a queueing system with bursty traffic. . Amsterdam. Reading. Bell. [93] K. [90] D. New York. [80] F. Cottrell. 2338. Area Commun. IEEE Trans. NorthHolland. [77] P. Z. [88] J. Nesbitt (1986) Actuarial Mathematics . Stockholm. II. Bucklew ( 1990) Large Deviation Techniques in Decision.J. [85] W. Control AC28.W. Herzog (1977) The phase concept : approximations of measured data and performance analysis. 749763. Sel. Fort & G. Itasca. [78] L. Fox & L.J. Boxma & J. 177204. Philos.A. 313319. Chandy & M. Adv. [84] D. Tech. Syst.I. 907920.Y. Cramer (1930) On the Mathematical Theory of Risk. 62. Wiley.R. Jones & C.C. Veraverbeke (1990). Geb.R. Bowers . Resnick & R. Oper. Schrage (1987) A Guide to Simulation. AddisonWesley.). Stockholm. D. IEEE J.) NorthHolland. Cambr. Cohen (1999) Heavytraffic analysis for the GI/G/1 queue with heavytailed service time distributions . Aut. 51. 34. [92] H.L. 127130. The Jubilee volume of Forsakringsbolaget Skandia. Wahrscheinlichkeitsth. S. [76] N. Burman & D . 93121. Burman & D. Qinlar (1972) Markov additive processes. Bratley. Breiman ( 1968) Probability. Tweedie (1982) Storage processes with general release rule and additive inputs . Smith ( 1986) An asymptotic analysis of a queueing system with Markovmodulated arrivals . Boxma & J. [79] P.. Broeck .BIBLIOGRAPHY 367 [74] O. 392433. Simulation and Estimation. Springer. Amsterdam. Bux & U. Reiser eds.A. Biihlmann ( 1970) Mathematical Methods in Risk Theory.L. [82] H. De Vylder (1986) Ordering of risks and ruin probabilities. Proc. 16.U. The Society of Actuaries . Appl. [83] D.W. [91] H. 24. Brockwell . Cox (1955) Use of complex probabilities in the theory of stochastic processes. Skandia Jubilee Volume. Academic Press. J. SpringerVerlag. 14. [81] J. [75] O. Soc. Probab. Illinois. Insurance : Mathematics and Economics 5. Goovaerts & F. [89] M.).C. 3539. Nonparametric estimators for the probability of ruin. Cramer (1955) Collective risk theory. New York. J. J. 14331453.L. 105119. M. Cohen (1998) The M/G/1 queue with heavytailed service time distribution.W. Jr. verve.
[98] A. Embrechts (1989). Aktuar. Act. [108] F. Haezendonck (1985) Inversed martingales in risk theory. J. 114119. Teugels ( 1990) Empirical Laplace transform and approximation of compound distributions . J. Scand. [99] R. N. [101] C. Martingales and insurance risk.B. Insurance: Mathematics and Economics 10. Jones and Bartlett. Steinebach (1990) On some alternative estimators of the adjustment coefficient in risk theory. [104] F. Springer. Insurance: Mathematics and Economics 6. [111] F. Daykin. English version published in Skand. Tidskr. Stockholm. [96] D. 17. 435436. 9 .M.P. [109] F. Probab. Laird & D. Csorgo & J. Haezendonck (1987) Classical risk theory in an economic environment. Delbaen & J. Suppl. Rolski (1991) Light traffic approximations in queues. Insurance: Mathematics and Economics 7. New York. 181217. Math. Deheuvels & J. Comp. De Vylder (1978) A practical solution to the problem of ultimate ruin probability. [105] A. Devroye (1986) Nonuniform Random Variate Generation. Math. van Dawen (1986) Ein einfacher Beweis fur Ohlins Lemma. Rubin (1977) Maximum likelihood from incomplete data via the EM algorithm. Insurance: Mathematics and Economics 3. Goovaerts (1984) Bounds for classical ruin probabilities. Soc. Csorgo & J. . Zeitouni (1992) Large Deviations Techniques and Applications. 135159.. Dembo & O. 1990 . [97] D. 201206. De Vylder (1996) Advanced Risk Theory. [102] P.368 BIBLIOGRAPHY [94] M. J.L. Delbaen & J. 227229.. Dempster. 624628. De Vylder & M. Actuarial J. Dassios & P. Davidson (1946) On the ruin problem of collective risk theory under the assumption of a variable safety loading (in Swedish). [103] F. Rolski (1984) A light traffic approximation for a single server queue. Steinebach ( 1991 ) On the estimation of the adjustment coefficient in risk theory via intermediate order statistics . 3750. Pesonen (1994) Practical Risk Theory for Actuaries. Oper. Scand. 7083 (1969). [100] A. T. 3. 583602. [112] F. 22. 85116. Res. Res. De Vylder & M. 121131. 88101.D. Pentikainen & E. Stochastic Models 5. Forsiikringsmatematiska Studier Tillagnade Filip Lundberg. J. Daley & T. Blotter der deutschen Gesellschaft fur Versicherungsmathematik XVII. Chapman & Hall. Goovaerts (1988) Recursive calculation of finitetime ruin probabilities. [95] M. Appl. 16. 5771. Boston. Insurance: Mathematics and Economics 4. De Vylder (1977) A new proof of a known result in risk theory. Roy. Editions de 1'Universite de Bruxelles. Daley & T. [107] L. Appl. Oper. [106] A. Math. [110] F. Statist. 27.
