This action might not be possible to undo. Are you sure you want to continue?
Ruin Probabilities
Seren Asmussen
World Scientific
Ruin Probabilities
ADVANCED SERIES ON STATISTICAL SCIENCE & APPLIED PROBABILITY
Editor: Ole E. BarndorffNielsen
Published Vol. 1: Random Walks of Infinitely Many Particles by P. Revesz Vol. 2: Ruin Probabilities by S. Asmussen Vol. 3: Essentials of Stochastic Finance : Facts, Models, Theory by Albert N. Shiryaev Vol. 4: Principles of Statistical Inference from a NeoFisherian Perspective by L. Pace and A. Salvan Vol. 5: Local Stereology by Eva B. Vedel Jensen Vol. 6: Elementary Stochastic Calculus  With Finance in View by T. Mikosch Vol. 7: Stochastic Methods in Hydrology: Rain, Landforms and Floods eds. O. E. Barndorff Nielsen et al. Vol. 8: Statistical Experiments and Decisions : Asymptotic Theory by A. N. Shiryaev and V. G. Spokoiny
Ruin P robabilities
Soren Asmussen
Mathematical Statistics Centre for Mathematical Sciences Lund University
Sweden
World Scientific
Singapore • NewJersey • London • Hong Kong
Published by World Scientific Publishing Co. Pte. Ltd. P O Box 128, Fatter Road , Singapore 912805 USA office: Suite 1B, 1060 Main Street, River Edge, NJ 07661 UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE
Library of Congress CataloginginPublication Data Asmussen, Soren
Ruin probabilities / Soren Asmussen. p. cm.  (Advanced series on statistical science and applied probability ; vol. 2) Includes bibliographical references and index. ISBN 9810222939 (alk. paper) 1. InsuranceMathematics. 2. Risk. I. Tide. II. Advanced series on statistical science & applied probability ; vol. 2. HG8781 .A83 2000 368'.01dc2l 00038176
British Library CataloguinginPublication Data A catalogue record for this book is available from the British Library.
First published 2000 Reprinted 2001
Copyright ® 2000 by World Scientific Publishing Co. Pte. Ltd. All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means, electronic or mechanical, including photocopying, recording or any information storage and retrieval system now known or to be invented, without written permission from the Publisher.
For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy is not required from the publisher.
Printed by Fulsland Offset Printing (S) Pte Ltd, Singapore
Contents
Preface I ix
Introduction 1 1 The risk process . . . . . . . . . . . . . .. . . . .. .. . . . . 1 2 Claim size distributions .. . . . . . . . .. . . . . . . . . . . . 5 3 The arrival process . . . . . . . . . . . . . . . . . . . . . . . . 11 4 A summary of main results and methods . . . . .. . . . . . . 13 5 Conventions . .. . .. .. . . . . . . . . . . . . . . . . . . . . 19
II Some general tools and results 23 1 Martingales . .. . .. .. . . . . . .. . . . . . . . . . . . . . 24 2 Likelihood ratios and change of measure . . .. . . . . . .. . 26 3 Duality with other applied probability models . . .. . . . . . 30 4 Random walks in discrete or continuous time . . . . . . . . . . 33 5 Markov additive processes . . . . . . . .. . . . . . . . . . . . 39 6 The ladder height distribution . . . .. . .. .. . . . . . . . . 47
III The compound Poisson model 57 1 Introduction . . . . . . . . .. .. .. . .. .. . . . . . . 58 . . . . . . . . . . . . . . . 61 3 Special cases of the PollaczeckKhinchine formula . . . . . . . 62 4 Change of measure via exponential families . . . .... . .. . 67 5 Lundberg conjugation . .. . . . . . . . . . . . . . . . . . . . . 69 6 Further topics related to the adjustment coefficient .. . . . . 75 7 Various approximations for the ruin probability . . . . . . . . 79 8 Comparing the risks of different claim size distributions . . . . 83 9 Sensitivity estimates . . . . . . . . . . . . . . . . . . . . . . . 10 Estimation of the adjustment coefficient . . . . . . . . . . . . 86 93 2 The PollaczeckKhinchine formula
v
vi
CONTENTS
IV The probability of ruin within finite time 97 1 Exponential claims . . . . . . . . . . . . . . . . . . . . . . . . 98 2 The ruin probability with no initial reserve . . . . . . . . . . . 103 3 Laplace transforms . . . . . . . . . . . . . . . . . . . . . . . . 108 4 When does ruin occur? . . . . . . . . . . . . . . . . . . . . . . 110 5 Diffusion approximations . . . . . . . . . . . . .. . . .. . . . 117 6 Corrected diffusion approximations . . . . . . . . . . .. . . . 121 7 How does ruin occur ? . . .. . . . . . . . . . . . . . . . . . . . 127 V Renewal arrivals 131 1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . 131 2 Exponential claims. The compound Poisson model with negative claims . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 3 Change of measure via exponential families . . . . . . . . . . . 137 4 The duality with queueing theory .. .. .. . . . .. . . . . . 141 VI Risk theory in a Markovian environment 145 1 Model and examples . . . . . . . . . . . .. . .. . . . . . . . 145 2 The ladder height distribution . . . . . . . . . .. . . . . . . . 152 3 Change of measure via exponential families ........... 160 4 Comparisons with the compound Poisson model ........ 168 5 The Markovian arrival process . . . . . . .. .. . . ... . . . 173 6 Risk theory in a periodic environment .. . . . .. . . . . . . . 176 7 Dual queueing models .... ... ................ 185 VII Premiums depending on the current reserve 189 1 Introduction . . . . . . . . . . . . . . . . . . . .. . . . . . . . 189 2 The model with interest . . . . . .. . . . . . . . . . .. . . . 196 3 The local adjustment coefficient. Logarithmic asymptotics . . 201 VIII Matrixanalytic methods 215 1 Definition and basic properties of phasetype distributions .. 215 2 Renewal theory . . . . . . . . . . . . . . . . . . . . . . . . . . 223 3 The compound Poisson model . . . . . . . . . .. . . . . . . . 227 4 The renewal model . . . . . . . . . . . . . . . .. . . . . . . . 229 5 Markovmodulated input . . .. . . . . . . . . . . . . . . . . . 234 6 Matrixexponential distributions . . . . . . . . . . . .. . . . 240 7 Reservedependent premiums . . . . .. . . . .. . . . . . . . 244
. . . . . . . . . . .. . . . . . . . . 323 6 Reinsurance . 281 2 Simulation via the PollaczeckKhinchine formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290 5 Regenerative simulation . 285 3 Importance sampling via Lundberg conjugation . 336 A3 Matrixexponentials . . . 297 2 Further applications of martingales . . . . . . . . . . . . . . . . 326 Appendix 331 Al Renewal theory . . . . . . . . . . .. . . . . . . . . . . . . . . . . .. . . . . . . .. . . . . . . . . . . . 331 A2 WienerHopf factorization . . . . . . . .. . . . . . . . . . . . . . .. . . .. . 259 3 The renewal model . . . . . . . . . . . 340 A4 Some linear algebra . . The twobarrier ruin problem . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . .. . . . . . . . . . 344 AS Complements on phasetype distributions . . . . . . . . . . . . . . . . . . . . . .CONTENTS vii IX Ruin probabilities in the presence of heavy tails 251 1 Subexponential distributions . 287 4 Importance sampling for the finite horizon case . . . . . . . . . . . . . . 316 5 Principles for premium calculation . . .. . .. . .. . .. . . . . . . . . . . . .. . . 264 5 Finitehorizon ruin probabilities . . . . . . . . . . . . . . . .. . 292 6 Sensitivity analysis . . . . . 306 4 The distribution of the aggregate claims . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . 251 2 The compound Poisson model . .. . . . . . .. . . 350 Bibliography Index 363 383 . . . . 279 X Simulation methodology 281 1 Generalities . . . . . . . . . 261 4 Models with dependent input . .. .. . . . . . . .. . . . . . . . . .. . .. . . . . . . . . . . . . . . . . . . . . . . . 304 3 Large deviations . . . . . . . 271 6 Reservedependent premiums . . . . . . .. . . 294 XI Miscellaneous topics 297 1 The ruin problem for Bernoulli random walk and Brownian motion. . . . . . . . .
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. and my belief was that this could be done rather quickly. 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. One reason for writing this book is a feeling that the area has in the recent years achieved a considerable mathematical maturity. Let me take this opportunity to thank above all my publisher World Scientific Publishing Co. Thus. As an excuse: many of these projects were related to the book. the book is basically mathematical in its flavour. and other projects absorbed my interest. Since I was to produce some handouts for the students anyway. and the series editor Ole BarndorffNielsen for their patience. Apart from these remarks. that it can only say something about very simple models and questions. the idea was close to expand these to a short book on the subject. The course was never realized. I was invited to give a course on ruin probabilities at the Laboratory of Insurance Mathematics. it is not by intention. this applies to longrange dependence which is intensely studied in the neighboring ix . I have deliberately stayed away from discussing the practical relevance of the theory. However. and has been an active area of research from the days of Lundberg all the way up to today. University of Copenhagen. if the formulations occasionally give a different impression. It has obviously not been possible to cover all subareas. But the pace was much slower than expected.Preface The most important to say about the history of this book is: it took too long time to write it! In 1991. and the result is now that the book is much more related to my own research than the initial outline. 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!). which has in particular removed one of the standard criticisms of the area. Risk theory in general and ruin probablities in particular is traditionally considered as part of insurance mathematics. In particular.
http:// www. IX.13. Willinger et al. [381]).lth.lth. The book does not go into the broader aspects of the interface between insurance mathematics and mathematical finance.se Lund February 2000 Soren Asmussen . Asmussen.g. In addition. IV.13 and IX.15. X. see also Schmidli [325] and the references in Asmussen & Taksar [52]. it has not been possible to incorporate more numerical examples than the few there are. the first part of 11.89. A book like this can be organized in many ways. Another interesting area which is not covered is dynamic control. More recently. 111. IV. One is by model. The main motivation comes from statistical data for network traffic (e. Resnick & Samorodnitsky [303] and references therein. IV. see e. for which I apologize to the reader and the authors of the many papers who ought to have been on the list. some papers not cited in the text but judged to be of interest are included in the Bibliography.g.2. Chapters IIIVII introduce some of the main models and give a first derivation of some of their properties. Hojgaard & Taksar [206].4a. The rest is up to your specific interests.13. For a brief orientation. 111. VI. VIII. In the classical setting of CramerLundberg models.g. see in particular Michna [259]. Good luck! I have tried to be fairly exhaustive in citing references close to the text. For a second reading. incorporate 11. some basic discussion can be found in the books by Biihlmann [82] and Gerber [157]. VII.5. read Chapter I. Here is a suggestion on how to get started with the book. the standard stochastic control setting of diffusion models has been considered. Finally.13 and XI. for the effects on tail probabilities.2 more properly). an area which is becoming increasingly important. Concerning ruin probabilities.maths . I intend to keep a list of misprints and remarks posted on my web page. VII.x PREFACE field of queueing theory. I regret that due to time constraints. Hojgaard & Taksar [35] and Paulsen & Gjessing [284]. 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). It is obvious that such a system involves a number of inconsistencies and omissions.14.3. e.2.1. another by method.se/matstat / staff/asmus and I am therefore grateful to get relevant material sent by email to asmusfmaths .6 (to understand the PollaczeckKhinchine formula in 111. The present book is in between these two possibilities.45.
4 from Asmussen. of which there are not many at this stage . IV. were produced by Lone Juul Hansen . Lund September 2001 Soren Asmussen Acknowledgements Many of the figures . Parts of II.1 and X.2 by Rafal Kulik . .6.6 by my 1999 simulation class in Lund. 5. 5 from Asmussen & Kliippelberg [36] with the permission from Elsevier Science . Parts of X. Aarhus. A number of other figures were supplied by Christian Geisler Asmussen .5 from Asmussen [21] with permission from CRC Press.6 is reprinted from Asmussen & Schmidt [49] and parts of IX.1 by Bjarne Hojgaard and the table in Example 111. not least the more complicated ones. 111 .8 . More substantial remarks. 3 is reprinted from Asmussen & Nielsen [39] and parts of IX. Section VII . Fig. Section VIII. Schmidli & Schmidt [47] with the permission from Applied Probability Trust . supported by Center for Mathematical Physics and Stochastics (MaPhySto).PREFACE xi The second printing differs from the first only by minor corrections. 1 is almost identical to Section 2 of Asmussen [26] and reprinted with permission of Blackwell Publishers.3 are reprinted from Asmussen & Rubinstein [46] and parts of VIII. Fig. as well as some additional references continue to be at the web page. many of which were pointed out by Hanspeter Schmidli .
This page is intentionally left blank .
we introduce some general notation and terminology.Rt. A risk reserve process { Rt}t>o. respectively.T) = P inf Rt < 0 I . We denote throughout the initial reserve by u = Ro. Letting T(u) = inf {t > 0 : Rt < 0} = inf It > 0 : St > u}. and give a very brief summary of some of the models. as defined in broad terms . (1.1) We also refer to t/) ( u) and 0(u.Chapter I Introduction 1 The risk process In this chapter .i(u. M = (1. is a model for the time evolution of the reserves of an insurance company. MT = sup St. t/i(u) = P (infRt < 0) = P (infR t < 0 t>0 t>0 The probability of ruin before time T is t.4) O<t<oo O<t<T 1 . (1. They are the main topics of study of the present book. The probability O(u) of ultimate ruin is the probability that the reserve ever drops below zero. T) as ruin probabilities with infinite horizon and finite horizon . (1.3) sup St. For mathematical purposes. results and topics to be studied in the rest of the book.2) (O<t<T Ro=ul. it is frequently more convenient to work with the claim surplus process {St}t>0 defined by St = u .
We denote the interarrival times of claims by T2. ..i(u. and T1 is the time of the first claim.. the ruin probabilities can then alternatively be written as . respectively. Figure 1. say. (1. However. 1. T3.T) = F (MT > u) = P(r(u) < T). St = E Uk . • Premiums flow in at rate p. the time of arrival of the nth claim is an = T1 + • • • + Tn. the number Nt of arrivals in [0.7) k=1 k=1 The sample paths of {Rt} and {St} and the connection between the two processes are illustrated in Fig.5) i. That is. (1.b(u) = P (r(u) < oo) = P(M > u). the following setup will cover the vast majority of the book: • There are only finitely many claims in finite time intervals.pt. Putting things together.6) Sofar we have not imposed any assumptions on the risk reserve process. t] is finite.1. (1.2 CHAPTER I. we see that Nt Nt Rt = u + pt . and Nt = min {n > 0 : 0rn+1 > t} = max {n > 0: Un < t}• The size of the nth claim is denoted by Un. Thus. INTRODUCTION be the time to ruin and the maxima with infinite and finite horizon. per unit time.E Uk.1 .
The models we consider will typically have the property that there exists a constant p such that Nt a E Uk k=1 p. one may well argue that Brownian motion in itself could be a reasonable model.) V 0. of course. .1. If 77 < 0. immaterial.b(u) = 1 for all u. we shall.8) holds. VII. It would appear obvious. We shall discuss Brownian motion somewhat in Chapter IV. 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. If 77 > 0. For the purpose of studying ruin probabilities this distinction is. that the insurance company should try to ensure 77 > 0. We study this case in Ch..1 Assume that (1.. Thus. and in fact: Proposition 1. and hence O(u) < 1 for all sufficiently large u. • Brownian motion or more general diffusions. but as an approximation to the risk process rather than as a model of intrinsic merit. though many results are straightforward to generalize from the compound Poisson model. and the basic ruin probabilities are derived in XI.1 the slope of {Rt} should depend also on the level). However. 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. then M = oo a. then M < oo a. say 10% .1.s. for example. since any modeling involves some approximative assumptions. Some main examples of models not incorporated in the above setup are: • Models with a premium depending on the reserve (i. (1.1. A further basic quantity is the safety loading (or the security loading) n defined as the relative amount by which the premium rate p exceeds p. • General Levy processes (defined as continuous time processes with stationary independent increments) where the jump component has infinite Levy measure. one could well replace Rt by Rtnr(u) or RtA.s. 1. We shall not deal with this case either. not discuss whether this actually corresponds to practice.(.20%. and hence . t * oo. however. however.8) The interpretation of p is as the average amount of claim per unit time.e. on Fig. allowing a countable infinity of jumps on Fig.
_ St __ k =1 Uk pt a4.T) = i. Then the connection between the ruin probabilities for the given risk process {Rt} and those ^(u). Thus p may well be random for such processes.i. namely that M = oo a. corresponding to the Pdlya process. .10) is a property which we will typically encounter. Here it is easy to see that p = .8) is given by ^t p = EU • lim it (3(s) ds t. (1. U2. St In concrete models..i(u..4 CHAPTER I.T) for {Rt} is given by V)(u) = t/i (u). k=1 (1.b(u) < 1 for all u when rl > 0. INTRODUCTION Proof It follows from (1..10) Again. are i. then similarly limSt/t < 0.11) . where {Nt} is a Poisson process with rate . zP(u .. . and here (1.s. in connection with risk processes in a Markovian or periodic environment (Chapter VI).6EU (on the average. not all models considered in the literature have this feature: Example 1. However.v.i. (1. If 77 < 0.s. it is not too difficult to show that p as defined by (1. and that .Q (say) and U1. M < oo a.s.2 (Cox PROCESSES) Here {Nt} is a Poisson process with random rate /3(t) (say) at time t.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. This case is referred to as the mixed Poisson process. and independent of {(0(t). Nt)}. (1. However.8). The simplest example is 3(t) = V where V is a r . If u oo. and independent of {Nt}. are i. then this limit is > 0 which implies St a$ oo and hence M = oo a.10) hold with p constant. with the most notable special case being V having a Gamma distribution.d. if {(3(t)} is nonergodic.8) that F N. this needs to be verified in each separate case. U2.. If U1.oo t 0 J (provided the limit exists). . 0(u. The simplest concrete example (to be studied in Chapter III) is the compound Poisson model. we obtain typically a somewhat stronger conclusion. rl > 0. tb(u) = 1 for all u holds also when rl = 0.Tp).3 Assume p 54 1 and define Rt = Rt1p. Proposition 1. 0 We shall only encounter a few instances of a Cox process. namely.d. t t p  p' t ^ oo.
U2. CLAIM SIZE DISTRIBUTIONS 5 The proof is trivial. We roughly classify these into two groups . the recent survey by Grandell [173] and references therein.. Heilmann [191]. and we do not get near to the topic anywhere in this book. Sundt [354].. Gerber [157]. the role of the result is to justify to take p = 1. Note that life insurance (e. Straub [353]. Some of the main general ideas were laid down by Lundberg [250]. [76]. see e . many results and methods in random walk theory originate from there and the area was ahead of related ones like queueing theory. the research literature is often published in journals like Astin Bulletin . in particular. [330]. 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.. An idea of the additional topics and problems one may incorporate under risk theory can be obtained from the survey paper [273] by Norberg.g. Notes and references The study of ruin probabilities. Besides in standard journals in probability and applied probability. the claim arrivals are Poisson or renewal at the same time). was largely initiated in Sweden in the first half of the century. Some main later textbooks are (in alphabetical order) Buhlmann [82]. often referred to as collective risk theory or just risk theory. Embrechts et al. Pentikainen & Pesonen [101]. Insurance: Mathematics and Economics. which is feasible since in most cases the process { Rt } has a similar structure as {Rt} (for example. see also Chapter XI.g. in a number of models. but in probability and applied probability as a whole. another important early Swedish work is Tacklind [373]. Daykin et al.2. Cox processes are treated extensively in Grandell [171]. lighttailed distributions (sometimes the term . Hipp & Michel [198]. Buhlmann [82]. the assumption > 0 is equivalent to p < 1. while the first mathematically substantial results were given in Lundberg [251] and Cramer [91]. Schmidt & Teugels [307] and Seal [326]. Note that when p = 1. [134]. De Vylder [110]. For mixed Poisson processes and Polya processes. Taylor [364]. Daykin. [101]. Mitteilungen der Verein der Schweizerischen Versicherungsmathematiker and the Scandinavian Actuarial Journal. Gerber [159]) has a rather different flavour. Grandell [171]. Schmidli. we shall be able to identify p with the traffic intensity of an associated queue.. Some early surveys are given in Cramer [91]. Since { Rt } has premium rate 1. The term risk theory is often interpreted in a broader sense than as just to comprise the study of ruin probabilities. The Swedish school was pioneering not only in risk theory. Rolski. some main texts (typically incorporating some ruin theory but emphasizing the topic to a varying degree) are Bowers et al. Segerdahl [334] and Philipson [289]. and in fact p < 1 is the fundamental assumption of queueing theory ensuring steadystate behaviour (existence of a limiting stationary distribution). In the even more general area of nonlife insurance mathematics.
g. 6 has density r(p)xPleax b(x) P and m. a fact which turns out to contain considerable information.1) The parameter 6 is referred to as the rate or the intensity. and heavytailed distributions. i. Example 2 .g.2 and /LB is the mean of B.3) .2) = 0. P B[s]= (8Is ) . then the conditional distribution of X . Here lighttailed means that the tail B(x) = 1 . the exponential distribution is by far the simplest to deal with in risk theory as well. In contrast.e.x given X > x is again exponential with rate b (this is essentially equivalent to the failure rate being constant).8. but different more restrictive definitions are often used: subexponential.O(u) can be found in closed form. for the compound Poisson model with exponential claim sizes the ruin probability . and can also be interpreted as the (constant) failure rate b(x)/B(x).f. INTRODUCTION 'Cramertype conditions' is used). if 1 °O AB Jbos x B(dx) > 0.f. On the more heuristical side. s<8. (2. B[s] is finite for some s > 0. The crucial feature is the lack of memory: if X is exponential with rate 6. regularly varying (see below) or even regularly varying with infinite variance. a simple stopping time argument shows that this implies that the conditional distribution of the overshoot ST(u) . where B(bo. 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'. For example in the compound Poisson model. the m. As in a number of other applied probability areas.B(x) satisfies B(x) = O(e8x) for some s > 0.6 CHAPTER I.1 (THE EXPONENTIAL DISTRIBUTION) Here the density is b(x) = beax (2. B is heavytailed if b[s] = oo for all s > 0.2 (THE GAMMA DISTRIBUTION) The gamma distribution with parameters p. In particular.u at the time of ruin given r(u) is again exponential u with rate 8. Equivalently. 2a Lighttailed distributions Example 2.
An important property of the hyperexponential distribution is that its s. the Gamma density (2. among others. .) VarX1 (EX )2 p is < 1 for p > 1. p) = J tPletdt. 0. i = 1. by Grandell & Segerdahl [175] and Thorin [369].c. if p is integer and X has the gamma distribution p.. the squared coefficient of variation (s. The exact form of the tail B(x) is given by the incomplete Gamma function r(x.1) (or the 1/pth root if p < 1). is > 1.v.2. u . 2. An appealing feature is its simple connection to the Poisson process: B(x) = P(Xi + • • • + XP > x) is the probability of at most p . x] so that B(x) = r` e.. > 1 for p < 1 and = 1 for p = 1 (the exponential case). Ruin probabilities for the general case has been studied.y i=1 where >i ai = 1. p. This special case is referred to as the Erlang distribution with p stages.. or just the Erlang(p) distribution. we develop computationally tractable results mainly for the Erlang case (p = 1. CLAIM SIZE DISTRIBUTIONS 7 The mean EX is p/b and the variance Var X is p/b2. In particular. 0 < ai < 1. p).c.. .2) can be considered as the pth power of the exponential density (2. p) °° where r (x..i. where X1. . u Example 2 .d.). then X v Xl + • • • + X..ate (b2 ): L• i=o In the present text. In particular. P b(x) = r` aibiea. are i. X2.3 (THE HYPEREXPONENTIAL DISTRIBUTION) This is defined as a finite mixture of exponential distributions.. one has r(bx. JP 1 B(x) r(p ) XP ie ax In the sense of the theory of infinitely divisible distributions.1 Poisson events in [0. B(x) = r(p) Asymptotically.v. and exponential with rate d.
Equivalent characterizations are that the density b(x) has one of the forms q b(x) j=1 = cjxienbx. it is notable from a practical point of view because of reinsurance: if excessofloss reinsurance has been arranged with retention level xo. See XI. The parameters of a phasetype distribution is the set E of transient states. there exists a xo < oo such that B(x) = 0 for x > xo.f.6. This class of distributions plays a major role in this book as the one within computationally tractable exact forms of the ruin probability z/)(u) can be obtained. resp.. but the current trend in applied probability is to restrict attention to the class of phasetype distributions. the restriction T of the intensity matrix of the Markov process to E and the row vector a = (ai)iEE of initial probabilities. T) is called the representation. Important special cases are the exponential. (or. INTRODUCTION Example 2 .. T) or sometimes the triple (E.8) j=1 j=1 j=1 where the parameters in (2. We give some theory for matrixu exponential distribution in VIII.xo)+ is covered by the reinsurer). Example 2 . The couple (a.4 (PHASETYPE DISTRIBUTIONS) A phasetype distribution is the distribution of the absorption time in a Markov process with finitely many states. B(x) = aeTxe where t = Te and e = (1 . .(2.7) q1 b(x) = cjxieWWx + djxi cos(ajx)ea'x + > ejxi sin(bjx)e`ix . This class of distributions is popular in older literature on both risk theory and queues.8 CHAPTER I.8) are realvalued.5 (DISTRIBUTIONS WITH RATIONAL TRANSFORMS) A distribution B has a rational m. 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 . However.e. which is slightly smaller but more amenable to probabilistic reasoning. are b(x) = aeTxt. The density and c. q2 q3 (2.1 and defer further details to u Chapter VIII.g.6 (DISTRIBUTIONS WITH BOUNDED SUPPORT) This example (i. equivalently. the Erlang and the hyperexponential distributions.7) are possibly complexvalued but the parameters in (2.f. a rational Laplace transform) if B[s] _ p(s)/q(s) with p(s) and q(s) being polynomials of finite degree. B(x) > 0 for x < xo) is of course a trivial instance of a lighttailed distribution. We give a more comprehensive treatment in VIII. of which one is absorbing and the rest transient. Example 2 . a.6. 1)' is the column vector with 1 at all entries.d.
13) u . a)/A)a+1' x > a. in practice one may observe that b(x) is either decreasing or increasing and may try to model smooth (incerasing or decreasing) deviations from constancy by 6(x) = dx''1 (0 < r < oo). There are various variants of the definition around. the tail is B (x ) 2 x. we obtain the Weibull distribution B(x) = eCx'. or equivalently as the distribution of a°U+µ where U . one being B(x) (1 + X)b(x) (1 + x)a+1' x > 0. x < a. u Example 2 . a2). It follows that the density is 't (1ogX . the mean u is eµ+a /2 and the second moment is e2µ+2o2.u l b(x) = d dx or J ax lor 1 exp Asymptotically.7 (THE WEIBULL DISTRIBUTION) This distribution originates from reliability theory.p a 1 (2.8 (THE LOGNORMAL DISTRIBUTION) The lognormal distribution with parameters a2.pl = 1 W (logx . b(x) _ A(1 + (x a The pth moment is finite if and only if p < a .1). (2. (2. (2.12) Sometimes also a location parameter a > 0 and a scale parameter A > 0 is allowed. CLAIM SIZE DISTRIBUTIONS 9 2b Heavytailed distributions Example 2. However.10) The loinormal distribution has moments of all orders.11) ex log logx 2r p 1 1 2 ( a ) f 1 (lox_P)2} (2. Here failure rates b(x) = b(x)/B(x) play an important role. and then b(x) = 0. the exponential distribution representing the simplest example since here b(x) is constant.N(0.1. In particular.9) which is heavytailed when 0 < r < I. Writing c = d/r.N(p.2. Example 2 . b(x) = crx''le`xr. All moments are finite.9 (THE PARETO DISTRIBUTION) Here the essence is that the tail B(x) decreases like a power of x. p is defined as the distribution of ev where V .
15) x2 + 16x3 ) a3x/2) 3 (1 . INTRODUCTION Example 2. in particular. in particular. 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). For p = 1.x6+lr(p) (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) . oo) is slowly varying . another standard example is (log x)').17) where L (x) is slowly varying.2).13). x + 00.11 (PARETO MIXTURES OF EXPONENTIALS) This class was introduced by Abate. A = 1 and X is standard exponential. Choudhury & Whitt [1] as the class of distributions of r.14) The pth moment is finite if p < 5 and infinite if p > 5. The density is 8p(log x)pi b(x) . where Y is Pareto distributed with a = (p . u Example 2 . i. the loggamma distribution is a Pareto distribution.(1 + 2x + 2x2)e2x) p = 2 (2.10 CHAPTER I.16) 11 Example 2. (2.1)/p. In general. B(x) = O(xP). the density is { 3 (1 . x 4 oo (any L having a limit in (0.'s of the form YX.12) (here L (x) * 1) and ( 2. u . the loggamma distribution (with exponent 5) and a Pareto mixture of exponentials. The simplest examples correspond to p small and integervalued.(1 + Zx + $ p = 3.e. { s () 1s+3s29s3log(1+2s I p=3. satisfies L(xt)/L(x) 4 1.10 (THE LOGGAMMA DISTRIBUTION) The loggamma distribution with parameters p. examples of distributions with regularly varying tails are the Pareto distribution (2.L( x ). (2. 6 is defined as the distribution of et' where V has the gamma density (2. Thus.v.
one may argue that this difficulty is not resticted to ruin probability theory alone. which each have a ( timehomogeneous) small rate of experiencing a . (2. From a practical point of view. We return to a closer study in IX.3. At least as important is the specification of the structure of the point process {Nt } of claim arrivals and its possible dependence with the claims.4) or even to completely different applied probability areas like extreme value theory: if we are using a Gaussian process to predict extreme value behaviour. When studying ruin probabilities. THE ARRIVAL PROCESS 11 Example 2. the knowledge of the claim size distribution will typically be based upon statistical data. the subexponential class of distributions provide a convenient framework for studying large classes of heavyu tailed distributions. The reason is in part mathematical since this model is the easiest to analyze. However. for example the lognormal distribution is subexponential (but not regularly varying). Namely. though the proof of this is nontrivial. but can never be sure whether this is also so for atypical levels for which far less detailed statistical information is available. Similar discussion applies to the distribution of the accumulated claims (XI.1) that any distribution with a regularly varying tail is subexponential. By far the most prominent case is the compound Poisson (CramerLundberg) model where {Nt} is Poisson and independent of the claim sizes U1.13 (THE SUBEXPONENTIAL CLASS OF DISTRIBUTIONS) We say that a distribution B is subexponential if xroo lim B `2^ = 2. this phenomenon represents one of the true controversies of the area.1. We give some discussion on standard methods to distinguish between light and heavy tails in Section 4f. Also.. 3 The arrival process For the purpose of modeling a risk process . and based upon such information it seems questionable to extrapolate to tail behaviour. Thus.. it will be seen that we obtain completely different results depending on whether the claim size distribution is exponentially bounded or heavytailed.. the claim size distribution represents of course only one aspect (though a major one).18) B(x) It can be proved (see IX. but the model also admits a natural interpretation : a large portfolio of insurance holders . U2. we may know that such a process (with a covariance function estimated from data) is a reasonable description of the behaviour of the system under study in typical conditions. and so is the Weibull distribution with 0 < r < 1..
with a common term {Nt} is a Markovmodulated Poisson process . such that 8(t) = . where {/3 (t)}too is an arbitrary stochastic process . the negative binomial distribution. getting away from the simple Poisson process seems a crucial step in making the model more realistic. see 11. so that . T2. the first extension to be studied in detail was {Nt } to be renewal (the interarrival times T1 . when Jt = i. to be studied in Chapter V.12 CHAPTER I. In order to prove reasonably substantial and interesting results . gives rise to an arrival process which is very close to a Poisson process. Historically..8 (t) is a periodic function of t. not many detailed studies of the goodnessoffit of the Poisson model in insurance are available . radioactive decay (a huge number of atoms each splitting with a tiny rate ) and many other applications.. The compound Poisson model is studied in detail in Chapters III.e. but with a general not necessarily exponential distribution ).g. in Chapter VII). 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. A more appealing way to allow for inhomogeneity is by means of an intensity .. Nevertheless . To the author 's knowledge . Another one is Cox processes. Mathematically. in particular to allow for certain inhomogeneities. This applies also to the case where the claim size distribution depends on the time of the year or . its basic feature is to allow more variation (bursty arrivals ) than inherent in the simple Poisson process. e. are i. epidemics in life insurance etc. which facilitate the analysis. Cox processes are. with the extension to premiums depending on the reserve. I. it is more questionable whether it provides a model with a similar intuitive content as the Poisson model. However .d.6. In others. found the Poisson distribution to be inadequate and suggested various other univariate distributions as alternatives . This model can be intuitively understood in some simple cases like { Jt} describing weather conditions in car insurance . 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 . An obvious example is 3(t) depending on the time of the year (the season). in just the same way as the Poisson process arises in telephone traffic (a large number of subscribers each calling with a small rate). Some of them have concentrated on the marginal distribution of NT (say T = one year ). and also that the ruin problem may be hard to analyze . This model . the periodic and the Markov modulated models also have attractive features . The one we focus on (Chapter VI) is a Markovian environment : the environmental conditions are described by a finite Markov process {Jt }too. has some mathematically appealing random walk features .i. IV (and. however. INTRODUCTION claim ..(3. too general and one neeed to specialize to more concrete assumptions . The point of view we take here is Markov dependent random walks in continuous time (Markov additive processes ). we study this case in VI . 5.3(t) fluctuating over time.
this gives only f0 O°i (u)du which is of limited . In fact. it is a recurrent theme of this book to stress this connection which is often neglected in the specialized literature on risk theory. point processes and so on. (4. with Poisson arrivals and constant release rule p(x) = 1. others are quite different. More generally. genetics models. The study of the steady state is by far the most dominant topic of queueing and storage theory.4. stochastic differential equations. and which seems well motivated from a practical point of view as well. this amounts to Vo having the stationary distribution of {Vt}). methods or modeling ideas developed in one area often has relevance for the other one as well. and the limit t 4 oo is the steadystate limit. it is desirable to have a set of formulas like (4. ruin probabilities for risk processes with an input process which is renewal.1) where V is the limit in distribution of Vt as t + oo. that quite often the emphasis is on computing expected values like EV.T) = P(VT > u). Similarly.v. A stochastic process {Vt } is said to be in the steady state if it is strictly stationary (in the Markov case.6) .1) holds as well provided the risk process has a premium rule depending on the reserve.'s like V is available. however.1) permitting to translate freely between risk theory and the queueing/storage setting. A general release rule p(x) means that {Vt} decreases according to the differential equation V = p(V) in between jumps. It should be noted. extreme value theory. and here (4. Thus. dam/storage processes. The M/G/1 workload process { Vt } may also be seen as one of the simplest storage models. In the setting of (4. 0(u) = P(V > u). queueing theory. A SUMMARY OF MAIN RESULTS AND METHODS 13 the environment (VI. 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 . Markovmodulated or periodic can be related to queues with similar characteristics. stochastic geometry. Mathematically. and a lot of information on steadystate r. time series and Gaussian processes. reliability. R = p(R) in between jumps.0 (u.1). Some of these have a certain resemblance in flavour and methodology. 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. The ones which appear most related to risk theory are queueing theory and dam/storage processes. others being branching processes. interacting particle systems.
1). The infinite horizon (steady state ) case is covered by letting T oo.3. which can be expanded into a sum of exponential terms by diagonalization (see. 3. Here Vi(u) is given in terms of a matrixexponential function ( Corollary VIII. see Corollary VII. 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. p = 0/8 and y = 8 .8. • The compound Poisson model with premium rate p(x) depending on the reserve and exponential claim size distribution B. as is typically the case. The qualifier 'with just a few phases ' refers to the fact that the diagonalization has to be carried out numerically in higher dimensions.'s like the environmental process {Jt} in a Markovmodulated setting.3.1) in the setting of a general premium rule p(x): the events {VT > u} and {r (u) < T} coincide when the risk process and the storage process are coupled in a suitable way (via timereversion ). the two areas.p(y) y^ Jo p(x) can be written in closed form. Thus . INTRODUCTION intrinsic interest . 4b Exact solutions Of course . the ideal is to be able to come up with closed form solutions for the ruin probabilities 0(u). which gives a sample path version of (4. A prototype of the duality results in this book is Theorem 11. • The compound Poisson model with constant premium rate p = 1 and B being phasetype with a just few phases .3. • The compound Poisson model with some rather special heavytailed claim size distributions.1 .T). B(x) = ebx. e . Vi(u.3. Example VIII..2).14 CHAPTER I. though overlapping. Here ?P(u) is explicit provided that . The cases where this is possible are basically the following for the infinite horizon ruin probability 0(u): • The compound Poisson model with constant premium rate p = 1 and exponential claim size distribution B.1 is a sample path relation should be stressed : in this way the approach also applies to models having supplementary r. Here O(u) = peryu where 3 is the arrival intensity.1.6. 3.g.v. the functions w x f d 1 exdx () . have to some extent a different flavour. Similarly. . The fact that Theorem H. see Boxma & Cohen [74] and Abate & Whitt [3]. • The compound Poisson model with a claim size distribution degenerate at one point. see Corollary III.
T) du dT 0 TO 00 in closed form than the ruin probabilities z/'(u). Here are some of the main approaches: Laplace transform inversion Often. Embrechts. However. Abate & Whitt [2]. We don't discuss Laplace transform inversion much. T) can then be calculated numerically by some method for transform inversion.1) are so complicated that they should rather be viewed as basis for numerical methods than as closedform solutions. where Furrer [150] recently computed ii(u) as an infinite series involving the Mittag. say the fast Fourier transform (FFT) as implemented in Grubel [179] for infinite horizon ruin probabilities for the renewal model. [s. Given this can be done.2) is the natural scale. Grubel & Pitts [132] and Grubel & Hermesmeier [180] (see also the Bibliographical Notes in [307] p. f {eXp U LX 2. esuTb( u. T) themselves. but are somewhat out of the mainstream of the area . . 1).7.b(u)du . a2 (x): Ip (u) = where S(u) = f °O exp {. • An astable Levy process with drift .4. O(u. the second best alternative is a numerical procedure which allows to calculate the exact values of the ruin probabilities.S(u) 1S(oo) f °D exp {. Also Brownian models or certain skip free random walks lead to explicit solutions (see XI .Lef$er function.1) is the explicit form of the ruin probability when {Rt} is a diffusion with infinitesimal drift and variance µ(x). relevant references are Grubel [179]. see VIII. 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. 191).ff 2µ(y)/a2(y) dy} dx . it is easier to find the Laplace transforms = f e8 . (u.f f 2µ(y)/a2(y) dy} dx  (4.u(y)/a2(y) dy} 4c Numerical methods Next to a closedform solution. 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 . T). A notable fact ( see again XI. the formulas ( IV. Ab(u). For the finite horizon ruin probability 0(u.
However. as the solution of linear differential equations or by some series expansion (not necessarily the straightforward Eo U'u/n! one!). whereas for the renewal arrival model and the Markovian environment model U has to be calculated numerically. U is explicit in terms of the model parameters. T) as the solution to a differential. most often it is more difficult to come up with reasonably simple equations than one may believe at a first sight.and integral equations The idea is here to express 'O(u) or '(u. and in particular the naive idea of conditioning upon process behaviour in [0. Differential.3) in the compound Poisson model which is an integral equation of Volterra type. 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. . One example where this is feasible is the renewal equation for tl'(u) (Corollary III.7.1) and y > 0 is the solution of the Lundberg equation (4.3) where C = (1 .3.or integral equation. In the compound Poisson model with p = 1. 0(u) is then given in terms of a matrixexponential function euu (here U is some suitable matrix) which can be computed by diagonalization. it states that i/i(u) . which can equivalently be written as f3 [7] = 1 +13 . For the compound Poisson model with p = 1 and claim size distribution B with moment generating function (m.p)/(13B'[ry] . and in quite a few cases (Chapter VIII). (4. see VIII.Ce"u.f. u * oo.g. dt] most often leads to equations involving both differential and integral terms.4) 00['Y]1)'Y = 0. and carry out the solution by some standard numerical method.) B[s]. 4d Approximations The CramdrLundberg approximation This is one of the most celebrated result of risk theory (and probability theory as a whole). 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.16 CHAPTER L INTRODUCTION Matrixanalytic methods This approach is relevant when the claim size distribution is of phasetype (or matrixexponential).
but typically not very precise in their first naive implementation. in such cases the evaluation of C is more cumbersome. a Markovian environment or periodically varying parameters. some further possibilities are surveyed in 111 . In fact. . In the case of heavytailed distributions. Diffusion approximations Here the idea is simply to approximate the risk process by a Brownian motion (or a more general diffusion) by fitting the first and second moment. the exact solution is as easy to compute as the CramerLundberg approximation at least in the first two of these three models. for the compound Poisson model ^(u) p pu In fact . the claim size distribution should have an exponentially decreasing tail B(x). Approximations for O(u) as well as for 1(u. See Chapter IX. (4. T) for large u are available in most of the models we discuss.2. and use the fact that first passage probabilities are more readily calculated for diffusions than for the risk process itself. However. For example. However. In particular. corrected diffusion approximations (see IV. often for all u > 0 and not just for large u. u > oo. Large claims approximations In order for the CramerLundberg approximation to be valid. 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. Diffusion approximations are easy to calculate. This list of approximations does by no means exhaust the topic. The CramerLundberg approximation is renowned not only for its mathematical beauty but also for being very precise. in some cases the results are even more complete than for light tails.4.7 and IV. T). other approaches are thus required. when the claim size distribution is of phasetype. incorporating correction terms may change the picture dramatically. It has generalizations to the models with renewal arrivals.6) are by far the best one can do in terms of finite horizon ruin probabilities '(u. J B dx.6) 4e Bounds and inequalities The outstanding result in the area is Lundberg's inequality (u) < e"lu.
obtained say by observing the risk process in [0. For example.3). . in the compound Poisson model.8.. When comparing different risk models.x U > x] = B(x) f '(yx)B(dx) typically has a finite limit (possibly 0) in the lighttailed case and goes to oo in the heavytailed case. In practice. say degenerate at m. they have however to be estimated from data.i. given NT. T]. How do we produce a confidence interval? And. This procedure in itself is fairly straightforward. which is a standard statistical problem since the claim sizes Ui. the difficulty comes in when drawing inference about the ruin probabilities. lower bounds etc.U(k)) i =k+ i . This is proved for the compound Poisson model in 111.k (U(`) .) at various places and in various settings. However.d. For example.. fitting a parametric model to U1. empirical evidence shows that the general principle holds in a broad variety of settings. one may question whether it is possible to distinguish between claim size distributions which are heavytailed or have an exponentially decaying tail. e. . can we trust the confidence intervals for the large values of u which are of interest? In the present author's opinion.18 CHAPTER I. The standard suggestion is to observe that the mean residual life E[U .. of being somewhat easier to generalize beyond the compound Poisson setting. one expects a model with a deterministic claim size distribution B. . UNT are i. 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. and to plot the empirical mean residual life 1 N . to have smaller ruin probabilities than when B is nondegenerate with the same mean m. more importantly. this is extrapolation from data due to the extreme sensitivity of the ruin probabilities to the tail of the claim size distribution in particular (in contrast. as a general rule. However.g. INTRODUCTION Compared to the CramerLundberg approximation (4. 4f Statistical methods Any of the approaches and results above assume that the parameters of the model are completely known. it is a general principle that adding random variation to a model increases the risk. . UNT may be viewed as an interpolation in or smoothing of the histogram). . though not too many precise mathematical results have been obtained. . it has the advantage of not involving approximations and also. We return to various extensions and sharpenings of Lundberg's inequality (finite horizon versions.
Simulation may be used just to get some vague insight in the process under study: simulate one or several sample paths. and of course the method is relevant in risk theory as well. but is not very satisfying. to observe whether one or the other limiting behaviour is apparent in the tail. respectively. (or a functional of the expectation of a set of r.2.3) or Section 3 of Chapter VI are referred to as Proposition VI. Thus Proposition 4. Truncation to a finite horizon has been used. and also discuss how to develop methods which are efficient in terms of producing a small variance for a fixed simulation budget.. this is a straightforward way to estimate finite horizon ruin probabilities.. We look at a variety of such methods in Chapter X.3 (or just VI. For example. . < U(N) are the order statistics based upon N i. 5 Conventions Numbering and reference system The basic principles are just as in the author's earlier book Applied Probability and Queues (Wiley 1987.e. claims U1. However.d. in all other chapters than VI where we just write .2.v. UN. reference [14].3). . in this book referred to as [APQ]).. good methods exist in a number of models and are based upon representing the ruin probability zb(u) as expected value of a r. formula (5. and in fact methods from that area can often be used in risk theory as well . formula VI.i. See further Embrechts. Still.4. having small probability) and that therefore naive simulation is expensive or even infeasible in terms of computer time. the more typical situation is to perform a Monte Carlo experiment to estimate probabilities (or expectations or distributions) which are not analytically available. and look at them to see whether they exhibit the expected behaviour or some surprises come up. because it appears to require an infinitely long simulation.3) and Section VI. Klnppelberg & Mikosch [134]. CONVENTIONS 19 as function of U(k). A main problem is that ruin is typically a rare event (i.v's) which can be generated by simulation. where U(1) < . The chapter number is specified only when it is not the current one...(5.5. The infinite horizon case presents a difficulty. The problem is entirely analogous to estimating steadystate characteristics by simulation in queueing/storage theory. 4g Simulation The development of modern computers have made simulation a popular experimental tool in all branches of applied probability and statistics.
2. B[s] the m.20 CHAPTER L INTRODUCTION Proposition 4.h. oo).The same symbol B is used for a probability measure B(dx) = P(X E dx) and its c. with respect to w. . i. (moment generating function) fm e82B(dx) of the distribution B.d. moment generating function.g.t.s. References like Proposition A. w.s. B(x) the tail 1 .29) refer to the Appendix.d.v.f.f. cumulant generating function. B(dy). mation. and for a defective probability distribution IIGII < 1.e. EX2/(EX)2. h + 0. with probability Mathematical notation P probability.g.g.d. (A. E expectation. n i oo.p. A different type of asymptotics: less precise. i. random variable s. then for 1s < 5).f. if B(x) . Abbreviations c. as for typical claim size distributions. b[s] is defined always if Rs < 0 and sometimes in a larger strip (for example.B(x) = P(X > x) of B.g. for a probability distribution IIGII = 1. infinitely often l.ceax.h.f.f. say a heuristic approxi1 + h + h2/2. cumulative distribution function P(X < x) c.r. In particular.4. formula (5. The Laplace transform is b[s]. independent identically distributed i. E. see under b[s] below.g. B is concentrated on [0.f.c. or a more precise one like eh . B(x) = P(X < x) = fx. squared coefficient of variation.o. log E[s] where b[s] is the m. n! 27r nn+1/2en. . r. IIGII the total mass (variation ) of a (signed ) measure G .v.Used in asymptotic relations to indicate that the ratio between two expressions is 1 in the limit. right hand side (of equation) r. If.i.3) or Section 3. left hand side (of equation) m.
.e. Thus. 21 D [0. In the Frenchinspired literature. a2) the normal distribution with mean p and variance oa2. i.. (the dimension is usually clear from the context and left unspecified in the notation).5.e. N(it. row vectors have lowercase Greek letters like a. E[X. and column vectors have lowercase Roman letters like t. Notation like f3i and 3(t) in Chapter VI has a similar . 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. the ith entry is 1 and all other 0. Usually. Xt_ the left limit limstt X8f i. the ith unit row vector is e'i. I(A) the indicator function of the event A. Matrices and vectors are denoted by bold letters. Usually. the processes we consider are piecewise continuous. of numbers.A] means E[XI(A)]. Unless otherwise stated. an example or a remark. . matrices have uppercase Roman or Greek letters like T. In particular: I is the identity matrix e is the column vector with all entries equal to 1 ei is the ith unit column vector. a. i. the value just before t. 7r. intensity interpretation. oo) the space of Rvalued functions which are rightcontionuous and have left limits. (xi)diag denotes the diagonal matrix with the xi on the diagonal (xi)row denotes the row vector with the xi as components (xi). all stochastic processes considered in this book are assumed to have sample paths in this space. xa. F o r a given set x1. CONVENTIONS {6B the mean EX = f xB(dx) of B ABA' the nth moment EXn = f x"B(dx) of B. only have finitely many jumps in each finite interval.e..oi denotes the column vector with the xi as components Special notation for risk processes /3 the arrival intensity (when the arrival process is Poisson). though slightly more complicated. R(s) the real part of a complex number s. often the term 'cadlag' (continues a droite avec limites a gauche) is used for the Dproperty. . 0 marks the end of a proof. A.
I. or quantities with a similar time average interpretation. cf. p the net amount /3pB of claims per unit time. EL the probability measure and its corresponding expectation corresponding to the exponential change of measure given by Lundberg conjugation.1.5.g. e. Notation like BE and B(t) in Chapter VI has a similar. FL. ry The adjustment coefficient.1. cf. cf. . J the rate parameter of B for the exponential case B(x) = eby. I.5. VI. though slightly more complicated. 111. 'q the safety loading .22 CHAPTER L INTRODUCTION B the claim size distribution. interpretation.
The topic is. a parallel selfcontained treatment is given of the facts needed there. strictly speaking. More precisely. The reader should therefore observe that it is possible to skip most of the chapter. The duality results in Section 3 (and. the relevance for the mainstream of exposition is the following: The martingale approach in Section 1 is essentially only used here. 5 on random walks and Markov additive processes can be skipped until reading Chapter VI on the Markovian environment model. used in Chapter VI on risk processes in a Markovian (or periodic) environment. however. fundamental ( at least in the author' s opinion) and the probability involved is rather simple and intuitive. The general theory is. Sections 4. 5) are. The likelihood ratio approach in Section 2 is basic for most of the models under study. Sections 4. When encountered for the first time in connection with the compound Poisson model in Chapter III. in most cases via likelihood ratio arguments. Due to the generality of the theory. not crucial for the rest of the book. All results are proved elsewhere . in particular at a first reading of the book. in part. somewhat more advanced than in the rest of the book. 23 . however. however.Chapter II Some general tools and results The present chapter collects and surveys some topics which repeatedly show up in the study of ruin probabilities. the level of the exposition is.
SOME GENERAL TOOLS AND RESULTS The ladder height formula in Theorem 6. We get 1 = Ee7So = E e'Y S(.u denote the overshoot. StUit. f1 . and in the limit (1.QµB < 1. claim size distribution B and p = .2) takes the form 1 = E [e'ys(). The more general Theorem 6. the time to ruin r(u) is inf It > 0 : St > u}.1 is basic for the study of the compound Poisson model in Chapter III.T(u) < cc] = e7uE {e7f(u) I T(u) < cc] z/. 1 Martingales We consider the claim surplus process {St} of a general risk process.)AT = E [e7ST(°). and the ruin probabilities are ip(u) = P (T(u) < oo). Proposition 1. Our first result is a representation formula for O(u) obtained by using the martingale optional stopping theorem .(u).. Thus N. (b) St a$ oo on {T(u) = oo}.2 Consider the compound Poisson model with Poisson arrival rate . however. T(u) < T] + E [eryST .0. T(u) < oo] + 0 = eryuE [e7Vu).5 can be skipped. (1. V) (u. As usual. Example 1 . Let e(u) = ST(u) . T) = P(T(u) < T).24 CHAPTER II. the second term converges to 0 by (b) and dominated convergence (e7ST < eryu on {r(u) > T}). using r(u) A T invokes no problems because r(u) A T is bounded by T).1 Assume that (a) for some ry > 0.2) As T > oo. {e'YS° }t>0 is a martingale. T(u) > T] . Then e7u (u) = E[e74(u)j7(u) < oo] Proof We shall use optional stopping at time r(u)AT (we cannot use the stopping time T(u) directly because P(T(u) = oo) > 0 and also because the conditions of the optional stopping time theorem present a problem.
ap.5 For the compound Poisson model with B exponential.Ft = a(S" : v < t). B(x) _ edx.g. 1.r" where y = S . of the normal distribution. u Corollary 1. the martingale property now follows just as in Example 1. with common distribution B (and independent of {Nt}). By standard formulas for the m. and (b) follows from p < 1 and the law of large numbers (see Proposition III. From this it is readily seen (see III.1) shows that Eels.i.1. Example 1 .3/6 < 1. are i. u Corollary 1. Since {St} has stationary independent increments. Under the conditions of Proposi Proof Just note that C(u) > 0. the ruin probability is O(u) = pe.2 and drift p > 0. the conditions of Proposition 1. Eeas° = e"(°) where n(a) = a2a2/2 ./3. Since {St} has stationary independent increments.(„)_ = x is that of a claim size U given U > u . Proof Since c(a) = /3 (B[a] .Q.Q(B[a] .1 is satisfied. O(u ) < e7".2(c)).x. and p =.u + x is again just exponential with rate S.1 are satisfied. Thus 00 E [e'rt (") I T(u) < oo] = I e5e  dx = f 5edx . it follows that E [e7st+v I J] = e"rstE [e7(st+vSt) I Ft] = e7StEe"rs° = elst where .f. A simple calculation (see Proposition III. Thus.a. Then {St } is Brownian motion with variance constant o2 and drift p < 0. Thus.2. condition (a) of Proposition 1.6a for details) that typically a solution to the Lundberg equation K(y) = 0 exists.1. MARTINGALES 25 where {Nt} is a Poisson process with rate .a = a . the conditional distribution of the overshoot e(u) = U .Q and the U. and thus Ee7s° = 1.1) . The available information on this jump is that the distribution given r(u) = t and S.1) . Now at the time r(u) of ruin {St} upcrosses level u by making a jump ..3 Assume that {Rt} is Brownian motion with variance constant o. and thus Ee'rs° = 1.a it is immediately seen that y = S .4 (LUNDBERG ' S INEQUALITY ) tion 1 .= e"(') where K(a) = .1. From this it is immediately seen that the solution to the Lundberg equation ic(y) = 0 is y = 2p/a2.d. and thus by the memoryless property of the exponential distribution .
Ft}too and the Borel afield F. Grandell [171]. A E Ft.6 N S = { lim Nt I t +00 t gJ are both in F. oo)..26 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Corollary 1.1 Let F. F(S) = P(S) = 1. and by the law of large numbers for the Poisson process . Example 2 . P correspond to the claim surplus process of two compound Poisson risk processes with Poisson rates /3. and is further exploited in his book [157]. then z/'(u) = e7" where 'y = 21A/a2. which we equip with the natural filtration {. P on (DE. Theorem 2. More recent references are Dassios & Embrechts [98]. (2. P are then singular (concentrated on two disjoint measurable sets). cf. B.F).6 If {Rt} is Brownian motion with variance constant a2 and drift p > 0.3 below). Two such processes may be represented by probability measures F. The number Nt F) of jumps > e before time t is a (measurable) r.e. F).1) 'though not always: it is not difficult to construct a counterexample say in terms of transient Markov processes. The interesting concept is therefore to look for absolute continuity only on finite time intervals (possibly random. hence so is Nt = limfyo N2`i. 0 and claim size distributions B. A somewhat similar u argument gives singularity when B $ B. I. Thus the sets S = I tlim +oot t =. on (DE. and in analogy with the theory of measures on finite dimensional spaces one could study conditions for the RadonNikodym derivative dP/dP to exist. Grandell & Schmidli [131].v. u Notes and references The first use of martingales in risk theory is due to Gerber [156]. the parameters of the two processes can be reconstructed from a single infinite path. Proof Just note that ^(u) = 0 by continuity of Brownian motion. [172]. . Embrechts. But if a $ ^ . A]. Delbaen & Haezendonck [103] and Schmidli [320]. 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. then S and S are disjoint . However. and F. we look for a process {Lt} (the likelihood ratio process) such that P(A) = E[Lt. as shown by the following example this setup is too restrictive: typically'.
if for some probability measure P and some {. .r. P be as in Proposition 2.2(i).F).2) Proof Assume first G C {T < T} for some fixed deterministic T < oo. u The following likelihood ratio identity (typically with r being the time r(u) to ruin) is a fundamental tool throughout the book: Theorem 2 . G ] = E [LT .3 Let {Lt}.1) holds. Lt < 0] can only be nonnegative if P(A) = 0.Y) such that P(A) = Pt(A). implies that E[LtI. P) such that LLt = 1. ELt = 1 follows by taking A = DE in (2. Then P(A) = E[Lt.A] = EE[LtI(A)IF8] = EI(A)E[LtIFB] = EI(A)L8 = PS(A). Proof Under the assumptions of (i). Hence E [_ . (i) If {Lt}t> o is a nonnegative martingale w. F) such that ELt = 1.F such that (2.e.r. then {Lt} is a nonnegative martingale w.2 Let {Ft}t>o be the natural filtration on DE. This proves (i). then there exists a unique probability measure Pon . under the assumptions of (ii) we have for A E rg and s < t that A E Ft as well and hence E[L8. we have E [ LTIFT]1 = LT on {T < T}. Conversely. F the Borel o•field and P a given probability measure on (DE. If r is a stopping time and G E PT. Ft). 1 J (2. G C {T < oo}. The truth of this for all A E Y.t.1) and nonnegativity by letting A = {Lt < 0}. Proposition 2. then { 1 P(G) = EG . _. Hence the family {Pt} is t>o consistent and hence extendable to a probability measure F on (DE.t. A].1) holds. Then Lt > 0 and ELt = 1 ensure that Pt is a probability measure on (DE. ({Ft} . By the martingale property. A E F. Then Ft (A) = E[Lt. Finally. define P by Pt (A) = E[Lt. A]. ({. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 27 (i. G l ] = E [_I(G)E[LTIFT] ] = E { _I(G)Lr ] = P(G). (ii) Conversely.Ft}.t.2. A] = E[Lt.. the restriction of P to (DE.r. A E Ft .Tt) is absolutely continuous w. Lets < t.F8] = L8 and the martingale property. .Pt)) The following result gives the connection to martingales. using the martingale property in the fourth step. that the restriction of P to (DE. A E F8.Pt}adapted process {Lt}t>o (2.
A change of measure is performed by finding a process {Lt} which is a martingale w.3) to G of{r < T} we get 1111 F(Gn {r <T}) = E[ 1 . where {Rt} is the risk reserve process. .r.s.Gn {r <T} . Lr(u) 11 The advantage of (2. T(u) < oo]. {St} = {u . r(u) < oo] occuring there than with the (conditional) expectation E[e'r{(u ) Jr(u) < oo] in (1. The crucial step is to obtain information on the process evolving according to F.4 Under condition (a) of Proposition 1. we have F(G) = V )(u).1) in Proposition 1. and this problem will now be studied. the natural filtration {. is nonnegative and has Ey Lt = 1 for all x. we assume for simplicity that {Xt} has Dpaths. we need the concept of a multiplicative functional. first in the Markov case and next (Sections 4. (2. Rt) or Xt = (Jt.t. Consider a (timehomogeneous) Markov process {Xt} with state space E.2) by noting that 1 = ersr(„) = e1'ue7Ou). First we ask when the Markov property is preserved.1). The problem is thus to investigate which characteristics of {Xt} and {Lt} ensure a given set of properties of the changed probability measure.O(u) = eryuE[e'YC(u).4) compared to (1. is Markov w.Rt} the claim surplus process and {Jt} a process of supplementary variables possibly needed to make the process Markovian. both sides are increasing in T.t.1) is that it seems in general easier to deal with the (unconditional) expectation E[eryVu). St). (2.r. To this end. in continuous time (the discrete time case is parallel but slightly simpler)..4) Proof Letting G = {r(u) < oo}..Ft} . Now just rewrite the r. say. u From Theorem 2. and letting T * oo. t. each F. applying (2. one would typically have Xt = Rt.1: Corollary 2. SOME GENERAL TOOLS AND RESULTS In the general case . 5) for processes with some randomwalklike structure.1. Xt = (Jt. Xt = St.3 we obtain a likelihood ratio representation of the ruin probability V) (u) parallel to the martingale representation (1. In the context of ruin probabilities. of (2. 1 Since everything is non negative.28 CHAPTER II. For the definition.2) follows by monotone convergence.h.
v.Ftmeasurable. (2. the Markov property can be written E.v. Then the family {Px}xEE defines a timehomogeneous Markov process if and only if {Lt} is a multiplicative functional. The converse follows since the class of r. or.2. (2.Ft].t. Y8. let {Lt} be a nonnegative martingale with Ex Lt = 1 for all x. t and let Px be the probability measure given by t. s. (2. Indeed.F8measurable r.'s of the form fl' f.(Xtitl) with all t(i) < t + s.7). t] * [0.Ft+8measurable.v.5) Pxa. 0 . and then L. this in turn means Ex[Lt+8Zt(V8 oet)] = Ex[LtZtExt[L8Y8]]. Similarly. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 29 on DE and define {Lt} to be a multiplicative functional if {Lt} is adapted to {. A]. o 9t = V.Ft }. Proof Since both sides of (2. Lt has the form Lt = 'Pt ({x }0<u<t) for some mapping cot : DE[O.T9measurable Y8.7).Ft] = Ex.6) for any . o 9tI.v.r..Y8f t < s.. which is the same as Ex[Zt(Y8 o Bt)] = E8[ZtEx.7) for any Ftmeasurable Zt and any .[Y. Ex[Lt+8Zt(Y8 o et)] = Ex[LtZt(Y8 o 0t)(L8 o Bt)].(A) = Ex [Lt. t. oo).5 Let {Xt} be Markov w. ({Xt+u} 0<u<8) Theorem 2. since Ext [L8Y8] = E[(Y8 o et)(L8 o 8t)I. nonnegative and Lt+8 = Lt•(Lso9t) (2.Pt+8measurable r. By definition of Px.8) which is the same as (2. (2. for all x.s. The precise meaning of this is the following: being . since Zt • (Y8 o Ot ) is . the natural filtration {Ft} on DE.5) are Tt+e measurable.6) implies (2.YB] for any Ftmeasurable r.5) is equivalent to Ex[Lt+8Vt+8] = E8[Lt • (L8 o 91)Vt+8] for any . which in turn is the same as Ex[Lt+8Zt • (V8 o Bt)] = Ex[Lt • (L8 o 91)Z1 • (Y8 o et)] (2. Vt+e. where Ot is the shift operator. Zt.v.'s of the form Zt • (Y8 o 0t) comprises all r.
(using the Markov property in the second step) so that the martingale property is automatic. A more elementary version along the lines of Theorem 2.e. The result is a sample path relation. the premium rate is p(r) when the reserve is r (i. The formulation has applications to virtually all the risk models studied in this book. We work on a finite time interval [0. with a proof somewhat different from the present one. R = p(R)). . T] in the following setup: The risk process {Rt}o<t<T has arrivals at epochs or. Indeed. SOME GENERAL TOOLS AND RESULTS to define a timehomogeneous Markov process. In between jumps. u Notes and references The results of the present section are essentially known in a very general Markov process formulation.1) The initial condition is arbitrary. it xEE suffices to assume that {Lt} is a multiplicative functional with Ex Lt = 1 for all x.At where At = k: vk <t U... < aN < T. then Remark 2.. and thus for the moment no parametric assumptions (on say the structure of the arrival process) are needed.30 CHAPTER H. reflection at zero and initiar condition Vo = 0. . More precisely . t. Thus R = Ro + f p(R8) ds . see Dynkin [128] and Kunita [239]... and just after time or* {Vt} makes an upwards jump of size UU = UN _k+l. ..6 For {u . .. The storage process {Vt }o<t<T is essentially defined by time reversion. UN. 0 < vl < . Ro = u (say). {Vt} .. t] = LtExt L8 = Lt. aN where or* = T UN_k+l. and the time to ruin is 7(u) = inf {t > 0: Rt < 0}.5 can be found in Kuchler & Sorensen [240]. In between jumps. the arrival epochs are Qi. (3.. The corresponding claim sizes are Ul. } E[Lt+B I.Ft] = LtE[L8 o 9t I. we shall establish a general connection between ruin probabilities and certain stochastic processes which occurs for example as models for queueing and storage... 3 Duality with other applied probability models In this section. . CN.
instead of (3.____•_.T) = P(VT > u). 3.e.. That is.AT_t..2) k: ok <t and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0. Note that these definitions make {Rt} rightcontinuous (as standard) and {Vt} leftcontinuous.3) (u) Proof Let rt' denote the solution of R = p(R) subject to r0 = u. .x._: 1} 0 011 =T01N ^N3 To 0 011 014 01N Figure 3.. :. The sample path relation between these two processes is illustrated in Fig.. (3.1. (3.__. {Vt} remains at 0 until the next arrival)._. Then rt°) > rt°) for all t when u > v. V)(u... DUALITY WITH OTHER APPLIED PROBABILITY MODELS 31 decreases at rate p(r) when Vt = r (i.T) = inf Rt < 0 P (O<t<T P(r(u) < T) be the ruin probability.. V = p(V)).1 Define r(u) = inf It > 0: Rt < 0} (r(u) = oo if Rt > 0 for all t < T) and let ii(u.11 4..1) we have Vt = At  f P(Vs)ds where A= U= AT .3. Theorem 3. In particular.1 The events {T(u) < T} and {VT > u} coincide..___ .__.
if nothing else n = N). and so on. Historically. u A basic example is when {Rt} is the risk reserve process corresponding to claims arriving at Poisson rate .3 and being i. Some further relevant more general references are Asmussen [21] and Asmussen & Sigman [51].32 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Suppose first VT > u (this situation corresponds to the solid path of {Rt} in Fig. Suppose next VT < u (this situation corresponds to the broken path of {Rt} in Fig.2 Consider the compound Poisson risk model with a general premium rule p(r).T l . say V. If VaN > 0. and a general premium rule p(r) when the reserve is r. see Siegmund [344]. We get: Corollary 3.Ul < roil  Ul = RQ„ Va1V_1 < RQ2.U1 = Rol. Then the storage process {Vt} has a proper limit in distribution.1 with Ro = u = u2).1 with Ro = u = ul). Theorem 3. the distinction between right. Then similarly VVN = r0. Then the time reversibility of the Poisson process ensures that {At } and {At } have the same distribution (for finitedimensional distributions. . The arrival epochs correspond to rainfalls. if and only if O(u) < 1 for all u. 3. say of water in a dam though other interpretations like the amount of goods stored are also possible. and in between rainfalls water is released at rate p(r) when Vt (the content) is r. 3. and then '0 (u) = P(V > u). with distribution B. the connection between risk theory and other applied probability areas appears first to have been noted by Prabhu [293] in a queueing context. Historically. The results can be viewed as special cases of Siegmund duality. one may feel that the interaction between the different areas has been surprisingly limited even up to today. we have RQ„ < 0 so that indeed r(u) < T. we can repeat the argument and get VoN_1 > Ra2 and so on. u Notes and references Some main reference on storage processes are Harrison & Resnick [187] and Brockwell.d. Corollary 3.1 and its proof is from Asmussen & Schock Petersen [50]. Nevertheless.i. Hence if n satisfies VVN_n+1 = 0 (such a n exists. Hence RQ„ > 0 for all n < N. Resnick & Tweedie [79]. we have r(u) > T.2 from Harrison & Resnick [188]. Thus we may think of {Vt} as having compound Poisson input and being defined for all t < oo.U1 > roil . this represents a model for storage.3). and since ruin can only occur at the times of claims.and left continuity is immaterial because the probability of a Poisson arrival at any fixed time t is zero). Proof Let T ^ oo in (3. = r(VT) . Then Vo.
. Z2 = YN_1 i . Z2.min n=0.. . Most often.1 Let r(u) = inf In: u + Y1 + • • • + Yn < 0}.2) (for a rigorous proof..d.....Yl min (YN .4. WN be the Lindley process generated by Z1 = YN. Theorem 4. R valued sequence Z1.... just verify that the r . .. .. there is an analogue of Theorem 3. .1. N min (Y1 + + Yn). Then the events {r(u) < N} and {WN > u} coincide.. ZN = .. .2) satisfies the same recursion as in (4. = Xo + Y1 + • • • + Y.2). evolves as a random walk with increments Z1i Z2. Xo = 0.h.w. RANDOM WALKS IN DISCRETE OR CONTINUOUS TIME 33 4 Random walks in discrete or continuous time A random walk in discrete time is defined as X.. WN = YN ..i.. For discrete time random walks . the Lindley process Wo.. 0 Corollary 4... Z2 = Y2. Let further N be fixed and let Wo... of (4.. if Wo = 0 then (Z1+•••+Zn) WN = Zl+•••+ZN.2 The following assertions are equivalent: (a) 0(u) = P(r(u) < oo) < 1 for all u > 0. where the Yi are i ..N From this the result immediately follows. i.. is defined by assigning Wo some arbitrary value > 0 and letting Wn+1 = (Wn + Zn+1)+• (4. as long as the random walk only takes nonnegative values. can be viewed as the reflected version of the random walk with increments Z1. 0).. W1.1)). with common distribution F (say)...... W1... n=0.. W2. (b) 1/i(u) = P(•r(u) < oo) > 0 as u * oo.... In particular..1.. generated by Z1. . N min (Y1 + • • • + YNn) n=0. ....1.1) Thus {Wn}n=o. Z2 .s.1 in terms of Lindley processes .. I. {Wn}n=0. (c) The Lindley process {WN} generated by Zl = Y1....YNn+1) n=0. hits (oo.Y1 according to Wo = 0. has a proper limit W in distribution as n + oo.1.1. Z2. Here F is a general probability distribution on R (the special case of F being concentrated on {1.d.. 1} is often referred to as simple random walk or Bernoulli random walk). Proof By (4.N (4.e. . and is reset to 0 once the r.1. .... For a given i..
.l. ..s ...34 CHAPTER II.5 does not necessarily lead to a random walk: if.... w. the Lindley processes in Corollary 4. ..1. (d) #.3 The i.ooa. 0 By the law of large numbers.2... Similarly. YN in Theorem 4.. e. a Markovian change of measure as in Theorem 2. assumption on the Z1. equivalently.=o. Thus the assertion of Theorem 4.. there is a more general version of Corollary 4. either M = oo a.) and defines Zn = Yn.s... 176) but appears to be rather intractable. W v m and P(W > u) = P (m > u) = 0(u)... For a random walk.1 is actually not necessary . In general. The converse follows from general random walk theory since it is standard that lim sup (Y1 + • • + Yn) = oo when Y1 + • • • + Yn 74 oo. (Yi + • • • + Yn) > oo a. YN).. . (e).d. . . . . ..1 have the same distribution for n = 0.s.s. + Z. By Kolmogorov's 01 law. .l. Proof Since (YN.. ZN or. the Y1. doubly u infinite (n'= 0. ±1.) sup n=0..1 is equivalent to WN D MN = (Z1 + . Clearly.g.1. The following result gives the necessary and sufficient condition for {Ln} to define a new random walk: ..1. a sufficient condition for (e) is that EY is welldefined and > 0.N so that WN _P4 M = supra=0. . (e)Yi+•••+Yn 74 ..1. In that case . the condition 00 F(YI+•••+ Yn<0)<00 n=1 is known to be necessary and sufficient ((APQ] p. then the restrictions of Fx. (Z1 + • • • + Zn) = m and P(W > u) = P(M > u) = i (u ). One then assumes Yn to be a stationary sequence..i.g. Y1) has the same distribution as (Y1..the result is a sample path relation as is Theorem 3.. SOME GENERAL TOOLS AND RESULTS (d) m = inf. Px to Fn are equivalent (have the same null sets) so that the likelihood ratio Ln exists. or M < oo a..o. 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. Remark 4 . ±2. Next consider change of measure via likelihood ratios.2 and Theorem 4. . Combining these facts gives easily the equivalence of (a)(d). N.
where h (y) = g(0. Y ) f (Y)] = E[g(O..5 Consider a random walk and an a such that c(a) = log F[a] = log Ee° ' is finite.3) 1Pxa. Breiman [78] p. and so onforn =3. For n = 2. Then the change of measure in Theorem 2.3) holds for n = 1. In particular. the random walk property implies Ex f (Y1) = Eo f (Y1 ). Y1).4) ({Ln} is the familiar Wald martingale ..s.5 corresponds to a new random walk if and only if Ln = h(Y1) . Then the change of measure in Theorem 2.g...nrc(a )} (4. Proof If (4. of F). for some function h with Eh(Y) = 1.. = 1 for all n and x. u A particular important example is exponential change of measure (h(y) = e°y'(") where r..g.4. In that case. and define Ln by (4. 100 ). We get: Corollary 4. RANDOM WALKS INDISCRETE OR CONTINUOUS TIME 35 Proposition 4. h(Yn) (4. The corresponding likelihood ratio is Ln = exp {a (Y1 + • • • + Yn) . . the changed increment distribution is F(x ) = E[h(Y).. 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.Y2) = h(1'i)h(I'a). e. cf. implying g (x.s. Y < x]. this means E(g(x. y ).3) holds.4. Y) = h(Y ) a. Conversely.4 Let {Ln} be a multiplicative functional of a random walk with E_L.f.5 corresponds to a new random walk with changed increment distribution F(x) = e'(a) Jr e"'F(dy) . Since L1 has the form g (Xo. Y) f (Y)] for all f and x.4). (a) = log F(a] is the c. we get L2 = L1 (L1 o91 ) = h(Y1)g(X1. (4.
a positive measure on R with the properties e J x2v(dx) < oo. i. Roughly. or imbedded into continuous time processes . however. The appropriate generalization of random walks to continuous time is processes with stationary independent increments (Levy processes).Xn < x).7) for all e > 0. with premium rate p..36 CHAPTER II. they arise as models for the reserve or claim surplus at a discrete sequence of instants. Xt =Xo+pt+oBt+Mt. (4. v2 = 0 and v = 3B).1). < t(n) and with Xt( i)_t(i_l) having distribution depending only on t(i) . a Brownian component {Bt} (scaled by a variance constant) and a pure jump process {Mt}. SOME GENERAL TOOLS AND RESULTS Discrete time random walks have classical applications in queueing theory via the Lindley process representation of the waiting time . see Chapter V. Xt(2)_t(l). . In discrete time. given by a the increment distribution F(x) = P(Xn+l . but we omit the details ). the claim surplus process for the compound Poisson risk model .e. However. we are . In continuous time. say the beginning of each month or year . In risk theory. the interpretation is that the rate of a jump of size x is v(dx) (if f of Ixlv (dx) = oo. this description needs some amendments. (4..Xt)I.t(i . {Xt} can be written as the independent sum of a pure drift {pt}..5) Note that the structure of such a process admits a complete description. the tradition in the area is to use continuous time models. e f x:IxJ>e} v(dx) < oo (4. say by recording the reserve or claim surplus just before or just after claims (see Chapter V for some fundamental examples). which corresponds to the compound Poisson case: here jumps of {Mt} occur at rate 0 and have distribution B = v/0 (in particular .).6) More precisely. . An equivalent characterisation is {Xt} being Markov with state space R and E [f (Xt+e . corresponds to a process with stationary independent increments and u = p. The simplest case is 3 = JhvMM < oo. {Xt} is a random walk.Ft] = Eof (X. The traditional formal definition is that {Xt} is Rvalued with the increments Xt(1)_t(o). the pure jump process is given by its Levy measure v(dx).. 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 (n)t(n1) being independent whenever t(O) < t(1 ) < .
v.8) O<t<T (assuming Wo = Xo = 0 for simplicity).Q and distribution B of the service times of the arriving customers. Then the storage process {Vt} has constant release rate 1.3 and decreases linearly at rate 1 in between jumps.6). then Ee'(xtxo) = Eoeaxt = etx(a).s.1. virtual waiting time refers to Vt being the amount of time a customer would have to wait before starting service if he arrived at time t (this interpretation requires FIFO = First In First Out queueing discipline: the customers are served in the order of arrival). i. 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. Chapter III. 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. defined as a system with a single server working at a unit rate. has upwards jumps governed by B at the epochs of a Poisson process with rate . and b(u) = P(V > u). 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.e. WT = XT .6 In the compound Poisson risk model with constant premium rate p(r) .min Xt (4.] Processes with a more complicated path structure like Brownian motion or jump processes with unbounded Levy measure are not covered by Section 3. Corollary 4. cf.2). is easily seen to be f3pB < 1. jxJ v(dx) < oo.7 If {Xt} has stationary independent increments as in (4. where c(a) = ap + a2a2/2 + f 00 provided the Levy measure of the jump part {Mt} satisfies f". Now consider reflected versions of processes with stationary independent increments. VT + V for some r.4.10) . Furthermore. RANDOM WALKS IN DISCRETE OR CONTINUOUS TIME 37 almost solely concerned with the compound Poisson case and shall therefore not treat the intricacies of unbounded Levy measures in detail. O(u. oo]. (ex . V E [0. having Poisson arrivals with rate .1)v(dx) (4. and the reflected version is then defined by means of the abstract reflection operator as in (4. Proposition 4. [The condition for V < oo a. A different interpretation is as the workload or virtual waiting time process in an M/G/1 queue. T) = P(VT > u).
8 Assume that {Xt} has stationary independent increments and that {Lt} is a nonnegative multiplicative functional of the form Lt = g(t.10) is the LevyKhinchine representation of the c. Then the Markov process given by Theorem 2.Xt)I'Ftl = E [f(Xt+B . e. then the changed parameters in the representation (4.1 that E eaMt = exp fmoo In the general case . Q2 = v2. we use the characterization (4.Xt)L8 o 0tIFt] = E [f (Xt+s .Xo is necessarily infinitely divisible when {Xt} has stationary independent increments. we show in the compound Poisson case ( IlvIl < oo) in Proposition III. .Xt)I Ftl = Eof (X8)g(s. if Lt = e9(xt . X8) = Eof (X8)L8 = Eof (X8)• For the second. Eea(µt + QBt) = et{aµ +a2OZ/2}. v(dx) = e9xv (dx). This is of course no coincidence since the distribution of Xl .1.1)eexv(dx). Xt +B .1 ) v(dx) .g.xo)tk ( e).. By explicit calculation .f.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 . Proof For the first statement . Chung [86]).Xt)g(s. (4. Theorem 4 . In particular.6) are µ = µ + Oo2 . use the representation as limit of compound Poisson processes.g. u Note that (4.38 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Proof By standard formulas for the normal distribution.5) and get E [f(Xt+B .11) (eax . 5 has stationary independent increments as well. Xt Xo) with E2Lt = 1 for all x. 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) . let e" (a ) = Eoeaxl. t. of an infinitely divisible distribution (see.