Hipp (1999) Ruin problems for phasetype(2) risk processes. 251262. 147167. Act. Dufresne. Insurance: Mathematics and Economics 19. Math. [116] D. Schmidli (1993) Finitetime Lundberg inequalities in the Cox case . 1993 .M. Dufresne & H.BIBLIOGRAPHY 369 [113] D.M. 1984 . Dickson & J. Regterschot (1988) Conditional PASTA. 363374.C. Gerber & E. Scand.C. and the amount of claim causing ruin. H. J. Soc. Duffield & N. Dickson & H. 193199. 105115. Berlin Gottingen Heidelberg. van Doorn & G.G. 1975 . J. Dickson (1992) On the distribution of the surplus prior to ruin . J. Waters (1996 ) Reinsurance and ruin . Hipp (1998) Ruin probabilities for Erlang (2) risk processes. Act. [122] B. Waters (1999) Ruin probabilities wih compounding. 1994 . 2000 . J.M. Insurance : Mathematics and Economics 7. Embrechts .J. [114] D. 6180. [125] F. Oper. 1741.R.U. 269274. Dickson & C. 4259.M.C. Shiu (1991) Risk theory with the Gamma process.U. 131138.C. Ruin estimates for large claims . [118] D. [126] F. Insurance : Mathematics and Economics 10.U. Grandell & H. O'Connell (1995) Large deviations and overflow probabilities for the general singleserver queue. Act.R. 4962. Emanuel . [119] D.M. [121] D. J. Astin Bulletin 21. Scand. Gerber (1988) The surpluses immediately before and at ruin. Gray (1984) Exact solutions for ruin probability in the presence of an upper absorbing barrier. Taylor ( 1975) A diffusion approximation for the ruin probability with compounding assets. 1984 . Insurance: Mathematics and Economics 25.M. 3745. 174186. [130] P. Res. [123] E. Gray (1984) Approximations to the ruin probability in the presence of an upper absorbing barrier.K.C. Philos. Camb. J. Dynkin (1965) Markov Processes I. Embrechts ( 1988 ). Dufresne & H. [127] F. British Actuarial J.S. Insurance : Mathematics and Economics 22. [128] E. [117] D. [120] D.C. Act. Dickson & J.M. 191207. Letters 7. Insurance : Mathematics and Economics 11. Scand. Scand.M. J. Dickson (1995) A review of Panjer's recursion formula and its applications.B. [131] P. Dickson & C. 118.J.C. Scand. Insurance: Mathematics and Economics 7. [115] D.C. 107124. . Act. Djehiche (1993) A large deviation estimate for ruin probabilities. Act. 1. Springer . Proc. 229232.M.C. Act. [129] D . [124] N.W. 5159. Harrison & A. 177192. Scand.R. J. Scand. with applications. Gerber (1991) Risk theory for the compound Poisson process that is perturbed by a diffusion.R.M. J. Dickson & H. 1993 .C. Dickson (1994) An upper bound for the probability of ultimate ruin.
Ann.). 616624. [142] W. Appl. Veraverbeke (1982) Estimates for the probability of ruin with special emphasis on the possibility of large claims . [134] P. Villasenor (1988) Ruin estimates for large claims. Math. 3339. Franken. Nyt Tidsskrift for Matematik B20. 695712.L. Adv. New York. F.J.A. Embrechts. Frenz & V. Insurance: Mathematics and Economics 10. Approximations for compound Poisson and Polya processes. Sci. R. [150] H. Insurance: Mathematics and Economics 7. Esscher (1932) On the probability function in the collective theory of risk. Wiley. Insurance : Mathematics and Economics 1. Springer. Rel. Schmidli (1994) Ruin estimation for a general insurance risk model.K. 404422. Embrechts & N.K. Danish Academy Tech. Scand. 175195. D. [147] M. [149] C. J. Adv. first passage times and moments of ladder variables in Markov random walks. 623637. Fuh (1997) Corrected ruin probabilities for ruin probabilities in Markov random walks. Fuh & T.D. Heidelberg. J. Statist. Probab. Fields 75. 5974. [144] P. Konig. [137] P. Pitts (1993) Some applications of the fast Fourier transform in insurance mathematics . 5975.). Trans. Statistica Neerlandica 47. [145] P. On the excursions of Markov processes in classical duality. Arndt & V.W.W. Erlang 2. [135] P. Appl. J. Th. Probab. C. Probab. [140] J.D. 17. Astin Bulletin 16.L. J. Furrer (1998) Risk processes perturbed by astable Levy motion. Probab. [139] A. Appl. Skand. Mikosch (1997) Modelling Extremal Events for Finance and Insurance. [143] W.370 BIBLIOGRAPHY [132] P. Erlang (1909) Sandsynlighedsregning og telefonsamtaler . 566580. Walkup (1967) Association of random variables. Schmidt (1982) Queues and Point Processes. [148] C. Embrechts. Embrechts & J. [141] F. Jensen. Probab. 131137. Feller (1971) An Introduction to Probability Theory and its Applications II (2nd ed. Fitzsimmons (1987). Teugels (1985).M. Kliippelberg & T. Lai (1998) Wald's equations. U. A bootstrap procedure for estimating the adjustment coefficient. 29. 159178. Adv. [146] E. Akt. Esary. [133] P. Schmidt (1992) An insensitivity property of ladder height distributions. Wiley. 8190. 14661474. Griibel & S. 5572. . Frees (1986) Nonparametric estimation of the probability of ruin. Probab. Embrechts & H.D. 181190. Embrechts. Embrechts & T. [138] P. Wiley. 1998 . with applications. 26. M. Maejima & J. 269274. Appl. Tidsskr. Reprinted 1948 as 'The theory of probabilities and telephone conversations' in The Life and work of A. Mikosch (1991). Act. Proschan & D. 35. 29. Appl. 38. [136] P. Feller (1966) An Introduction to Probability Theory and its Applications I (3nd ed. New York.