(3B(dx).3B[B].St)g(Jt+s)I.3 and claim size distribution B # B. Example 4 . .3 and claim size distribution B. the structure of MAP's is completely understood when E is finite: 2and only there . (5. Ei. u 5 Markov additive processes A Markov additive processes.0. B have densities b. then the martingale {eex(t)tk(e)} is the continuous u time analogue of the Wald martingale (4.0 in the following. As for processes with stationary independent increments .. where . it is then easily seen that Lt = H dB(Ui) i:o. we write Pi. is defined as a bivariate Markov process {Xt} = {(Jt..3 =. U2. and let the Px refer to the claim surplus process of another compound Poisson risk process with Poisson rate. corresponding to p = 1.g. 0.10 Let Xt be the claim surplus process of a compound Poisson risk process with Poisson rate . one reason is that in parts of the applied probability literature.5.(3 = . are the arrival times and U1. B(dx) = B[9] B(dx). Recalling that U1. b = a = 0) the changed process is the claim surplus process of another compound Poisson risk process with Poisson rate . v(dx) _ .. a = 0.2. 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. . abbreviated as MAP in this section2.Ft] = Ejt..11 For an example of a likelihood ratio not covered by Theorem 4. <t whenever the RadonNikodym derivative dB/dB exists (e.8.l3 and claim u size distribution B. Thus (since µ = p = 1.1) For shorthand .4). the corresponding claim sizes . Ei instead of P2.o[f (S8)g(J8)]. Example 4 . MARKOV ADDITIVE PROCESSES 39 Remark 4. Then we can write v(dx) _ /3eOxB(dx) = / (dx). MAP stands for the Markovian arrival process discussed below.9 If X0 = 0. b with b(x) > 0 for all x such that b(x) > 0). dB/dB = b/b when B. St)} where {Jt} is a Markov process with state space E (say) and the increments of {St} are governed by {Jt} in the sense that E [f (St+8 ..
o(Ji = j.6) depending on i. v. 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. An alternative description is in terms of the transition matrix P = (piA.1 For a MAP in discrete time and with E finite. oo). If all Fij are concentrated on (0. As a generalization of the m.40 CHAPTER H.jEE (here pij = Pi(J1 = j)) and the probability measures Hij(dx)=P(Y1 EdxlJo=i. Y1 E dx) where Y„ = S„ . this means that the MAP can be simulated by first simulating the Markov chain {J„} and next the Y1. a MAP is the same as a semiMarkov or Markov renewal process. a jump of {Jt} from i to j # i has probability qij of giving rise to a jump of {St} at the same time.f. As an example..i. the converse requires a proof. In addition. vi(dx) in (4.... with the Y„ being interpreted as interarrival times. (That a process with this description is a MAP is obvious. t+s) where Jt . {St} evolves like a process with stationary independent increments and the parameters pi.it = A.g. In continuous time (assuming Dpaths). the distribution of which has some distribution Bij. SOME GENERAL TOOLS AND RESULTS In discrete time.[a) = (Ei[easl. a MAP is specified by the measurevalued matrix (kernel) F(dx) whose ijth element is the defective probability distribution Fij(dx) = Pi.9 EE = (iii&ij[a])i j EE . Proposition 5. consider the matrix Ft [a] with ijth element least Ei . {Jt} is specified by its intensity matrix A = (Aij)i. Y2.Sr_1. let {Jt} be standard Brownian motion on the line. by generating Yn according to Hij when J„_1 = i.) If E is infinite a MAP may be much more complicated.jEE• On an interval [t. . Jn = j. Fn[a] = F[a]n where P[a] = P ..J1=j)= Fij (dx) Pij In simulation language. which we omit and refer to Neveu [272] or cinlar [87]. 1 J1 ='^])iJEE = (Fij[a])i .
In matrix formulation . we infer that in the discrete time case the . Proof Let {Stt) } be a process with stationary independent increments and pa rameters pi . up to o( h) terms. MARKOV ADDITIVE PROCESSES Proof Conditioning upon (Jn. Sn ) yields Ei[easn+ '.1)) . 00 r(i) (a) = api + a2ot /2 + f (e° . aSt h = (1 + Ajjh) Ei [east . kEE Jn = k]Ek[e"Y" . pi. Jn+1 = A] = 41 Ei[ e 5„. qij. Then. this means that F't+h [a] = Ft[a] II+h(rc(i)(a)) +hA+h(Aijgij(Bij[a]1)) I. 013 . Bij (i.(')(a)) diag + ().qkj + k?^j qkj Bkj [a] } = Ei [east. J1 = A which in matrix formulation is the same as Fn+1 [a] = Fn[a]F[a]. Jt = j] is given by etK[a].2 Let E be finite and consider a continuous time Markov additive process with parameters A. a= . u Proposition 5.5. vi(dx). j E E) and So = 0. vi(dx) (i E E).4c). where K[a] = A+ (r. By PerronFrobenius theory (see A. Jt = k] { xk kEE j la] . \ diag Ft[a] = Ft[a]K.1 )v(dx). Jt = k] { 1 . u In the following. which in conjunction with Fo[a] = I implies Ft[a] = etK[a) according to the standard solution formula for systems of linear differential equations. Jt = j] Ejesh'^ + E Ak j hEi [ease .1) } (recall that qjj = 0). assume that the Markov chain/process {Jt} is ergodic. Then the matrix Pt[a] with ijth element Ei [east.ijgij(Bij[a] . Jt = j] (1 + htc (j) (a)) j + Ak j qk j (Bk +h E Ei [east .
(5. Since v("). its derivatives are 'asymptotic cumulants'.t. The corresponding left and right eigenvectors v("). .tK(a)h(a) J jj it L o is a martingale.f. u Let k(a) denote the derivative of h() w. We also get an analogue of the Wald martingale for random walks: Proposition 5. and appropriate generalizations of the Wald martingale (and the associated change of measure) can be defined in terms of . Corollary 5.4. Furthermore. cf. Jt = j] .e=e°tk.4 Eie"sth(a) = h=a)et?("). Yrh(a ) = 1.7. Proof For the first assertion. Eie"sth^a) = e'Pt[a]h( a) = e. just note that [a]h(a) = eietK (a)h(a) = etK(a)h(a). Proof By PerronFrobenius theory (see A. Jeast.3 Ei [east. Corollary 5. In particular.5 EiSt = tK'(0) + ki . Proposition 5. The function ic(a) plays in many respects the same role as the cumulant g. h(") are only given up to a constants. we are free to impose two normalizations. cf.etx It then follows that E feast+^(t+v)K(a)h(a) I ^tl l .r.c(a) (and h(")).2) where 7r = v(°) is the stationary distribution. of a random walk. and we shall take V(a)h(a) = 1. a.h(a)vva)etw(a).42 CHAPTER II. and write k = k(°). Then h(°) = e. h(") may be chosen with strictly positive components. Corollary 5.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).4c).Eikjt = ttc'(0) + ki . as will be seen from the following results. 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.
. ] = t2tc (0)2 + 2tK'(0)vk + ttc"(0) + O(1) . More precisely. u The argument is slightly heuristic (e. t im v^"St = '(0) Proof The first assertion is immediate by dividing by tin Corollary 5.4. Squaring in Corollary 5. Since it is easily seen by an asymptotic independence argument that E„ [Stkjt] u = trc'(0) E„kjt + O(1). MARKOV ADDITIVE PROCESSES Proof By differentiation in Proposition 5. the distribution of Jo). 8 Also for E being infinite (possibly uncountable ). In the same way.St]2 = t2/c'(0 ) 2 + 2ttc (0)vk . one obtains a generalization of Wald's identity EST = Er • ES. Remark 5 . For the second .Eikjr .2ttc (0)Evkjt + 0(i).+ k. t a oo.3) Let a = 0 and recall that h(°) = e so that 0=°) = h(o) = 1. E=ST = tc'(0)E7.4) .a) + ttc (a)2hia ) Multiplying by v=. (5.3) to get Ej [St a " st h i(a ) + 2Ste"st k(a) + e"st k^a) J etI(a) (kia )' + ttc (a)ki") + t {ic"(a)h. we differentiate (5. [E.6 For any stopping time T with finite mean. summing and letting a = 0 yields E„ [St + 2Stkj.5 yields + W (a)k.. tam E tSt a (0). there is typically a function h = h(") on E and a ^c(a) such that Ey a"st t" (") * h(x). . (5. subtraction yields Vary St = tic"(0) + O(1).7 No matter the initial distribution v of Jo.5. for a random walk: Corollary 5.") }) . the existence of exponential moments is assumed ) but can be made rigorous by passing to characteristic functions. Corollary 5.g. Ee"st typically grows asymptotically exponential with a rate ic(a) independent of the initial condition (i.e.5. 43 Ei [Steast h(a) + east k^a)1 = et"(a) (kia) + tic (a)hia)) .
5) is a martingale can be expressed via the infinitesimal generator G of {Xt } = { (Jt. see.6. 1) (i. in particular that f is bounded. G is defined as Gf (x) = lim Exf (Xt) . First.5) is a martingale . and the family {f LEE given by Theorem 2.6. Usually. where {Jt} is deterministic period motion on E = [0. xEE . this is.s. 0 Proposition 5. however.(9) {Lt}t>o = . V. 0) = h(i )( 1 + ttc(a)). From (5.. For t small . St)} be a MAP and let 0 be such that h(Jt) OStt. An example beyond the finite case occurs for periodic risk processes in VI.3b and Remark VI. In view of this discussion . h(Jo) Lo is a Px martingale for each x E E.6) We shall not exploit this approach systematically. u forsEE). this leads to h(i) + tcha( i.5. Jt = (s+t) mod 1 P8a.9 The condition that (5.. inconvenient due to the unboundedness of ea8 so we shall not aim for complete rigour but interpret C in a broader sense.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. s) = ea8h(i). 0) = n(a) h(i).44 CHAPTER IL SOME GENERAL TOOLS AND RESULTS for all x E E. Given a function h on E. We then want to determine h and x(a) such that Ejeasth (Jt) = etK(a)h(i).5 defines a new MAP. gha(i. we take the martingale property as our basic condition below (though this is automatic in the finite case). some extra conditions are imposed.f (x) tyo t provided the limit exists. let ha(i. St) } as follows. Then {Lt } is a multiplicative functional.10 Let {(Jt.for the present purposes. (5.e. however. Remark 5.4 that { h(Jt) easttK(a) L o (5.
1 + q(b . qij = r. .10 is given by P = eK(e) Oh e) F[e]Oh('). (5. Bi.(0)j.12 The expression for A means h(e) Aij = hie) Aij [1 + gij(Bij[0] i 0 j.Qi < oo and Bi a probability measure.7) In particular. In the infinite case . Here Oh(e) is the diagonal matrix with the h=e) on the diagonal. That the rows sum to 1 follows from Ae = Oh(e) K[O]h(B) .7(dx) Bij [0] Bij(dx) in the continuous time case . then also vi (dx) is compound Poisson with e Ox ^i = /3iBi[0].tc(0)e = 0 .. 0 < qij < 1 and Bij [0] > 0. In particular.11 below in the finite case.St)sl(e) h(Jt) 45 The proof that we have a MAP is contained in the proof of Theorem 5.ic(0)e = ic(0)Oh e) h(e) .1) . Bi [0] Remark 5. Bi(dx) = Bi(dx). MARKOV ADDITIVE PROCESSES Proof That { Lt} is a multiplicative functional follows from L8 ogt = h(Jt+s) es(St+ .1) eft ea' f ij (dx) = Hij (dx) Hij [0] . Then the MAP in Proposition 5..c(0)e = tc(0)e . if vi(dx) is compound Poisson. u Theorem 5. one can directly verify that (5. We omit the details. and by A = Oh(°) K [0]Oh(e ) vi(dx) = e"xvi (dx). Ai = µi + 0Q. ^i = of qij Bij [0] 1 + qij ( Bij [0] . this gives a direct verification that A is an intensity matrix: the offdiagonal elements are nonnegative because Aij > 0.11 Consider the irreducible case with E finite.1) holds for the P. vi (dx) = f3 Bi(dx) with . That 0 < qij < 1 follows from the inequality qb <1. 0<q<1. 0<b<oo. in the discrete time case.5.
Jl = j] :(Yi E dx. (dx) of a process with stationary independent increments follows from Theorem 4.11. Ji = j) h(e) eeyK(B)p h(8) h(e) eexK ( h=e) e)Fi. this implies k[a] = A 1 ) (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' . a = 0 in (5. v= . .8).tc (') (0) corresponds to the stated parameters µ...8) yields et'[a] = Ohie )et (K[a +e]K(e)I)Oh(°) By a general formula (A. this means that Ft[a] = etw ( e)Ohc) Ft[a + 9]oh (e) (5. Here the stated formula for P follows immediately by letting t = 1. This shows that F. it follows that indeed the normalizing constant is H1 [0].e) Consider first the discrete time case ..tc(0)I.t. v. Yi E dx. F:j with a density proportional to eei . Letting a = 0 yields the stated expression for A. Further Fib (dx ) = P=(YI E dx. SOME GENERAL TOOLS AND RESULTS Proof of Theorem 5. . Now we can write K[a] =A+A ) ( K[a + 0] .K [O])Oh(e) (0) l + ( A + (tc(') (a + 0) .tc(') (8)/ d)ag h 7 Aiiii (Bii[a + 0] .8) h(.tc(0)' )Ah() = Oh(o) K[a + 0]Oh() .Bay [0]) That k(') (a + 0) . since Hij. Hence the same is true for H=j and H. is absolutely continuous w. are probability measures .r. Jl = j) = Ei[Lt. Similarly. in continuous time ( 5. Jt = j] = hie) .13) for matrixexponentials . In matrix notation . Jt = A.46 CHAPTER II. (dx).8. H1.
3 for an infinite E are given by Ney & Nummelin [266]. an extensive bibliography on aspects of the theory can be found in Asmussen [16]. is slightly less general than the present setting. Much of the pioneering was done in the sixties in papers like Keilson & Wishart [224]. [226] and Miller [260].1) = Aij4ij(Bij[a] ..)Ajjgij(Bij[a+0] . which. has no mass on (oo. [262] in discrete time.Bij[0]) = hjel)ijgijBij[0](Bij[a] . [261]. i. however.6. < x) = 11 (S. Notes and references The earliest paper on treatment of MAP's in the present spirit we know of is Nagaev [265]. Write r+ = T(0) and define the associated ladder height ST+ and ladder height distribution by G+(x) = 11 (S. 6 The ladder height distribution We consider the claim surplus process {St } of a general risk process and the time 7. 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). Though the literature on MAP's is extensive. 0].e.7). hardly a single comprehensive treatment. 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. Note that G+ is concentrated on (0. however.1). see also Fuh & Lai [149] and Moustakides [264]. Conditions for analogues of Corollary 5.6.(u) = inf {t > 0 : St > u} to ruin in the particular case u = 0 . there is. and is typically defective. the literature on the continuous time case tends more to deal with special cases. .. h. [225]. THE LADDER HEIGHT DISTRIBUTION 47 Finally note that by (5.+ < x. 7+ < oo). oo).
1. Also. On Fig. Recall that B(x) = 1 . has no mass on ( 0.2) . see Fig. In any case. R+ is concentrated on (oo. and the maximum M is the total height of the ladder. For the proof of Theorem 6. oo). a fact which turns out to be extremely useful.1. at present we concentrate on the first ladder height.48 CHAPTER K.d. the second ladder point is ST+(2) where r+(2) is the time of the next relative maximum after r+(1) = r+.e. To illustrate the ideas.2.1 The term ladder height is motivated from the shape of the process {Mt} of relative maxima. where basically only stationarity is assumed. The main result of this section is Theorem 6. the ladder heights are i. 0 f T+ (6. o 00 (6.1) The interpretation of R+(A ) is as the expected time {St} spends in the set A before T+. we shall first consider the compound Poisson model in the notation of Example 1. G+ is given by the defective density g + (x) =..i. 6. Thus. = ST+(1) Figure 6.B(x) denotes the tail of B. which gives an explicit expression for G+ in a very general setting.00 ). 1 For the compound Poisson model with p = 01LB < 1. there are only finitely many). the sum of all the ladder steps (if rl > 0. The first ladder step is precisely ST+. g(y)R+(dy) = E f g(St)dt.T+ > t)dt = E f 0T+I(St E A) dt. by approximation with step functions . 6. 0]. In other cases like the Markovian environment model.e. Theorem 6 . In simple cases like the compound Poisson model. SOME GENERAL TOOLS AND RESULTS M ST+(2) Sr. i. the second ladder height (step) is ST+(2) . they have a semiMarkov structure (but in complete generality. the dependence structure seems too complicated to be useful)..ST+(1) and so on.1. it follows that for g > 0 measurable.5 below. define the prer+occupation measure R+ by R+(A) = E f o "o I(St E A.(3B(x ) = pbo(x) on (0. i. Here bo(x) _ B(x)/µB.
0 < t < T.2(b): r+ < t Thus. {St }o<t<T is constructed from {St}o<t<T by timereversion and hence. 49 Proof Let T be fixed and define St = ST .2(a): T+ > t Figure 6.ST<St.St<0. 0]. 6.T+>T) = P(STEA.6.O<t<T) = P(STEA. THE LADDER HEIGHT DISTRIBUTION Lemma 6 .0<t<T) = F(ST E A. 0 < t < T) P(STEA.2.ST<St. see Fig. has the same distribution as {St}o<t<T.O<t<T). P(STEA. . St S* t a Figure 6.ST_t. ST < ST_t. That is. since the distribution of the Poisson process is invariant under time reversion.2 R+ is the restriction of Lebesgue measure to (00.
G+(A) = Q f 0 B(A . The probability of this given { Su}u<t is B(A . oo). U + St_ E A.50 CHAPTER II. we get G+ (A) = f 00 /3 dt E[B(A .St _).3 G+ is the restriction of /3R+*B to (0. and since the jump rate is /3. oo).2) in the last). for A C (0. cf. T+ > t] 0 _ /3 f E[B( A ..t dT.St). and (6. Figure 6. SOME GENERAL TOOLS AND RESULTS Integrating w. this is just the Lebesgue measure of A. 6.St _)I(r+ > t).y) (here we used the fact that the probability of a jump at u t is zero in the second step. Fig.. it follows that R+ (A) is the expected time when ST is in A and at a minimum at the same time .T+ > t] dt 0 T+ _ /3E f g( St) dt = 0 f g(y) R+(dy) 0 00 where g(y) = B(A .r. That is.3 where the bold lines correspond to minimal values.y)R+(dy) 00 Proof A jump of {St} at time t and of size U contributes to the event IS. . E A} precisely when r+ > t. s. But since St 4 oo a.3 Lemma 6 .
i. The points in the plane (marked by x on Fig. assuming basically stationarity in time and space.4) are (ak. 4 (the points in the plane are (ak .. we define the arrival rate as E# { k : ak E [0 .M o 08 shifted by s is defined the obvious way. i.T+ < oo). .6. this is equivalent to the risk process {St*} being stationary in the sense of (6. The traditional representation of the input sequence {(TT. Fig. Lemma 6.:T1 +•••+Tk <t}. h]} /h (by stationarity. We call M * stationary if M* o B8 has the same distribution as M* for all s > 0. . we consider the claim surplus process {St }t>o of a risk reserve process in a very general setup.3 yields g+ (x) = .s.. THE LADDER HEIGHT DISTRIBUTION 51 Proof of Theorem 6.1 With r+(y) = I(y < 0) the density of R+.z)B(dz) _ f I(x < z)B(dz) _ f (x). as a point process on [0.Q f r+(x .) where ak = Ti + • • • + Tk .1. The sample path structure is assumed to be as for the compound Poisson case: {St*} is generated from interclaim times Tk and claim sizes Uk according to premium 1 per unit time.. The first ladder epoch r is defined as inf It > 0 : St > 0} and the corresponding ladder height distribution is * G+ (A) = P(S** E A) = P(ST+ E A. 6 . In the stationary case. Nt St =>Uk k=1 t where Nt = max{k = O. obviously. 0 Generalizing the setup. Uk) for those k for which ak . oo). 6 .. {St+8 ..* ) and the second the mark (the claim size Uk ). oo) x (0. cf. this does not depend on h).s > 0).e. 2. the first component representing time (the arrival time o. U k)} k=1 a is as a marked point process M *.4). Uk) (k = 1. The marked point process .e.S8 )t> o = {St }t>o for all s > 0.
. letting h J. V(M* o eak ). and let T = T2 denote the first proper interarrival time . We represent M by the sequence (Tk.g. o. h] Eco(M*) = 1 E f co(M o Bt)dt..5) represents the conditional distribution of M* given vi = 0.2.QiBi(dx).5) does not depend on h. of (6.. Note also that (again by stationarity) the Palm distribution also represents the conditional distribution of M* o Ot given an arrival at time t. Sigman [348] for these and further aspects of Palm theory. where T is the first arrival time > 0 of M and h > 0 an arbitrary constant (in the literature. Oh becomes the approximate probability F(ri < h) of an arrival in [0. h. Example 6 . SOME GENERAL TOOLS AND RESULTS M* U. The two fundamental formulas connecting M* and M are Eco(M) = aE E. This more or less gives a proof that indeed (6. where TI = 0.e. i.4 Given a stationary marked point process M*. vi(dx) = . Section 5) which has pure jump structure corresponding to pi = a = 0. See. = 0 . i 1 U2 Us 1_ 0 or Q2 $ U3 *1 L 0 7 X I 11 1 Figure 6. Uk) k=1. h] and the sum approximately ^o(M*)I(ul < h). As above . most often one takes h = 1). e. 0. .s. Assume {Jt} irreducible so that a stationary distribution 7r = (1i)iGE exists. k: vk E [0.4 Consider a finite Markov additive process (cf.52 CHAPTER II. the r.. we define its Palm version M as a marked point process having the conditional distribution of M* given an arrival at time 0 .
OF(x). Jo) w. 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 . After that.e. It follows that we can describe the Palm version M as follows . we note: Corollary 6. First choose (Jo_. the ruin probability .oo a. .p. Note in particular that the Palm distribution of the mark size (i.O for i # j. we get a marked point process generated by Poisson arrivals at rate /3i and mark distribution Bi when Jt = i. qij when {Jt} jumps from i to j and have mark distribution Bij.O for i = j and iriAijgij/. Then the ladder height distribution G+ is given by the (defective) density g+(x) = . Before giving the proof. j) and let the initial mark Ul have distribution Bi when i = j and Bij otherwise.5.s. This follows by noting that iP*(0) = IIG+JI = J0 "o g+(x)dx = .O fo "o F(x)dx = . and by some additional arrivals which occur w.p. 5 Consider a general stationary claim surplus process {St }t>o. let U0 be a r.6 Under the assumptions of Theorem 6. THE LADDER HEIGHT DISTRIBUTION 53 Interpreting jump times as arrival times and jump sizes as marks. aij for (i.*(0) with initial reserve u = 0 is p = /3EU0. an arrival for M* occurs before time t + dt w. the distribution of Ul) is the mixture B = E aii Bi + aij Bij J = j#i !i J. A stationary marked point process M* is obtained by assigning Jo distribution Tr. Assume that St * .6.6iBi + Aijgij Bij j#i iEE iEE 0 Theorem 6 . Jt = j is iri(3i /.. let the arrivals and their marks be generated by {Jt} starting from Jo = j. having the Palm distribution of the claim size and F (x) = F(Uo < x) its distribution ./.p. and that p = 0EU0 < 1. v.OEU0. the probability aij of Jt .= i. If Jt_ = i.
Then clearly * G+ (A) = P(ST+ E A) = Consider a process { f p(t)f3dt. h. . oo) x (0 .Su_ <0.s. are point processes on (oo . 6. .(left limit) when 0 < it < t and is illustrated on Fig .o.Mt). in (oo...5)..0<u<t) = P(StEA. oo)). Proof of Theorem 6.l. We then represent M by the mark (claim size ) Uo of the arrival at time 0.e. The result is notable by giving an explicit expression for a ruin in great generality and by only depending on the parameters of the model through the arrival rate 0 and the average ( in the Palm sense) claim size EU0.5. the arrival times 0 < 0'1 < Q2 < . The last property is referred to as insensitivity in the applied probability literature.. that M* and M have doubly infinite time (i. A standard argument for stationary processes ([78] p. T+ = t given the event At that an arrival at t occurs .. 105) shows that one can assume w. It follows that for A C (0. and the kth preceding claim arrives at time t . SOME GENERAL TOOLS AND RESULTS V` (0) = E E Uk k: ak E [0..o<u<t where a claim arrives at time t and has size Uo.St<Su. (k = St}t>o 1. moves down linearly at a unit rate in between jumps and starts from S0 = U.0<u<tIAt) = P(St EA.5. CHAPTER H.$St_ u. oo) p(t) = P(St EA. Now conditionally upon At . { Su}0<u<t is distributed as a process {Su} . oo ) and the arrival times 0 > 0_1 > a_2 > .Su<0.54 By (6.1] here the r . in (0.o.0<u<tIAt) = P(St EA. 0).. The sample path relation between { Su } and { Su } amounts to S„ = St .g. 2. Let p(t) be the conditional probability that ST+ E A. which makes an upwards jump at time .Q_k and has size U_ k.). has a very simple interpretation as the average amount of claims received per unit time . the mark at time Qk is denoted by Uk.St*_ u. 0<u<t) = P(St EA..Su< 0.St <. .A.0<u<t) = P(St EA.
6. 2 therefore immediately shows that L(dy) is Lebesgue measure on (oo. 0 < u < t } is the event that { Su } has a relative minimum at t . Thus. 6.5. the left endpoint of the support is oo.. NIt)dt . G' (A) = 3 f P(St E A. and since by assumption St * oo a. Since So = U0.5 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. the support of L has right endpoint U0. and we let L(dy) be the random measure L(A) = fo°° I(St E A.s. THE LADDER HEIGHT DISTRIBUTION 55 { A Su}0<u<t U0 U0 \t tt u>0 N U_1 Figure 6. Uo]. Mt)dt = i3EL(A) o"o .5 where it = { St < Su. In Fig. 6. A sample path inspection just as in the proof of Lemma 6 . Fig. t a oo. cf. time instants corresponding to such minimal values have been marked with bold lines in the path of { St}.
.6 is Bjork & Grandell [67]. [147].2.1).56 CHAPTER II. [263] (a special case of the result appear in Proposition VI. A further relevant reference related to Corollary 6. SOME GENERAL TOOLS AND RESULTS = OE f 0 I(Uo>y)I (yEA)dy = Q f IP (Uo>y)dy A 0o a fA P(y) dy• 0 Notes and references Theorem 6.5 is due to Schmidt & coworkers [48].
and assume that • { Nt}t>o is a Poisson process with rate j3.6) and simulation methods ( Chapter X).e. U2. It is worth mentioning that much of the analysis of this chapter can be carried over in a straightforward way to more general Levy processes . A common view of the literature is to consider such processes as perturbed compound Poisson risk processes . 3).. Panjer's recursion ( Corollary XI. and independent of {Nt}.d. with common distribution B.4 below . . i=1 i=1 An important omission of the discussion in this chapter is the numerical evaluation of the ruin probability. Thus . i.. Some possibilities are numerical Laplace transform inversion via Corollary 3. • the premium rate is p = 1. say.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. • the claim sizes U1. being of the form Rt = Rt+Bt + Jt where {Rt } is a compound 57 . 4. are i. i. see Chapter IV. For finite horizon ruin probabilities . St = uRt = EUi t. {Rt} and the associated claims surplus process {St} are given by Nt Nt Rt = u+t EUi. exact matrixexponential solutions under the assumption that B is phasetype (see further VIII.
.Rt. THE COMPOUND POISSON MODEL Poisson risk process. For (c). We do not spell out in detail such generalizations.t = fltpB .1)..1) = t(p . (b) Var St = t. Dufresne & Gerber [126]. A more formal proof goes as follows: Nt r Nt ESt = E > U k . Furrer [150]. we get Ee8st = 00 e8t c` Ee8 (U1+. 1 Introduction For later reference. for (d) just note that the kth cumulant of St is tic(k) (0).1) . cumulants .u . 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). (c) Ee8St = et" (8) where c(s) = f3(B[s] . [324].t = E E [ U k k=1 k=1 Nt . and that B(k)[0] = Pak). See e. and Schlegel [316]. 0 . Proof It was noted in Chapter I that p . where K(k) (0) is the kth derivative of is at 0. and this immediately yields (a).s.)3t (fit' k t} = etk(8) exp {st '3t + B[s]f Finally. Schmidli [319].1 (a) ESt = t(13µ$ . of the claim surplus St .t = t(p .1). P = PAB = 1/(1 + rl) Proposition 1. e .g. we shall start by giving the basic formulas for moments. (d) The kth cumulant of St is tf3p(k) for k > 2. {Bt} a Brownian motion and {Rt} a pure jump process.g.t = E[Ntµs] .f. m. say stable Levy motion.'s etc.+Uk)P(Nt = k) k=O e8t k=O B[s]k . Write pB^) = EUn' YB = Pali = EU.6pBa).58 CHAPTER III.1 is the expected claim surplus per unit time.
We return to this approach in Chapter V.3) is proved similarly. The right hand inequality in (1. rather to view {St} directly as a random walk in continuous time. For example.1)th claim. u + v]. and there are at least two ways to exploit this: Recalling that ok is the time of the kth claim. St = oo.h < St < S(n+1)h + h. (c) If 77 > 0. where Tk is the time between the kth and the (k . (d) If 17 = 0. (b) If 77 < 0. 2.. The point of view in the present chapter is.Tk are i.1 is the same as if {St} was a random walk indexed by t = 0.. meaning that the increments are stationary and independent. v > 0. Proof We first note that for u. then Snh . we have Sok . lim supt. however. 1.S„ attains its minimal value when there are no arrivals in (u.d.. the Uk . we need the following lemma: Lemma 1. (a) No matter the value of 77. INTRODUCTION 59 The linear way the index t enters in the formulas in Proposition 1.3 If nh < t < (n + 1)h.Sok_l = Uk . we get a discrete time random walk imbedded in the claim surplus process {St}. Here is one immediate application: Proposition 1. obviously 0(u) = F(maxk Sok > u). Obviously.1. St = oo. In this way. which is often used in the literature for obtaining information about {St} and the ruin probabilities. cf.Tk.i.3EU01 = 1µs where rt is the safety loading.2 (DRIFT AND OSCILLATION) St/ta3'p1 ast >oo. S„+V > S„ .V. .. then lien inft. Sn+0 . For the proof. The connections to random walks are in fact fundamental. and the value is then precisely v. so that {Sok } is a random walk with mean EUET = EU. then St 00. II. then St> SnhV>Snhh.1 = . then St 4 co. In particular.4. Indeed. if t = nh + v with 0 < v < h.
.5 The limiting distribution of St .o.1) as t 4 oo is normal vtwith mean zero and variance )3µsz) Proof Since {St}t>o is a Levy process (a random walk in continuous time).2: Proposition 1. Corollary 1. If rl > 0. and (b).1. Considering the next downcrossing (which occurs w. THE COMPOUND POISSON MODEL Proof of Proposition 1. where the size of the portfolio at time t is M(t). 2h. h A similar argument for lim sup proves (a). 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. p. and < 1 for all u when 77 > 0.2. hence by induction i. This contradicts u St400.1. and hence it folz lows from standard central limit theory and the expression Var(St) = tf3pB (Proposition 1. Snh/n a4' ESh = h(p . {Snh}n=o.60 CHAPTER III. 0 Snh = 00. 1 since St 4 oo) and repeating the argument. it suffices to prove 4'(0) = F(M > 0) < 1.1. we get lim inf St t>oo t nroo nh<t<(n+1)h t = lim inf inf St h l++m of Sn 7t h = ESh = p . However._.2. Snh u = 00 (the lemma is not needed for (d)). is a discretetime random walk for any h > 0.. Proof The case of 17 < 0 is immediate since then M = oo by Proposition 1. The general case now follows either by another easy application of Lemma 1. 169) stating that lim infra.4 The ruin probability 0(u) is 1 for all u when 77 < 0. For any fixed h. at least once. Part (d) follows by a (slightly more intricate) general random walk result ([APQ].t . h.s. ... Remark 1 .1(b)) that the assertion holds as t 4 oo through values of the form t = 0. or by a general result on discrete skeletons ([APQ] p.. There is also a central limit version of Proposition 1.3. Assuming that each risk generates claims at Poisson intensity /3 and pays premium 1 per unit time... is a discrete time random walk.3. it is seen that upcrossing occurs at least twice. {Snh}n=o.. Thus using Lemma 1. u 307).1).. (c) are immediate consequences of (a).p. lim supn_. if P(M > 0) = 1.6 Often it is of interest to consider size fluctuations. and hence by the strong law of large numbers.. then {St} upcrosses level 0 a. Notes and references All material of the present section is standard.
we may view the ladder heights as a terminating renewal process and M becomes then the lifetime. cf.6. and we further get information about the joint conditional distribution of the surplus and the deficit. The decomposition of M as a sum of ladder heights now yields: 00 Theorem 2 . p < 1. the formula for the distribution of M follows . IV. 0 Alternatively.just before ruin is again B0. Theorem 2.1. i. nevertheless. We assume throughout rl > 0 or. and we shall here exploit the decomposition of the maximum M as sum of ladder heights. d. Fig. 1e.1) is not entirely satisfying because of the infinite sum of convolution powers. B(x)/aB.34 or A. 11. the ladder heights are i. This follows simply by noting that the process repeats itself after reaching a relative maximum. that r(0) < oo) is Bo: taking y = 0 shows that the conditional distribution of (minus) the surplus ST(o).1) representing the distribution of M as a geometric compound. which we henceforth refer to as the PollaczeckKhinchine formula. Thus . Here bo(x) _ Proof The probability that M is attained in precisely n ladder steps and does not exceed x is G+ (x)(1 . [APQ] Ch. The expression for g+ was proved in Theorem 11. oo ).e.1 provides a representation formula for 0(u). n=0 (2.2.IIG +II)EG+ . Note that the distribution B0 with density bo is familiar from renewal theory as the limiting stationary distribution of the overshoot (forwards recurrence time )..IIG+II) (the parenthesis gives the probability that there are no further ladder steps after the nth ). Summing over n. 1 The distribution of M is (1..1.P) E PnBon(u) . (2. Combined with i/i(u) = P ( M > u). Note that this . but we shall be able to extract substantial information from the formula. we can rewrite the PollaczeckKhinchine formula as 00 (u) = P (M > u) = (1 . where G+ is given n=0 by the defective density g+ (x) = 3B (x) = pbo(x) on (0. As a vehicle for computing tIi(u).6. The following results generalizes the fact that the conditional distribution of the deficit ST(o) just after ruin given that ruin occurs (i. 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}. It is crucial to note that for the compound Poisson model. equivalently.
62 CHAPTER III. 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). [62]. and the conditional distribution of ST(o)_ given ST(o)_ = z is Bo z) The proof is given in IV. 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.2(a) is from Dufresne & Gerber [125]. For the study of the joint distribution of the surplus ST(u)_ just before ruin and the deficit ST(„). the form of G+ is surprisingly insensitive to the form of {St} and holds in a certain general marked point process setup. the PollaczeckKhinchine formula is often referred to as Beekman 's convolution formula. Theorem 2 .i. see Schmidli [323] and references there. Beekman [61].d. 2 The joint distribution of (ST(o )_. (d) the marginal distribution of ST(o)_ is B0. . f +b (b) the joint distribution of (ST( o). Again. see for example [APQ]. ST(o) > y. in this setting there is no decomposition of M as a sum of i. Theorem 2. there is a general marked point process version. 7r(0 ) < oo) = Q 3 Special cases of the PollaczeckKhinchine formula The model and notation is the same as in the preceding sections.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. Asmussen & Schmidt [49]. ST(o)) is given by the following four equivalent statements: B(z) dz. ladder heights so that the results do not appear not too useful for estimating 0(u) for u>0. ST(o )) given r (0) < oo is the same as the distribution of (VW.1 is traditionally carried out for the imbedded discrete time random walk. Feller [143] or Wolff [384]. In the risk theory literature.just after ruin. (1 . The proof of Theorem 11. We assume rt > 0 throughout. 1) and W has distribution Fw given by dFyy/ dB(x) = x/µB.5. cf. cf. where it requires slightly more calculation.6. W are independent.5. As shown in Theorem 11. (c) the marginal distribution of ST(o)_ is Bo . cf.6.V)W) where V. Theorem A1. V is uniform on (0. However. (a) 11 (ST(o)_ > x.
p) = S . use Laplace transforms. 0 .1)1 00 ( 1 . B0 is exponential with rate S and the result can now be proved from the Pollaczeck Khinchine formula by elementary calculations . Let r ( x) be the failure rate of M at x > 0. Thus .p.p.p) E pn S n x n.p)pSe a ( l v)x = p( S . the current ladder step must terminate which occurs at rate S and there must be no further ones which occurs w. the formula for P(O) holds in a more general setting. and hence this overshoot has the same distribution as the claims themselves . The result can.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.O)e(b0)x. As shown in 11.2 If B is exponential with rate S. 1 . then.6. a further relevant reference is Bjork & Grandell [67]. Alternatively.1 e ax = n1 (n .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.. however .3. also be seen probabilistically without summing infinite series . hence without memory. 3b Exponential claims Corollary 3. the result follows . For a failure at x. Thus r(x) = S(1 . But claims are exponential .e. Bon is the Erlang distribution with n phases and thus the density of M at x > 0 is (1 . I. 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. Integrating from u to oo.
then 24 1 V.3).g.S.y)G+(dy) For the last identity in (3.1) For a heavytailed B. T+ <00) (3. The case of (3. (3.3) Equivalently. E. 2 is one of the main classical early results in the area.2). 0 Proof Write o (u) as P(M>u) = P(S.4) can be derived by elementary algebra from (3.3.y)f3 (y) dy. THE COMPOUND POISSON MODEL In VIII. we use the PollaczeckKhinchine formula in Chapter IX to show that b(u) .+ <U.y)G+(dy ) = f U V(u .1 p pBo(u). 3c Some classical analytical results Recall the notation G+(u) = f^°° G+(dx). and weights 1/2 for each.p + G+ * Z(u) = 1 .s. (3. A variety of proofs are available . cf. II.2) Notes and references Corollary 3. (u) 35eu + 35e6u.p + f u Z(u .3.S.+ >u.T+ <oo). and conditioning upon S. if 3 = 3 and B is a mixture of two exponential distributions with rates 3 and 7.4) zu P(M > u .+ = y yields P(M>u.y)/3B (y) dy.3) below. (3. u .T+ <oo)+P(M> u. u + oo. Corollary 3. (Example VIII. (b) use stopped martingales .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 . just insert the explicit form of G+.+ <u.64 CHAPTER III. (3. the survival probability Z(u) = 1 .1.i(u) satisfies the defective renewal equation Z(u) = 1 .h.3)).4) is similar (equivalently. we show that expression for /'(u) which are explicit (up to matrix exponentials) come out in a similar way also when B is phasetype. Then the first term on the r. is ?7+ ( u). We mention in particular the following: (a) check that ip (u) = pe (60)u is solution of the renewal equation (3.