van Heerwarden & T.BIBLIOGRAPHY 371 [151] H. 1977. contained in the authors PhD thesis. J. Insurance : Mathematics and Economics 7. Amsterdam. Act. [168] J. Ver. 2336. 7778. Gerber (1973) Martingales in risk theory. eds.). Appl. J. F. Whitt (1994) Logarithmic asymptotics for steadystate tail probabilities in a singleserver queue. Vers.W. Aktuarietidskrift 1970 . Mitt. 131156. Springer. Appl.). Basel. Ver. Gerber (1981) On the probability of ruin in the presence of a linear dividend barrier. [152] H. 205216. [160] H. Math. Vers. [159] H.J. 5.S. Glasserman (1991) Gradient Estimation via Perturbation Analysis. Mitt. Appl. [155] H. Haezendonck (1984) Insurance Premiums. 6370. 205210. Michna & A. [156] H. 1978 . 105115. NorthHolland. ETH Zurich. University of Pennsylvania. Gerber (1979) An Introduction to Mathematical Risk Theory. J. Kluwer. Gerber (1986). Math. NorthHolland. . Insurance: Mathematics and Economics 15. Scand. de Vylder & J. A. 73. [169] J. Schweiz.U. Goovaerts. Z. Probab. [163] P. Kovalenko (1989) Introduction to Queueing Theory (2nd ed.U. Act. Furrer & H. Act. R.M. Schmidli (1994) Exponential inequalities for ruin probabilities of risk processes perturbed by diffusion. Boston. 424445. Goldie & R. [167] C. 97114. [161] P.J. Furrer.. London. Scand. 3752. 1523.U. Gnedenko & I. Glynn & W.U. Probab. Huebner Foundation Monographs. Dordrecht.U. Goovaerts. [157] H. Grandell (1978) A remark on 'A class of approximations of ruin probabilities'. 28. J. Ann. Gerber (1971) Der Einfluss von Zins auf die Ruinwahrscheinlichkeit. Life Insurance Mathematics.71..U. Grandell (1977) A class of approximations of ruin probabilities. Galambos & J. [166] M. Suppl. Schweiz. Glasserman & S. [154] H. S. 463480. Gani.G. Bauwelinckx (1990) Effective Actuarial Methods.N. Unpublished. Adv.U. Amsterdam. Probab. Kaas. In Studies in Applied Probability (J. [165] M. [153] H. Furrer (1996) A note on the convergence of the infinitetime ruin probabilities when the weak approximation is astable Levy motion. Gerber (1988) Mathematical fun with ruin theory. Griibel (1996) Perpetuities with thin tails. 31A. Gerber (1970) An extension of the renewal equation and its application in the collective theory of risk. [158] H. Kou (1995) Limits of first passage times to rare sets in regenerative processes. 1981 . [162] P.V. [164] B. Weron (1997) Stable Levy motion approximation in collective risk theory. Birkhauser .E. Scand. Skand. Insurance: Mathematics and Economics 20.
Grigelionis ( 1992) On Lundberg inequalities in a Markovian environment. 8. Grnbel (1991) G/G/1 via FFT. Grandell (1997) Mixed Poisson Processes. 3041. Math. M.I. Res.S. [181] D.V. 1. 197214. Skand.A. Gusak & V.I. [189] A. 6779. [188] J. KTH. 566576. [175] J. Korolyuk (1969) On the joint distribution of a process with stationary increments and its maximum . [178] B. [182] A. 143158. Informatica 2. M. Aktuar Tidskr. 14. Amsterdam. [180] R. 332. Harrison (1977) Ruin problems with compounding assets . M. M. 5 . New York.and Sozialforschung 6. Appl. [186] J. [173] J. Probability Theory and Mathematical Statistics (I. [177] B. Proc.O. Applied Statistics 40. van Heerwarden (1991) Ordering of Risks: Theory and Actuarial Applications. Appl. Hermesmeier (1999) Computation of compound distributions I: aliasing errors and exponential tilting. 355365. Harrison & S. [187] J. Grandell ( 1992) Finite time ruin probabilities and martingales . E. [172] J. [184] H . 5766. Resnick (1976 ) The stationary distribution and first exit probabilities of a storage process with general release rule . Gyllenberg & D. 14. [171] J. Rink. Grigelionis ( 1993) Twosided Lundberg inequalities in a Markovian environment. Grandell (1999) Simple approximations of ruin functions. Th. 167176. Statistical Algorithm 265. SpringerVerlag. Matem. 7590. Grandell (1979) Empirical bounds for ruin probabilities.243255. [183] M. Theory and Applications. Stoch. J. 3 . 33. ed.131135. Math. AkademieVerlag. Winter School on Stochastic Analysis and Appl . Opns . Griibel & R. [179] R. Proc . Grandell & C. Harrison & S. 347358. Res.372 BIBLIOGRAPHY [170] J. SpringerVerlag. Opns . [185] J. Hadwiger (1940) Uber die Wahrscheinlichkeit des Ruins bei einer grossen Zahl von Geschiiften . [176] B. Preprint. 400409. Gut (1988 ) Stopped Random Walks . Appl. Grigelionis (1996) Lundbergtype stochastic processes. Silvestrov (2000) Cram€rLundberg approximations for nonlinearly perturbed risk processes .. Tinbergen Institute Research Series 20. Liet. mathematische Wirtschaft . Appl. Berlin. Grandell (1990) Aspects of Risk Theory. Chapman & Hall. 54.). Stoch. Ibragimov. Segerdahl (1971) A comparison of some approximations of ruin probabilities. [174] J. Insurance: Mathematics and Economics 26. Lemoine (1977) Limit theorems for periodic queues.J. Astin Bulletin 29. Probab. Harrison & A. Archiv v. Probab. Proc. . Resnick (1977) The recurrence classification of risk and storage processes .
Klugman (1984) Loss Distributions.L. 25.V. 29. eds. Wiley. Huskova. Statist. Hoglund (1974) Central limit theorems and statistical inference for Markov chains. Scand. Mathematical Statistics.A. [204] T. Ann. Asmussen & 0. Z. Hipp & R. [206] B.V. Probab. Heidelberger (1995) Fast simulation of rare events in queueing and reliability models. Willekens (1986) Estimates for the probability of ruin starting with a large initial reserve. Verlag Versicherungswirtschaft e. Herkenrath (1986) On the estimation of the adjustment coefficient in risk theory by means of stochastic approximation procedures. [192] U. Scand. New York. [193] U. Foruth Prague Symp. 12981310. Wahrscheinlichkeitsth. Insurance: Mathematics and Economics 5. [196] C.. Astin Bulletin 19. 1998 .BIBLIOGRAPHY 373 [190] P. Hesselager (1990) Some results on optimal reinsurance in terms of the adjustment coefficient. 6. J. 11631174. 1990 . 19. . [205] T. Staalhagen (1999) Waiting time distributions in M/D/1 queueing systems. Nachrichten 30. Appl. Studies in Statistical Quality Control and Reliability 1992 :4. 285293. Appl. 3. No. Hermann (1965) Bin Approximationssatz fiir Verteilungen stationarer zufalliger Punktfolgen. [208] V. on Asymptotic Statistics (P. [191] W. Ann. Hoglund (1991). Prob. [194] 0. Insurance: Mathematics and Economics 5. Horvath & E. 377381. Mandl & M. Proc. J. 259268. Hojgaard & M. Hogan (1986) Comment on corrected diffusion approximations in certain random walk problems. Hill (1975) A simple general approach to inference about the tail of a distribution. 378389. Probab. Taksar (2000) Optimal proportional reinsurance policies for diffusion models. 18. [200] M. 166180.R. S. Karlsruhe. Hogg & S. Karlsruhe. Michel (1990) Risikotheorie: Stochastische Modelle and Statistische Methoden. 285292. Ann. Hipp (1989b) Estimators and bootstrap confidence intervals for ruin probabilities. verve. [198] C. Nerman (1992) EMPHT . [202] 0. Act. Probab. [197] C. The ruin problem for finite Markov chains.).L. Versicherungswirtschaft. [207] D.a program for fitting phasetype distributions. [195] B. Act. 123151. 5770. [201] L. Iglehart (1969) Diffusion approximations in collective risk theory. 8095. Electronic Letters 35. Heilmann (1987) Grundbegriffe der Risikotheorie. ACM TOMACS 6. 305313.M. [203] T. Iversen & L. 8996.B. J. 4385. Geb. Hoglund (1990) An asymptotic expression for the probability of ruin within finite time. Hipp (1989a) Efficient estimators for ruin probabilities. 23. [199] R. Haggstrom. J. Chalmers University of Technology and the University of Goteborg. Math.