206207)../3B[s] which is the same as (3. it is not surprising that such arguments are more cumbersome since the ladder height representation is not used./3B[s] . g. [APQ] pp.s . SPECIAL CASES OF POLLACZECKKHINCHINE Corollary 3. 191).Ps s(. Corollary 3. eau B(u) du = f PB 3PB SPB 0 o (3.)3B[s]) (3.7) s +.p = (1 . Bo of B0 as m e8u B(du) = B[s] .f. for example. [APQ] pp. by analytical manipulations (L'Hospital's rule) from (3. Griibel [179] and Thorin & Wikstad [370] (see also the Bibliographical Notes in [307] p.(3B[s] 1 .g.P)PB 2(1 . which yields the survival probability as 00 f u }t Z(u) = f f3eRtdt 0 from which (3.7) and Corollary 3. In view of (3.g.PPB2) EM2 = PPB) + QZPBl 2(1 .3 .3 .5).(3 . Embrechts.p) E p"Bo[s]" = 1 .8) Proof This can be shown.p)2 3(1 .Ee8M) f ao e8' ( u)du = a8uP (M > u)du = 0 o 1 ( 1+ (1 .3 is standard . numerical inversion of the Laplace transform is one of the classical approaches for computing ruin probabilities. 0 Notes and references Corollary 3.4) can be derived by elementary but tedious manipulations.3.5 can be found in virtually any queueing book. either of these sets of formulas are what many authors call the PollaczeckKhinchine formula.p)s s /3 . In fact. see e. Of course. e. Some relevant references are Abate & Whitt [2].5 The first two moments of M are 2 EM .P)pB' (3.s . The approach there is to condition upon the first claim occuring at time t and having size x .pBo[s] no (1 .1 Bo[s] = f oc. .4 The Laplace transform of the ruin probability is 65 fo Hence Ee8M 00 e8uiP(u)du . 111112 or Feller [143]. Also (3. (3.5) Proof We first find the m.p)s .7). Griibel & Pitts [132].6) 00 = (I . We omit the details (see..5).
of (3.1). THE COMPOUND POISSON MODEL 3d Deterministic claims Corollary 3.u)]k d 1 u) _ a) n ( du ( k! (1  .u)]k k! (1 L3) 1: e_O(ku) NIN (k (k . differentiation yields Z'(u) _ /3Z(u) which together with the boundary condition Z(0) = 1/3 yields Z(u) _ (1/3) eAu so that (3.4) for Z( u) means f lhu Z(u) = 1. we may assume p = 1 so that the stated formula in terms of the survival probability Z(u) = 1 .u/p)]k ko k! Proof By replacing {St} by {Stu/p} if necessary.u + 1 )]k = QZ(u) .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.u) [N(k . differentiation yields Z(u) _ /3Z(u) .3I( 0<y<1)dy Z(y)/3I(0<u.u) a)Qea" + (1 .6 If B is degenerate at p. .Q (k 1 k= n  [O(k .)3(1 .h.9) follows for 0 < u < 1.Q) 3e.u) [p(k .3+ 18+ J0 Z(uy).1)! k=1 u1 . then p) 1: ep(k u/.66 CHAPTER III./3Z(u .1 < u < n and let Z(u ) denote the r./32(u . Z^ =eR(k.s. eO('u) [)3(k .z/'(u) takes the form Z(u) L^J L.Q) k=0 k! E e0( = /32(u) .9). Assume (3. For n < u < n + 1.9) shown for n .u)]k1 ku+1) [/3( k .1).u)]k k! k0 The renewal equation (3.
3B = . 00 the standard definition of the exponential family {F9} generated by F is FB(dx) = e°xK(e)F(dx).a. We could first tentatively consider the claim surplus X = St for a single t. (4. (4. co(a) = rc(a + 9) .Qe(Bo[a] .6 is identical to the formula for the M/D/1 waiting time distribution derived by Erlang [139].g.f.2) (Here 9 is any such number such that r. The question then naturally arises whether ie is the c.g.f. say t = 1: recall from Proposition 1. Formalizing this for the purpose of studying the whole process {St}.2) shows that the solution is Ox [O]0].4) . See also Iversen & Staalhagen [208] for a discussion of computational aspects and further references. B9(dx) = B[9] B(dx).3) by t.f.1) or equivalently.4) works as well.4. 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) . we just have to multiply (4.2). The answer is yes: inserting in (4.(9). CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 67 Since Z(n) = 2(n) by the induction hypothesis. 4 Change of measure via exponential families If X is a random variable with c.1) .rc(9) = .a. we set up . in terms of the c.f. K(a) = logEe'X = 109f 00 eaxF(dx) = logF[a]. F and c. 0 Notes and references Corollary 3.d. but will now be repeated for the sake of selfcontainedness.3B[9]. and thus (4. or equivalently BB[a] = B[^+ Repeating for t 54 1.) The adaptation of this construction to stochastic processes with stationary independent increment as {St} has been carried out in 11. (4. and define rce by (4.1 that c(a) = /3(B[a] .(9) is welldefined.4.g.r. it follows that Z(u) = 2(u) for n<u<n+1. of F9.1) .
r. v(Xi.5) for the density of n i.68 CHAPTER III. .r.. .t. Ti(a)/n.7) now follows by taking Z = eBST+TK(e)I(G) u Theorem 4 . t < T.Tic (0)} . in particular the expression (4.t. EeeBSt + tk(B) = 1.g. G]. n) for a given n.10) . (4.3 and claim size distribution B. Then FB denotes the probability measure governing the compound Poisson risk process with arrival intensity. Proposition 4.nr. with T taking the role of n) is the analogue of the expression exp{8(x1 + • • • + xn) . Xn). BB by (4.i. for G E FT.0e and claim size distribution Be.4). and PBT) the restriction of PB to FT. . The identity (4.6) F(G) = Po (G) = EB [exp {BST + Ttc(0)} . The following result (Proposition 4.7) Proof We must prove that if Z is FTmeasurable. Then the Xk are i.8) By standard measure theory. Z is measurable w..5) for the density. (4.3 Let T be any stopping time and let G E FT.i. and thus (4. .1. Then P(G) = Fo(G) = EB [exp {BST + TK(O)} . then EBZ = E [Ze9ST _T"(9)I .9) Proof We first note that for any fixed t. But let Xk = SkT/n .1 Let P be the probability measure on D[0. G].f.d.FTn) = Q(SkTIn : k = 0.1) and multiply from 1 to n). = exp {BST . with common c.(9)} (4.2.2 For any fixed T.S(k_1)Tln.d. THE COMPOUND POISSON MODEL Definition 4. oo) governing a given compound Poisson risk process with arrival intensity. and dP(T) dP^T) That is. it suffices to consider the case where Z is measurable w.FT. G C {T < oo}.. Let FT = o(St : t < T) denote the o•algebra spanned by the St. . .8) follows by discrete exponential family theory. (4. (4. the PBT) are mutually equivalent on. (4. replications from Fe (replace x by xi in (4. the corresponding expectation operator is E9. and define 09.
Ee [exp { BST +Trc(9)} I(G) FT)] = 1.7 1 Some discussion further supporting this statement is given in the next section. the typical shape of rc is as in Fig. LUNDBERG CONJUGATION 69 Now assume first that G C Jr < T} for some deterministic T. Thus.1(a).1 It is seen that typically) a ry > 0 satisfying 0 = r. Given FT.r is deterministic. The behaviour at zero is given by the first order Taylor expansion c(a) r.(Y) = 13(B['Y] . GT C_ Jr < T}. .9) holds with G replaced by GT.r)rc(9)}I . (0) + rc'(0)a = 0 + ES1 a = a (p . according to what has just been proved.. 5. Letting T t oo and using monotone convergence then shows u that (4.7) holds.f. so that PG = EeE0 [exp { 9ST+Trc(9)}I(G)I FT)] = Ee [exp { BST + rrrc(O)} I(G)EB [ exp {9 (ST . 5 Lundberg conjugation Being a c. Then G E FT. and hence (4.9) holds for G as well.1) _ 1 + a. 77 Thus. Thus by (4. (a) rc (a) (b) KL(a) 'Y 'Y Figure 5.5.FT]] = EB [exp { BST + Trc(9)} I(G)] .g. Now consider a general G. t = T . c(a) is a convex function of a.10). Then GT = G n Jr < T} satisfies GT E FT. (4. subject to the basic assumption ij > 0 of a positive safety loading.1) .ST) + (T .
5. 5. the claim surplus process has positive drift > 0. u It is a crucial fact that when governed by FL.3. An established terminology is to call y the adjustment coefficient but there are various alternatives around. Equation (5.1(b). Taking T = r(u). the Lundberg exponent. we write FL instead of F7. b[s] = 5/(b .2)) is 7 = 5/3.3) cf. Thus B[7] = 6/.2 s As support for memory.1(b).1 Consider the case of exponential claims. .1) is precisely what is needed for one of the terms in the exponent .QL instead of /37 and so on in the following .g.4) ELS1 = #L(0) cf.a = i(a + 7). we further note that ( 5. Example 5 .4) yields /3L = b and that BL is again exponential with rate bL =.2 is B(7) = 1 + ^.3. e. and (4. G = {T(u) < oo} in Theorem 4. 5.1) (or (5.70 CHAPTER III. (5. Lundberg conjugation corresponds to interchanging the rates of the interarrival times and the claim sizes. Thus.s). (5.3.1) . THE COMPOUND POISSON MODEL exists . It is then readily seen that the nonzero solution of (5. Fig. (5. an equivalent version illustrated in Fig. Fig.1) is known as the Lundberg equation and plays a fundamental role in risk theory .2) 7 Figure 5. Note that KL (a) = /L (BL [a] .
(5.4).5.7) 0 and all that is needed to check is that ( 5.t. we can rewrite this as 0(u) = e"ELe7^(u).5) Theorem 5 .3.3 takes a particular simple form. LUNDBERG CONJUGATION 71 to vanish so that Theorem 4.5). T = T+. 0 Theorem 5 . T(u) < oo] .1p . Then P(ST+ E A) = EL [exp { 7S?+} . e(u) has a limit i. G = {S.1.(u)} .7) is the same as (5. To this end.+ E A} in Theorem 4. (5.ascending ladder height distribution and µ+ its mean.1 (5. where C .G+ L) (x) G+L) (x) IL(+) µ+L) L) where G+L) is the FL.e7x)G+(dx). Since a7' is continuous and bounded.3 (THE CRAMERLUNDBERG APPROXIMATION) i'(u) .r.6 ). V)(u) < e7u.8) .P Y j o' xeryxOB (x) dx /3k [Y] .u be the overshoot and noting that PL(T(u) < oo) = 1 by (5.(oo) (in the sense of weak convergence w.G+L)(x)) dx ry^+L) J 00 f 0 (1 . take first 0 = ry. see A . Letting e(u) = ST(u) . PL ) with density 1 .6) Proof By renewal theory. we therefore have ELe7t(u) + C where C ELe7 (00) = µ+) f e7(1 . ST+ E A] .1e.Ce7u as u 4 oo. (5.2 (LUNDBERG'S INEQUALITY) For all u > 0. Proof Just note that e(u) > 0 in (5. which shows that G(L) (dx) = e7xG +(dx) = e7x /3 (x) dx. V) (u) = P(T(u) < oo) = EL [exp {ryS.
u . Noting that SIG(L)II = 1 because of (5.12) Example 5 .10) VW = JI c* e° (x) dx = a (B[a] .11) so that I 7B ['Y](B[7]1) BI [7]Q VP (7) 72 7 (using (5.72 CHAPTER III.8) yields +L) J0 xel'B ( x) dx (5. (5.1) (5.4 Consider first the exponential case b(x) = Seax. of course. that 7 = S . we get L where 00 (1 .4). THE COMPOUND POISSON MODEL In principle. but some tedious (though elementary) calculations remain to bring the expressions on a final form. Then 0(u) = pe(a_Q)u where p = /3/S. From this it follows. .1 .1 above) and that C = p.")G + (dx ) = 1  J0 00 3B(x) dx = 1p. 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. this solves the problem of evaluating (5.7).1 = ^7 The accuracy of Lundberg's inequality in the exponential case thus depends on how close p is to one.3 (this was found already in Example 5. or equivalently of how close the safety loading 77 is to zero.1)) and 7µ+L) = 'y/3 [7] 7 1/0 = /3B ['y] .e. Using (5.
5. LUNDBERG CONJUGATION Remark 5.5 Noting that PL  1 = ,3LIBL  1 = #ci (0 ) = k (ry) _ ,QB' ['Y]  1 ,
73
we can rewrite the CramerLundberg constant C in the nice symmetrical form G, _'(0)1  1  p K'(7) PL1
(5.13)
In Chapter IV, we shall need the following result which follows by a variant of the calculations in the proof of Theorem 5.3: 1  aB[ry  a]  1 Lemma 5 . 6 For a # ry, ELea^ (°°) = 7 aK'(7) 7  a Proof Replacing 7 by a in (5.7) and using ( 5.8), we obtain 1 (I 1  ^ e('ra) x,3 (x)dx) (L ) ELea^*) = a \\\ f
using integration by parts as in (3.6) in the last step . Inserting (5.12), the result follows. u
Notes and references The results of this section are classical, with Lundberg's inequality being given first in Lundberg [251] and the CramerLundberg approximation in Cramer [91]. Therefore, extensions and generalizations are main topics in the area of ruin probabilities, and in particular numerous such results can be found later in this book; in particular, see Sections IV.4, V.3, VI.3, VI.6.
The mathematical approach we have taken is less standard in risk theory (some of the classical ones can be found in the next subsection). The techniques are basically standard ones from sequential analysis, see for example Wald [376] and Siegmund [346].
5a Alternative proofs
For the sake of completeness, we shall here give some classical proofs, first one of Lundberg's inequality which is slightly longer but maybe also slightly more elementary:
74 CHAPTER III. THE COMPOUND POISSON MODEL
Alternative proof of Lundberg 's inequality Let X the value of {St} just after the first claim , F(x) = P(X < x). Then , since X is the independent difference U  T between an interarrival time T and a claim U, ,3+ry F'[7} = Ee7 ( UT) = Ee7U • Ee7T = B['Y] a = 1' where the last equality follows from c(ry) = 1. Let 0( n) (u) denote the probability of ruin after at most n claims. Conditioning upon the value x of X and considering the cases x > u and x < u separately yields
,0(n +1) (u) = F(u) +
Ju
0 (n) (u  x) F(dx).
We claim that this implies /,(n) (u) < e 7u, which completes the proof since Vi(u) = limniw 1/J(n) (u). Indeed , this is obvious for n = 0 since 00)(u) = 0. Assuming it proved for n, we get
„/, (n+1)(u) <
F(u) + e7u
00
Ju
e7(u=) F(dx)
00
<
f
e7x F(dx)
+ fu
e  7(u z) F(dx)
u
o0
= e 7uE[ 'Y] = e 7u.
Of further proofs of Lundberg's inequality, we mention in particular the martingale approach, see II.1. Next consider the CramerLundberg approximation. Here the most standard proof is via the renewal equation in Corollary 3.3 (however, as will be seen, the calculations needed to identify the constant C are precisely the same as above): Alternative proof of the CramerLundberg's approximation Recall from Corollary
3.3 that
(u) = )3
J OO B(x) dx + J U Vi(u  x)/3 (x) dx.
u 0
Multiplying by e7u and letting Z(u) = e7" O(u), we can rewrite this as
u Z(u) =
z(u) = e7u/
J
B(x)dx, F(dx) = e7x,QB(x)dx,
u
z(u)
f +
J
e7(ux ),Y' 1 • l•(u  x) • e7'/B(x) dx,
0
= z(u) +
J0 u Z(u  x)F(dx),
6. MORE ON THE ADJUSTMENT COEFFICIENT 75
i.e. Z = z+F*Z. Note that by (5.11) and the Lundberg equation, ry is precisely the correct exponent which will ensure that F is a proper distribution (IIFII = 1). It is then a matter of routine to verify the conditions of the key renewal theorem (Proposition A1.1) to conclude that Z (u) has the limit C = f z(x)dx/µF, so that it only remains to check that C reduces to the expression given above. However, µF is immediately seen to be the same as a+ calculated in (5.10), whereas
L
00
z(u) du =
f
J
/3e7udu "o
J "o B(x) dx = J "o B(x)dx J y,0eludu
u 0 0
B(x)^ (e7x  1) dx = ^' (B[7]  1)  As] [0 µs] = l y P^
using the Lundberg equation and the calculations in (5.11). Easy calculus now gives (5.6). u
6 Further topics related to the adjustment coefficient
6a On the existence of y
In order that the adjustment coefficient y exists, it is of course necessary that B is lighttailed in the sense of I.2a, i.e. that b[a] < oo for some a > 0. This excludes heavytailed distributions like the lognormal or Pareto, but may in many other cases not appear all that restrictive, and the following possibilities then occur: 1. B[a] < oo for all a < oo. 2. There exists a* < oo such that b[a] < oo for all a < a* and b[a] = 00 for all a > a*. 3. There exists a* < oo such that fl[a] < oo for all a < a* and b[a] = 00 for all a > a*. In particular , monotone convergence yields b[a] T oo as a T oo in case 1, and B[a] T oo as a f a* in case 2 (in exponential family theory , this is often referred to as the steep case). Thus the existence of y is automatic in cases 1 , 2; standard examples are distributions with finite support or tail satisfying B(x) = o(eax)
76 CHAPTER III. THE COMPOUND POISSON MODEL
for all a in case 1, and phasetype or Gamma distributions in case 2. Case 3 may be felt to be rather atypical, but some nonpathological examples exist, for example the inverse Gaussian distribution (see Example 9.7 below for details). In case 3, y exists provided B[a*] > 1+a*/,3 and not otherwise, that is, dependent on whether 0 is larger or smaller than the threshold value a*/(B[a*]  1). Notes and references Ruin probabilities in case 3 with y nonexistent are studied, e.g., by Borovkov [73] p. 132 and Embrechts & Veraverbeeke [136]. To the present authors mind, this is a somewhat special situation and therefore not treated in this book.
6b Bounds and approximations for 'y
Proposition 6.1 ry <
2(1  aps) 2µs
OMB PB)
Proof From U > 0 it follows that B[a] = Eea' > 1 + µsa + pB2)a2/2. Hence 1 = a(B[7]  1) > Q (YPB +72µs)/2) = 3µs + OYµa2) 2 (6.1) 7 'Y from which the results immediately follows. u
The upper bound in Proposition 6.1 is also an approximation for small safety loadings (heavy traffic, cf. Section 7c): Proposition 6.2 Let B be fixed but assume that 0 = ,3(77) varies with the safety loading such that 0 = 1 Then as 77 .0, µB(1 +rl) 2) Y = Y(77) 277 PB Further, the CramerLundberg constant satisfies C = C(r1)  1. Proof Since O(u) + 1 as r7 , 0, it follows from Lundberg's inequality that y * 0. Hence by Taylor expansion, the inequality in (6.1) is also an approximation so that OAY]  1) N Q (711s + 72µB2) /2) = p + 3,,,(2) B 'y 7 2 2(1  p) _ 271µB
QPB PB)
6. MORE ON THE ADJUSTMENT COEFFICIENT 77
That C 4 1 easily follows from y 4 0 and C = ELe7V°O) (in the limit, b(oo) is distributed as the overshoot corresponding to q = 0 ). For an alternative analytic proof, note that C  1P = rlµB 73B' [7]  1 B' [ry)  1/0 711µB µB +7µB2 )  µB(1 +77 ) 'l = 1. 277q
77
7PBIPB
 77
13 Obviously, the approximation (6.2) is easier to calculate than y itself. However, it needs to be used with caution say in Lundberg's inequality or the CramerLundberg approximation, in particular when u is large.
6c A refinement of Lundberg 's inequality
The following result gives a sharpening of Lundberg 's inequality (because obviously C+ < 1) as well as a supplementary lower bound:
Theorem 6 .3 C_eryu < ,)(u) < C+ eryu where
= B(x) = C_ x>o f °° e7( Yx)B(dy )' C+
B(x) xuo f 0 e'r( vx)B(dy)
Proof Let H(dt, dx ) be the PLdistribution of the time r(u) of ruin and the reserve u  S7(„)_ just before ruin . Given r(u) = t, u  ST (u) = x, a claim occurs at time t and has distribution BL(dy)/BL(x), y > x. Hence ELe7£(u) 0
J
°o
H(dt, dx)
fX
eY(Y x) 00 f°° B(dy) x
BL dy
BL(x)
o
f
f H(dt, dx)
L ^ H(dt, dx) f e7B( x)B(dy) Jo oc, < C+
J0 0 o" H(dt, dx) = C. o" J
The upper bound then follows from ik(u) = e7uELeVu), and the proof of the u lower bound is similar.
78 CHAPTER III. THE COMPOUND POISSON MODEL
Example 6.4 If B(x) = eax, then an explicit calculation shows easily that B(x) _ e6X fz ° e7(Yx)B(dy) f x' e(6,6)(Yx)8esydy = 5 = P. Hence C_ = C+ = p so that the bounds in Theorem 6.3 collapse and yield the exact expression pey" for O(u). u The following concluding example illustrates a variety of the topics discussed above (though from a general point of view the calculations are deceivingly simple: typically, 7 and other quantities will have to be calculated numerically). Example 6.5 Assume as for (3.1) that /3 = 3 and b(x) = 2 .3e3x + 2 .7e7x, and recall that the ruin probability is 24 5su 5eu + 3e *(u) = 3 Since the dominant term is 24/35 • e", it follows immediately that 7 = 1 and C = 24/35 = 0.686 (also, bounding aS" by a" confirms Lundberg's inequality). For a direct verification, note that the Lundberg equation is
7 = /3(B['Y]1)
= 3\
2.337
+2.7771
which after some elementary algebra leads to the cubic equation 273  1472 + 127 = 0 with roots 0, 1, 6. Thus indeed 7 = 1 (6 is not in the domain of convergence of B[7] and therefore excluded). Further, 1P = B [7] 181B = 13 2.3+2.71 = 1 3 1 7 I 7'
_ 17
2 (3 a )2 + 2 (7  a)2 «=7=1 2 1p _ 7 _ 24
36 '
3.171 35* 36 For Theorem 6.3, note that the function QB[Y]1 f°°{L 3e_3x+
u
• 7e7x 1 dx
J
3 + 3e4u
f 0c, ex .
I 2 . 3e3x + 2 . 7e7x l dx
l J
9/2 + 7/2e4u
7. VARIOUS APPROXIMATIONS FOR THE RUIN PROBABILITY 79
attains its minimum C_ = 2/3 = 0.667 for u = oo and its maximum C+ = 3/4 = 0.750 for u = 0, so that 0.667 < C < 0.750 in accordance with C = 0.686.
Notes and references Theorem 6.3 is from Taylor [360]. Closely related results are given in a queueing setting in Kingman [231], Ross [308] and Rossberg & Siegel [309]. Some further references on variants and extensions of Lundberg's inequality are Kaas & Govaaerts [217], Willmot [382], Dickson [114] and Kalashnikov [218], [220], all of which also go into aspects of the heavytailed case.
7 Various approximations for the ruin probabil
ity
7a The BeekmanBowers approximation
The idea is to write i (u) as F(M > u), fit a gamma distribution with parameters A, 6 to the distribution of M by matching the two first moments and use the approximation
0(u)
f
u
Sa
r(A)
xa  leax dx.
According to Corollary 3.5, this means that A, 8 are given by A/S = a1, 2A/52 = a2 (2) PIB3) ^ZP(B)2 __ PPB a2 al 2(1  P)PB 3(1  P)µ8 + 2(1  p)2' i.e. S = 2a1 /a2, A = 2a2 1/a2.
Notes and references The approximation was introduced by Beekman [60], with the present version suggested by Bowers in the discussion of [60].
7b De Vylder's approximation
Given a risk process with parameters ,(3, B, p = 1, the idea is to approximate the ruin probability with the one for a different process with exponential claims, say with rate parameter S, arrival intensity a and premium rate p. In order to make the processes look so much as possible alike, we make the first three cumulants match, which according to Proposition 1.1 means p=AUB1=P1,
2N
(2) 6^= =OP
,
/3,4)
.
or equivalently that /3 is only slightly smaller than /3max = 1/µ8. Mathematically.3 and Corollary 3.8µBo  Ss' u where 6 = µB/µBo = 2µa/µB 2) . the approximating risk process has ruin probability z.g. 7c The heavy traffic approximation The term heavy traffic comes from queueing theory.1 As /3 f Nmax./3)] 1 . we have according to the PollaczeckKhinchine formula in the form (3.(3)2 P PB 2µB 2µB Letting /3* = /3/P. Letting Bo be the stationary excess life distribution.1. That is.Ps(/3max .7) that Ee$(Amex /j)M _ 1p _ 1p Eo [s (0max 1 . THE COMPOUND POISSON MODEL These three equations have solutions 9/3µB2)3 30µa2)2 3µa2) (3) P+ (3) ' 0 .b(u) = p*e./3)PBo PB .3 )1 } _ 1p 1 .(bA*)". cf. we shall represent this situation with a limit where /3 T fl but B is fixed.2) was suggested by De Vylder [109]. p* _ . [174]) shows that it may produce surprisingly good results.p = (/3max 0)µB.p + p { 1 1p ti 1 .s(/3max .80 CHAPTER III. Grandell [171] pp. and hence the ruin probability approximation is b(u) e(bAln)u. Though of course it is based upon purely empirical grounds.p . Proposition 7. the premiums exceed only slightly the expected claims. Proposition 1. Notes and references The approximation (7./3)M converges in distribution to the 2a exponential distribution with rate S = B' Proof Note first that 1 .PBo [s (/3max .P . 1924. (/3max . but has an obvious interpretation also in risk theory: on the average. numerical evidence (e.)3 )PBo µB  . heavy traffic conditions mean that the safety loading q is positive but small.3* /S.1.
[APQ] Ch. . or equivalently that 0 is small compared to µB . and hence 2µ2B 1 . In the setting of risk theory./3)u).4) suggested by the Cramer Lundberg approximation and Proposition 6.l3)M > (/3max . VARIOUS APPROXIMATIONS FOR THE RUIN PROBABILITY 81 Corollary 7.Q T /3max.g. Numerical evidence shows that the fit of (7. However . The present situation of Poisson arrivals is somewhat more elementary to deal with than the renewal case (see e . 2 provides the better mathematical foundation.p _ 2rl11B PB p. 7d The light traffic approximation As for heavy traffic . VIII). It is worth noting that this is essentially the same as the approximation (2) z/i(u) Ce.p. light traffic is of some interest as a complement to heavy traffic . light traffic conditions mean that the safety loading rl is positive and large . Of course. u * oo in such a way that (3max . obviously Corollary 7. we shall represent this situation with a limit where 3 10 but B is fixed. That is .3) is reasonable for g being say 1020% and u being small or moderate.0)u. the term light traffic comes from queueing theory.2. then P(u) 4 e6„ Proof Write z'(u) as P((/3max .7. Notes and references Heavy traffic limit theory for queues goes back to Kingman [230].ryu . the first results of heavy traffic type seem to be due to Hadwiger [184]. Mathematically.ze a2unµB laB (7.2 If . 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 . This follows since rl = 1/p .B AB ) 6()3max _'3) = However .1 1 . These results suggest the approximation Vi(u) e6(0_. in risk theory heavy traffic is most often argued to be the typical case rather than light traffic . the premiums are much larger than the expected claims . as well as it is needed for the interpolation approximation to be studied in the next subsection. while the approximation may be far off for large u./3)u * v.
e.T > u) = J o" B(x + u)/3eax dx . 0(u) /3 J B(x)dx = /3E[U . 10 ( u The alternative expressions in (7.= 1 aJ 1 a 0+ 1 = = p. z/' (u) convergence P(U . (7. 7e Interpolating between light and heavy traffic We shall now outline an idea of how the heavy and light traffic approximations can be combined. by monotone time T of the first claim .5) follow by integration by parts. Asmussen [19] and references there.5) u Proof According to the PollaczeckKhinchine formula.(3 10.3 As . and hence 00 (U) /3pBBo (u) = 0 / B(x)dx.3 is the same which comes out by saying that basically ruin can only occur at the F(U . cf. U > u] = /3iE(U . 0 u Notes and references Light traffic limit theory for queues was initiated by Bloomfield & Cox [69]. see Daley & Rolski [96].u. Another way to understand that the present analysis is much simpler than in these references is the fact that in the queueing setting light traffic theory is much easier for virtual waiting times (the probability of the conditioning event {M > 0} is explicit) than for actual waiting times . u Note that heuristically the light traffic approximation in Proposition 7. the Poisson case is much easier than the renewal case. THE COMPOUND POISSON MODEL Proposition 7. Again. For a more comprehensive treatment. Indeed. Sigman [347].82 CHAPTER III. ( 3 J O B dx. ao n=1 00 n=1 (u) P) anllBBon(U) onPaBon(u) • Asymptotically.u)+.T > u). Light traffic does not appear to have been studied in risk theory. The crude idea of interpolating between light and heavy traffic leads to 0 (u) C1 . [97]. . En'=2 • • • = O(/32) so that only the first terms matters. i.Q limIP ( u) + Q lim z/'(u) Amax &0 amax ATAm. Omax max m.
[84].O(E)(u) 1 (1 . f / Qmax B(x)dx 00 eQmaxxdx 4/ Qmax 00 QmaxQ amaze" and the approximation we suggest is J B(x) dx = cLT(v) (say).8. _(E) (u). that is. 8 Comparing the risks of different claim size distributions Given two claim size distributions B(1). . the idea of interpolating between light and heavy traffic is due to Burman & Smith [83 ]. Substituting v = u(.Wmax f(x ) dx + pee6mQ. 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. "/Qmex Cu) CLT(u ( /3max 0) + O16 CHT( U(Qmaz . with rate 1/µB = /3max.6) is . Instead. (U). Another main queueing paper is Whitt [380]. to get nondegenerate limits . available. B(2).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. Thus .3).O0 M. Let OLT) (u) denote the light traffic approximation given by Proposition 7./3)) . ) M. Al . ^ LT Q maxQ m"^ Qlo V LT) ( CHT(v) (say). ^IE) exist: 1 (B) HT QmsxQ hm J e e6" 2µE/µE2)'" = e(1 6)" =  Q1Qm. (7.3n.x . we see that the following limits HT) (u'). .VHT) ( ax QmQ ) h (B) ( . where further references can be found . COMPARISONS OF CLAIM SIZE DISTRIBUTIONS 83 which is clearly useless . one may hope that some correction of the heavy traffic approximation has been obtained. 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. The adaptation to risk theory is new. even if the safety loading is not very small. however.3 and use similar notation for %(B) (u) = (u). no empirical study of the fit of (7. z/i(E) (u) = pe(QmaxQ)u. Notes and references In the queueing setting .
THE COMPOUND POISSON MODEL To this end. .2 If B(') <j. we shall need various ordering properties of distributions. B(') <i. B(2)) if f fdB(1) < f fdB(2) for any convex function f. u Of course. whereas (consider x2) B(2) has the larger variance. B(') <d B(2)) if B(1)(x) < B(2)(x) for all x. Here convex ordering is useful: Proposition 8. or the existence of random variables U(l). this implies St T(l)(u) > r(2)(u) for all u so that 17(I) (U) < oo} C_ {T(2)(u) < oo}. an equivalent characterization is f f dB(') < f f dB (2) for any nondecreasing convex function f. B(' <. Proposition 8. for more detail and background on which we refer to Stoyan [352] or Shaked & Shantikumar [337].' 1)(u) < V)(2) (U) for all u. A weaker concept is increasing convex ordering: B(1) is said to be smaller than B(2) (in symbols. B(2) and PB(1) = µB(2). Finally.6. we have the convex ordering. 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). In particular (consider the convex functions x and x) the definition implies that B(1) and B(2) must have the same mean. one can interpret f x°° B(y) dy as the net stoploss premium in a stoploss or excessofloss reinsurance arrangement with retention limit x. Proposition 8. cf. Taking probabilities. this ordering measures difference in variability. B(2)) in the increasing convex order if f BM (y) dy < f 00 Bi2i (y) dy x x for all x.84 CHAPTER III. Rather than measuring difference in size. the proof is complete. Recall that B(') is said to be stochastically smaller than B(2) (in symbols.ill(u) < V)(2) (U) for all u. we can assume that 1) < St 2l for all t. then i.1 is quite weak. XI. U(2) distribution B(2) and U(1) < U(2) a. In terms of the time to ruin.s. Bill is said to be convexly smaller than B(2) (in symbols. In the literature on risk theory. equivalent characterizations are f f dB(') < f f dB (2) for any nondecreasing function f. Proof According to the above characterization of stochastical ordering. U(2) such that U(l) has distribution B(').1 If B(') <d B(2). then . most often the term stoploss ordering is used instead of increasing convex ordering because for a given distribution B. then Bill = B(2).
B(2). and here is one more result of the same flavor: Corollary 8. we have Bol) (x) f ' B(1) (y) dy < ' f' B(2) (y) dy = Bo2) (x)• µ 85 I.p ) E /3"µ"Bo2)* n(u) _ V(2) (u) n=1 = Corollary 8.2 is the following: Proposition 8.5 If '0(1)(u) < p(2) (U) for all u and a. it is seen that asymptotically in heavy traffic larger claim size variance leads to larger ruin probabilities.. then /'(')(u) < 0(2)(u) for all u. 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. and consider the following claim size distributions: B1: the standard exponential distribution with density ay.3 provides another instance of this. (D) (u) < O(B) (U ) for all u.4) certainly supports this view: noting that. Hence by the PollaczeckKhinchine formula .6 below is that (in a rough formulation) increased variation in B increases the risk (assuming that we fix the mean).8. Then V.4 Let D refer to the distribution degenerate at 'LB . we have by Jensen 's inequality that E f (U) > f ( EU).u) = (1 _ P) E /3npnBo( 1):n(u) n=1 00 < (1. larger variance is paramount to larger second moment.3 If B(1) <. A partial converse to Proposition 8. with fixed mean. Bo1) <_d Bo2) which implies the same order relation for all convolution powers. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS Proof Since the means are equal.. The problem is to specify what 'variation' means. B. A general picture that emerges from these results and numerical studies like in Example 8. from which the result immediately follows. Proof If f is convex. . say to p. then B(1) <. A first attempt would of course be to identify 'variation' with variance.1 and µB at 1 so that the safety loading 11 is 10%. This u implies that D <.1. B(2).6 Fix /3 at 1/1. Corollary 8. Example 8.e.(1) (. The heavy traffic approximation (7. Proof Consider the light traffic approximation in Proposition 7.
11 Notes and references Further relevant references are Goovaerts et al.9A2e'2r where A. Kluppelberg [234]. Let ua denote the a fractile of the ruin function. 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). In terms of variances o2.001 u0.0' U0. 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. in comparison to B2 the effect on the ua does not show before a = 0.e'\1x + 0. the behaviour of which is governed by a parameter 9. we have 0r3 = 2 < or2 = 1 < 02 = 10 < 04 = 00 so that in this sense B4 is the most variable. A standard example from queueing theory is .. B3 the comparison is as expected from the intutition concerning the variability of these distributions. However. 1/)(u.e. A2 = 3. One then obtains the following table: U005 U0. B3: the Erlang distribution with density 4xe2x.1%. and this is presumably a consequence of a heavier tail rather than larger variance.4142. i. = 0. 0. 9 Sensitivity estimates In a broad setting.01%. which appears to be smaller than the range of interest in insurance risk (certainly not in queueing applications!). all distributions have mean 1. Note to make the figures comparable. with the hyperexponential distribution being more variable than the exponential distribution and the Erlang distribution less.01%. B4: the Pareto distribution with density 3/(1 + 2x)5/2. 1%.lA.1358. We return to ordering of ruin probabilities in a special problem in VI.) = a.000. THE COMPOUND POISSON MODEL B2: the hyperexponential distribution with density 0.86 CHAPTER III. B. [166]. 0. Pellerey [287] and (for the convex ordering) Makowski [ 252].4. For B1i B2. and consider a = 5%. van Heerwarden [189].
(u) for large u.19P a/ . For example. with 0 the vector of service rates at different nodes and routing probabilities.e. if = a e(6A)u. obtained say in the natural way as the empirical arrival rate Nt/t in [0. the premium rate p and the claim size distribution B.. u Proposition 9. Let R(P) = Rtli. the distribution of %3 0 is approximatively normal N(0„ Q/t). In particular . say estimated from data. t]. the standard deviation on the normalized estimate ^/1' (the relative error ) is approximatively . Thus.1. where the partial derivatives are evaluated at p = 1. Then a p ao = 00 Qa/. SENSITIVITY ESTIMATES 87 a queueing network. while /3 = j3 is an estimate.3. 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. or we may be interested in aV)/0/3 as a measure of the uncertainty on '0 if 0 is only approximatively known. Thus at p = 1. Then the arrival rate /3(P) for { R(P) } is )31p. Example 9. and hence a _ e(60)u + u e(60)u = ( i + which is of the order of magnitude uV.1 Consider the case of claims which are exponential with rate 8 (the premium rate is one). and hence the effect of changing p from 1 to 1 + Ap corresponds to changing /3 to /3/(1 + Op) /3(1 . where Q2 = fl ( l2 1113 / _ Ou2v)2. it follows that ' is approximatively normal N(0.Ap).01/2u. Then ib = Pe(613)u. i. a/3 0 . Then if t is large .9. a2/t). In the present setting. Assume for example that 8 is known. increasing in u. Similar conclusions will be found below. Proof This is an easy time transformation argument in a similar way as in Proposition 1. we may be interested in a'/ap for assesing the effects of a small change in the premium. and s(9) the expected sojourn time of a customer in the network.2 Consider a risk process { Rt} with a general premium rate p. a0 as ao 80 19P .