Wishart (1964). Kalashnikov (1997) Geometric Sums: Bounds for Rare Event with Applications. [215] J. [221] E. Soc. Technometrics 30. Academic Press.J. Taylor (1981) A Second Course in Stochastic Processes. 11. Act. Janssen & J. [219] V. [225] J. J.M.C. 187193. 165167. 63.L. ibid. [214] A. Oxford. Reinhard (1985) Probabilities de ruine pour une classe de modeles de risque semiMarkoviens. [216] M. [226] J. Astin Bull. 6. Jagerman (1991) Analytical and numerical solution of Volterra integral equations with applications to queues. [212] J. Munksgaard. 123133. N. Copenhagen. Wishart (1964) Boundary problems for additive processes defined on a finite Markov chain. Keilson (1966) A limit theorem for passage times in ergodic regenerative processes. 6. Cambridge Philos.A. 325247. [210] D.R. Johnson & M. [217] R. [213] P. Ann. 173190. Janssen (1980) Some transient results on the M/SM/1 special semiMarkov model in risk and queueing theories. 4151. Kaas & M. Appl. NEC Research Institute. 561563.B. Manuscript. Clarendon Press. [223] J.374 BIBLIOGRAPHY [209] D. [211] J. Proc.H.M. Probab. 118. 1996. 61. Insurance: Mathematics and Economics 5. 31. Stochastic Models 1. Amer. New York. Goovaerts (1986) General bound on ruin probabilities. Soc. 8793. N. J. Keilson & D.G. 283306. Keilson & D. Statist. 259281.M. Astin Bull. Lazar (1998) Subexponential asymptotics of a Markovmodulated random walk with a queueing application. Addenda to for processes defined on a finite Markov chain. Jagerman (1985) Certain Volterra integral equations arising in queueing.P. 3. Act. [228] J. 37. Probab. Math. Princeton. Kao (1988) Computing the phasetype renewal and related functions. Taaffe (1989/90) Matching moments to phase distributions. Karlin & H. Jensen (1953) A Distribution Model Applicable to Economics. Jelenkovic & A. 35. Proc. University of Chicago Press. 239256. Stochastic Models 5. Wishart (1964) A central limit theorem for processes defined on a finite Markov chain. [224] J. 711743. Keilson & D.M. Kluwer. Cambridge Philos. 15. [227] J.J. [220] V. 547567. Soc. J. Appl. J. Jensen (1995) Saddle Point Approximations. Chicago. 116129. Kemperman (1961) The Passage Problem for a Markov Chain. Kalashnikov (1996) Twosided bounds of ruin probabilities. Kalashnikov (1999) Bounds for ruin probabilities in the presence of large claims and their comparison. Cambridge Philos. 866870.M.G. Proc. . 60.G. Kennedy (1994) Understanding the WienerHopf factorization for the simple random walk. [222] S. ibid. [218] V. Scand.
F. SIAM. 154168.F. Cambr.C. 146. 6075. Scand. Soc. J. Insurance : Mathematics and Economics 12. Probab. J. Appl. Roy.a Linear Algebraic Approach. Appl. Lucantoni (1991) New results on the single server queue with a batch Markovian arrival process. (235] C.F. 279285. Penev & A.P. Probab. Nyrhinen (1992b) On asymptotically efficient simulation of ruin probabilities in a Markovian environment. Lecture Notes in Mathematics 511. 4477. Klnppelberg & T.C. Bernoulli 1 .BIBLIOGRAPHY 375 [229] J. 132141. Probab. 25. Lemoine (1989) Waiting time and workload in queues with periodic Poisson input. SpringerVerlag. [246] D. Nyrhinen (1992a) Simulating levelcrossing probabilities by importance sampling. Kingman (1964) A martingale inequality in the theory of queues. Kuchler & M. Seminaire de Probabilties X . Mikosch (1995) Explosive Poisson shot noise with applications to risk retention. 259264. [239] H. Macmillan. J. Mikosch (1995) Modelling delay in claim settlement. B24. [242] A. [247] L. Scand. Camb. Actuarial J.C. 390397.S. 60 . Act. [245] T. Prob. Appl. Lipsky (1992) Queueing Theory . J.J.J. [241] G. [230] J. 133135 (in Russian). Kunita (1976) Absolute continuity of Markov processes. [244] T. [237] C. [248] D. Klnppelberg (1993) Asymptotic ordering of risks and ruin probabilities. J. [234] C. Turbin (1973) An asymptotic expansion for the absorption time of a Markov chain distribution.F. Philos. New York. Kingman (1962) On queues in heavy traffic. [243] A. 125147. 4958. Quart. [238] V. Lehtonen & H. Cybernetica 4. 26. 48. Kingman (1961) A convexity property of positive matrices . I. 1995 . Quart. 283284. Stadtmiiller (1998) Ruin probabilities in the presence of heavytails and interest rates . Scand. 383392. Klnppelberg & T. Act. [240] U. 858874. Soc. Latouche & V. Adv. Ramaswami (1999) Introduction to MatrixAnalytic Methods in Stochastic Modelling. 24. Kliippelberg (1988) Subexponential distributions and integrated tails. [236] C. Sorensen (1997) Exponential Families of Stochastic Processes. Proc. 277289. 1998 . [231] J. Math.. Klnppelberg (1989) Estimation of ruin probabilities by means of hazard rates. 18. Appl. Lehtonen & H. [232] C. Philos. Kliippelberg & U. Statist. Lemoine (1981) On queues with periodic Poisson input. . SpringerVerlag. [233] C. Soc. Korolyuk. 889900. Insurance: Mathematics and Economics 8. Stochastic Models 7. J. Proc. J. Lindley (1952) The theory of queues with a single server. 359361. Oxford 12.
[250] F. Probab. MartinL6f (1983) Entropy estimates for ruin probabilities . 834841. 124147. [261] H. Meier ( 1984). 36. Miyazawa & V. [258] K . Probability and Mathematical Statistics (A. F. 1986 . Scand. Appl. Proc.). Hoist eds. 223235.D. Act. A fitting algorithm for Markovmodulated Poisson processes having two arrival rates . Englunds Boktryckeri AB. [256] A. Ann. 268285. J. 11 . Appl. Probab . Schmidt ( 1993) On ladder height distributions of general risk processes . 31. Miller (1962 ) A matrix factorization problem in the theory of random variables defined on a finite Markov chain . Probab. [251] F. 4859. 2939. Probab. Anal.D. 763776. [262] H . Neuts ( 1990) A single server queue with server vacations and a class of nonrenewal arrival processes . [257] A. Ann. [263] M . Cambridge Philos. [252] A. Soc. Michna ( 1998) Selfsimilar processes in collective risk theory. Ney & E. Stoch. Math. 29 . Malinovskii ( 1996) Approximation and upper bounds on probabilities of large deviations of ruin within finite time. Lundberg (1926) Forsdkringsteknisk Riskutjdmning.K. Mammitzsch ( 1986 ) A note on the adjustment coefficient in ruin theory. [255] V. Lucantoni .V. with an application . 378406. Moustakides ( 1999) Extension of Wald 's first lemma to Markov processes. Appl. 3. 561592. 1996 . Appl.V. 1261 (0?)1270. [266] P.K. Probab. Stockholm. Nagaev (1957) Some limit theorems for stationary Markov chains.M. [253] V. Cambridge Philos. [265] S. 429448. Makowski ( 1994) On an elementary characterization of the increasing convex order . [254] V . J. Lundberg (1903) I Approximerad Framstdllning av Sannolikhetsfunktionen. 2. MartinL3f (1986) Entropy. J. Probab. 161174. [260] H. J. J. 676705.376 BIBLIOGRAPHY [249] D. Appl. 22. Statist. Appl. MeierHellstern & M . 15.D. Scand. Scand. Math. Miller ( 1962 ) Absorption probabilities for sums of random variables defined on a finite Markov chain . Nummelin (1987) Markov additive processes I. Opns. J. Insurance : Mathematics and Economics 5. Gut & J. Act. Almqvist & Wiksell. a useful concept in risk theory. Adv. 370377. Ann. Soc.S. . 58 .S.F. J. K. Miller ( 1961 ) A convexity property in the theory of random variables defined on a finite Markov chain. Malinovskii ( 1994) Corrected normal approximation for the probability of ruin within finite time. Europ. 286298. Proc. 1994. Eigenvalue properties and limit theorems . Res. Uppsala. Th. 32. Act. 58. [264] G. II Aterforsdkring av Kollektivrisker . . 147149. [259] Z .
Probab. [281] J. Johns Hopkins University Press. Neveu ( 1961 ) Une generalisation des processus a accroisances independantes. 555564. Stoch. Ohlin (1969) On a class of measures for dispersion with application to optimal insurance. 12551265. 273299. 157. Probab. [272] J . Commun. Baltimore . Probab. 30. [284] J. Proc. Abh.BIBLIOGRAPHY 377 [267] M . 12. Willekens (1986) Second order behaviour of the tail of a subordinated probability distributions. Probab. Stochastic Models 3 . Adv. . [277] J. Appl. 135148. Astin Bulletin 5 . Omey & E. 339353. London. 10081026. Proc.F.G. Neuts ( 1981 ) MatrixGeometric Solutions in Stochastic Models. J. [274] H . Stoch. Stoch.A.F. [275] H. Gjessing (1997b) Present value distributions with applications to ruin theory and stochastic equations. Neuts (1977) A versatile Markovian point process. Appl. [280] A. 764779. Appl. [276] C. IEICE Trans. Paulsen (1998) Ruin theory with compounding assets . Neuts ( 1978) Renewal processes of phase type. Stoch. 36.F. 16. [270] M . Hamburg. Paulsen ( 1998) Sharp conditions for certain ruin in a risk process with stochastic return on investments. E75B . Statistics 21. 75. 445454. Math.F. 327361. Pakes (1975) On the tail of waiting time distributions. Insurance: Mathematics and Economics 22.K. [269] M . 311342. Stoch. Gjessing (1997a) Optimal choice of dividend barriers for a risk process with stochastic return on investments. Nyrhinen (1999) Large deviations for the time of ruin. 46. Paulsen & H.F. [279] E. Paulsen (1993) Risk theory in a stochastic economic environment. Marcel Dekker. 71. Naval Research Logistics Quarterly 25.K. [285] J. 123144. 733746. Appl. Neuts ( 1989) Structured Stochastic Matrices of the M/G/1 Type and their Applications. Sem. Neuts (1992) Models based on the Markovian arrival process. J. Models 6.a survey. [271] M. Appl. Omey & E. 316. [278] E. New York. Appl. [268] M . Appl. Nyrhinen ( 1998 ) Rough descriptions of ruin for a general class of surplus processes. [282] J . Norberg ( 1990) Risk theory and its statistics environment . Proc. Appl. 21. Insurance: Mathematics and Economics 20. J. O'Cinneide (1990) Characterization of phasetype distributions. Willekens (1987) Second order behaviour of distributions subordinate to a distribution with finite mean . Paulsen & H. [273] R. Proc. 249266. [283] J. 215223.