() = log(1 + y//3). ()} p(dx) .88 CHAPTER III. (9.3 70 = 'Ye = = 7 /3(1 we(e +'y.4). so that heuristically we obtain '00 50ryu = Coe"u . However .()wC(e. namely that of a twoparameter exponential family of the form Bo.(/3 + y)we(9 + 7. Differentiating w.w(O. u Now consider the ruin probability 0 = 0 (u) itself. 4) (9 . and write yp = 8y/8/3 and so on . Consider first the case of 8/8/3: . ^) . THE COMPOUND POISSON MODEL As a consequence. 9.6) As will be seen below. Similar notation for partial derivatived are used below. Of course.3. Consider first the adjustment coefficient y as function of 3.g. (9.Owe (9. In the case of the claim size distribution B. this intuition is indeed correct.5) are similar.2) (see Remark 9.()(0 +'0) ' (9 .3.r.6 below for some discussion of this assumption). Viei '0(. we cannot expect in general to find explicit expressions like in Example 9.^)] 1(/3+y)we(9+'y.()^ 1 . (.3) follows by straightforward algebra.uypCe7u urypO. it suffices to fix the premium at p = 1 and consider only the effects of changing . but we shall concentrate on a special structure covering a number of important cases. () Proof According to (9. 3) ( 9 . various parametric families of claim size distributions could be considered.0 = t/'(u) and the CramerLundberg constant C.1 or Proposition 9. mathematically a proof is needed basically to show that two limits (u * oo and the differentiation as limit of finite differences) are interchangeable.w(6. Proposition 9. for the ruin probabilities . e. we can rewrite the Lundberg equation as w(9+ y.()YC = 1 +y/ /3 \ Q2 From this (9. /3 yields w e(e + Y.10) below. 5) (Q+'Y)[we(0+7. and the proofs of (9.3 or/and B.((dx ) = exp {Ox + (t(x) .t. The most intuitive approach is to rely on the accuracy of the CramerLundberg approximation . x > 0 (9. but must look for approximations for the sensitivities 0. (3+'y)PC (0+7.
2 of the Appendix ). BarndorffNielsen [58]). By dominated convergence.8) Letting cp = e0/e/3 and differentiating (9. (9.3 (see in particular (5. Z(u)/u a C//3PF where PF is the mean of F.(e"U = = wS(O.9) (9.9.x)B(x) dx + J U W(u . () . we get p(u) = J "O B(x) dx + J U O(u . () exp {w(9 + a. it holds that 89 a ue ryu a/3 Q(1 P) 7C2 Proof We shall use the renewal equation (3.QB(x) dx. we note the formulas Ee. Further write de = [we (9 +'y.([a] = exp {w(9 + a.C).x)B(x)dx. But from the proof of Theorem 5.x). Hence by a variant of the key renewal theorem (Proposition A1. ()} . SENSITIVITY ESTIMATES Proposition 9. () .w(9.3(x) dx.8) (Section 5). we multiply by e7" and let Z(u) = elt" cp(u). PF = (1 . and alsoo zl(u) + 0 because of B['y ] < oo.3) for z/'(u). Z= zl + z2 where zl (u) = e7u J m B(x)dx. 0(u) = /3 Ju"O B(x) dx + f 0 0(u .w(9. Be.w(O.g.x) F(dx ) f u J C F(dx) = C as u 4 oo. () .10) (9. ()] exp {w (O + y.x).4t (U)e°`U = which are wellknown and easy to show (see e. F(dx) = e'yy/3B(x)dx. Combining these estimates . z2(U) = e7" J u b(u . z2(u) _ 1 ^ e'ri`i7i( u .11) Ee. the proof is complete.8). ()} .St (U) Ee. u 0 Proceeding in a similar way as in the proof of the CramerLundberg approximation based upon (9.12)). w((9 + a.p)/C'y. O} (9.4 As u oo. 11 For the following. () .we(9 . u 0 Then Z = z + F * Z and F is a proper probability distribution .
()} 1z(dy) = f [t(y) .x)f3 f ^[t(y) .9)(9. ())B(dy) dx.8) that cp(u) . 2 z 07P N ue7u (3C de .6C do 89 1p 8( 1p Proof By straightforward differentiation.12) f exp {O y + (t(y) . ^) . this implies Z = z + F * Z.6e7u f "o f[t(y) .90 CHAPTER III. Then as u > oo.w(e.w(0. C) . ()](e7v .w( (0.2) holds. 8^ ue7u.QB(x)dx. 0 x Multiplying by e7" and letting Z(u) = e"uV(u). C)] (1 + 7 ) Proposition 9. z = zl + z2. )}B(dy)• Letting cp it thus follows from (9.wc(9. 01 (i+) do = +'Y.5 Assume that (9. THE COMPOUND POISSON MODEL [we(e+7.e7x/3 f 00 [t(y) . F(dx) = e7x. ()]B(dy) dx.11). By dominated convergence and (9. ^)} [wc (0 + 7. oo z2 (u) f C .lB(x) dx = e7uzl(u) + e7°zz(u) + V(u T where zl (u) = . ()]e7vB(dy) 'fCd 7 c .we (0.w (9.x).wc(O.1) B(dy) 'f '[t(y) . ()]B(dy) dx x 0 0C T ON O .wc (O. u Z2(U) = e7° f u ^/i(u . 8 8() 8( (9.w( (0.
by inserting in the above formulas. Here (9. we (9. SENSITIVITY ESTIMATES as u 3 oo. Z(u) /3C 91 o c'o e11(t (y) .6 Consider the gamma density b (x ) = Sa xa.Sry a/32 + a/37 + /37 . We get w( (0. that C = a. w(e.12) follows. and also zj (u) 4 0 because of f Hence.2) holds with p(dx) = xldx. U 7µF from which the second assertion of (9.15) (9.14) de = d( 7!3 76 = 7e = log ( \ ( \5a_ / \SSry ) 72 . Example 9.pa+1 . 9 = S.. ( 9.12) takes the form y) alp a .18) (05 + 57 _'3_y ./35' a/i'y + aryl 625ry. t(x) = logx.9.17) (9. ..16) (9.3ary tog('Finally.w((9. () ='I'(t.rye) S 5rya.Y)a+1 ' (9. It follows after some elementary calculus that p = a)3/5 and.a log S = log r(c) .QS 1 .ry) 5a1 cry (5 .1 . () = C/9 = a/S. < = a.13) (9. ())B(dy) < oo.C log(9).) log(9) = %F(a) logs where %1 = F'/]F is the Digamma function. ue_Yu 'C2d( 8a 8( 1 p .a/35a&y' ' (9.1edz = 1 exp {Sx + a log x .. () = log r(a) . and the proof of the first one u is similar.yu/3C2do u86 89 1p' az/) = 8z/. a /(S .(log r(a) a log S)} • r(a) 1.
2) with µ(dx) = 2x3zrdx. w(e.2 log (0.92 CHAPTER III.w(9.([Y] .3. C) . Be. C = . for a < a* = z (. Straightforward but tedious calculations . C) = B = Yc = de = do = .log c = 2 In particular.21og 2.2a) } Thus the condition B[a*] > 1 + a* /.l3 of Section 6a needed for the existence of ry becomes e^Q > 1+62 / 2.1 16 +ry c C22ry 2( = + 70 We (e.1 = eXP {c(C .9) (() . further yield .22. 9 = . which we omit in part .CZ try)} 1 C C2 try . t(x) _ .7 Consider the inverse Gaussian density ( b(x) Zx37 exp This has the form (9. () = Cc .3Ee.S[a] = exp {w (9 + a.2 . THE COMPOUND POISSON MODEL Example 9."62 . ()} = exp {c (C .
However . Notes and references The general area of sensitivity analysis (gradient estimation) is currently receiving considerable interest in queueing theory. by the LLN both F (NT = 0) and F (PT > 1) converge to 0 as T . B are assumed to be completely unknown. Comparatively less work seems to have been done in risk theory. Note that if NT = 0.1) .cue_7u)3C P Remark 9. That it is no restriction to assume k < 2 follows since if k > 2. to our knowledge.2 of the parameters. . BT [a]= NT ^` e"U.3C2de 1p' z a = c .7 and references there. we have assumed k = 2 and ti (x) = x.8 The specific form of (9. In general. Thus. To this end. the results presented here are new. in u which case the extension just described applies. and hence explicit or asymptotic estimates are in general not possible.. in which case we can just let t(x) = 0. 10 Estimation of the adjustment coefficient We consider a nonparametric setup where /3.a. [379] consider a special problem related to reinsurance. Thus.oo.. (9. Also. kT (a) = /T (BT [a] .12) takes the form a = a 93 ar.10.+UNT) > 1.. or Ct(x). the models there (e. then BT and hence ryT is undefined.2) is motivated as follows. Van Wouve et al. let NT 16T = ^T . the main tool is simulation. queueing networks) are typically much more complicated than the one considered here. ESTIMATION OF THE ADJUSTMENT COEFFICIENT Finally. thus. ae t 1lEY u S _ . Finally if k = 1. However. if 1 PT = /3TNT(U1+. 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). the exponent of the density in an exponential family has the form 01 tl (x) + • • • + 9ktk (x).g. for which we refer to X. then ryT < 0. we can just fix k . the exponent is either Ox. sj=1 and let YT be defined by IKT('ryT) = 0. and we estimate y by means of the empirical solution ryT to the Lundberg equation.
1) . 16T where V1. B [7]2 (10.v.B[7]) 0+ Iv/o(b[y].(27)/K'(7)2. N ( n[7]. it is easy to see that we can write \ V1 1 l _ . (10./^ B[27] .T y . More generally.'s.B[7]) + B [7] .: N 0.1 As T 4 oo. since NT /T . B[2'Y]  /3T ) .b[Yp'V21 T { (E[7] . then (10. THE COMPOUND POISSON MODEL Theorem 10 .. 7T a4' 7. If furthermore B[27] < oo.B[7]2 V2 . Hence KT(7) = (F' + (OT a(B[7l 0))((BT [7] .)vl+ N CO. V2 are independent N (0. Lemma 10 . a2 where a2 = /3r.2 As T * oo.Q and Anscombe 's theorem. For the proof.3/ T).7 + (. vfoVFB[2y].2) follows from NT/T a4' . 1) r. B[27] .'Y .2) rT(7) N N (0.3) Proof Since Var(eryU) = we have B[7]. we need a lemma.94 CHAPTER III.: N ()3.a BT[7] I B[7] I + .i3)(B[7] 1) + (3(BT[7]  .If .1) .3T .1) 'YT . .B[7]2 n Hence ( 10.1)2 + E[27] .B[7]2 }) ( T 0 .
(ry + e) and hence KT(7 . lcT(a ) 4 /c(a).Q. and the truth of this for all e > 0 implies ryT at 'y.e.10.E ) < 4T(7T) < (7 +0' which implies 'T(ry4) a$' r.'T(a) = 1 E Uie°U' a$' EUe "u = B'[a].'(y). BT[a] 3 B[a].4) + E).e) < 0 < r. To this end . OT a 95 u 4 /3.(ry .KT(7) kT(7) K'(7) . it follows that 7T7 KT(7T) . NT i =1 n'(a) for all a so that for all sufficiently large T K7 . °7IT) . I.e. If ryT E (7  we have KT(7 .E) < 4T(7T) < 4T(7 + E). where ryT is some point between ryT and ry. Theorem 10. ESTIMATION OF THE ADJUSTMENT COEFFICIENT which is the same as (10.2. 6"Y (10. By the law of large numbers.4) and Lemma 10..3).E) < 0 < kT(7 + E) for all sufficiently large T .1 can be used to obtain error bounds on the ruin probabilities when the parameters . Now write KT(7T)  kT(7) = 4T(7T)( 7T 7). 0 are estimated from data . Combining ( 10. 7T E (y . Then r.c'(7) N (0' T (2(7) / N (0. first note that e7TU N (e7U u2e27Uo'2/T) 7 . Let 0 < E < ry. NT BT [a] Hence r. Proof of Theorem 10. y + E) eventually.1 By the law of large numbers.
it means 2 (8 . = 1. A major restriction of the approach is the condition B[2ry] < oo which may be quite restrictive.ueryuU ".i. 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. > 0 for some t E [Wn_ 1. wn = inf{t > W.info< „< t S..1 is from Grandell [170]. and the known fact that the Y„ = max Vt tE[W„1.96 CHAPTER III. i. V. Hipp [196].. For example . Griibel & Embrechts [292]. t]}. ft.f. Letting Wo = 0.. Herkenrath [192].e.g.1 : Vt = 0. various alternatives have been developed. Asmussen [23]) can then be used to produce an estimate of ry.C1e"a ( see e.96 if a = 2. Frees [146].Wn) are i . the nth busy cycle is then [Wn1. Wn). . Deheuvels & Steinebach [102].. Further work on estimation of y with different methods can be found in Csorgo & Steinebach [94]. if B is exponential with rate 8 so that ry = 8 .Q.T = 3TKT ( 21T)IKT (^T)2 is the empirical estimate of vy and fc. satisfies b(. U2.3 or equivalently p > 1/2 or 11 < 100%.) = a (e. with a tail of the form P(Y > y) .T VIT where r7ry. Mammitzsch [253] and Pitts. Vt = St .0) < 5. i .d. Embrechts & Mikosch [133]. Notes and references Theorem 10.5%). For this reason . 6 < 2. [197].. This approach in fact applies also for many models more general than the compound Poisson one..g.e.. THE COMPOUND POISSON MODEL Thus an asymptotic upper a confidence bound for a7' (and hence by Lundberg's inequality for 0(u)) is e"TU + f. Csorgo & Teugels [95]..
exists. defined as solution of c(ry) = 0 where ic(s) _ /3(B[s] . 0. 97 . In particular. the premium rate is 1. The safety loading is q = 1/p . T) = P( /r(u) <T) \ = PI inf Rt <OIRo=u1 /\0<t<T PI sup St>ul 0<t<T Only the compound Poisson case is treated.1 (the role of ryy will be explained in Section 4b).f.Chapter IV The probability of ruin within finite time This chapter is concerned with the finite time ruin probabilities 0(u. 'y) where c(a) attains it minimum value. Further let 'Yo be the unique point in (0. the Poisson intensity is 0 and the claim size distribution is B with m. The notation is essentially as in Chapter III. B[•] and mean AB.s. it is assumed that i > 0 and that the adjustment coefficient (Lundberg exponent) y.g. generalizations to other models are either discussed in the Notes and References or in relevant chapters.1) .1 where p = 13µB. Unless otherwise stated. See Fig.
Var[T(u) I T(u) < 00] = VarL T( U) .2) Proof Let as in Example 111. By the likelihood identity III.98 CHAPTER IV. . the time of ruin is T(u) and ^(u) = ST(t&) .t. 1 FL.5 .) = e7u ELe'Y^(u) ELT(U)k = e'Yu b ELT(u)k = O(u)ELT(u)k.r.(4. FL and independent of T(u).u is the overshoot. PROBABILITY OF RUIN IN FINITE TIME Figure 0. In particular.(. 7. 1 Exponential claims Proposition 1. we have for k = 1. E[T(u) I T(u) < 00 ] = ELT (U).(U) < 00] = ELT(u)ke'YS. PL = 6/0 = 1/p > 1).. (u) is exponential with rate 0 w. EL refer to the exponentially tilted process 3 with arrival intensity S and exponential claims with rate / (thus .1 The claims surplus is {St}.1) (1.9). using that the overshoot l. 2 that E [T(u)k.1 In the compound Poisson model with exponential claims with rate S and safety loading 77 > 0. the conditional mean and variance of the time to ruin are given by E[r(u) I T (u) < oo] Var [T ( u) I T( u) < oo] /3u+1 J )3 _ 2/3Su+/3+S (S)3)3 (1.
the 1.2 In the compound Poisson model with exponential claims with rate 6 and safety loading rl > 0.h.1//32 (6/)3 1)2 26(/3u + 1)/(6 . Let 0 > yo be determined by ^c(0 ) = a.6 + a)0 .1)T(u)) = VarLe(u) + (PL . 1). Since Sr (u) and (PL .1) .2) is aLELT( u) .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 .(yo) = 2V .12 Thus the l. the Laplace transform of the time to ruin is given by Eea7( u) = E [eaT (u)."(ry) = 26//32. T(u) < oo] fora > r. u + ELe(u) _ PL .I (1.1.B = a. 0 Proposition 1.2).h. This means that /3(6/(6 ./3) .1)) ELST(u) ELT(u) (PL .6a = 0 with solution 0 (the .1)2VarLT(u) + 2 Ca 1I VarLT(u). is V1rLSr( u) +VarL ((PL .s.(PL . we have by Wald's identity that (note that ELSt = t(pL .h. Wald's second moment identity yields 2 EL (Sr(u) . of (1. EXPONENTIAL CLAIMS For (1 .1)T(u) are independent with QL the same mean .1 /3u + 1 u + 1 //3 = 6/3 6/01 For (1.s. . where = eBu I 1 . which leads to the quadratic 02 + (/3 ./3 .s.1)ELT(u).1)T(u))2 = UL where = s.6.0) .V/ is as asserted.1 (6)3)2 which is the same as the r.
the result follows.v.Y1 Y2 Figure 1..'s with rate 5.3 that we can write EeaT( u) = eeuEe 017(o).. Fig. M(u) T(u) = T + E Tk k=1 where T = T(0) is the length of the first ladder segment . T2 . St Ti F.. and M(u)+1 is the index of the ladder segment corresponding to T(u).9ST(u) +T(u)!c(0)} . More precisely.v. 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 > 0.0.3.1 where Y1. ... Note that it follows from Proposition 1.. But by the fundamental likelihood ratio identity ( Theorem 111. (1.. Ti.100 CHAPTER IV. . are the ladder heights which form a terminating sequence of exponential r. are the lengths of of the ladder segments 2..4. T(u) < oo] = EB [exp {aT(u ) . Using 5 = 6 .4) The interpretation of this that T(u) can be written as the independent sum of T(0) plus a r. PROBABILITY OF RUIN IN FINITE TIME sign of the square root is + because 0 > 0).3) we have E [e«T(u ). Y2.1 . Y(u) belonging to a convolution semigroup . T(u) < oo] = e. Cf.T+ Ti a t U T I 1 a i F.
.T. UN.ST). T) to be evaluated by numerical integration: Proposition 1.0. If QT = N > 0. .k + 1). UN.i (u.i.3 sin0 + 29) f3(0) = 1+/32/cos9. cf. where U1.T are conditionally i. Note that the case 6 # 1 is easily reduced to the case S = 1 via the formula V. .T is the residual service time of the customer being currently served and U2 .6(u) = Vfl/j l(Su. 1).4.3 Assume that claims are exponential with rate b = 1.T. the conditional distribution of VT given QT = N is that of EN where the r.1.cos (u/..T) 1 I fl(O)h(0) fdO where (1. including the customer being currently served). Let {Qt} be the queue length process of the queue (number in system. [4]) 00 (x/2)2n+3 Ij (x) OnI(n+j)! . Since U1 . then VT = U1. Corollary 11.d.6. 2. and exponential with rate S = 1. .T + • • • + UN. T..v.. Proof We use the formula .T the service times of the customers awaiting service ... For j = 0.T) = P(VT > u) where {Vt } is the workload process in an initially empty M/M/1 queue with arrival rate 0 and service rate S = 1. ..1(u.e. Then V(u.6) fl(9) f2(0) = = fexp {2iTcos9(1+/3)T+u(/cos91) cos (uisin9) . 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 . U2.I ex cos B cos j O dB fo " . EN has an Erlang distribution with parameters (N. EXPONENTIAL CLAIMS 101 For numerical purposes . the following formula is convenient by allowing t.1 )!.1.. let (cf.T. density xN lex/(N . i.
and define tj = e(1+R)Taj/2Ij(2vT T).ie . f3(0) .(31/2 cos (( k + 2)9) .)3k+1 = e(1+0)T e201/2Tcos 7r 0 e )3(k +l)/2 [31/ 2 cos ( kO) . (1. k k2 + $k+1 E bj 00 t j . PROBABILITY OF RUIN IN FINITE TIME denote the modified Bessel function of order j. let I _ j (x) = Ij (x). similar formulas are in [APQ] pp.102 CHAPTER IV. 00 E '3j/2 cos(je) j=k+1 00 _ j=k+1 ^j/ zeij = .)3k +1 tj g'(QT >.38).i(k +1)e R [/3( klal/2e:0 (01 /2 e .31 /2eie L 1)] 1 I/31/2eie .(31/2eie .112 l 1( k +1)/2 [ 31/ 2 cos(kO) .cos (( k + 1)0)] f3(9) Hence the integral expression in (1.cos((k + 2)9)] d9.1 R [. 8789) 00 E aj j= 00 = 1.1)] L _112 /(k+1)/2 [.8 ) yields F(QT > k + 1) . Then (see Prabhu [294] pp. in particular equations (1.3(k +1)/ 2ei(k + l)6 (.13(k +l)/2ei(k +1)9 R E .k + 1) = 1 k +1 + bj j=00 j=00 00 j=kk+1 j=k1 By Euler 's formulas.44).1 00 ok+lR 00 j=k1 +1)/2e . 912.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 .
or.0(u.2. u Notes and references Proposition 1. T). We first prove two classical formulas which are remarkable by showing that the ruin probabilities can be reconstructed from the distributions of the St. however.. t )). equivalently. The first formula. 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. however.7) that _ [^ au ak+l (30 k L. it follows as in (1. expresses V)(0. t) = P . Seal [327] gives a different numerical integration fomula for 1 . 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). Ui < x I / (note that P(St < x ) = F(x + t. F(x. We allow a general claim size distribution B and recall that we have the explicit formula z/i(0) _ P(7(0) Goo) = p. we are concerned with describing the distribution of the ruin time T(0) in the case where the initial reserve is u = 0. The rest of the proof is easy algebra. k! k=O k0 i/z Co Uk ate" o'/z e . Related formulas are in Takacs [359]. going back to Cramer. the numerical examples in [12] are correct). oo (u)31/2e^e)k = )3k z cos(k9) = R k.e = e' COS a cos(uf31/2 sin 0). E Fk. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Since P(QOO > k + 1) = flk+1. there are several misprints in the formula there. is numerically unstable for large T. T) in terms of F(.T) which. 2 The ruin probability with no initial reserve In this section .3 was given in Asmussen [12] (as pointed out by BarndorffNielsen & Schmidli [59]. from the accumulated claim distribution N. .
S„ 0 <t<Tv STS„+St_T+v Tv<t<T as the event that IS. See Fig. T T o where the second equality follows from II.3) with A = (0. ") } is at a minimum at time t.T) T F(x.T]. Proof For any v E [0. f T lStv)} 0<t<T by a 'cyclic translation'. [v. v]. 1 1 .1 In formulas.T))dv. Then 1 .104 CHAPTER IV.t)= {Stv) < SM. we define a new claim surplus process St StM NJ Figure 2.0<w<t} St+v . meaning that we interchange the two segments of the arrival process of {St}o<t<_T corresponding to the intervals [0.(6. T]. . T) = P(Tr(0) > T) = P(M(0.T)) 1 fT P(M(v.1. co ). resp.b (0. 2.T)) does not {Stv)} depend on v.i.T))dv E^T I(M(v.(.T)dx.(0. PROBABILITY OF RUIN IN FINITE TIME Theorem 2 . 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. Stv^ _ Define M(v.
Indeed.. T) as {ST<St+ vS.T) occurs or not as long as ST < 0. T) = M(0.2. then i fT I(M(v. v).T)f(I z /)(0. Fig 2. T Theorem 2 . v < t < T} n M(0. cf. Proof The event {ST < u} = { Ei T Ui < u + T j can occur in two ways: either ruin does not occur in [0. T].T)) dv f T I(M(0.v<t<T} = {ST<StSv.2 10(u. v<t<T}n{ST<STSv+St. T) occurs. there exist v such that M(v.2. . we can write M(v. Obviously.xdx. It is then clear from the cyclical nature of the problem that this holds irrespective of whether M(0. or it occurs. in which case there is a last time o where St downcrosses level u. v). T)) dv = TEST = T fP(ST < x) dx T T NT 1 f P(ST < x) dx = 1 f P Ui T .Tt))f(u+t. For example. THE RUIN PROBABILITY WITH NO INITIAL RESERVE 105 Now consider the evaluation of fo I(M(v. this integral is 0 if STv) . 0<t<v} = {ST < St . T) occurs. 0<t <Tv}n{ST<ST Sv+St T+v. t). w) for some small E.T) occurs. we can take v E (w E. T T o i =1 Let f (•.T) = F(u+T.Sv. ST > 0.t)dt.. v). where the last equality follows from ST < St on M(0. v)) dv = ST T T o (note that the Lebesgue measure of the v for which {St} is at a minimum at v is exactly . T. letting w = inf It > 0 : St_ = mino<w<T Sw}. T)). v) = M(0. Hence T TE f I( M(v. If ST < 0. We claim that if M(0. T)) dv. t) denote the density of F(•.T) and Sv < 0 on M(0. It follows that if M(0 . then M(v.ST on M (0.
Proof of Theorem 111. O(T . The proof is combined with the proof of Theorem 111. u which is the same as the assertion of the theorem.T) = {St < 0. 0 < t <T .T) = . {S t > z. z > 0.ST_ t_ and let A(z. For a fixed T > 0.p. E [t. define St = ST .z. ST_ _ z} . which occurs w.b(u.u+dt]). PROBABILITY OF RUIN IN FINITE TIME u Q II T Figure 2.T)+ J0 T (1V. The following representation of T(0) will be used in the next section. Proposition 2.T) = C(z. ST_ _ z}.2.106 CHAPTER IV. ST_ _ z}. 0 < t < T.v. Let Z be a r.t).2 . 0 < t <T. t + dt] occurs if and only if St E [u.Tt))P(StE[u. Hence P(ST<u) = 1 . 2.2. u + dt] and there is no upcrossing of level u after time t. C*(z.2 Here o.3 Define r_ (z) = inf It > 0 : St = z}. Then P(T(0) E • I T(0) < oo) = P(T_ (Z) E •).(0. {St > . which is independent of St and has the stationary excess distribution B0.
ST(o)_ E [z. 2.T))dT = Off(z) dz P(T_ (z ) < oo) = 3B(z) dz. A(z. T(0) < oo) = OR(z) dz in (2. z + dz]) = P(A(z. z + dz].T + dT]. z + dz].3 But by sample path inspection (cf.2. r(0) < oo) = 3R(z) dz JP(C(z. 7( 0) < oo) = P (C(z)) dT. z + dz].2.T(0)<oc) = f x F(U > y + z U > z) P(Sr(o)_ E [z. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Then 107 P(r(0) E [T. u which is the assertion of Theorem 111. {St }o<t<T have the same distribution .3). Hence integrating (2.ST(o) >y.1) z T . Thus P(Sr(o)_>x. T + dT] I S7(o)_ E [z. z + dz].T)). Figure 2.T).T) = C*(z.T)) = P(Cx. and since {St}o<t<T. we therefore have P(A(z.2. Fig. Proof of Proposition 2.1) that P(T(0) E [T.T))f3B(z) dz dT. .1) yields P(ST(o)_ E [z. T(0) < oo) B(y B(z) + z) f3B(z) dz = 3 f °^ B(y + z) dz = f3 + x v f B(z) dz. (2.3. It follows by division by P(ST(o)_ E [z.
r(a) denotes the solution < 'Yo of the equation a = ic(r (a)) = .2. [329]. who instead of the present direct proof gave two arguments. some relevant references are Shtatland [338] and Gusak & Korolyuk [181]. a martingale proof is in Delbaen & Haezendonck [103]. I L Let ga(x) be the density of the measure E[ear(°). T(0) < oo) 0 = dT f 0 P(C(z))P(Z E [z. Notes and references For Theorems 2.(yo). see in addition to Prabhu [293] also Seal [326].1.(3(B[r( a)] . Lemma 3.1) where a > r.T+dT]).5a). z + dz]. because of77>0.r(a). Theorem 2. ^(0) E dx] (recall that ^(0) = Sr(o)) and write ga[b] = f OD ebxga(x) dx. (3. Proposition 2. Lemma 3 .108 Hence CHAPTER IV. In the setting of general Levy processes.1 Eear( y) = eyr(a). Note that T_ (y) < oo a.3. z + dz]. T(0) < oo) = dTP(T_(Z) E [T. 3 Laplace transforms Throughout in this section. PROBABILITY OF RUIN IN FINITE TIME ]P(7(0) E [T.s. one based upon a result of Asmussen & Schmidt [49] generalizing Theorem 11.3 was noted by Asmussen & Kl(ippelberg [36].1) . Let T_ (y) be defined as Proposition 2.1 and the present proof is in the spirit of Ballot theorems.c(r(a)) l = l er( a)se+at } u yields 1 = eyr(a)Eear(y). r(0) < oo. Tak'ecs [359]. cf. 2.T + dT] T(0) < oo) dT f ' P(C(z))P(Sr( o)_ E [z.2 ga(x) = Qexr(a) f "o eyr(a)B(dy) x . Proof Optional stopping of the martingale I er (a) 9 t.6.5 and one upon excursion theory for Markov processes (see IX.
(Laplace transform) of the ruin Corollary 3. rr(0) < oo) = 1_ r(a) Proof Let b = 0. It is then easily seen that Za(u) is the solution of the renewal equation Za (u) = za (u) + fo Z. Corollary 3.2 P(Z E [y.4 E[eaT (o). b .3.r(a) = a [B[b] B[r(a)]] . LAPLACE TRANSFORMS 109 Proof Let Z be the surplus .°° ga(x)dx. £(0) E dx) = /3B(x + dy) dx and hence ga(x) = f e r)/3B(x + dy) _ /3 f x e(v. Z = y] = EeaT. T(u) < oo] du = Proof Define Za(u) = E [eaT(" ).f.1] evr(a)B(dy)[ b . Further by Theorem 111. the result follows after simple algebra.T(0) < oo] = 20[b] = za[b] (9a[b] 9a[0])/b 1 .r(a) The result now follows by inserting /3B[s] = ic( s) +/3+ s and ic(r(a)) =a. r(u) < oo). Hence eb"du E[eaT(").ic(b)/b x(b) + a eb"E[eaT(" ). (u . y + dy].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)) . time T(u): u u Here is a classical result : the double m.5 f 00 o a/r(a) . u .ga [b] 1 .r(a) b . E[ear (o) I T(0) < oo .ST(o)_ just before ruin .x)(a) B(dy)• Lemma 3 .3 ga[b] = c(b) Proof + b + a .3.ga [b] 0 TO Using Lemma 3.x)ga (x) dx where za(u) = f.(v) = ev''(a). Then by Proposition 2.g.2.3.
the known results are even less explicit than for the exponential claims case. i.1) i. Later results then deal with more precise and refined versions of this statement. The first main result of the present section is that the value umL.1)Er(u) . we need the following auxiliary result: Proposition 4. = (p . PROBABILITY OF RUIN IN FINITE TIME 4 When does ruin occur? For the general compound Poisson model. St/t 1 1/m.mL > E T(u) < 00 ) 40. Then given r(u) < 00. (4. That is. Then as u * oo...1. T(u) a. uoo u using e. For the second .00 St = lim . cf. note that by Wald's identity u + EC(u) = ES.r(u) = Er(u) • ES.s.(u ) = o(u) a. Theorem 4 .s. T(u)/u mL as u + oo. and hence a. By Proposition 111.mu D 2 4 N(0.2.2 Assume ri < 0. mu ) ( 0 m < ML '(u) 1 m > rL.h(u. Proposition A1. for any c > 0 P( Further. and take basically the form of approximations and inequalities. u 1 ET(u) 1 p1 u where Pw2 = 311B)m3• 7(u) .UProof The assumption 11 < 0 ensures that P(T(u) < oo) = 1 and r(u) a4' oo. where _ 1 _ 1 1 C ML w(ry) 6B'[7J 1 . t T(u) T(u) T(u) t m = lim = lim = lim Utioo u + Sr(u) u+oo S.e.6. This proves the first assertion of (4.1 Assume 77 > 0.3LELU 1 1p' is in some appropriate sense critical as the most 'likely' time of ruin (here C is the CramerLundberg constant). For the proof. P = /3µB > 1.w ) v/.3). (u) t. for any m T(u) u .110 CHAPTER IV.
and (4.6µB2) Z v m (3µB2) Z. 1'r(U) .g. the same conclusion holds with t replaced by r(u).1) is T (u)  U mL P( T (u) < I > E.1).N(0. For (4 . cf.3. .2.mu m .7 6  11 Proof of Theorem 4. again Proposition A1. Tu) T( u) .4.1 is standard.1). If Z . T) for T which are close to the critical value umL).3).mu (2) '• m3/2 µB 7 .1 (by considering 0(u.t/m D (2) 111 . though it is not easy to attribute priority to any particular author. 4).1.s. of (4. note first that ( Proposition 111. According to Anscombe' s theorem (e. and as a timedependent version of the CramerLundberg approximation. proving (4.6.h. the result comes out not only by the present direct proof but also from any of the results in the following subsections.2 of [86]) and (4. T(u) < oo f / 00) e7uE L [e_7 (t1). 4a Segerdahl's normal approximation We shall now prove a classical result due to Segerdahl.r(u)/m T(u) ti µB2) Z.^ N (o.1).2) follows immediately from u (4. implying T(u) . Thus. this can be rewritten as u + 1(u) . which may be viewed both as a refinement of Theorem 4.mL U > E. WHEN DOES RUIN OCCUR? and that Ee(u)/u a 0. T (u) < 00 J 0(u) e7'PL U \ I T u) .1 The l.5) St . PL (•)+ 0. Notes and references Theorem 4. apB ) . Theorem 7.mL >E By Proposition 4.
Let h(u) = E f (^(u)).t.)mu \ h(oo)Eg (r(ul) .T(u') given F.mul h(oo)Eg(Z).r.3). with w2 as in (4. Then for any y.^(T(u')).w2) r. Then h(u) 4 h(oo) = E f (6(oo)). and thus in (4. we get E[ T (u) . one has 9 (r(u)_rnu) Ef (^(u)) * E. oo). Hence Ef (Vu )) 9 (T(u. we can replace T(u) by r(u').) is readily seen to be degenerate at zero if ST(u•) > u and otherwise that of T(v) with v = u .v.f ( (oo)) . S( u ) < ul/4] < ET(ul / 4) = O(ul/4).ST( u') = u1/4 . letting Z be a N(0. Then the distribution of T(u) . Proof Define u' = u .ul/4. then e(u) and r(u) are asymptotically independent in the sense that. PROBABILITY OF RUIN IN FINITE TIME Corollary 4.l:(oo) (recall that rt < 0).))I h(ul /4  ^(u)) I(6 (u') C ) f < ul /4 + f(e(u') . oo ).e(u') oo w . E9(Z) (4.u1/4)I(S(u') > u1 /4) h(oo) + 0.L+YWLV'U) .a C4'(y )• ( 4.5) For the proof.6). e'°'/b (u.112 CHAPTER IV. we need the following auxiliary result: Proposition 4. (oo.4 (SIAM'S LEMMA) If 71 < 0. Using ( 4. using that ul/4 .3 (SEGERDAHL [333]) Let C be the CramerLundberg constant and define wL = f3LELU2mL = f3B"[ry]mL where ML = 1/(pL1) = 1/($B'[ry]1). and similarly as above we get E[f(^(u)) I Fr(u. P because of ^(u') . resp .VU T. O .(u.6) whenever f.T ( u')] = E[ T ( ul /4 .um. g are continuous and bounded on [0.4).
z/)(u . CL Fig. see also von Bahr [55 ] and Gut [182]. oo ) as u * oo. where we used Stain's lemma in the third step and (4.7) to be good. Theorem 4. however .5 '(u . . Segerdahl 's result suggests the approximation b(u. For refinements of Corollary 4. 4b Gerber's time.4) in the last. 10) '5(u) .(ay) = 17 7y = ay .4. T(u) < umL + ywL f.umL wI V"U u (4. y u) < .yK(ay)• (4. e7v" y < ^'(7) (4 . yy by 1 K. PL(T(u ) < umL + ywL) 113 4 C4(y). For practical purposes .9) ( 4 .T) Ce7"4 (T . 0.7) to be valid is that T varies with u in such a way that y(T) has a limit in (. that for the fit of (4.8) Note that ay > 7o and that 7y > •y (unless for the critical value y = 1/ML). Notes and references Corollary 4 .3 in terms of Edgeworth expansions . define ay. Thus .5) and solve for y = y(T). umL + ywL f) = e"P(T (u) < umL + ywL) = EL [e7V "). see Asmussen [12] and Malinovskii [254].1. Cf. also Hoglund [204]. WHEN DOES RUIN OCCUR? Proof of Corollary 4.7) To arrive at this . 3 is due to Segerdahl [333].dependent version of Lundberg's inequality For y > 0.7) whenever u is large and ly(T)l moderate or small (numerical evidence presented in [12 ] indicates . The present proof is basically that of Siegmund [342]. The precise condition for (4.oo. in practice one would trust (4. just substitute T = umL + ywL in (4. u needs to be very large). y > k'(7) . ELe7E (") . y u) < e 7v" .3 ery"z/i(u .
h(u. Numerical comparisons are in Grandell [172 ].8). yu < T (u) < oo 1 l e ayuEav [eT ( u)K(ay). see MartinLM [257] .b (u. .7 i. T(u) < yu] < eayu + yUr(ay) Y < eayuEav [ eT(u)K(av )L T(u) < yu} Similarly. yy is sometimes called the timedependent Lundberg exponent. and generalizations to more general models are given in Chapter VI. f Some urther discussion is given in XI. However. which may be understood from Theorem 4. the point is that we want to select an a which produces the largest possible exponent in the inequalities. a.v"U.9): Proposition 4. From the proof it is seen that this amounts to that a should maximize ayic(a). Hoglund [203] treats the renewal case.5. PROBABILITY OF RUIN IN FINITE TIME Proof Consider first the case y < 1/K'(y). the bound a7y° turns out to be rather crude . we have rc(ay) < 0 and get (u) . In view of Theorem 4.6.t. and hence t. yu 11 < T(u) < oo j < eayu +Y UK(ay) Remark 4.r.2. For a different proof.Y' (u. yu) < C+(ay)e7a„ where l C+(ay) = sup f 00 eayR(xy)B( .8).1). u Differentiating w. yu) is e 'Yyu/ .8 below . Then ic(ay) > 0 (see Fig .3 yields easily the following sharpening of (4. yu ) = < eayuEay [eay^ ( u)+T(U)K ( ay). 0.yu ) = eayuEav [e .ay4(u)+ T(u)K(ay ).114 CHAPTER IV. 5 is due to Gerber [156 ]. we arrive at the expression in (4. dy) Notes and references Theorem 4 . if y > 1/ic'(y). An easy combination with the proof of Theorem 111.6 It may appear that the proof uses considerably less information on ay than is inherent in the definition (4. who used a martingale argument. which shows that the correct rate of decay of tp(u.
c(&) = ic(ay) is < 0. This idea is precisely what characterizes the saddlepoint method.ayuEay f eay^ ( u)+T(u)K(ay). the choice of ay.i(u.8 If y < 1/ic'(ry). (4.ay and get Ea e ayf (00) y _ 'Ya( ayKal lay C 1 .. u 4 oo. and in case of ruin probabilities the approach leads to the following result: Theorem 4 .e. (4.5. WHEN DOES RUIN OCCUR? 115 4c Arfwedson's saddlepoint approximation Our next objective is to strengthen the timedependent Lundberg inequalities to approximations.e.yyu y l ay I 21ry/3B" [ay] V fU_ u + 00. As a motivation. then the solution &y < ay of . (0) r1 (a) ' I.6 with P replaced by Pay and FL by Pay. For any a > yo. we have ryas = ay . yu ) ayay e ryyu ayay 27ry/3B"[ay] u Proof In view of Stam 's lemma.ay y 'Yay  ay . T(u) < yu] .2 yields EaT(u) u u r. [eT(u )K( ay).11) ' If y > 1/ r .(u... T(u) suggests heuristically that l t/. then the relevant choice is precisely a = ay where y = T/u.4.yu) c ay .ay a. Ea . yu) = e. not inequalities. yu ) eaauEaye .^3 ]1/ Bay [lay .13) . then ay > 0. We thereby obtain that T is 'in the center' of the Padistribution of T(u).ayC() . the formula 0(u. Using Lemma 111. i. and b(u. and ii(u) .: T. Proposition 4.12) < yu] Here the first expectation can be estimated similarly as in the proof of the CramerLundberg ' s approximation in Chapter III.z. The traditional application of the saddlepoint method is to derive approximations.'(y ). if we want EaT(u) . it is instructive to reinspect the choice of the change of measure in the proof. (4.
(ay) _ y(ay .13) rigorously.7ruw2 Inserting these estimates in (4.13).c'(a) _ /3a/(8 .a) .13). 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). T(U) < yu] = eyuk (ay)E''ay (ek(ay )"1/2WV.3(5/(S .1B[ay]1 ) y(ay . Then ic(a) = .c(ay)ul/2W p 2ir = eyu(ay) dz 1 rc(ay ) 2.1) under Pay mation (4.l'B)y /(Pay . it seems tempting to apply the normal approxiyu + ul/2wV. we get heuristically that Eay Ler (u)r(ay).(j (1 .ay + ayl /BLay] . (4.ay ) r.a.1)3 = (jB"[ay]l (Pay . and the equation ic'(a) = 1/y is easily seen to have . PROBABILITY OF RUIN IN FINITE TIME ry I i . a nr=.a)2 .B[ay] /ay &y y(ay . . Writing r(u) and W2 = I3ay{.9 Assume that B(x) = eay.12) is 0 entirely similar. i B[7ay .ay) ay +. Example 4. (ay) J0 1 K(ay )u 1 00 c2(x) dx /2 w 1 ezcp(z /( k(ay)u1 /2w)) dz /O° _ 1 1 J e Z .4). The difficulties in making the proof precise is in part to show (4.11) follows. where V is normal(0.I ay &y a ^c'(ay) a (1 +. V < 01 Ir 00 er(ay)"1'2"'x eyur.1)3 = y3/3B"[ay].1.ay)K(ay) ay ayI&YI For the second term in (4.1) . The proof of (4.116 CHAPTER IV.
A related result appears in BarndorffNielsen & Schmidli [59].g. 2 = Var(Si ) the variance. and next to note that such an approximation in particular implies that the first passage probabilities are close. DIFFUSION APPROXIMATIONS solution ay=5 117 V 1 (the sign of the square root is . y) a''y" L '3 _ fl ) 51 /4(1 +1IY)3/4 \..f./4 ^y for 1/i (u. .because the c. It follows that 5^y =5ay = /«y =f3+ay=l3+d 1+1/y' V 1+^1/y /35 1+1/y /3' ay ay =Qay say =. 0 Notes and references Theorem 4.. (5.1) .8 is from Arfwedson [9].11) gives the expression '31/4 ( .5.ay)3 0 3/2 and (4. 5 Diffusion approximations The idea behind the diffusion approximation is to first approximate the claim surplus process by a Brownian motion with drift by matching the two first moments.i )( v s vc ('3 + s _2 / . yu) when y < 1/ic'('y) = p/1 . then { __ .3+52 1+/351/y' sy 7 B ii[ay] 25 _ 251/2(1 + y)3/2 (5 .= (s.p.tcp) Lo {Wo ( t)}t>0 .1.. The mathematical result behind is Donsker's theorem for a simple random walk {Sn}n=o. c a 00. is undefined for a > 5). in discrete time: if p = ES. is the drift and o.
1 below). such that the claim size distribution B and the Poisson rate a are the same for all p (i..2) t>o where p = pp = p .3.7c). + {Wo(t ) . this is an easy consequence of (5. Mathematically.t} _ {W_1(t)} .p.3) whenever c = cp f oo as p 1 p.1 As p J. However. cf. This is the regime of the diffusion approximation (note that this is just the same as for the heavy traffic approximation for infinite horizon ruin probabilities studied in III. It is fairly straightforward to translate Donsker's theorem into a parallel statement for continuous time random walks (Levy processes).tcpp) y = { WC (Sct) pct) } {Wo( t)}t>o (5. Lemma 111.1. for the purpose of approximating ruin probabilities the centering around the mean (the tcp term in (5. p. oo)). of which a particular case is the claim surplus process (see the proof of Theorem 5. we have o {i!t s: .z } {W_1(t )}t>o (5.e. where p is the critical premium rate APBTheorem 5 .1)) is inconvenient. PROBABILITY OF RUIN IN FINITE TIME where {W( (t)} is Brownian motion with drift S and variance (diffusion constant) 1 (here 2 refers to weak convergence in D = D[0. and consider the limit p j p. Indeed .118 CHAPTER IV. n/c < t < (n + 1)/c.1) with S. We want an approximation of the claim surplus process itself. a2 =/3µB2) Proof The first step is to note that { WC (St P) . we shall represent this assumption on 77 by a family {StP) L of claim surplus processes indexed by the premium rate p.p/c < St(p) < S((n+l)/ c + Pp/c.a = Snp) and the inequalities Sn )C . Letting c = a2/pp. 0 . St = EN` U= . (5..tp). and this can be obtained under the assumption that the safety loading rt is small and positive.3) takes the form LI S(P) { a2 to2/µ2 + t LI S (P) { a2 ta2/µ2 {W0(t)}.
and the r.2 suggests the approximation u 0(u. 199.1 I 7= . 119 It is wellknown (Corollary XI. ulpI/a2).u).(u) ti IG(oo. we omit the details .s. TS(u)=inf{t>0: WW(t)>u}. ulpl /a2) = e2"1µl / or2. 1. [169] or [APQ] pp.e. u)..t.4) Note that IG(.h. Corollary 5 . is IG(T. (ua2 To2 op \ IPI > IG ( T .h.1 .1. u) of r( (u) (often referred to as the inverse Gaussian distribution) is given by IG(x.T) IG(Tp2/ a2).2 As p j p. u) =PIT( (u) < x) = 1 ..7c.s. is 1/ip (ua2 /IpI.h.5) Note that letting T * oo in ( 5.6) This is the same as the heavy traffic approximation derived in III.5. has a continuous distribution. ^ p2 Proof Since f 4 SUP0<t<T f (t) is continuous on D a.f I \\\ J \ (5. the continuity argument above does not generalize immediately. and in fact some additional arguments are needed to justify (5.( ^ I + e2( \ I .1. (5. . ('. C. (5. see Grandell [ 168]. 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. Because of the direct argument in Chapter III.r. the continuous mapping theorem yields sup W Sz2 to lP 4 sup Wi(t)• O<t<T O<t<T a2 Since the r.8 or [APQ] p. w. since ti(u) has infinite horizon .Ta2 /p2). For practical purposes . any probability measure concentrated on the continuous functions. 196.5). 263) that the distribution IG(•. u) is defective when < 0.. Corollary 5. DIFFUSION APPROXIMATIONS Now let Tp(u) = inf{t>0: S?)>u}. (.6) from Theorem 5.s. we obtain formally the approximation V. However.
Theorem 5. See for example Billingsley [64]. However. and two further standard references in the area are Grandell [168]..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.t. The proof is a straightforward combination of the proof of Theorem 5.1. that 00 4090.3 Consider a family {Ste) } oc claim surplus processes indexed by a parameter 9. for more general models it may be easier to generalize the diffusion approximation than the CramerLundberg approximation.g.6 of [APQ]. We conclude this section by giving a more general triangular array version of Theorem 5. the simplicity of (5. pe . the B9.6) are presented. . in particular for large u.5) and (5. we have ^A. The first application in risk theory is Iglehart [207]. as an example of such a generalization we mention the paper [129] by Emanuel et al. 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. Further relevant references in this direction are Furrer [151] and Boxma & Cohen [75]. such that the Poisson rate Oe. Assume further that 039µB6 < pe. In contrast. in Asmussen [12]. 0) { 2 StQ2 /µ2 D { W_ i(t)}t>o t>o D 2 where p = pe = pe . All material of this section can be found in these references.120 CHAPTER IV. [169]. However.1 and Section VIII. and which is much more precise. in the next subsection we shall derive a refinement of (5. a2 = ae = 00µa6 Notes and references Diffusion approximations of random walks via Donsker's theorem is a classical topic of probability theory.5) for the compound Poisson model which does not require much more computation. (5.00µB6 + 0. B0 * Boo. For claims with infinite variance. Then as 0 _+ 90.Pe.r. Furrer. e. pt? 4 peo. on the premium rule involving interest.6) therefore does not appear to of much practical relevance for the compound Poisson model. The picture which emerges is that the approximations are not terribly precise. PROBABILITY OF RUIN IN FINITE TIME Checks of the numerical fits of (5. as 0 * 00 and that the U2 are uniformly integrable w.Po = 09µB6 . In view of the excellent fit of the CramerLundberg approximation.
90) .1 > 0. risk process with safety loading 77 > 0 correspond to 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. P9(r (u) < oo) = 1 for 9 > 0.s and p = /3µB < 1. . this means the following: 1. Then r. Let PO refer to the risk process with parameters e9oz Qo = QB[90]. However .'(yo) = 0 and let 90 = 'Yo. it is more convenient here to use some value 9o < 0 and let 9 = 0 correspond to n = 0 (zero drift). 3. claim size distribution B . In terms of the given risk process with Poisson intensity .6. PB('r(u ) < oo) < 1 for 9 < 0. The objective of the corrected diffusion approximation is to take this and other deficits into consideration.1) . let P9 refer to the risk process with parameters Q9 = QoB0[9] = QB[9 9o]. Bo(dx) = B[eo]B(dx). this is because in the regime of the diffusion approximation .6. which we have seen to play an important role for example for the CramerLundberg approximation . 9o T 0. The setup is the exponential family of compound risk processes with parameters ( B9 constructed in III. 0(0) = 0.Q (B[s] . For each 9.ao (0) _ /c(s + 9 .T) = Peo(r(u) < T) for 90 < 0. and we want to consider the limit 77 10 corresponding to Oo f 0. 77 is close to zero. and we are studying b(u. 2. Since Brownian motion is skip free./c(9 . this idea ignores (among other things) the presence of the overshoot e(u). whereas there we let the given 3B.90) and the given risk process corresponds to Poo where 90 = 'yo. Then EOU' = Boki[0] = Biki[eo]/E[9o] and "(s) = k(sBo)k(9o). In this setup.90] B(dx).c(s) = .4. Determine yo > 0 by r.9(s) = Ico ( s + 9) . 77 = 1/p . B9(dx) =Bale] Bo(dx) e9z keo)z = B[9 . .
C) = 2A + (2 .() The idea of the proof is to improve upon this by an O (u1) term (in the following. 9otc0" (0) = 0061 = ul.1) . 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.122 CHAPTER IV.. The first step in the derivation is to note that µ = k (0) = r0 (00) . C .. . Vargo S. PROBABILITY OF RUIN IN FINITE TIME Recall that IG(x.e.2' where as ususal ry > 0 is the adjustment coefficient for the given risk process. .. bl IG(t81. for brevity.Varo S1 = f30Eo U2 = S1. i.1) IG(x. u) = IG(x/u2. (. (U. () where h (A. The corrected diffusion approximation to be derived is (u.C.(y) = 0. Theorem 5. u) = euh(a . tu2 ) i IG (t. (. (6. and Si = QoEoU2 = Q B"'['Yo Eo U3 ]. IGu+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 > 0.7(u)/u2} eh(A. 1) • Since L eatIG (dt. 0o to. C.T) 1+u2 (6. _ ^(u) = ST . means up to o(u1) terms): ..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.2) .S. (01.3) this implies (take u = 1) Ego exp { . One has (6. the solution of r. write r = T(u).
2).2 ). that the saddlepoint approximation of BarndorffNielsen & Schmidli [59] is a serious competitor and is in fact preferable if 77 is large] .1 below is exact.2) is indeed o(u1). 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.h.h. the formal Laplace transform inversion is heuristic: an additional argument would be required to infer that the remainder term in (6.5) according to (6. But the Laplace transform of such a r.s.4.1 + 629. and the dotted line the corrected diffusion approximation (6. distributed as Z . we get by formal Laplace transform inversion that C 2 u. that whereas the proof of Proposition 6. CORRECTED DIFFUSION APPROXIMATIONS 123 Proposition 6.7.3.v. . To arrive at (6.f. In ( 1) and (2). we have p =. 9o T 0 in such as way that C = Sou is fixed. however. is the c. . 1% in (2) and (4).d. which is based upon exponential claims with mean µB = 1.v. u is Eeazead2/++ Eeaz[1 + ab2/u] where the last expression coincides with the r. just replace t by Tb1/u2.6. in (3) and (4).1 + u2 I Indeed.s. (6. Note. .3 = 0. it holds for any fixed A > 0 that Ego exp { Ab1rr(u)/u2} .1 As u + oo.z . The solid line represents the exact value . however . 1. of (6.5) Once this is established .ry2 . . the r. bl I IG I t +2 .52/u where Z has distribution IG (•. The initial reserve u has been selected such that the infinite horizon ruin probability b(u) is 10% in (1) and (3). 6 . of a (defective) r. A numerical illustration is given in Fig. p = 0.exp { h(A.'yu /2)(1 + b2/u)} + Aug 1I J .3). calculated using numerical integration and Proposition 1.
T) 0. The proof of Proposition 6.19)2 11 20 20 i0 T 1n0 Figure 6. For further numerical illustrations. and all of the numerical studies the author knows of indicate that its fit at p = 0.4 may not be outstanding but nevertheless.1 W IU. Note that the ordinary diffusion approximation requires p to be close to 1 and '0 (u) to be not too small. A51 7(SAT 3 3 h(X.07 0. see Asmussen [12].^) .08 a.u2 2u3 (e .01 0. OM 0.. it gives the right order of magnitude and the ordinary diffusion approximation hopelessly fails for this value of p.0 0.T1 00. (Inc 0s 0.() Lemma 6. BarndorffNielsen & Schmidli [59] and Asmussen & Hojgaard [34].TI CHAPTER IV.011 L1 60 T IM 11.T) 111 0.7.124 0.1 It is seen that the numerical fit is extraordinary for p = 0. PROBABILITY OF RUIN IN FINITE TIME 0.OOIi O.111 W(U. Similarly.2 e.114 0.00 0.aa1 .7 or at values of Vi(u) like 1% is unsatisfying.W21 0. the fit at p = 0. .199 0.08 0..1 proceeds in several steps.EB 0 p ex p ( 7 S h ^)u .05{ 0.(061 0.02 I 90 120 160 2W A0 Z WT 40 80 120 160 100 240 280 T 111 WI.
co ((/u)) } Let 8 = (2a + (2)1/2 = h().4) that the r.. in Lemma 6. (6.6) u U3 Lemma 6 .T (co (8/u) .3 EoU2 + 103OoEoU3 + " 2 6 Using d2 .. CORRECTED DIFFUSION APPROXIMATIONS Proof For a>0.7) 2 2 .+ h (A.. + a1b2 + .(3)Eea LauT exp i 3J .C)C/u . (6.s.00)(u +C)  'r (.6.2u (B3 .h.co (e) . 3 lim Eof (u) = EoC(oo) = a2 Ep = 3EoU2 uroo Proof By partial integration .C2 = 2). 1 = PB(T < oo) = Eo0 exp 125 {(B . the result follows. () 62 Eeo exp u u2 J . exp ue } al 1J 3 exP I [2).61a2T (B3 .1) h(A..4 Ea. 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.(3) J t _ aa1T l + eh(A. the formulas Po(C(0) > x) Po(C(co) > x) imply 1 °° Po(ST(o) > x) = EIU fIP0 (U>y)dy .r0 (00)) } Replacing B by 8/u and Bo by C/u yields e(B() = E eo exp { (e .2 behaves like C l Eeo eXp r _ ^81T 1 Sl u2 1 u 2u3 [1+h(AC) S . () + C and note that 2 KO (0) = 102. 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 .
PROBABILITY OF RUIN IN FINITE TIME The last term is approximately (e 3 (3) 27.x.() . 0 The last step is to replace h(A. we get h(A. () . 2 + 00 = .6) and 7co (Oo) = ico('y + Bo) to get 0 = 21 (^/2 + 2y90) + 1112 (_Y3 + 3_Y200 + 3y9o) + O(u4).2u [2A+ (2 3 . C) ( 1+ u2 The result follows by combining Lemma 6 . 2 and (6. we get the correct asymptotic exponential decay parameter ^/ in the approximation ( 6.\+ (2 (3 e 2u [ (2. 5 exp { _h(A) (1 + / y u J)) exp 1.6  d h(A. yields +90 62 0 + O(u 3) 2u2 +O(u 3). [2+ (2 .7) and using eh(a. yu/2). There are two reasons for this : in this way. Thus a2 y = 290 + O (u2).h.e h(aS)h (^^ 262 exp {_h(.() I 1 + u2 ) y . () by h(\. yu/2) h(A.s. letting formally T * oo yields 7/)(u) C'e7u where C' = e7a2). and inserting this and 9o 2 = S/u on the r. yu/2) 11+ 62 I} S 1 \\\ u/11 l 62 (3 2u 2A Proof Use first (6. l Lemma 6 .4.\ + () 1 2 / . and the correction terms which need to be added cancels conveniently with some of the more complicated expressions in Lemma 6. Thus by Taylor expansion around ( = 90u.(2A + ()1/21 exp S h(A.1 (y/2 + Oo)u .126 CHAPTER IV..2.S) d e 62 .2) for O(u) (indeed .h (A.2 (^/2 + 3y9o + 390) + O(u3).
()} 3 h (A. Hogan [200] considered a variant of the corrected diffusion approximation which does not require exponential moments.(i+ 62 exP{ h(A. In Siegmund's book [346]. 'yu/2) 127 ( i+ M pz^ exP { h (A. the approach to the finite horizon case is in part different and uses local central limit theorems. () (i+a ) 2A + (2 . this case is in part simpler than the general random walk case because the ladder height distribution G+ can be found explicitly (as pBo) which avoids the numerical integration involving characteristic functions which was used in [345] to determine the constants.T) has not been carried out and seems nontrivial. with the translation to risk processes being carried out by the author [12].1: Just insert Lemma 6. Fuh [148] considers the closely related case of discrete time Markov additive processes.7. 7 How does ruin occur? We saw in Section 4 that given that ruin occurs.4. the 'typical' value (say in sense of the conditional mean) was umL.1 (y/2 + Oo)u )} 1 (i + U ) [2+ C2 2u 62 S Pt^ exP { J 62(2 exp { h(A. u Notes and references Corrected diffusion approximations were introduced by Siegmund [345] in a discrete random walk setting. . His ideas were adapted by Asmussen & Binswanger [27] to derive approximations for the infinite horizon ruin probability 'i(u) when claims are heavytailed. the same as for the unconditional Lundberg process. i. the analogous analysis of finite horizon ruin probabilities O(u. The adaptation to risk theory has not been carried out.5 in Lemma 6. ()} . HOW DOES RUIN OCCUR? exp { h (x. () I 1 + u 2 ) } S 1 . and to the Markovmodulated model of Chapter VI in Asmussen [16]. The corrected diffusion approximation was extended to the renewal model in Asmussen & Hojgaard [34]. () I 1 + u2 )I 2u L 2A+C2_(2 exp { _h.e. that is. The answer is similar: the process behaved as if it changed its whole distribution to FL. We shall now generalize this question by asking what a sample path of the risk process looks like given it leads to ruin. 0 1 Proof of Proposition 6.
+ TMOO ). {ST(u)+t .(u).3 to .vi(u) Ce'Yu Corollary 7. In the exponential case. (u) and satifying PL(F(u)) * 1. {St}0< t<T(u)) Proof Write e'rsr(u ) = e'rue'r£(u).3L and the claim size distribution from B to BL. PROBABILITY OF RUIN IN FINITE TIME changed its arrival rate from 0 to /3L and its claim size distribution from B to BL. the numerator becomes e'ruELe7^ (u)PL(F( u)t) = e7uCFL (F(u)°) when F(u) E . we give a typical application of Theorem 7. .r. u * oo.2 If B is exponential. Recall that 13L = (3B[ry] and BL(dx) = e'rxB(dx)/B[7]. so in the in the proof. Note that basically the difference between FT(u) and . then P(u) and FL coincide on . P(u) and rate = aL w.1 Let {F(u)}u>0 be any family of events with F(u) E F.FT(u) is the stopping time oalgebra carrying all relevant information about r(u) and {St}o<t<T(u)• Define P(u) = P(•IT(u) < oo) as the distribution of the risk process given ruin with initial reserve u.r. Then also P(u)(F(u)) + 1.(u) is not . r(u) < oo) .128 CHAPTER IV. . ..t.EL[e7S.F.t. FL As example.FT(u) = o' (T(u ). Recall that . Theorem 7 . In fact.(u)_ and ^(u) are independent .(u)_ and similarly the denominator is exactly equal to Ce7u.T. the Poisson rate changes from .1. F(u)c] P(r(u) < oo) ?P(U) < EL[e7u.(u)_ is that i.. F(u)c] ti e' ru]PL (F(u)`) > 0. stating roughly that under F(u). and let M(u) be the index of the claim leading to ruin (thus T(u) = Ti + T2 + .ST(u)}t> o is just an independent copy of {St}t>o). Proof P(u) (F(u)c) = F(flu)c. 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. ^(u) is exponential with rate 8 w.F.TT(u) _measurable.
eaLx. Proof For the first assertion. the queueing results are of a somewhat different type because of the presence of reflection at 0. take I(Tk < x) .(1 .3. From a mathematical point of view. . 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 . see further XI. the subject treated in this section leads into the area of large deviations theory. This is currently a very active area of research.eALx) M(u) k=1 u The proof of the second is similar. HOW DOES RUIN OCCUR? Corollary 7. Notes and references The results of the present section are part of a more general study carried out by the author [11]. 129 M(u) >2 I(Tk < x) M(tu) p(u) M(u) >2 I(Uk < x) BL(x).7. however.
This page is intentionally left blank .
the Tn are independent. Proposition 1. the distribution Al of T1 is A as well.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.d.1 Define p = !µ. T3.(1. of the risk process form a renewal process: letting Tn = Qn . U2. . with Nt = # {n: Un <t} the number of arrivals before t.Then no matter the distribution Al of T1i B.. are i. . the one corresponding to the stationary case by 00)(u). We use much of the same notation as in Chapter I. Var(St) = 11Ba2A + I4AaB lim t goo t PA 131 . and the one corresponding to T1 = s by 1/i8 (u). r(u) the time to ruin.1...Q. with the same distribution A (say) for T2. Thus the premium rate is 1. see A.. and M is the maximum of {St}.. D'2. A different important possibility is Al to be the stationary delay distribution A° with density A(x)/µA.1).i. The ruin probability corresponding to the zerodelayed case is denoted by 1/'(u).. AA t*oo lim St = lim ESt t tioo t = p .. with common distribution B.Chapter V Renewal arrivals 1 Introduction The basic assumption of this chapter states that the arrival epochs O'1. the claim sizes U1. In the socalled zerodelayed case.1 (T1 = a1). {St} is the claim surplus process given by I.
and (1 . The renewal model is often referedd to as the Sparre Andersen process. by the elementary renewal theorem (cf.0 > 0.1 gives the desired interpretation of the constant p as the expected claims per unit time. one could imagine that the claims are recorded only at discrete epochs (say each week or month) and thus each U. stating that E[Nt+a .1) follows .a is really the accumulated claims over a period u of length a.3 (SWITCHED POISSON ARRIVALS) Assume that the process has a random environment with two states ON. RENEWAL ARRIVALS lim E [St+a .132 Furthermore for any a > 0. CHAPTER V. t 4oo Proof Obviously. From this ( 1. However . If the environment is Markovian with transition rate A from on to off and u from OFF to ON. the definition 77 = 1/p .Nt] * a/PA.2 (DETERMINISTIC ARRIVALS) If A is degenerate. we get similarly by using known facts about ENt and Var Nt that Nt Var(St) = Var E Nt U.1 of the safety loading appears reasonable here as well. Proposition 1.St] = a(p . 3) follows similarly by Blackwell 's renewal theorem. For (1 . Nt ESt = E E UI Nt t = ENt•pB . but the arrival rate in the ON state is .1) ENt/t + 1/µA. the . OFF. such that no arrivals occur in the off state. Nt + EVar U. say at a. Example 1 . Here are two special cases of the renewal model with a similar direct interpretation: Example 1. A. Sparre Andersen whose 1959 paper [7] was the first to treat renewal assumptions in risk theory in more depth. after E.1). Nt = Var(PBNt) + E(4Nt) Q2 2 0`2 A tpB B + o(t). 2). This has a direct physical interpretation (a large portfolio with claims arising with small rates and independently). The simplest case is of course the Poisson case where A and Al are both exponential with rate 0. Thus. s + t µA PA 0 Of course.t.
if for nothing else then for the mathematical elegance of the subject. S o<t<oo n=0. (an arrival occurs necessarily in the ON state. U1 . 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 ..4) with phase space {oN. 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 .T between a claim U and an interarrival time T. arrival times. Proof The essence of the argument is that ruin can only occur at claim times. and for historical reasons. the relevance of the model has been questioned repeatedly.oFF}. More precisely. Therefore. in general the mechanism generating a renewal arrival process appears much harder to understand..4 The ruin probabilities for the zerodelayed case can be represented as 0(u) = P(M(d) > u) where M(d) = Max {Snd) : n = 0. (1. 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^. the first term represents the probability F(U1 ... However.d.1. For the stationary case.} with {S(d)} a discrete time random walk with increments distributed as the independent difference U .4) w. Ao.2.t. and then the whole process repeats itself)..s < u).r. u For later use.. The following representation of the ruin probability will be a basic vehicle for studying the ruin probabilities: Proposition 1. initial vector (1 0) and phase generator 11 However. . and the present author agrees to a large extent to this criticism. we have From this the result immediately follows. A is phasetype (Example 1. integrate (1.1.4) fo Indeed.s.y)B(dy). we feel it reasonable to present at least some basic features of the model.1. INTRODUCTION 133 interarrival times become i.s > u) of ruin at the time s of the first claim whereas the second is P(r(u) < oo.i. the fundamental connections to the theory of queues and random walks.
3* (say ) and the U.0* (u) = P (rr* (u) < oo) where rr* (u) = inf It > 0: Rt < 0} ) . b=1 !=1 where {Nt } is a Poisson process with rate . That is . i. resp .1) +ry.Ut. 00).d. we shall be able to compute the ruin probabilities i(i* (u) for this model very quickly (. the claims and the premium rate are negative so that the risk reserve process . the remaining part of the prepayment (if any ) is made available to the company. t. A simple sample path comparison will then provide us with the ruin probabilities for the renewal model with exponential claim size distribution. then 0*(u) = e 'r" where ry > 0 is the unique solution of 0 = k*(ry) = *(B*[ry] .1) .134 CHAPTER V.3* pB. St = t .1 r* (u) One situation where this model is argued to be relevant is life annuities. If . with common distribution B* (say) concentrated on (0. are independent of {Nt} and i.1. < 1. 2. At the time of death . RENEWAL ARRIVALS 2 Exponential claims. U Figure 2. the claim surplus process are given by Nt Nt Rt = u+^U. The initial reserve is obtained by prepayments from the policy holders. each of which receive a payment at constant rate during the lifetime . then 0 * (u) = 1 for all u > 0. Using Lundberg conjugation . A typical sample path of {Rt } is illustrated in Fig. (2.1 If. Theorem 2 .a*PB• > 1. The compound Poisson model with negative claims We first consider a variant of the compound Poisson model obtained essentially by signreversion .
then by Proposition 111. EXPONENTIAL OR NEGATIVE CLAIMS [Note that r.2. Then the c. 2.(u ) < oo) = E {e7sr_ (u). > 1 . St=Rtu=St. and thus 1 = P(T. Proof Define 135 St =u . cf. 0 Now return to the renewal model. 2.Rt.2 sup St = inf St = 00 t>o t>o and hence 0* (u) = 1 follows. Fig. Hence y exists and is unique.* (a) = log Ee'st I.2(a). T_ ( u) < 001 e7"P(T_ (u) < oo) = e"V)* (u). and the Lundberg conjugate of {St} is { St } and vice versa.0. Hence T_(u) < oo a. of {St} is c(a) = is*(a7).g. If I3*pB* < 1. (a) is*(a) (b) . B* [7] and let {St} be a compound Poisson risk process with parameters . Then { St } is the claim surplus process of a standard compound Poisson risk process with parameters 0 *. B*. Let B(dx) = ^e7x B*(dx).1. Then the function k* is defined on the whole of (oo. . 0) and has typically the shape on Fig. Since ic'(0) < 0.2 Assume now . the safety loading of { St} is > 0. Define T_ (u) = inf It > 0 : St = u} . B. T_ (u) = inf { t > 0 : St = u 'r* (u)..2(b).f.s.3*.(a) 7 Figure 2.UB.
which has the advantage of avoiding transforms and leading up to the basic ideas of the study of the phasetype case in VIII.+Tn U1 Un. A variant of the last part of the proof.2). To + M(d) in the notation of Proposition 1. 2.. and hence the failure rate ...'s and noting that V)*(u) = P(M* > u) so that Theorem 2..• • • .. and (2 . respectively.2 If B is exponential.Tr+. the distribution of M(d) is a mixture of an atom at zero and an exponential distribution with rate parameter ry with weights 1 ..u+ and lr+. we get Ee'M(d) = Ee°M* _ Y/(.Ui .4. alternatively termination occurs at a jump time (having rate 8)..1 it is seen that ruin is equivalent to one of these values being > u. and 5PA > 1. then .Tn} n=0. Taking m. Hence M* max {To + Ti + • • • + Tn . with rate S (say)....136 CHAPTER V.. T2 = U2.. According to Theorem 2.Un } = max St = t>0 n=0.7r+ 7r EeTo b/(Sa) + +. Now the value of {St*} just before the nth claim is To +T1* +. the failure rate of this process is y..Y] (2.Y a I. u Hence P(M(d) > u) _ 1r+e'r". 1) means that 8(A[ry] . Then B* = A.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.a) = 1 .1.e.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 . To + max {Ul+•••+UnTI. 3* = 6.g.•.1.)(u) _ 1r+e7" where ry > 0 is the unique solution of 1 = Ee'Y(uT ) = S 8 A[.Ti = U1. RENEWAL ARRIVALS Theorem 2 . However. and from Fig .f.1) + ry = 0 which is easily seen to be the same as (2.. with the probability that a particular jump time is not followed by any later maximum values being 1 .1.
7r+) and hence r+ = 1.2.7r+) = ry and hence P(M(d) > u) = P(M(d) > 0)e7u = 7r+e'r". we have ] A[a )3] E«d'efl' = Bad> [a] A ad> [Q] = B[a +. resp. 0 3 Change of measure via exponential families We shall discuss two points of view. 111.3 A[a] B[a] F( d) [a +)3] F(d) [a] = Fad) [^] Letting M(u) = inf in = 1.. However. 3a The imbedded random walk The key steps have already been carried out in Corollary 11. hence exponential with rate b. B^d) where Aad> (dt) = ^[ a] A(dt). Putting this equal to y.6.y/b. This follow since. It only remains to note that this change of measure can be achieved by changing the interarrival distribution A and the service time distribution B to Aad^. a ladder step is the overshoot of a claim size.7r+). : S(d) > u} ..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] .4.3. letting P(d) refer to the renewal risk model with these changed parameters .. consider instead the failure rate of M(d) and decompose M(d) into ladder steps as in II. Bads (dx) = . CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 137 is b(1.B(dx). we see that ry = 6(1. the imbedded discrete time random walk and Markov additive processes.5. Hence the failure rate of M(') is 6(1 . which states that for a given a. Furthermore. The probability that the first ladder step is finite is 7r+..2. Thus a ladder step terminates at rate b and is followed by one more with probability 7r+. the relevant exponential change of measure corresponds to changing the distribution F(d) of Y = U .
Corollary 3..e. (d) (7) _ 0.Ce"u where C = limu.C(°)eryu where C(O) = C0[7] . This is known to be sufficient for ^(O) ]p (d) ([APQ] Proposition 3. 00)(u) . in the easiest nonexponential case where B is phasetype.2 p. RENEWAL ARRIVALS be the number of claims leading to ruin and ^(u) u = SM(u) . Consider now the Lundberg case. For the stationary case.r.1).u the overshoot . ik. cf. For claim (b). i .138 CHAPTER V. In fact. 7µA . We have the following versions of Lundberg' s inequality and the CramerLundberg approximation: Theorem 3 . we get: Proposition 3. 187) and thereby for ^(u) to be nonlattice w. E(d)e 1' (u).p)/($B'[7] .t.1) is explicit given 7. (b) V)(u) . Proof Proposition 3. VIII.4. 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 .1 For any a such that k(d)' (a) > 0.. O(u) = eauE (d)ea{ (u)+M(u)K (d)(a) . and claim (a) follows immediately from this and e (u) > 0..C8e7u where Cs = Ce78B[7]. provided the distribution F of U .(u) .2 In the zerodelayed case. (a) '(u) < eryu. to converge in distribution since p(yd) (r(0) < oo) = 1 because of r (d)' (y) > 0.T is nonlattice.3 For the delayed case Tl = s. let 7 > 0 be the solution of r. just note that F7(d) is nonlattice when F is so . the evaluation of C is at the same level of difficulty as the evaluation of i/i(u) in matrixexponential form.1 implies Cu) = e«uE ( 7d)e«^(u) .
we get r u +8 e"8(u) 139 e7uB(u + s) + 4 0 + L 00 J e7(v8)e7(u+8v).a .Sdt] = Ee'uh(T) means 1 = f ' e°^B(dy) f ' h( s)A(ds).dt(ah ( s) + h'(s)) so that Gha ( s. 3b Markov additive processes We take the Markov additive point of view of II. (s.3. Let P8f E8 refer to the case Jo = s. E8h0 (Jdt.dt ) eadt = h ( s) . y) = e°yh(s).1) = C(O).St)} can be defined by taking Jt as the residual time until the next arrival.. To determine boundary 0. For s > 0. another use of dominated convergence combined with Ao[s] = (A[s] 1)/SPA yields 00 u) e7u iP8(u) Ao(ds) + f 0 = CB['Y](A[y] . B(x) = o(e7x) and dominated convergence.5. 0) = ah (s) . 0) = tc(a)h(s).(°) ( Ce8B[7] Ao(ds) similar manner. delayed version of Lundberg's inequality can be obtained in a e7u.1) (normalizing by h(0) = 1).y) B(dy) 0 For the stationary case.0 ) = Eo[ha ( Jdt.5. h(s) = e(a +x( a))8 (3. 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.4). St)} and h.h'(s). IPA 0 Of course. 0 0 .9. Equating this to tch(s) and dividing by h(s) yields h'(s)/h(s) _ . (u + s ./c. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES Proof Using (1. we invoke the behavior at the 1 = h«(0. According to Remark 11. The underlying Markov process {Jt} for the Markov additive process {Xt} = {(Jt. Here K.. Sdt) = h(s . The expressions are slightly more complicated and we omit the details.(s.
8.rc(a)] B[a] A[a .s is the probability measure governing a renewal risk process with Jo = s and the interarrival distribution A and the service time distribution B changed to Aa. (3. the determination of the adjustment coefficient ry where the defining equations rc(d) (ry) = 0 and rc(ry) = 0 are the same. .S„):0<v< )be Lt = eaSt tK(a) h(Jt) = east . Proposition 3. resp.140 CHAPTER V. St)}too by letting the likelihood ratio Lt restricted to Yt = a((J. [a + /3] A[b . . Ba(dx) = B(dx). = J8 = T2. U. 5 For the compound Poisson case where A is exponential with rate .e. Further... Ba where Aa (dt) . J n+1 u are independent with distribution Aa for the Tk and Ba for the Uk.c(a)] which shows that U1.2) means 1 = B[a]/3/(/3+a+rc (a)).a .13]A[b .1)a in agreement u with Chapter III. rc(a) = 0 (B[a] . T2 are independent with distributions Ba.rc(a)] = B = Ba[13]Aa[5]. An easy extension of the argument shows that U1.a . Remark 3 .5. .e(«+k(a))t esy A(dt). [e1U1 + 6T2ea ( U1s)stc ( a)e(a+K(a ))( T2s)I B[a +. resp.c(a)] B[a] Proof Pa. i.s governing {(Jt. ...4 The probability measure Pa.2).. since JT. Aa as asserted . .. T2.s(Jo = s) = 1 follows trivially from Lo = 1. A[a . however. we can now for each a define a new probability measure Pa.tK( a)e. Note that the changed distributions of A and B are in general not the same for Pa. RENEWAL ARRIVALS B[a]A[a .2) As in 11. Ease AU1+6T2 [ AU1+6T2 = Ea a LT.rc(a)] = 1.s and P(d). (3. ] = E. An important exception is. .( a+r' (a))(Jt s) h(s) where c(a) is the solution of (3.
[APQ] Ch. it is easily seen that ic(ay ) > 0. 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. yu). and define yy = ay .4. see e. the time from he arrives to the queue till he starts service. The virtual waiting time Vt at time t is the residual amount of work at time t. that is. and assume that T„ is the time between the arrivals of customers n . the amount of ..yx(ay). .t. A(ds).s eaysr(")+r(u ) K(ay) h (s) .)+1 e J j e(ay+w(ay))8 e .y yuAa y [ay + K(ay) . Then J(rr(u)) = TM(u)+1 and hence Ws(u. Notes and references 4 The duality with queueing theory We first review some basic facts about the GI/G/1 queue..g.5. u The approach via the embedded random walk is standard.(u. Then "^ e(ay+w(aY))8 Ys(u. T(u) < yu h(JT(u)) < eayu+yuk(ay ) ( Eia y Le(a(+k(ay))s v.4.r.6 Let y < let ay > 0 be the solution of ic'(ay) = 1/y. defined as the single server queue with first in first out (FIFO. The actual waiting time Wn of customer n is defined as his time spent in queue (excluding the service time).yu ) e7vu A[ay . which is the same as the asserted inequality for 0.ay+ray))TM(. Let M(u) be the number of claims leading to ruin . XII. The claim for the zerodelayed case follows by integration w. yu) = F'ay. that is.1 and n. 2. not after time T. say finite horizon ruin probabilities where the approach via the imbedded random walk yields results on the probability of ruin after N claims. Using the Markov additive approach yields for example the following analogue of Theorem IV. .rc( ay)] = e(aa+(r ))sb[a ]e7yu L y1 In particular. and U„ the service time of customer n. yu ) < e7yu. or FCFS = first come first served) queueing discipline and renewal interarrival times.4. . see in particular Dassios & Embrechts [98] and Asmussen & Rubinstein [45].5 Proposition 3. For the approach via Markov additive processes. for the zerodelayed case zp8(u. Label the customers 1. Proof As in the proof of Theorem IV.