U. [294] N. Pitman (1980) Subexponential distribution functions.U.G. Philipson (1968) A review of the collective theory of risk. Ann. [297] R. Res. Stat. Ann. [299] C. Appl. 965985. 215246. Regterschot & J. 46.J. 465483.J. Tidskr. Wiley.H. Appl. J. [303] S. Reinhard (1984) On a class of semiMarkov risk models obtained as classical risk models in a markovian environment. J. Oper. Math.L. 19. Berlin. [306] T. 139143. [301] G. Insurance: Mathematics and Economics 3. Appl. 222261. Adv. Insurance Risk. Probab. Queues. 32. Adv. Inst. Soc. Probab. 235243. New York. Math. [302] J. 29. R.378 BIBLIOGRAPHY [286] J.R.M.. [290] E. Samorodnitsky (1997) Performance degradation in a single server exponential queueing model with long range dependence. Math. 30. Res. Austr.M. [292] S. [288] S. 2343. Ramsay (1984) The asymptotic ruin problem when the healthy and sick periods form an alternating renewal process. 691707. Act. Pyke (1959) The supremum and infimum of the Poisson process. Pitts. Rao (1965 ) Linear Statistical Inference and Its Applications. 11. Insurance: Mathematics and Economics 16. Appl. Paulsen & H. 117133. Springer. Ramaswami (1980) The N/G/1 queue and its detailed analysis. 1989 . [287] F. Wiley. J. Aktuar. 390413.M. Ripley (1987) Stochastic Simulation . [300] C. Appl. Wiley. [295] N.K. New York. 757764. 2330. Pitts (1994) Nonparametric estimation of compound distributions with applications in insurance. [296] N. Probab. Math. Embrechts (1996) Confidence bounds for the adjustment coefficient. 4. Resnick & G. 337347. Gjessing (1997c) Ruin theory with stochastic return on investments. 147159. Ann. 61. Prabhu & Zhu (1989) Markovmodulated queueing systems. Probab. Statist. de Smit (1986) The queue M/G/1 with Markovmodulated arrivals and services. [289] C. 28. Ann. Rogers (1994) Fluid models in queueing theory and WienerHopf factorisation of Markov chains. 29A. Adv. and Dams.G. Prabhu (1965) Queues and Inventories. 12. Rolski (1987) Approximation of periodic queues. Math. 537555. [293] N.K. Statist.U. Prabhu (1980) Stochastic Storage Processes. 820827. Pellerey (1995) On the preservation of some orderings of risks under convolution. Heidelberg. 568576. Queueing Systems 5. Prabhu (1961). On the ruin problem of collective risk theory. Probab . . Astin Bulletin XIV. Schock Petersen (1989) Calculation of ruin probabilities when the premium depends on the current reserve.A. 4568. [298] V. Griibel & P. Scand. [291] S. Opns. [304] B. Skand.U. [305] C.
Wiley. 155188. J. Probab. 135149. Wiley. 131156.BIBLIOGRAPHY 379 [307] T. Rudemo (1973) Point processes generated by transitions of a Markov chain. Act. Proc. [311] R.Y. 5. Comp. Appi. 11. Ryden (1994) Parameter estimation for Markov modulated Poisson processes. Schmidli ( 1997a) Estimation of the Lundberg coefficient for a Markov modulated risk model . Schmidt & J. 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. 4857. [310] R. Ross (1974) Bounds on the delay distribution in GI/G/1 queues. [325] H . Schmidli ( 2001 ) Optimal proportional reinsurance policies in a dynamic setting . Schmidli ( 1999b) Perturbed risk processes: a review . 2001 . [319] H. 431447 [316] S. Insurance: Mathematics and Economics 16. 5.Y. Scand. Appl. Scand. Appl. 7. Data Anal. Stoch. [308] S. Rubinstein & B. Lecture Notes of the 8th Summer School on Probability and Mathematical Statistics ( Varna). Adv. Ann. Operationsforsch. [309] H. Schmickler ( 1992 ) MEDA : mixed Erlang distributions as phasetype representations of empirical distribution functions . [317] L. [321] H. [318] H.Y. Science Culture Technology Publishing . Statist. J.J. V. [320] H. Probab. 5. Melamed (1998) Modern Simulation and Modelling. [313] M. [326] H. Schmidli. Statist. Seal ( 1969) The Stochastic Theory of a Risk Business. Schmidli ( 1999a) On the distribution of the surplus prior and at ruin. Wiley. Teugels (1999) Stochastic Processes for Insurance and Finance. 93104. Schmidli ( 1996) Martingales and insurance risk. [323] H. Rubinstein & A. Astin Bulletin 29. [322] H. Probab .L. 1997. Wiley. 687699. [324] H .M. Rolski. Schmidli ( 1997b) An extension to the renewal theorem and an application in risk theory. [315] T. Insurance: Mathematics and Economics 22. Stochastic Models 10. 21. [314] T. 417421. 4068. Act. 365388. Schmidli ( 1994) Diffusion approximations for a risk process with the possibility of borrowing and interest. Rubinstein (1981) Simulation and the Monte Carlo Method. . Schlegel ( 1998) Ruin probabilities in perturbed risk models. Stochastic Models 6. 262286. [312] R. 795829. Schmidli ( 1995 ) CramerLundberg approximations for ruin probabilities of risk processes perturbed by a diffusion . Stochastic Models 10 . Math. 145165. 121133. Ryden (1996) An EM algorithm for estimation in Markovmodulated Poisson processes. 227244. Wiley. J. Shapiro (1993) Discrete Event Systems: Sensitivity Analysis and Stochastic Optimization via the Score Function Method. H. Singapore.
11.A. 914924. [335] B . . 191197. Siegmund (1975) The time until ruin in collective risk theory.the Harald Cramer volume (ed. Academic Press. Chapman & Hall. Appl. Weiss (1995) Large Deviations for Performance Analysis.L. Scand.W. 121139. [332] C : O. Siegmund (1976) Importance sampling in the Monte Carlo study of sequential tests. 61. 2236. Statist. Probab. Shwartz & A. Math. Ann. [330] H. Sengupta (1990) The semiMarkov queue: theory and applications. Math. 1974. Shtatland ( 1966) On the distribution of the maximum of a process with independent increments. 243249. 1955.G. [348] K . Segerdahl ( 1942) Uber einige Risikotheoretische Fagestellungen . Siegmund (1976) The equivalence of absorbing and reflecting barrier problems for stochastically monotone Markov processes . Appl. 4. [328] H. 4383. Segerdahl (1959) A survey of results in the collective theory of risk. [336] B. 171178.L. [340] E. 72. Verein Schweiz. Aktuar Tidsskr. Seal (1974) The numerical calculation of U(w. [331] G. Grenander).380 BIBLIOGRAPHY [327] H. Shin (1989) Ruin probability by operational calculus. [343] D.L. Chapman and Hall. [346) D .F. pp. Seal (1972) Numerical calculcation of the probability of ruin in the Poisson/Exponential case.W. Wiley. Sengupta ( 1989) Markov processes whose steadystate distribution is matrixgeometric with an application to the GI/PH/1 queue. Aktuar Tidsskr. Mitt. Seber (1984) Multivariate Observations. Mitt. Seal ( 1972). 159180. [339] A. Act. Ann. L. Shaked & J.S. Almqvist & Wiksell. [341] E. Skand. Sigman ( 1994 ) Stationary Marked Point Processes: An Intuitive Approach. 21. Versich. the probability of nonruin in an interval (0. Mitteil. 77100. Shantikumar (1993) Stochastic Orders and Their Applications. 279299. Scand. Adv. 673684. t). [337] M. 72. New York. Verein Schweiz. 429442. Verein Schweiz Versich. [338] E. . [342] D. Sigman (1992) Light traffic for workload in queues. Act. 11. Adv. [347] K. Shin (1987) Convolution of uniform distributions and ruin probability. Stochastic Models 6. In Probability and statistics .S. [345] D. Wiley. 701719. Seal (1978) Survival Probabilities. Siegmund ( 1985) Sequential Analysis . [333] C: O. 4. Siegmund (1979) Corrected diffusion approximations in certain random walk problems. J. Math. Th. SpringerVerlag. t).S. Segerdahl (1955) When does ruin occur in the collective theory of risk? Skand. Queueing Systems 11. 1987. 75. Stockholm. Probab. [334] C: O. Insurance: Mathematics and Economics 8. 383413. Probab. J. 157166. [344] D. Versich. Probab . [329] H. 483487. Risk teory and the single server queue . Appl.
Takhcs (1967) Combinatorial Methods in the Theory of Stochastic Processes. 722. New York.C. 5776. Teugels (1997) The adjustment coefficient in ruin estimates under interest force . . [359] L. 65102. Actuarial J. Taylor (1976 ) Use of differential and integral inequalities to bound ruin and queueing probabilities . [367] O. J. [369] O. 235254.L. Scand. J. Taylor (1978) Representation and explicit calculation of finitetime ruin probabilities. Teugels ( 1995 ) Ruin estimates under interest force . Daley ed. Smith ( 1953) Distribution of queueing times. Astin Bulletin 24. Proceeedings of the IEEE 77. Sundt ( 1993) An Introduction to NonLife Insurance Mathematics (3rd ed.and largedeviations probabilities in actuarial risk theory. 149162. Probab. Thorin (1977). [357] B. 2739. Scand.C. Act. Insurance: Mathematics and Economics 16.L. 21. Thorin (1986) Ruin probabilities when the claim amounts are gamma distributed. Taylor (1986) Claims reserving in NonLife Insurance. Appl. Taylor (1979) Probability of ruin under inflationary conditions or under experience ratings . J. [355] B . 118. 65102. [354] B. Wiley. [358] R. [350] W. 1978. 725741.V. [368] O.S.J. [366] O. Suri ( 1989) Perturbation analysis : the state of the art and research issues explained via the GI/G/1 queue.C. NorthHolland. Springer.J. Insurance: Mathematics and Economics 19. Cambridge Philos. [361] G. 163175.). J. Adv. Slud & C. Stroinski ( 1994) Recursive method for computing finitetime ruin probabilities for phasedistributed claim sizes. [352] D. Thorin (1974) On the asymptotic behaviour of the ruin probability when the epochs of claims form a renewal process. Unpublished manuscript. Scand. Act. 1980 . [351] D . Act. Scand. Act. J. Jewell ( 1981 ) Further results on recursive evaluation of compound distributions . Proc. 8199. Sundt & J. Astin Bulletin 12. Ruin probabilities prepared for numerical calculations. Insurance: Mathematics and Economics 1. Astin Bulletin 16 .L. [364] G. 114137. Scand. [353] E. Karlsruhe. [360] G. Verlag Versicherungswirtschaft e.. 1974. 197208. Teugels (1982) Estimation of ruin probabilities . Hoesman ( 1989) Moderate. Act. Straub (1988) NonLife Insurance Mathematics.BIBLIOGRAPHY 381 [349] E.C. New York. 1976. [365] J.L. [356] B . Stanford & K. 8594. John Wiley & Sons. Scand. Soc 49. Sundt & J. [363] G. Sundt & W. Stoyan (1983) Comparison Methods for Queues and Other Stochastic Models (D. Thorin (1982) Probabilities of ruin. 449461.C..). [362] G. 1982 . Taylor (1979) Probability of ruin with variable premium rate.A.
Goovaerts (1983) The influence of reinsurance limits on infinite time ruin probabilities. Stationarity and Regeneration. Waters (1983) Probability of ruin for a risk process with claims cost inflation. 1983 .R. [372] H. [381] W. [375] N. [377] V. [376] A. 936952. M. PrenticeHall. 231246. Whitt (1989) An interpolation approximation for the mean workload in a GI/G/1 queue. Willmot & X. 31. Van Wouve.E. [383] G. Wald (1947) Sequential Analysis. Insurance: Mathematics and Economics 15. [371] 0.E. Oper. [378] H. Veraverbeke (1993) Asymptotic estimates for the probability of ruin in a Poisson model with diffusion. Goovaerts. R. Appl. 148164. Probab. . Wolff (1990) Stochastic Modeling and the Theory of Queues. In: Premium Calculation in Insurance (F. Unpublished Ph. Insurance : Mathematics and Economics 13. J. Thorin & N. Taqqu. Sherman & D. Wilson (1995) Selfsimilarity in highspeed traffic: analysis and modeling of ethernet traffic measurements. Submitted. M.382 BIBLIOGRAPHY [370] 0. University of Michigan. D. Thorin & N. Lin (1994) Lundberg bounds on the tails of compound distributions. De Vylder & M. [380] W. SpringerVerlag. J.). Statistical Science 10. Willinger. 37. Aktuar. Willmot (1994) Refinements and distributional generalizations of Lundberg's inequality. thesis. [374] F. J. 6785.W. 142. 5762. [379] M. F. Tacklind (1942) Sur le risque dans les jeux inequitables. Wikstad (1977) Numerical evaluation of ruin probabilities for a finite period. Res. Reidel. 4963. Skand. Wikstad (1977) Calculation of ruin probabilities when the claim distribution is lognormal. [384] R. Scand. Astin Bulletin VII. Thorisson (2000) Coupling. 1942. Astin Bulletin IX. VazquezAbad (1999) RPA pathwise derivatives of ruin probabilities. Wiley. Haezendonck eds. 137153. Wallace (1969) The solution of quasi birth and death processes arising from multiple access computer systems. De Vylder. Dordrecht Boston Lancaster. Tidskr. [382] G. 743756. [373] S. Act.