The next result summarizes the fundamental duality relations between the steadystate behaviour of the queue and the ruin probabilities (part (a) was essentially derived already in 11.1 Wn+1 = (Wn + U. Vt converges in distribution to a random variable V. an+1) = [on. 0 Applying Theorem 11. Then: (a) as n + oo.. Then P(r(u) < T) is the probability z/iiNi (u) of ruin after at most N claims. (4. If W1 = 0. but interchanging the set . It then jumps to VQ„ . (4.1). and obviously z/'(u) = limN. The traffic intensity of the queue is p = EU/ET. Wn converges i n distribution to a random variable W.• • • Tk ). in which case {V} remains at zero until time on+1.1) The following result shows that {Wn} is a Lindley process in the sense of II.Tn)+ Proof The amount of residual work just before customer n arrives is VQ„ . but we shall present a slightly different proof via the duality result given in Theorem II.1. and combining with (4.2) (b) as t * oo.+ Un.3. and we have P(V > u) = ?/iiol(u). then Wn v M. we get: Corollary 4.4..Tn)+.3 Assume rl > 0 or. Thus.2 Let Mnd) = maxk=o. Let the T there be the random time UN.. p < 1. since customer n arrives at time on. equivalently. the waiting time a customer would have if he arrived at time t. and we have P(W > u) = V.1. .4."^ Vi(N) (u). the proposition follows.3) Proof Part (a) is contained in Theorem 11. we have Wn = Van(left limit). 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.n1 (U1 +• • •+Uk Tl .4. equivalently. on + Tn) the residual work decreases linearly until possibly zero is hit.1 and Corollary 11. whereas in [On . Thus Vos}1 _ = (Wn + Un . Also {Zt}o<t<T evolves like the leftcontinuous version of the virtual waiting time process up to just before the Nth arrival.4: Proposition 4.4): Proposition 4.2. (u). (4.142 CHAPTER V..
Hence for x > 0. In fact ..2). and we get: .2.. u Now return to the Poisson case ... which implies the convergence in distribution and (4.e.T*)+.T* < x) fK(x_y)F(dy) (x > 0 is crucial for the second equality!).T* are independent and distributed as U1.4 The steadystate actual waiting time W has the same distribution as M(d). i. However.. (4..Ao in (b). we obtain: Corollary 4. convergence in distribution hold for arbitrary initial conditions .oo in Proposition 4.. THE DUALITY WITH QUEUEING THEORY 143 (T1.5) Proof Letting n . For part (b). T1 . by an obvious reversibility argument this does not affect the distribution .le) the same is true for the timereversed point process which is the interarrival process for { Zt}o<t < T• Thus as before .T* = y yields K(x) = P ((W + U* .5 (LINDLEY'S INTEGRAL EQUATION) Let F(x) = P(U1T1 < x). T] form a stationary renewal process with interarrival distribution A. but this requires some additional arguments (involving regeneration at 0 but not difficult) that we omit.4. K(x) = P(W < x). { 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). Then K(x) = J x00K(x . hence (since the residual lifetime at 0 and the age at T have the same distribution . (4. Corollary 4. TN) with (TN..1. resp ..y)F(dy). conditioning upon U* . It follows that P(WN > u) =. Then the arrivals of {Rt} in [0.(N)(u) has the limit tp(u) for all u. where U*.. Letting n oo in Corollary 4.T*)+ < x) = P(W + U* . we let T be deterministic . we get W = (W + U* . Ti) and similarly for the U. as WN. T1. x > 0. and hence in particular ZT is distributed as the virtual waiting time just before the Nth arrival. A. Then the corresponding queue is M/G/1. namely W1 = 0 in (a) and Vo = 0.4) 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. cf.
. That is. Asmussen [24] and references there. 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.g. see e. W v V. the zerodelayed and the stationary renewal processes are identical.g. implying P(W > u) = P(V > u) for all u.5) to hold for all x E R and not just x > 0). Hence '(u) = Ali(°)(u). Cohen [88] or [APQ] Ch.5) is in fact a homogeneous WienerHopf equation.5) looks like the convolution equation K = F * K but is not the same (one would need (4. Some early classical papers are Smith [350] and Lindley [246].144 CHAPTER V. VIII).6 For the M/G/1 queue with p < 1. The equation (4. the actual and the virtual waiting time have the same distribution in the steady state. RENEWAL ARRIVALS Corollary 4. 0 Notes and references The GI/G/1 queue is a favourite of almost any queueing book (see e . Note that (4. Proof For the Poisson case.
and can be computed as the positive solution of WA = 0. • Claims arriving when Jt = i have distribution Bi.. • The premium rate when Jt = i is pi.)iJEE and its stationary limiting distribution by lr. t St = E Ui . As in Chapter I. Ire = 1. The intensity matrix governing {Jt} is denoted by A = (A.f pi. N. i=1 0 and r(u) = inf It > 0: St > u}. . The ruin probabilities with initial environment i are '+ki(u) = pi(T(u ) < oo) = Pi (M > u). {St} denotes the claim surplus process. here it exists whenever A is irreducible which is assumed throughout. Thus. dv.(3i when Jt = i.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 . M = supt>o St.T) = Pi (T(u) < T). 145 Oj( u. {Jt} describes the environmental conditions for the risk process.
1 implicitly assumes that the sojourn times of the environment in the normal and the icy states are exponential. cf. = iii when j E E(i). we could distinguish between normal and icy road conditions.1 Consider car insurance. which is clearly unrealistic.T(n)).14. r^ = P (1. we can approximate A(i) with a phasetype distribution (cf. and p is the overall average amount of claims per unit time.5 below. Example 1. Proposition 1. in blockpartitioned form. one expects that 3i > on and presumably also that Bn # Bi.. and we have f3. Assume similarly that the sojourn time in the normal state has distribution A(n) which we approximate with a phasetype distribution with representation (E('). Unless otherwise stated.4) with representation (E(i). Then the state space for the environment is the disjoint union of E(n) and E(i). a(i). u .11 below. 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. and assume that weather conditions play a major role for the occurence of accidents. the intensity matrix is A OW) T(i) T(n) t(n)a(i) where t(n) = T(n)e. f3i and claim size distributions Bn. For example. the operational time argument given in Example 1. Bi.1) iEE Then pi is the average amount of claims received per unit time when the environment is in state i.a('). P = E 7riPi. leading to E having two states n.146 CHAPTER VI. i and corresponding arrival intensities Qn.2. According to Theorem A5. we shall assume that pi = 1. An example of how such a mechanism could be relevant in risk theory follows. Example 1 . with rates Ani and Ain. assume that the sojourn time in the icy state has a more general distribution A(i). say. MARKOVIAN ENVIRONMENT where as usual Pi refers to the case Jo = i. t(i) = T(')e are the exit rates. Thus. /3 = Nn when j E E(n). this is no restriction when studying infinite horizon ruin probabilities. respectively.2 (ALTERNATING RENEWAL ENVIRONMENT) The model of Example 1. say. meaning that accidents occuring during icy road conditions lead to claim amounts which are different from the normal ones. T(=)). cf. We let p Pi = /ji/AB.
i8f n1. Example VIII.tEH is a transition matrix. w. but assume now that the arrival intensity changes during the icy period.. St = SB=(t). A('^). such that the icy period is of two types (long and short) each with their sojourn time distribution A('L). T(9) +wggt(9)0. and . Approximating each A('?) by a phasetype distribution with representation (E('l).a(n).>. is the probability that a long icy period is followed by a short normal one.5 (MARKOVMODULATED PREMIUMS) Returning for a short while to the case of general premium rates pi depending on the environment i. The simplest model for the arrival intensity amounts to . .Q. say it is larger initially.1.3i/pi.3.. Qi = . one could for example have H = {i1.. the state space E for the environment is { ('q.2. n8}. This amounts to a family (A(")) ?CH Of sojourn time distributions.. we assume again pi = 1 so that the claim surplus is Nt St = ?Ui_t... it = Jel(t). t(n) = T("i)e. say. Indeed.. such that a sojourn time of type rt is followed by one of type c w.4) with states i1. u Example 1 . 4 (SEMIMARKOVIAN ENVIRONMENT) Dependence between the length of an icy period and the following normal one (and vice versa) can be modelled by semiMarkov structure.. iq (visited in that order) and letfOil >. where W = (w. One way to model this would be to take A(') to be Coxian (cf.. 0 Then (by standard operational time arguments) {St } is a risk process in a Markovian environment with unit premium rate. 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) ..n. let T 9(T) = f pi. . resp.1. and similarly for the normal period.T(n)). the parameters are ^ij = aid/pi.3 Consider again the alternating renewal model for car insurance in Example 1. depending only on 77.p. (9) where q = CHI.. 1 . In the car insurance example. dt. i ) : n E H.3i. u From now on. i E E(n) }.j = . u Example 1. MODEL AND EXAMPLES 147 Example 1 .. and 1/ii(u) = t/ii(u).J017. Then for example wi.
MARKOVIAN ENVIRONMENT We now turn to some more mathematically oriented basic discussion.7 The Pidistribution of T1 is phasetype with representation (ei.148 CHAPTER VI. Pi (Ti E dx.6 The claim surplus process {St} of a risk process in a Markovian environment is a Markov additive process corresponding to the parameters µi = pi. More precisely.(3iBi(dx). o = 0. the empirical distribution of the claims is B*. In particular. A remark which is fundamental for much of the intuition on the model consists in noting that to each risk process in a Markovian environment. we put )3* = E 7fi/3i. The key property for much of the analysis presented below is the following immediate observation: Proposition 1. .A . vi(dx) = . 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). )3*. B* = 1 /^* Bi. . the Markov additive structure will be used for exponential change of measure and thereby versions of Lundberg's inequality and the CramerLundberg approximation. N > 1(Ul < x) a4 B*(x).e(A(Oi)d'sg)xe.(Qi)diag)• More precisely. Nt Nt a . JT1 = j) = Qj • e. iEE iEE )3 These parameters are the ones which the statistician would estimate if he ignored the presence of Markovmodulation: Proposition 1.5. dx.8 As t oo. one can associate in a natural way a standard Poisson one by averaging over the environment. qij = 0 in the notation of Chapter 11. Next we note a semiMarkov structure of the arrival process: Proposition 1. 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. t l=1 Note that the last statement of the proposition just means that in the limit.
this converges to the exponential distribution with rate 0* as a * 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$. N + oo Hence 1 Nt 1 N`+) Nits Nt E I ( Ut <. y Ni) i Nti) t a.aA .. In particular.4. {St} to the compound Poisson model with parameters 0 *. The next result shows that we can think of the averaged compound Poisson risk model as the limit of the Markovmodulated one obtained by speeding up the Markovmodulation.1. Bi. iEE 13 A different interpretation of B* is as the Palm distribution of the claim size.9 Consider a Markovmodulated risk process {St} with param eters Ni. Bi. the Fidistribution of T1 in {St(a ) } is phase type with representation (E. zli( (u) . cf. given {Jt}0<t<0.. Bi(x). However .* (u) for all u. and furthermore in the limit JT. Conditioning . Proof According to Proposition 1.. oo) as a 4 oo.(/3i)aiag). 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 . ^j 7riNi.7.2. Proposition 1. the limiting distribution of the first claim size U1 is B*.. and let {St °i} refer to the one with parameters Pi.6. B*. aA. Example 11.x) = Nt E > I (Uk) X) Nt Bi(x) 1=1 iEE k=1 iEE 1: t5 Bi( x) = B*(x). Nt a' t t iEE Also. e. has distribution (7ri)3i //3*)iEE and is independent of Ti. By Proposition A5. Then it is standard that ti lt '4' iri as t > oo. A. In particular. i. Hence Nt'> a . we may view Nt`i as the number of events in a Poisson process where the accumulated intensity at time t is Niti.
. T) for all u and T..150 CHAPTER VI..1.g. we first get that 3 (3* = 2.=1. U. marked by thin.. from Theorem 3. 5 5 where E5 denotes the exponential distribution with intensity parameter 5 and a > 0 is arbitrary. Continuing in this manner shows that the limiting distribution of (T. with T2 being exponential with rate . we may imagine that we have two types of claims such that the claim size distributions are E3 and E7.2}. s 5 in state 2. thick.10 Let E_{1. T) + ?P* (u.). On Fig.s = 1o in state 2. Claims of type E3 arrive with intensity 2 .. the paths of the surplus process will exhibit the type of behaviour in Fig. those of type E7 with intensity z s = 5 in state 1 and with intensity z . resp.s = o in state 1 and with intensity 1 .1 of [145]. state 1 appears as more dangerous than state 2. From this the convergence in distribution follows by general facts on weak convergence in D[0. B1=3E3+2E7.1 with periods with positive drift alternating with periods with negative drift. U2) are independent of . MARKOVIAN ENVIRONMENT upon FT. is as in {St }. That is. oo). Thus. the company even suffers an average loss. which also yield O(a) (u. B2=1E3+4E7.2... and in fact P1 = 31AB1 = 9 3 1 2 (5 3 3 1 1 2 1 5 7 1 81 70 ' _ 19 4 5 P2 = . e.2 +2 2 = 3. 132=2. A= (  a a ) \ a a 5 5 J 9 3 2 a1=2. lines in the path of {St}.FT. since E3 is a more dangerous claim size distribution than E7 (the mean is larger and the tail is heavier).2. The fact that indeed 0(a) (U) 3 0* (u) follows.31µB 2 = 2 5 3 7 70 Thus in state 1 where p. 0 Example 1. > 1. 1. and (at least when a is small such that state changes of the environment are infrequent). shows similarly that in the limit (T2. Computing the parameters of the averaged compound Poisson model. the overall drift is negative since it = (2 2) so that p = 71P1 + 112P2 = 7..l3* and U2 having distribution B*.. 1. 9 . there are p = 2 background states of {Jt}.
1).8. (b) St/t * p .. we have E[St + t I (t(i))iE EI = E t(i)OW = iEE t(i)Pi• iEE Taking expectations and using the well known fact Et(i)/t * 7ri yields (a). For (b). MODEL AND EXAMPLES 151 Figure 1.1 a.1 Thus. 01 /. t * oo.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. Hence B* = 415E3+5E7/ +4 ( 51E3 +5 E7) = 1E 3 +2E7.11 (a ) ESt/t * p .1. the averaged compound Poisson model is the same as in III.3* = 3/4 of the claims occur in state 1 and the remaining fraction 1/4 in state 2. Proof In the notation of Proposition 1. note first that EN Uk')/N a4' µgi. t + oo. 0 The definition (1. Hence (i) Nti) 1 U(i) k' N(i)k=1 E t 4 St + t = iEE Nt t 1: 7ri Qi µs. That is.1. iEE . a fraction r. = P.(3.s.
. and a more comprehensive treatment in Asmussen [16]. 2 The ladder height distribution Our mathematical treatment of the ruin problem follows the model of Chapter III for the simple compound Poisson model. + Xn SWn ](1 a . Proof The case 77 < 0 is trivial since then the a. [302]. The case 77 > 0 is similarly easy.. and so on. also + . The proof of Proposition 1. with some important improvements being obtained in Asmussen [17] in the queueing setting and being implemented numerically in Asmussen & Rolski [43]. [212]..g. In risk theory. The mainstream of the present chapter follows [16]. limit p . then M = 00 a. with X2. Now let r) = 0. then M < oo a.2(a) p. see the Notes to Section 7.. There seems still to be more to be done in this area.1 of St / t is > 0. dt . Now obviously the w. Eiw. If 77 > 0.\ i and EiX1 Ei f 13 J.Jt=i}.1 jEE = (p . [APQ].1 and the Corollary are standard. and hence wn /n a4. Since the X„ are independent . and hence 1/ii(u ) = 1 for all i and u.Jt=i}. X2 =SW2 So.s. EiX = 0.0i(u) < 1 for all i and u... let some state i be fixed and define w=wl=inf{t >0:Jt_#i.1(b) is essentially the same as the proof of the strong law of large numbers for cumulative processes.a form a renewal process .. Statistical aspects are not treated here. s. and involves a version of the . Theorem II. and hence oscillates between 0o and oo so that also here M = oo. .. 0 Notes and references The Markovmodulated Poisson process has become very popular in queueing theory during the last decade. X 1 =Sty. see [APQ] p. and . 38) Eiw1 = 1/ir.ld. PB. MARKOVIAN ENVIRONMENT Corollary 1. Proposition 1.. 136 or A. [315].4. and hence M = 00.. X3.152 CHAPTER VI.s. Then by standard Markov process formulas (e. n n Thus {SWn l is a discrete time random walk with mean zero. .1)Eiw = 0. w2=inf {t>w1:Jt_#i. having the Pidistribution of X. some early studies are in Janssen & Reinhard [211]. See Meier [258] and Ryden [314].Eiw o'o Eiw • E ^ifjµs.12 If 77 < 0.
j.dx). let G+(i.x). For measurevalued matrices. which represents a nice simplified form of the ladder height distribution G+ when taking certain averages : starting {Jt} stationary.g. j. we define the convolution operation by the same rule as for multiplication of realvalued matrices.1 irG+(dy)e =. The form of G+ turns out to be explicit (or at least computable).x.1) 0 (b) G+ (y. •) II = JG+(i. •). j. .2. and S (dx) the measure valued diagonal matrix with /3 Bj(dx) as ith diagonal element.i. oo)). T R(i. (y.4) we obtain the following result .2 (a) The distribution of M is given by 00 1 . cf.3*B *(y)dy. G+ is the matrix whose ijth element is E G +(i. for i. Let further R denote the preT+ occupation kernel. but is substantially more involved than for the compound Poisson case .5.A) = Pt(ST+ E A. the definition of . n=0 (2.IIG +II)e. we get the same ladder height distribution as for the averaged compound Poisson model. •) * G +(k. only with the product of real numbers replaced by convolution of measures.2) R(dx)S((y . That is.6*. However ..Jt=j)dt. dx)/jBj(y . k. by specializing results for general stationary risk processes (Theorem II .Jr+ =j. e. oo) = J ao 0 G+(i.A) =ZI(St E. 6. THE LADDER HEIGHT DISTRIBUTION 153 PollaczeckKhinchine formula (see Proposition 2.6. oo)) = f R(i.a/i.j E E. •)• kEE Also.EA. j. T+ < oo) and let G+ be the measurevalued matrix with ijth element G+(i. Proposition 2. 00 (2. j. see also Example II. Proposition 2.2(a) below ) where the ladder height distribution is evaluated by a time reversion argument. Define the ladder epoch T+ by T+ = inf It : St > 0} = r(0). IIG+ II denotes the matrix with ijth element IIG+(i.j. Thus.(u) = Pi(M < u) = e' E G+ (u)(I . B* in Section 1. j.
MARKOVIAN ENVIRONMENT Proof The probability that there are n proper ladder steps not exceeding x and (x)ej. {mx} is a non terminating Markov process on E. To this end . and that the environment is j at the nth when we start from i is e . III to bring R and G+ on a more explicit form . That is.2 useful .IIG+II)e. thick. we need as in Chapters II. mx = j when for some (necessarily unique) t we have St = x.2) is just the same as the proof of Lemma 11.1) follows by summing over n and j.154 CHAPTER VI. lines in the path of {St}. see Figure 2.1 The following observation is immediate: Proposition 2. To make Proposition 2. The u proof of (2. the intensity matrix A* has ijth element * 7r ^i3 7ri and we have Pi(JT = j) = 7rj P2(JT = i)7ri (2. hence uniquely specified by its intensity matrix Q (say). 0  x Figure 2.1 for an illustration in the case of p = 2 environmental states of {Jt}. and let further {my} be the Evalued process obtained by observing {Jt } only when {St*} is at a minimum value.3 When q > 0.6.3) We let {St*} be defined as {St}. G+ the probability that there are no further ladder steps starting from environment j is e^ ( I . marked by thin. St < S* for u < t. . From this (2. only with {Jt} replaced by {Jt } (the /3i and Bi are the same ).3. resp. we need to invoke the timereversed version {Jt } of {Jt} . JJ = j.
An excursion of {St*} above level x starts at time t if St = x. and the excursion ends at time s = inf {v > t : S. s]. In general.(/3i)diag + T S(dx) eQx. For example the excursion of depth 2 has one subexcursion which is of depth 1. and S(dx) is the diagonal matrix with the f3iBi(dx) on the diagonal.2. THE LADDER HEIGHT DISTRIBUTION 155 Proposition 2. Otherwise each jump at a minimum level during the excursion starts a subexcursion. Furthermore. = x}. {S.(/3i) diag. Q( n+l) _ ^. Proof The argument relies on an interpretation in terms of excursions. The definitions are illustrated on Fig. corresponding to two subexcursions of depth 0.0. and the excursion is said to have depth 1 if each of these subexcursions have depth 0. we say that the excursion has depth 0. Figure 2. If there are no jumps in (t.4 Q satisfies the nonlinear matrix equation Q = W(Q) where 0 co(Q) = n* .. 0 mms1   ^O \ T.2 where there are three excursions of depth 1.2 .and a jump (claim arrival) occurs at time t.2.*. 2. the sequence {Q(n)} A* defined by Q(O) = . 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). we recursively define the depth of an excursion as 1 plus the maximal depth of a subexcursion. } is a minimum value at v = t. ( Q( n)) converges monotonically to Q.
.5 R(i.4).St EA. By considering minimum values within the excursion. A) = L' U(j.s. we first compute qij for i $ j. it becomes clear that pij = r [eQh] 0 ij Bi (dy) • (2. h. Suppose mx = i.4) To show Q = cp(Q). 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. j. Writing out in matrix notation . of the definition to make U be concentrated on (co. e. 7rE Proof We shall show that Fi(Jt=j.u< t). A). Similarly. i. the subintensity matrix of {min+i ) } is cp (Q(n)) = Q(n +l) which implies that qgj +1) = \!.Qi + )%pij) Now just note that t pij and insert (2. StEA . Theorem 2 . It is clear that { mini } is a terminating Markov process and that { mio) } has subintensity matrix A* .j +/3ipij. 0)). Now let {m ( n) } be {mx } killed at the first time i7n (say) a subexcursion of depth at least n occurs . mx+dx = j) occurs in two ways .5) A (note that we use A = {x : x E Al on the r.(01)diag = Q. The proof of Q = W(Q) then immediately carries over to show that the subintensity matrix of {mil) } is cp (Q(o)) = Q(l).6) .St <S*. Q = W(Q) follows.. p1^) Define a further kernel U by f U(i. either due to a jump of {Jt } which occurs with intensity A= j.T+>t) _ ^iF 7ri (JJ =i. A) = f Pi(mx = j) dx eie4xej dx A u (2. Fi(mh =i ) = 1 + =hflh+Qihpii+o(h) implies qii = 'iii /i +)3ipii. Similarly by induction . or through an arrival starting an excursion terminating with J.156 CHAPTER VI. = j. (2. Then a jump to j (i. It follows that qij = A.j.
St <Su. Remark 2. From Qe = 0. we shall see that nevertheless we have enough information to derive.4]. x < 0. consider stationary versions of {Jt}..S„<0.0<u<t) = P. u It is convenient at this stage to rewrite the above results in terms of the matrix K = 0'Q'A.. We may then assume Ju=Jtu. 0 < u < t) = 7rjPj(Jt =i. 0<u<t). dt.r. {Jt }. e. (b) for z > 0.6(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. Jt = i. the CramerLundberg approximation (Section 3).StEA. where A is the diagonal matrix with 7r on the diagonal: Corollary 2. and we let k be the corresponding right eigenvector normalized by Irk = 1. `` {K(n)} [the W(•) here is of course not the same as in Proposition 2. oo))dx.2. THE LADDER HEIGHT DISTRIBUTION 157 from which the result immediately follows by integrating from 0 to oo w.Jo=i.=StSt.St EA. 0 +1) = cp (K( n)) defined by K(o) = A . K( n (d) the sequence converges monotonically to K. it is readily checked that 7r is a left eigenvector of K corresponding to the eigenvalue 0 (when p < 1).0<u<t. (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.. To this end. St EA. and get irPi(Jt =j. and this immediately yields (2. (Jo = j. St < St U..6 (a) R(dx) = eKxdx.(Jt=j.Qi)diag.7 It is instructive to see how Proposition 2. and to obtain a simple solution in the .6 is hardly all that explicit in general.1 can be rederived using the more detailed form of G+ in Corollary 2. oo)) = f o' eIXS((x + z. G+((z.6).g..t.z+>t) = P.(. St E A.
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
(3.(ay)}. we have ic(ay) < 0 and get 'i(u) . 1 Similarly. for a vector <p(u) = (cpi (u))iEE of functions . We further write G(u) for the vector with ith component Gi(u) = EiEE G+(i. av 'i [h.8 ) Then for all i E E and all u > 0. as in the classical case (3. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 165 hiav)e _avuE.y)G+(z. 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. hj P . if y > 1lk'(ry).5).j) * coj)(u) _ f u ^Pj(u .i [eT(u)K(av ). dy)• o iEE jEE .i I (a) exp {aye(u) + r(u)r. (u.3. we shall need the matrices G+ and R of Section 2. yu < r(u) < 00] < hiav)C+o)( y)eavu+yuw(av) 0 Note that the proof appears to use less information than is inherent in the definition (3.9) For the proof. Chie ryu < Vi(u ) < C+hie 7u. exp {e() + r(u))} . oo)) and. r(u) yu o)(y)eavuEav.7. yu) f h(av)e v avuE«v. However.i [e*(u)K(av). we let G+ * W(u) be the vector with ith component E(G+(i.j.00 su e7( ( 3. yu < r(u) < 00 h 4(u) < h(av)C+o)(y)eavuEav .V)i(u. r(u) < yu] hiay)C+ h=av)C+ o) (y)eayu+yuw(av).11 Let Bj (x) C_ = min 1 • inf jEE hj x>o f2° e'r( vx)Bj(dy) ' C+ _ mE 1 Bj(x) J Y x)Bj (dy)..5) will produce the maximal ryy for which the argument works.
dy).j.12 Assume sup1. dx) f e7( vu)Bj (dy . _ To see that the ith component of U * G(u) equals ?Pi (u).(0) ] (u) < sup Jt t.x ) R(i. 0 G+(i. Then iterating the defining equation ip(n+1) = G + G+ * V(n) we get W(N+1) = UN * G + G+N+1) * ^(o) However. we have G *(N +1) * ^. MARKOVIAN ENVIRONMENT Lemma 3 . 00 Thus C+ > hj f"o e7(Yu)G +(i.u IMP:°) (u) I < oo. = Eo G+ G. 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. jEE u 0 j.3jhj // f 00 R(i.x) jEE 00 u 0 //^^ C+E. dy) 00 C+ ijhj f R(i.7. Lemma 3 . j. dy) : 1(u) < C+ > hj u e(1tL)G+(i. dx) 100 C . dx). j. and define W(n+1) (u) = G(u) + (G+ * tp(n))(u). j. j. 00 f C_ hj f e(Y)G+(i. dy) = aj f Bj(dy .166 CHAPTER VI.x ) = Gi(u). U = U". Hence lim cp(n) exists and equals U * G. Proof Write UN = EN G+ . Then cpin)(u) sit (u) as n + oo.x) x) jEE 0 E Qj f jEE R(i. dx ) Bj (u . if r+ (n) is the nth ladder epoch. °O .u Iv 2°)(u)I Pi(rr+(N + 1) < oo) + 0. n > oo.j.ery(&u+x)Bj (dy) Bj(u Bj (u .& (u).13 For all i and u.
11.M > u) = Pi(ST<u. j. letting MT = maxo<t<T St. T) = Pi (7.11) C_e7u 57 O+[i.10: Theorem 3 . 14 Let yo > 0 be the solution of 'c'(yo ) = 0. and let 8 = e'(70).tpi(u.ST). dy) (3. dy) jEE o (3. and assuming it shown for n.8) with y replaced by yo and hi by h=7o ). +i .13) Hence. y]hj = C_ e7uhi.MT<u. we get Wo n +1) (u) = ? 7 i ( U ) + E J u gyp. we have Vii (u) . taking cps°) (u) = 0. u The proof of the upper inequality is similar .u)G+(i. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES proving the upper inequality. Then 0< Vi (u )  0i(u.(u) < T ) to 0i (u) which is different from Theorem 3. 167 u Proof of Theorem 3.10 ) by Lemma 3 . 9 for the last equality in (3. (3.13 Let first cp=°)(u) = C_ hie"u in Lemma 3. it follows that Vi(u) < C_(yo) h=70)e7ou. j. MT < u. T) < C+(')' o)hi7u)e7ou8T . dy) jEE u U +C_ hje7( yu)G jEE"" +(i. this is obvious if n = 0.12) Proof We first note that just as in the proof of Theorem 3. We claim by induction that then cpin) (u) > C_ hie7u for all n. from which the lower inequality follows by letting n * oo.Pi(MT > u) = Pi(MT < u. Indeed. let C+(yo) be as in (3. and the proof of the lower one is similar.y)G+(i. j.10) C_ 1 f hje7(y.T) = Pi(M > u) . jEE estimating the first term in (3. and using Lemma 3 .n) ( u .M>u) = Ei [VGJT (u . ST < u] < C+(yo)e7ouEi [h^7o)e70ST1 l T J = C h(7o)e7ou8T . (3.3. j.11).13. Here is an estimate of the rate of convergence of the finite horizon ruin probabilities 'i (u. 13 and the second by the induction hypothesis .
Further related discussion is given in Grigelionis [176]. where it has been observed repeatedly that Markovmodulation increases waiting times and in fact some partial results had been obtained. It was long conjectured that 0* Vi.o.o. we define the stochastic ordering by 0' < s. < . (4. The Markov process {Jt} is stochastically monotone (4.0.o.168 CHAPTER VI. where o*(u) is the ruin probability for the averaged compound Poisson model defined in Section 1 and .. Bp.3) to B = Bi does not depend on i.5) Note that whereas (4..4) To avoid trivialities.o. and finally in part from queueing theory. The conditions which play a role in the following are: ... u > 0. (4. this is not the case for (4.. in part from the folklore principle that any added stochastic variation increases the risk. <s. The motivation that such a result should be true came in part from numerical studies. (4. this correponds to the usual stochastic ordering of the maxima M'. 0"(u) = P(M" > u)) Now consider the risk process in a Markovian environment and define i' (u) _ >iEE irioi(u). V)" if z/i'(u) <'c"" (u).2) alone just amounts to an ordering of the states. [177].1) Obviously. B2 <_s. Occasionally we strengthen (4. MARKOVIAN ENVIRONMENT Notes and references The results and proofs are from Asmussen and Rolski [44].3).2) (4. For the notion of monotone Markov processes. but that in general the picture is more diverse.31:5)32 . ". we also assume that there exist i # j such that either /3i <... The results to be presented show that quite often this is so.. M" of the corresponding two claim surplus proceses (note that 0'(u) _ P(M' > u).3) Bl <_s. we refer to . is the one for the Markovmodulated one in the stationary case (the distribution of J0 is 7r).33 or Bi 0 Bj. 4 Comparisons with the compound Poisson model 4a Ordering of the ruin functions For two risk functions 0'.3p.
note that (4. dx (4.7) 7ri. The first is a standard result going back to Chebycheff and appearing in a more general form in Esary. also Proposition 2.7) and Lemma 4. the second follows from an extension of Theorem I1.. Section 4. < a.9) follows by considering the increasing functions 3iBi (x) and Oi (u .4) hold. Conditions (4.3iBi(x)YPi(u . Comparing (4.x) of i and using Lemma 4. (b) P. b1 < . T(0) < oo) = Bi(x) dx/tcai .1 Assume that conditions (4. E 7r i Wi(u . . (4. 2 If al < . 1:7riaibi > E 7riai i=1 i=1 j=1 The equality holds if and only if a1 = .* For the proof. p)... Theorem 4 . and it is in fact easy to show that Vii(u ) < t/j(u) (this is used in the derivation of (4.2)(4.r (JT(o) = i.9 ) below).6).2.x)B*(x) dx.r (Sr(o) E dx Jr(o) = i.13* J0 u 0*(u ....4) say basically that if i < j ..x)dx _ /3*B*(u) + f u / ^ t=1 > 3 * B* ( ) + f (4.6) 7r= fl*B*(u) + p> s=1 +) fu 0 b (u  x)Bt (x) /pB..9) (4. COMPARISONS WITH THE COMPOUND POISSON MODEL 169 Stoyan [352].1) which with basically the same proof can be found in Asmussen & Schmidt [49]..r(u x)dx. = b.5 (cf.3* f uB(x) z/^. Then V.. Proschan & Walkup [140]..10) Q*B*(u)+. then j is the more risky state .1 for the first term in (4. Lemma 4 . we need two lemmas.. Lemma 4 . 0 Here (4. < bp and 7ri > 0 (i = 1. then P P P 7rjbj..4. Conditioning upon the first ladder epoch. Proposition 2.2.1. 7(0) < oo) = pirf+).6. it follows by a standard .. = aP or b1 = . .3 for the second) *(u) _ /3 *B* (u) +.10) and (4.8) ^j Tri/iBd(x) . ^i 7ri = 1. where 7r2+) = QiµBilri/p. Proof of Theorem 4. we obtain (cf.x) dx u o i =1 i=1 (4. 3 (a) P.2)(4.4) is automatic in some simple examples like birthdeath processes or p = 2 .
8) we get P P '*' (0) = 3* + /3*1* (0) _ > lri'3qqi • E 7i/ipBi ..4 Assume that .1 and Proposition 4. µB2 = 104. Then i/i*(u) < . What is missing in relation to Theorem 4.0*• i=1 But it is intuitively clear (see Theorem 3. dominates the solution 0* to the renewal equation (4. except possibly for a very special situation . MARKOVIAN ENVIRONMENT argument from renewal theory that tk.1 of [145] for a formal proof) that z/ii(u) converges to the ruin probability for the compound Poisson model with parameters . i=1 i=1 7'r(0) _ EFioiwi(0) . Rolski & Schmidt [32]. Frey. of (4. As is seen. of order 101.4 is not vacuous../3*.4) is essential (the present author conjectures it is). (4.(0) < b *'(0) for e small enough.11) is of order 104 and the r. Notes and references The results are from Asmussen.4 is the understanding of whether the stochastic monotonicity condition (4. it will hold for all sufficiently large u.0. Q2 = 1. 0. = 102.h. this ruin probability is /3iPBi.6).. they are at present not quite complete. 01 = 103.2. Recall that .* (0).6). 4b Ordering of adjustment coefficients Despite the fact that V)* (u) < *. u To see that Proposition 4. let = ( 1/2 1/2 ) . For u = 0.11) i=1` and that A has the form eAo for some fixed intensity matrix A0.s. Then the l. it is sufficient to show that 0'. that P P /^ 1r1NiµBi /^2 /^ ^i/ji pBi < 1il3i i=1 i=1 (4..170 CHAPTER VI. µB. u Here is a counterexample showing that the inequality tp* (u) < V).. Using (4. (u) may fail for some u. and from this the claim follows. (u) is not in general true: Proposition 4.(0) = V.3i.s.3µi < 1 for all i.h. Bi as e J. Proof Since 0.r (u ) fails for all sufficiently small e > 0.
It is clear that the distribution of X.13) implies A(a) > 0 for all a.g. Further (see Corollary 11.1) . cf. 0 . (4. Now we can view {Xt} as a cumulative process (see A.14) is nonzero so that A"(0) > 0. (4. with strict inequality unless rci (y*) does not depend on iEE.a = E irirci(a). with strict inequality unless a = 0 or bi = 0 for all i E E. which in view of EiEE 1ibi = 0 is only possible if Si = 0 for all i E E.4(b) that the limit in (4.6 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. The adjustment coefficient y* for the averaged compound Poisson model is the solution > 0 of rc*(ry*) = 0 where rc*(a) _ 13*(B*[a] . 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] .5.12) iEE Theorem 4. (4. Proof Define X= f &ids.ld) with generic cycle w = inf{t>0: Jt_54 k. Lemma 4.5. is nondegenerate unless bi does not depend on i E E.Jt=kI A (the return time of k) where k E E is some arbitrary but fixed state.a. and by Proposition II. This implies that A is strictly convex.. e. Then {(Jt.7) )i is convex with A'(0) = lim EXt tioo t = iEE 70i = 0. Xt)} is a Markov additive process (a socalled Markovian fluid model.g.4. Jt = i])' EE = vA+n(6. in particular .. Hence if 5i 54 0 for some i E E.14) A„(O) iioo varXt t t By convexity.5 y < ry*.1) .(a) > 0 for all a 0 0.5.2 we have (Ei[e"X'. it follows by Proposition A1. Asmussen [20]) as discussed in 11.)a.13) (4.
4c Sensitivity estimates for the adjustment coefficient Now assume that the intensity matrix for the environment is Ae = Ao/ e. h(0) = e. Hence letting e = 0 in (4.Qi and Bi are fixed .e7r)1 (Ici(Y*))diage.5 is from Asmussen & O'Cinneide [40].) and rc (•). The corresponding adjustment coefficient is denoted by ry(e).5. the basic equation is (A + (rci(y))diag)h = 0. Frey. (4. In the case of e. we have 7rh' = 0. y. Thus y(e) * y* as e 10.16) Differentiating (4.. note that y(a) + mins=1.12) and rc*(y*) = 0. If rci(y* ) is not a constant function of i E E. Rolski & Schmidt [32].. Hence rc (y*) > 0. 0 = ((ri(Y))diag + ery (4{('Y))diag)h + (A0 + e(?i'Y))diag)h'.172 CHAPTER VI.15) once more and letting e = 0 we get . a = 1 in Lemma 4. MARKOVIAN ENVIRONMENT Proof of Theorem 4. Further a(1) = rc(y*) by definition of A(. improving upon more incomplete results from Asmussen. this implies that the solution y > 0 of K(y) = 0 must satisfy y < y*. multiply the basic equation by a to obtain 0 = (A0 + e(r£i(y))diag)h.p yi and compute 8y 8a a=0 In both cases. where A. Let bi = rci(y*). h depend on the parameter (e or a). h'(0) = (Ao .15) Normalizing h by 7rh = 0.6. whereas the . Here we put a = 1/e. Notes and references Theorem 4. (4. Since ic is convex with rc'(0) < 0 .. and our aim is to compute the sensitivity ay e a E=O A dual result deals with the limit a 4 oo. we get rc (y*) > 0 which in a similar manner implies that u y < y*.eir)h'(0).15) yields 0 = (Ii(y*)) diage + Aoh'(0) = (rci('Y*)) diage + (Ao . Then > risi = 0 because of (4..
18) 0 = 27'(0)p+27r(rs. Inserting (4.17) (4. We assume that 0 < y < 7i. p.19) holds. Rolski & Schmidt [32]. The analogue of Proposition 4.19) Then 'y ^ ryl as a ^ 0 and we may take h(0) = el (the first unit vector). and we have proved: Proposition 4.. .7 8ry aE = 1 7r(ci ('Y*))diag ( Ao e7r)1(Xi(Y*))diage *=0 P Now turn to the case of a. (4. Frey. (4.8 If (4..i(7' *))diagh'(0). and may have some relevance in risk theory as well (though this still remains to be implemented).16) yields Proposition 4.8 when ryi < 0 for some i is open.18). then 8a a=o All rci (0) Notes and references The results are from Asmussen. i = 2.5. 5 The Markovian arrival process We shall here briefly survey an extension of the model. We get 0 = (aAo + ( lc&Y))diag)h. . multiplying (4. 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.20) and multiplying by el to the left we get 0 = All + 7'(0)rci (0) + 0 (here we used icl (ry(0)) = 0 to infer that the first component of K[7(0)]h'( 0) is 0).20) Letting a = 0 in (4.17) by 7r to the left to get (4. the intensity for such a transition (referred to as marked in the following) is denoted by Aii l and the remaining intensity . (4. THE MARKOVIAN ARRIVAL PROCESS 173 0 = 27'(0)(ri(`Y *)) diage + 2(ci('Y* )) diag h' (0) + Aoh" (0) . which has recently received much attention in the queueing literature. 0 = (Ao + ry'(ii(Y)) diag )h + (aAo + (Ki(7'))diag)h'.