14.285292. 117127 corrected 121127 duality 1314. 226.178184. 17. 248 WienerHopf 144 interest rate 190. 180182.226. 227229. 80 81. 4851. 117128.203. 360 excursion 155156. 341. 89.328330. 308.135. 3334.182. 196201 inverse Gaussian distribution 76. 1415.314316.307312 compound Poisson model 4. 201 Brownian motion 3 . 271274.281. 7475.185187.150. 9496. 12 CramerLundberg approximation 1617.301 central limit theorem 60 . 111117. 141144. 201214.228229. 7179. 1819.121129.346349 383 . 9293.Index adjustment coefficient 17. 40. 3436. 316323 Bessel function 102.160167.6779. 283.287292.242.137141. 2526. 170173. 302303 diffusion approximation 17.100.86. 189. 3839. 332333 Volterra 192194.318320 change of measure 2630. 239. 323 Coxian distribution 147.203.and sum 221. 138139. 97. 245248. 162164. 207 heavytailed distribution 6.359 aggregate claims 103106. 3032.249250 integral equation 16 Lindley 143 renewal 64. 17.299.292293 Edgeworth expansion 113. 278 gamma distribution 67.217. 97129.251280 heavy traffic 76. 7079. 122. 39.9899.259261. 110113. 9396. 5. 37. 301 Kronecker product. 218 Cox process 4.249. 361 diffusion 3. 8283 hyperexponential distribution 7. 119. 79.4447. 91. 7879. 86. 2425.269. 1112. 5796.293294. 217. 318319 Erlang distribution 7. 205.272.200201.308 CramerLundberg model: see compound Poisson model cumulative process 334 dams: see storage process differential equation 16. 15. 135.
336339 Laplace transform 15. 137139. 44.298299. 185187 GI/G/1 141144 M/D/1 6667 equation 16. 144. 3639.275278. 171. 108 life insurance 5.134135. 245 M/G/1 13.269271. 133. 229 M/M/1 101 Markovmodulated 185187 periodic 187 martingale 2426. 6162.261264.152160. 271274. 157.234. 71.201. 149.302. 251.304 process 2830. see also sensitivity analysis phasetype distribution 8. 38. 44.384 ladder heights 4756.336339 . 5758.240244. 42. 154.285287 queue 14 . 3947. 227228. 213214.161164.287.339 large deviations 129.227230.234240. 2730. 32. 112113. 108109. 234 matrixexponential distribution 240244 matrixexponentials 14. 14.259261. 16. 203204. 141144. 267269 Panjer's recursion 320323 Pareto distribution 910.108.288290.139141. 142 likelihood ratio : see change of measure lognormal distribution 9.160161. 132133. 86 periodicity 12. 145187. 260 Lundberg conjugation 6979 .297299. 134135.146148. 37. 7576. 25. 178 modulation 12. 306316 Levy process 3. 176185 nonhomogeneous 60 PollaczeckKhinchine formula 6167. nonlinear 155. 9899. 38.218221. 39.148. 134.287291 INDEX matrix equation . 261264. 176185.161.238.123. 6970. 59.128129. 41. 113114. 295. 3947. 269 PerronFrobenius theory 4142. 230. 100. 203 Markov additive process 12. 80. 96.349 350 perturbation 172173.178182. 25. 35. 52 53. 175 light traffic 8183 Lindley integral equation 143 process 3334.340350 multiplicative functional 2830.348 terminating 215216. 257. 179 NP approximation 318320 Palm distribution 5253.215250.315 inequality 1718. 7179. 65. 162. 35. 304305 random walk 3336.350361 Poisson process Markovmodulated 12 periodic 12. 106108. 16. 4446. 15.174. 133.180. 99. 138.161.
131144. 238 saddlepoint method 115117. 152.244250. 37. 160. 11. 4950. 123.279280 Rouche roots 158. 251280 time change 4. 251. 294296 shotnoise process 314 simulation 19. 7475.336339 workload 13. 326330 Weibull distribution 9. 260 WienerHopf theory 144. 251. 5455.186. 168172 storage process 13. 256258. 307308. 244. 331336 equation 64. 8386. 107. 87. 186187 virtual: see workload rational Laplace transform 8. 333334 regular variation 10. 327 . 260 reinsurance 8. 60. 1819. 338 utility 324. 186187 renewal process 131. 281296 stable process 15. 189214. 3032. 89. 191192. 253.INDEX 385 waiting time 141. 233234. 257. 317318 semiMarkov 147. 147. 146. 213.359361 stochastic control x stochastic ordering 18. see also matrixexponential distribution regenerative process 264 268. 177 timereversion 14.314. 162. 172173. 31.273274. 335336 sensitivity analysis 8693. 12. 279280 subexponential distribution 11. 280. 223226.154157. 292294. 222. 141144.262263. 240. 332333 model 12. 261264 reservedependent premiums 14. 174. 120 statistics x. 9693. 233. 229234.
. "This book is a must for anybody working in applied probability." Short Book Reviews ISBN 9810222939 mi u inn i nun I I I I I I i in u www. y finite horizon ruin probabilities. It is a comprehensive treatment of the known results on ruin probabilities. Markovmodulation or periodicity. for heavytailed claim size distributions). phasetype distributions as a computational vehicle and the connection to other applied probability areas like queueing theory. worldscientific.g. exact solutions.. P'i yfliother approximations (e.Advanced Series on Statistical Science & Applied Probability .Vol.T [Ail i The book is a comprehensive treatment of  I i I \ classical and modern ruin probability theory. the ^W A l \ i l ' ''' CramerLundberg approximation. 2 A I 11 JjVb l' i  i Yj . I 1! Ruin Probabilities .com 2779 he 9 "789810ll22293211 . Special features of the book are the emphasis on change of measure techniques. Some i (l I JL I J r of the topics are Lundberg's inequality. extensions of the classical compound Poisson model to allow f o r reservedependent premiums.
This action might not be possible to undo. Are you sure you want to continue?