1 (PHASETYPE RENEWAL ARRIVALS) Consider a risk process where the claim sizes are i. refer to notation) { Jt k) }. we may let {Jt} represent the phase processes of the individual interarrival times glued together (see further VIII. with common distribution B.2 for details).4). is neither 0 or 1 is covered by letting Bij have an atom of size qij at 0. Jt2)) (2. The extension of the model can also be motivated via Markov additive processes: if {Nt} is the counting process of a point process. We then let (see the Appendix for the Kronecker E = E(1) x E(2). Note that the case that 0 < qij < 1.6i ) diag. Jt = (Jtl). A ( 2) = A (2`1 ) ® A. where qij is the probability that a transition i * j is accompanied by a claim with distribution.^) etc. A(1) = A . For i = j. Here are some main examples: Example 5 . j(2) } be two independent environmental processes and let E(k). let { Jt 1) }. A(l) = tv. we use the convention that a1i = f3i where 3i is the Poisson rate in state i.174 CHAPTER VI. B. the claim surplus is a Markov additive process (cf. that Bii = Bi .2 (SUPERPOSITIONS) A nice feature of the setup is that it is closed under superposition of independent arrival streams . u Example 5 . Thus . Bij = B. and that are determined by A = A(l ) +A(2) where A is the intensity matrix the governing {Jt}. II. Bii = Bi . A(l) = T.(13i )diag. This is the only way in which arrivals can occur.i. A(1'k) A(2 k1).2). and thus 1i = 0.2) A(1) = A(' 1) ® A(1.d. T). but the point process of arrivals is not Poisson but renewal with interclaim times having common distribution A of phasetype with representation (v. the Markovmodulated compound Poisson model considered sofar corresponds to A(l) = (. then {Nt} is a Markov additive process if and only if it corresponds to an arrival mechanism of the type just considered. Indeed. the definition of Bij is redundant for i i4 j. Again . and the marked transitions are then the ones corresponding to arrivals. . In the above setting. MARKOVIAN ENVIRONMENT f o r a transition i + j by A . the definition of Bi is redundant because of f3i = 0.
11. Example 5 .. superpositions of renewal processes. Hermann [193 ] and Asmussen & Koole [37] showed that in some appropriate . WIDOWED. This means that the environmental states are of the form i1i2 • • • iN with il..g.. and that the policy then expires.4 (A SINGLE LIFE INSURANCE POLICY ) Consider the life insurance of a single policy holder which can be in one of several states. say. claims occur only at state transitions for the environment so that AN2. In fact . i2i . THE MARKOVIAN ARRIVAL PROCESS Bij. The versatility of the setup is even greater than for the Markovmodulated model.. DEAD etc.... u Example 5 . DIVORCED. the kth policy enters a recovering state. However . where ik = 0 means that the kth policy has not yet expired and ik = 1 that it has expired. with rate ai.. E = { WORKING. Similarly. after which it starts afresh. iN. MARRIED.iN C17 AilO. all Al i2.... assume that there is a finite number N of policies.1i2..1i2 . or.. possibly having a general phasetype sojourn time. more recently.kl is redundant).iN. Thus.iN = a2.. INVALIDIZED. The individual pays at rate pi when in state i and receives an amount having distribution Bij when his/her state changes from i to j. Easy modifications apply to allow for • the time until expiration of the kth policy is general phasetype rather than exponential.kj = Bik) B13 4k = Bak) 175  (the definition of the remaining Bij.iil. Assume further that the ith policy leads to a claim having distribution Ci after a time which is exponential.}..5. as the Markovian arrival process ( MAP).iN.iN are zero and all Bi are redundant. u Notes and references The point process of arrivals was studied in detail by Neuts [267] and is often referred to in the queueing literature as Neuts ' versatile point process . e.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). E 10. In this way we can model. the idea of arrivals at transition epochs can be found in Hermann [193] and Rudemo [313]. • upon a claim..iN = C27 All other offdiagonal elements of A are zero so that all other Bii are redundant. Bilo. iN..iil... RETIRED... iN = all BOi2.
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 . where i f00 xB(°) (dx) _ . Obviously. for s E E = [0. we may assume that the functions /3(t). continuity would hold in presumably all reasonable examples. 1). B* = J f B(t) ((*) dt. [248]. p(t) and B(t) are defined also for t t [0.2) Note that p is the average net claim amount per unit time and µ* = p//3* the average mean claim size. from an application point of view. We denote throughout the initial season by s and by P(8) the corresponding governing probability measure for the risk process. Lucantoni et at. we talk of s as the 'time of the year'. p * = 0 p(t) dt. 0 < t < 1. Neuts [271] and Asmussen & Perry [42]. . • Claims arriving at time t of the year have distribution B(t). Some main queueing references using the MAP are Ramaswami [298]. • The premium rate at time t of the year is p(t). For the Markovmodulated model. Without loss of generality. Lucantoni [248]. let the period be 1.1) Then the average arrival rate is /3* and the safety loading rt is 77 = (p* . a claim arrives with rate /3(s + t) and is distributed according to B(8+0 . By periodic extension. 6 Risk theory in a periodic environment 6a The model We assume as in the previous part of the chapter that the arrival mechanism has a certain timeinhomogeneity. one limitation for approximation purposes is the inequality Var Nt > ENt which needs not hold for all arrival streams. one needs to assume also (as a minimum) that they are measurable in t. Sengupta [336].3*µs • p = f /3(v) dv 0 0 (6. 1). Let 1 1 /3* _ f /3(t) dt.p)/p.176 CHAPTER VI. but now exhibiting (deterministic) periodic fluctuations rather than (random ) Markovian ones. )3 t 1 J (6. The basic assumptions are as follows: • The arrival intensity at time t of the year is 3(t) for a certain function /3(t). Thus at time t the premium rate is p(s + t).
the conditional distribution . 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. Section 4b). not random. Many of the results given below indicate that the averaged and the periodic model share a number of main features.w(t). Thus.2 Define T 6(T) = p(t ) dt.1 As an example to be used for numerical illustration throughout this section. the average compound Poisson model is the same as in III.6.w(t)) dt). In contrast. We u assume in the rest of this section that p(t) . The behaviour of the periodic model needs not to be seen as a violation of this principle.3(t) = 3A(1 + sin 27rt). equivalently.1) and Example 1. of the periodic model as arising from the compound Poisson model by adding some extra variability.(3. and we recall from there that the ruin probability is 24 1 *(u) _ 3 5eu + 35e6u. one may think of the standard compound Poisson model with parameters 3*. St = SeI(t). RISK THEORY IN A PERIODIC ENVIRONMENT 177 In a similar manner as in Proposition 1. or. since the added variation is deterministic. 0 Then (by standard operational time arguments ) {St} is a periodic risk process with unit premium rate and the same infinite horizon ruin probabilities.9). In particular. let . It is easily seen that . for Markovmodulated model typically the adjustment coefficient is larger than for the averaged model (cf. Thus . Example 6 . (6.8. B*. In contrast.t. p* as an averaged version of the periodic model.3) Note that A enters just as a scaling factor of the time axis. The claim surplus process {St } two is defined in the obvious way as St = ^N° Ui . The arrival process {Nt}t>0 is a timeinhomogeneous Poisson process with intensity function {/3(s + t)}t>0 . u Remark 6 . in agreement with the general principle of added variation increasing the risk (cf.3* = 3A. the discussion in 111. respectively.10. and thus the averaged standard compound Poisson models have the same risk for all A.1. it turns out that they have the same adjustment coefficient. we shall see that for the periodic model increasing A increases the effect of the periodic fluctuations. p(t) = A and let B(t) be a mixture of two exponential distributions with intensities 3 and 7 and weights w(t) _ (1 +cos27rt)/2 and 1 .
3(s + t)dt[B(8+t)[a] ..f. given that the ith claim occurs at time t is B(8+t). Jt = (s + t) mod 1 P(8) . Dassios & Embrechts [98] and Asmussen & Rolski [43].8). see the Notes to Section 7).^8 [.e.(3(s + t)dt)e«St adt + /3(s + t)dt . 3 E(8)eaSt = h(s. 1).g.. with the underlying Markov process {Jt} being deterministic period motion on E = [0. As usual.al.5. 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. The claim surplus process {St} may be seen as a Markov additive process.4) At a first sight this point of view may appear quite artificial. of the claim surplus process. e. with some variants in the proofs.adt +.g.g. (6.1]) . Daykin et.1) dv .(8) [eaSt+dt I7t] = = (1 .T) = P(8)(r(u) <T). a) is periodic on R.Q(v) (B(„) [a] . we start by deriving formulas giving the m.a) Proof Conditioning upon whether a claim occurs in [t.f. r(u) _ inf It > 0 : St > u} is the time to ruin . 6b Lundberg conjugation Motivated by the discussion in Chapter II. and define h(s. a) etw*(a) h(s+t. i.a) = exp { . t + dt] or not. . 0 (5)(u. [44] (the literature in the mathematical equivalent setting of queueing theory is somewhat more extensive.a . of the averaged compound Poisson model (the last expression is independent of s by periodicity).s . The exposition of the present chapter is basically an extract from [44].tc* (a)] dv then h (. Notes and references The model has been studied in risk theory by.a.1) a = J8 . [101] . we obtain E.1) . let f 8+1 tc *(a) _ (B* [a] .178 CHAPTER VL MARKOVIAN ENVIRONMENT of U.east B(8+t) [a] east . To this end. J Theorem 6 . and the ruin probabilities are 0(8) (U) = P(s )(r(u) < 00).(1 ..5 (see in particular Remark 11.3(v)(B(vl [a] .a be the c.
(e) Let = h( h(Jo.1]. a) Thus E(8)east = h(s + t.. a) h(s + t. a) et.t. we can write Lo Jt.1]) .1)dv l og E(8) et where atetk•(a) h(t.3. 0) exist and are finite. it then suffices to note that E(8)Le. With g the infinitesimal generator of {Xt} = {(Jt. a) Corollary 6. RISK THEORY IN A PERIODIC ENVIRONMENT E(8)east+ dt d Et. St)} . a) = h(s.s. According to Remark 11. a). so that obviously {Lo. 0) P(8)a.1)dv  o h(t.5 The formula for h(s) = h(s.4 For each 0 such that the integrals in the definition of h(t .0(s + t)dt[B(8+t)[a] .2.log h(s.t} is a multiplicative functional for the Markov process { (Jt.adt +.t = 1 by Theorem 6. St)} and .* (a) h(s.(8)east 179 = = = = = E(8)east (1 . + v)(B([a] .6 . a) = exp I f t3(v)(kv)[a) .9) eastt.5. dt log E(8)east a + f3(s + t) [B(8+t) [a] . at + f log h(s + t.3(s + t)[D(8 +t)[a] . Proof In the Markov additive sense of (6. u Remark 6.1]) .6. E (8)east (a +. a) as well as the fact that rc = k` (a) is the correct exponential growth rate of Eeast can be derived via Remark 11. B) eoSt t.t}t>o = h(s. 9) is a P ( 8)martingale with mean one. a) .9 as follows.c* (e) {Le.4). h(s + t.
That rc = is*(a) then follows by noting that h(1) _ u h(0) by periodicity.a . the requirement is cha(i. When a = y. Lemma 6 .180 CHAPTER VI. yo is determined by 0 = k* (70) = QB*.y) = eayh(s).3.7 When a > yo. y solves n* (y) = 0.1) . it follows by Theorem II.3(s)B(s) [a]h(s).5 that we can define a new Markov process {(Jt. P(s) (T(u) < oo) = 1 for all u > 0. J s [. say. 0 < s < 1.0) = Kh(s).'y). we put for short h(s) = h(s. Equating this to rch (s) and dividing by h(s) yields h(s ) = h(s) = a + . B(s). Proposition 6. Now define 'y as the positive solution of the Lundberg equation for the averaged model. of St is as for the asserted periodic risk model.3(s)dt • B(s)[a]h(s) = gha(s. the restrictions of Plsi and Pest to Ft are equivalent with likelihood ratio Le. correspond to a new periodic risk model with parameters ex .g.60(t) = a(t)B(t)[0]. (iv) finally. 0) = h(s) + dt {ah(s) . MARKOVIAN ENVIRONMENT ha(s.f. Bet)(dx) = ^ B(t ) (dx).6 The P(s). (ii) use Markovmodulated approximations (Section 6c).4. ( iii) use approximations with piecewiese constant /3(s).3(s)ks)[a]h(s)} ah(s) 13(s)h(s) + h'(s) +.1. cf.3(v)( Bi"i [a] . A further important constant is the value yo (located in (0.tc] dv} (normalizing by h(0) = 1). For each 0 satisfying the conditions of Corollary 6. see [44] for 11 a formal proof. [70] . . However.2.(3(s)dt) +. That is.6 ( s ) exp { 0( s )&s) [a] + tc .3(s)h(s) + h'(s) +.T. as above E (s) ha(Jdt. Proposition 6. such that for any s and T < oo. That is. Sdt) = h(s + dt) eadt (1 . ry)) at which n* (a) attains its minimum. St)} with governing probability measures Fes). Proof (i) Check that m.
T) = h(s. which is not used elsewhere in the book. a)e«uE (a iP(s) (u) = h( s)e7uE(` ) h(O(u)) To obtain the CramerLundberg approximation from Corollary 3. have components with densities b(8)(x) satisfying inf sEI. RISK THEORY IN A PERIODIC ENVIRONMENT Proof According to (6. and no matter what is the initial season s. a) a > ry0 (6. has a unique stationary distribution. The proof involves machinery from the ergodic theory of Markov chains on a general state space. a) e«uE(8 ) e «^ .2). and we refer to [44]. Here and in the following.9(u))} u>0. the Markov process {(^(u).g.6. we need the following auxiliary result .7) h(B(u). we get: . e(cc)) Letting u > oo in (6.9 Assume that there exist open intervals I C [0. ^(u) = ST(u) . Wu). 9(u)) for any bounded continuous function (e.1. Corollary 6.8 The ruin probabilities can be computed as (u)+T(u)k'(a) ^/i(8) (u. considered with governing probability measures { E(8) }E[ .9) and noting that weak convergence entails convergence of E f (^(u). The relevant likelihood ratio representation of the ruin probabilities now follows immediately from Corollary 11.10) Then for each a. 0(u)) * (b(oo).1) the distribution of (l: (oo).8) (6.2.9) 0(')(u) = h(s. say s0. T(u) < (6. q) = eryx/h(q)). 1). s E I.6(v) dv Jo ' xe«xB (°) (dx) r^ xe«xB'(dx) = Q'B' [ a] = ^' J 0 = ^c"'(a) + 1. a) TI h(9(u).u is the overshoot and 9(u) = (T(u) + s) mod 1 the season at the time of ruin. xEJ 0 (s)b(8)(x) > 0.4. (6. u which is > 1 by convexity. f (x. the mean number of claims per unit time is p« 181 = Jo 1. J C R+ such that the B(8). B(oo)). Lemma 6 .
182 CHAPTER VI.11) Note that ( 6.16.1 In contrast to h. Among other things. which may provide one among many motivations for the Markovmodulated approximation procedure to be considered in Section 6c. Theorem 6 . (6. MARKOVIAN ENVIRONMENT Theorem 6. For our basic Example 6 . where C(o) = 1 + info < t<i h(t) . illustrating that the effect of seasonality increases with A.9). 10 shows that certainly ry is the correct Lundberg exponent. this provides an algorithm for computing C as a limit. A=1/4 A=1 A=4 0 Figure 6.ir) } Plots of h for different values of A are given in Fig.1. elementary calculus yields h(s) = exp { A C 2^ cos 2irs  4^ sin 21rs + 11 cos 41rs . Vi(8) (u) . where e. 11 7/'O (u) < C+°)h(s) ery".10 Under the condition (6.Ch(s)ery".11) gives an interpretation of h(s ) as a measure of how the risks of different initial seasons s vary.1. At this stage .6 for the Markovmodulated model: Theorem 6 . it does not seem within the range of our methods to compute C explicitly. 6.10) of Lemma 3. Noting that ^(u) > 0 in ( 6. 1.) C = E1 h(B(oo)) u + oo.W. we obtain immediately the following version of Lundberg ' s inequality which is a direct parallel of the result given in Corollary 3.
14) < C+)(y)h(s) e7yu.42 so that 183 tp(8) (u) < 1. whereas ay < y.3x + (1 .7x j dx _7x } _ 6w + 6(1 . 1 (6. #c( ay) < 0 when y > 1/tc'('y).7e .(8) (u. yu) 000 (u) .16 In order to apply Theorem 6. we obtain Co) = 1.g. we substitute T = yu in 0(u.0(8) (u+ yu) (6. r.w ) • 7e u{w • 3e3x + ( 1 . Consider first the timedependent version of Lundberg's inequality.w) .4. Just as in IV. (6. e.13) Elementary convexity arguments show that we always have ryy > Y and ay > ry.42 • exp {J_ cos 27rs . where ay is the unique solution of W(ay) =y• (6.47r sin 27rs + 167r cos 47rs .15) The next result improves upon the constant C+) in front of eryu in Theorem 6. in our basic example with A = 1.w)e4u . ay) • (6. Lundberg's inequality can be con siderably sharpened and extended. we first note that the function fu° exu {w • 3e .13 to our basic example.(ay) > 0 when y < 1/ic' (7).11 as well as it supplements with a lower bound. Theorem 6. e7 ( yx)B(t)(dy) > Then for all $ E [0.12) As for the Markovian environment model.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. (ay).(s)(u) < C+h(s)e7".yr.17) (6. RISK THEORY IN A PERIODIC ENVIRONMENT Thus. C_h(s)e7u < V. T) and replace the Lundberg exponent ry by ryy = ay . Theorem 6 ..w)e4u dx 9w + 7(1 . .167r I Cu. 1 ) and all u > 0. We state the results below.6.12 Let 00)(y) 1 Then info < t<i h(t. the proofs are basically the same as in Section 3 and we refer to [44] for details.
013.(8)(u. 1/i18 1 s (u) > 0.181 s(u) < 1.1 sin 2irs + 16_ cos 47rs .1 sin 27rs + 1 cos 47rs . we have the following result: Theorem 6 . .66. Finally.18) Notes and references The material is from Asmussen & Rolski [44]..g.66. MARKOVIAN ENVIRONMENT attains its minimum 2 /3 for u = oo and its maximum 6 /(7 + 2w) for u = 0. Thus C_ = 2 inf ex cos 2irs .cos 27rs . Thus. 14 Let C+('yo) be as in (6. Some of the present proofs are more elementary by avoiding the general point process machinery of [44].013.16) with 'y replaced by yo and h(t) by h(t.19 } 0 <8<1 8 + cos 21rs Thus e. 1). 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..16. C+ = 1.184 CHAPTER VI. with s the initial season. The idea is basically to approximate the (deterministic) continuous clock by a discrete (random) Markovian one with n 'months'. n}. Of course. the nth Markovian environmental process {Jt} moves cyclically on {1. yo). and let 8 = er' (Y0).19 I eu. completing a cycle . 1) for the environment).20). for A = 1 (where 3 e0.4^ sin 2irs + 16^ cos 41rs .\ 3 C+ = sup 6 exp { A (. but thereby also slightly longer. 0 <'p(8)(u ) . This observation motivates to look for a more formal connection between the periodic model and the one evolving in a finite Markovian environment. such a deterministic periodic environment may be seen as a special case of a Markovian one (allowing a continuous state space E = [0. and in fact.I eu.'Yo)e (6.T) < C+('Yo)h( s. 6c Markovmodulated approximations A periodic risk model may be seen as a varying environment model.9 3 0<8<1 p 27r 47r 167r 161r 2 _ _e.0.\ = 0 . Then 7oudT .L sin 27rs + 1 I cos 47rs . where the environment at time t is (s + t) mod 1 E [0.20 •exp { 2n cos 27rs . exp 2^ cos 21rs . . .
jEE of the risk process. . Thus. one simple choice is Oni = 0( i . (6. We let {Stn)} (6. but others are also possible. it is desirable to have formulas permitting freely to translate from one setting into the other. Let 0j. Bi.1 ((i 1)/n) ) and Bni = B .21) which serves as an approximation to 0(1)(u) whenever n is large and i/n s. We want to choose the /3ni and Bni in order to achieve good convergence to the periodic model. 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.20) be the claim surplus process of t>o the nth approximating Markovmodulated model. This queue is commonly denoted as the Markovmodulated M/G/1 queue and has received considerable attention in the last decade. To this end. since the settings are equivalent from a mathematical point of view. Notes and references See Rolski [306]. AE= Aii'r?/7ri• The arrival intensity is /3i when Jt = i. and the ruin probability corresponding to the initial state i of the environment is then Y'yn)(t) = F (M(n) > t). so that the intensity matrix is A(n) given by n n 0 ••• 0 0 n n ••• 0 A(n) _ (6.7. z/'i (u. 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. 7 Dual queueing models The essence of the results of the present section is that the ruin probabilities i/ (u). M(n) = Supt>o Stn).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. DUAL QUEUEING MODELS 185 within one unit of time on the average .
. (7. The actual waiting time process 1W1. JT = i) = 'P. ii (u) = it /3 P(W > u. JT = i) = P(V > u. and for (7. MARKOVIAN ENVIRONMENT • Customers arriving when Jt = i have service time distribution Bi. (7. J* = i) = P. For (7.1) over j.1) follows. 2 .2) Oi(u) = 1.1) 7ri In particular. T) = 7ri 1 P. . and the virtual waiting time (workload) process {Vt}too are defined exactly as for the renewal model in Chapter V.2).3.1 Assume V0 = 0.1). Proposition 7.=1 .2 The relation between the steadystate distributions of the actual and the virtual waiting time distribution is given by F(W > u. J*) is the steadystate limit of (Vt.3) 7ri where (V.0i (u . 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. J* = i). JJ = i). JT = Z). JT = j) = LjPj (VT > u. {Jt }o<t<T• Then we may assume that Jt = JTt. Jo = j.. (VT > u I JT = 2). JT = i} coincide.3).n(VT > u. (7. . The first conclusion of that result then states that the events {T(u) < T. let T . Now let In denote the environment when customer n arrives and I* the steadystate limit. Then Pi(T(u) < T.2) and use that limF (VT > u. JT = j} and {VT > u. just sum (7.T(V > u I J* = i).P(V > u.186 CHAPTER VI. In particular.. I* )3i P(V > u. Proposition 7. JT = j) = 7rjPj(VT > u. Taking probabilities and using the stationarity yields 7riPi(T(u) < T. Proof Consider stationary versions of {Jt}o<t<T. J* = i) for all j. • The queueing discipline is FIFO. Jo = i. Jt ). I* = i).4) where 0* = >jEE 7rj/3j. and (7.oo in u (7. (7.
P(1')(r(u) < oo) = P(')(00) > u). A more probabilistic treatment was given by Asmussen [17].7) of that paper. n=1 N However. and one has PI'>(rr(u) < T) = P('_T)(VT > u). we have 1: I(W. Lemoine [242].8) For treatments of periodic M/G/1 queue.I. P(. p < 1 then ensures that V(*) = limNloo VN+9 exists in distribution.l.1 is from Asmussen [16]. [243]. Taking the ratio yields (7.5) follows from (7. and further references (to which we add Prabhu & Zhu [296]) can be found there. that /3(t). I*) with the timeaverage . The first comprehensive solution of the waiting time problem is Regterschot & de Smit [301]. >u. a general formalism allowing this type of conclusion is 'conditional PASTA'.o. a paper relying heavily on classical complex plane methods.3).4) can be found in Regterschot & de Smit [301]. In the setting of the periodic model of Section 6.=i) a4. With {Vt} denoting the workload process of the periodic queue.6) (7. N * oo. on average /32TP(V > u. see in particular Harrison & Lemoine [186]. and of these. 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. see Regterschot & van Doorn [123].g.4). T]. and (7. B(t) have been periodically extended to negative t). if T is large. DUAL QUEUEING MODELS 187 Proof Identifying the distribution of (W. u Notes and references One of the earliest papers drawing attention to the Markovmodulated M/G/1 queue is Burman & Smith [84]. J* = i) see W > u.. Proposition 7.3) improving somewhat upon (2.4) and (7.7) (7. P(W >u. I* = i. with (7. (7. on average 0*T customers arrive in [0. and Rolski [306].7.I *=i).T)(T(u) <T) = P(8)(VT > u). The relation (7. .
This page is intentionally left blank .
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 . z/i(u) = F IinffRt< 0IRo=u 1 (1. 189 . resp .d.. and that the claim sizes U1.At + p(R8) ds. i&(u.i. Thus.1) (other terms are accumulated claims or total claims). and the evolution of the reserve may be described by the equation Rt = u .T) = F(T(u) < T). {Rt} moves according to the differential equation R = p(R). the premium charged is assumed to depend upon the current reserve Rt so that the premium rate is p(r) when Rt = r. U2. t] are Nt At = Ui (1. Zt As earlier. However . Thus in between jumps.2) tk(u. with common distribution B and independent of {Nt}.. finite horizon. the aggregate claims in [0. are i. .6. and T(u) = inf {t > 0 : Rt < u} is the time to ruin starting from Ro = u so that '(u) = F(T(u) < oo).T) = FloinfTRt< OIRo=u1 denote the ruin probabilities with/initial reserve u and infinite.
Assume 0(u) < 1 for some u. dividends are paid out at rate pi .2 (INTEREST) If the company charges a constant premium rate p u but invests its money at interest rate e. but assume now that the company borrows the deficit in the bank when the reserve goes negative. Another could be the payout of dividends: here the premium paid by the policy holders is the same for all r. the payout rate of interest is Sx and absolute ruin occurs when this exceeds the premium inflow p. No tractable necessary and sufficient condition is known in complete generality of the model. when x > p/S. where one would try to attract new customers as soon as the business has become reasonably safe. Thus at deficit x > 0 (meaning Rt = x). pi > p2 and p(r) = One reason could be competition. Now return to the general model.Vi(v) u > e(1 . rather than when the reserve itself becomes negative. but when the reserve comes above v. that {Rt} will reach level u before the first claim arrives. say e. Example 1. However. A basic question is thus which premium rules p(r) ensure that 'O(u) < 1. oo) is given by i (u + p/S). i.'(u)) > 0 so that V'(v) < 1.e. we can put Rt = Rt + p/S. we get p(r) = p + er. or o(u) < 1 for all u. If Ro = v < u.190 CHAPTER VII. there is positive probability. and the probability of absolute ruin with initial reserve u E [p/S. That is. it seems reasonable to assume monotonicity (p(r) is u . Hence in terms of survival probabilities. say at interest rate b.4 Either i. P(r) _ p + e(r .2.p2. Proposition 1. Proof Obviously '(u) < ilb(v) when u > v.3 (ABSOLUTE RUIN) Consider the same situation as in Example 1.1 Assume that the company reduces the premium rate from pi to p2 when the reserve comes above some critical value v. In this situation. 1 . RESERVEDEPENDENT PREMIUMS The following examples provide some main motivation for studying the model: Example 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.i(u) = 1 for all u. Example 1.
Let Op(u) refer to the compound Poisson model with the same 0. We next recall the following results. hence Rt < uo also for a whole sequence of is converging to oo. cf.6 For any T < oo. Then 0(u) = P(V > u). one can couple the risk process and the storage process on [0. and hence by a geometric trials argument. say V.5 (a) If p(r) < /. then ?(u) = 1 for all u. 296297): Theorem 1.2) for r sufficiently large so that p(oo) = limr.+ p(r) exists. {Vt} decreases at rate p(v) when Vt = v (i. B and (constant) premium rate p. which was proved in 11.1.2(d)). instead of (1. .o(uo) = 1 so that t/'(u) = 1 for all u by Proposition 1. Then if u > no. V = p(V)). [APQ] pp. the probability that Rt < uo for some t is at least tp(0) = 1 (cf. T] i n such a way that the events {r(u) <T} and {VT > u} coincide. we have z/i(u) <p(u .1. appealing to Proposition 111. if and only if V)(u) < 1 for all u. let uo be chosen such that p(r) > p = 0ILB + e for r > uo. (1. let uo be chosen such that p(r) < p = /3µB for r > uo.e.2 once more. In particular. Here {Vt}two is a storage process which has reflection at zero and initial condition Vo = 0.T) = P(VT > u).1.2) we have t Vt = At .f p(Vs) ds.I3IB requires a more detailed analysis and that µB < oo is not always necessary for O(u) < 1 when p(r) 4 oo. Starting from Ro = uo.4.4) 0 and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0. That is. and P(Rt + oo) > 0. obviously infu<uo z/'(u) > 0. Proposition I1I.. (1.3µB for all sufficiently large r.1 and increasing in Example 1. This is basically covered by the following result (but note that the case p(r) . (1. However. In case (b). (b) If p(r) > /3µB + e for all sufficiently large r and some e > 0. Theorem 1. INTRODUCTION 191 decreasing in Example 1. { Vt} remains at 0 until the next arrival). then l/i(u) < 1 for all u.2(d).b(u.3. that u zPp(u .1.uo) < 1. Hence ik(u) < 1 for all u by Proposition III.5) and the process {Vt} has a proper limit in distribution . In case (a). Proof This follows by a simple comparison with the compound Poisson model.uo) and.6) . In between jumps.
s. B(x) = e. say. u Define ^x 1 w(x) Jo p(t) dt. say. In view of the path structure of {V t }. of (1. this means that the rate of upcrossings of level x must be the same as the rate of downcrossings. x + p(x)dt]). Oeax f x e'Yg (y) dy } = p) eaxa(x) . x] to (x. Jo AX) (1. say when {Vt} is in state y. we thus need to look more into the stationary distribution G. It is intuitively obvious and not too hard to prove that G is a mixture of two components.Sx} dx. 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. An attempt of an upcrossing occurs as result of an arrival. Considering the cases y = 0 and 0 < y < x separately. oo) must be the same as the flow the other way.7 p(x)g(x) = tofB (x) + a f (x .h. yo ^ 1 + oo Q exp {. and the other being given by a density g(x) on (0.6x and that w(x) < oo for all x > 0.9) Proof We may rewrite (1. (1. t + dt] if and only if Vt E [x. and is succesful if the jump size is larger than x .8) is the rate of downcrossings (the event of an arrival in [t.y)g(y) dy.192 CHAPTER VII.8) as the rate of upcrossings. we arrive at the desired interpretation of the r.6 applicable. say if p(r) goes to 0 at rate 1 /r or faster as r j 0. for the storage process {Vt}.8) as g(x) = p 1 {yo13e_6x +.Sx}.7) Proposition 1.Qw(x) . of (1. the flow of mass from [0.s. RESERVEDEPENDENT PREMIUMS In order to make Theorem 1.8 Assume that B is exponential with rate b.8) Proof In stationarity. Note that it may happen that w (x) = oo for all x > 0. Corollary 1. Then the ruin probability is tp (u) = f' g(y)dy. It follows in particular that 0(u) = fg(Y)dy. one having an atom at 0 of size 'yo. t + dt] can be neglected so that a path of {Vt} corresponds to a downcrossing in [t.y. where g(x) = p( ^ exp {. the l. (1.h. Now obviously. oo).6w(x) .
1. INTRODUCTION
where c(x) = 1o + fo elyg(y) dy so that (x) = eaxg(x) _
193
1
p(x)
nkx).
Thus log rc(x) = log rc(0) + Jo X L dt = log rc(0) + /3w(x), p(t) c(x) = rc (0)em"lxl = Yoes"lxl, g(x) = eaxK' (x) = e6x ,Yo)3w'(x)e'6"lxl which is the same as the expression in (1.9). That 'Yo has the asserted value is u a consequence of 1 = I I G I I = yo + f g• Remark 1.9 The exponential case in Corollary 1.8 is the only one in which explicit formulas are known (or almost so; see further the notes to Section 2), and thus it becomes important to develop algorithms for computing the ruin probabilities. We next outline one possible approach based upon the integral equation (1.8) (another one is based upon numerical solution of a system of differential equations which can be derived under phasetype assumptions, see further VIII.7). A Volterra integral equation has the general form x g(x) = h(x) + f K(x, y)9(y) dy, 0 (1.10)
where g(x) is an unknown function (x > 0), h(x) is known and K(x,y) is a suitable kernel. Dividing (1.8) by p(x) and letting K(x, y) _ ,QB(x  y) _ 'YoIB(x) p(x) , h(x) p(x) we see that for fixed to, the function g(x) in (1.8) satisfies (1.10). For the purpose of explicit computation of g(x) (and thereby %(u)), the general theory of Volterra equations does not seem to lead beyond the exponential case already treated in Corollary 1.8. However, one might try instead a numerical solution. We consider the simplest possible approach based upon the most basic numerical integration procedure, the trapezoidal rule hfxN() dx = 2 [f ( xo) + 2f (xi) + 2f ( x2) + ... + 2f (XN1) + f (xN)1
p
194
CHAPTER VII. RESERVEDEPENDENT PREMIUMS
where xk = x0 + kh. Fixing h > 0, letting x0 = 0 (i.e. xk = kh) and writing 9k = 9(xk ), Kk,e = K(xk, xe), this leads to h 9N = hN + 2 {KN,09o+KN,N9N}+h{KN,191+'''+KN,N19N1},
i.e. 9 N=
hN+ ZKN ,ogo +h{KN,lgl+•••+KN,N19N1} 1  ZKNN
(
1.11
)
In the case of (1.8), the unknown yo is involved. However, (1.11) is easily seen to be linear in yo. One therefore first makes a trial solution g*(x) corresponding to yo = 1, i.e. h(x) = h*(x) = (3B(x)/p(x), and computes f o' g*(x)dx numerically (by truncation and using the gk). Then g(x) = yog*(x), and IIGII = 1 then yields f 00 g*(x)dx (1.12) 1= 1+ 'Yo from which yo and hence g(x) and z/'(u) can be computed. u
la Twostep premium functions
We now assume the premium function to be constant in two levels as in Example 1.1, p(r) _ J 1'1 r < v P2 r > v. (1.13)
We may think of the risk reserve process Rt as pieced together of two risk reserve processes R' and Rt with constant premiums p1, P2, such that Rt coincide with Rt under level v and with above level v. For an example of a sample path, Rt see Fig. 1.1.
Rt
V
Figure 1.1
1. INTRODUCTION
195
Proposition 1.10 Let V)' (u) denote the ruin probability of {Rt}, define a = inf It > 0 : Rt < v}, let pi ( u) be the probability of ruin between a and the next upcrossing of v (including ruin possibly at a), and let q(u) = 1  V" (u) Then
1  q(u) + q ( u)z,b(v) p1(v) u = 0<u<v v
0 < u < v. (1.14)
1 + pi (v )  '02 (0) pi (u) + (0, (u  v)  pi (u)) z/i(v ) v < u < oo.
Proof Let w = inf{ t > 0 1 Rt= v or Rt < 0} and let Q1 (u) = Pu(RC,, = v) be the probability of upcrossing level v before ruin given the process starts at u < v. If we for a moment consider the process under level v, Rt , only, we get Vil (u ) = 1  q, (u ) + g1(u),O1( v). Solving for ql (u), it follows that q1 (u) = q(u). With this interpretation of q(u) is follows that if u < v then the probability of ruin will be the sum of the probability of being ruined before upcrossing v, 1  q(u), and the probability of ruin given we hit v first , q(u)z'(v). Similarly, if u > v then the probability of ruin is the sum of being ruined between a and the next upcrossing of v which is pl (u), and the probability of ruin given the process hits v before ( oo, 0) again after a, (Pu(a < oo )  p1(u))''(v) = (Vi2(u  v)  p1 (u))''(v)• This yields the expression for u > v, and the one for u = v then immediately follows. u Example 1 .11 Assume that B is exponential, B(x) = e62. Then
01 (u)
_
0 e .yiu ,,2 (u) = )3 e 72u p1S P2S
1  ~ ery1u p1S 1  Q eryly P1S
where ry; = S  ,Q/p;, so that
q

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

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

u + Z =
(u + Zr ) + ( Z  Zr) = e
ET {e
(u + Zr)  f '* eE(tT )dSt] T
e
ET [
R( u)
+ Z`],
2. THE MODEL WITH INTEREST
199
where Z* =  K* eE(tT)dSt is independent of F, and distributed as Z. The last equality followed from Rt") = eEt(Zt + u), cf. Proposition 2.1, which also yields r < oo on {Z < u}. Hence H(u) = P(u + Z < 0) = P(RT + Z* < 0; r < oo) zb(u)E [P(RT + Z* < 0 I)7T, r < oo)] _ O(u)E [H(RT(")) I r(u) < oo] .
Corollary 2.4 Assume that B is exponential, B(x) = e6', and that p(x) _ p + Ex with p > 0. Then
. o€Q/E Ir, (8(p + cu);
V) (u)
aA/Epal Ee 6n1 E +^3E1 / E
1\ E E
1r
Cbp;
E El al
where 1'(x; i) = f 2°° tnletdt is the incomplete Gamma function. Proof 1 We use Corollary 1.8 and get
w(x) fo P + Etdt = g(x) = p +0x
e log(p + Ex)  e loge,
exp {  log(p + Ex)   log p  6x }
pal(p + ex)plE1e6^ J ryo)3 70 = 1 + J p) exp {Ow(x)  Sx} dx x r^ = 1+ ' /E (p + Ex)01'leax dx + 0
f J
= 1+
a
Epo/ E
f yI/ E 1e 6(Y P)/E dy
P (
1+ OEA/E 1e6 P /Er
60/e po/ e
,;,3 )
E E
lp(u) = to foo a exp {w(x)  bx} AX)
acO/E" 1 ePE l
Yo
50 1epolE
(
+ cu); 0)
5(p
E E
r (j3/E) In particular.3/E) By the memoryless property of the exponential distribution. The process {St} corresponds to {Wt} so that c(a) or2a2/2 .a) . then {Rt} is the diffusion with drift function p+Ex and constant variance a2. of Z is IogH[a] = f ytc(y)dy = e fa (0. RESERVEDEPENDENT PREMIUMS u from which (2.200 CHAPTER VII. From ic(a) = .2y +µ ) dy .x) dx e.pa.pa. where V is Gamma(b. assume that {Wt} is Brownian motion with drift µ and variance v2.3 is also valid if {Rt} is obtained by adding interest to a more general process {Wt} with stationary independent increments. 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 .01'E) + (p/E)al aO l febP/E } IF (0 /0 jF From this (2.e. /^ u Example 2 . i.3a/ (5 .3.2) follows by elementary algebra. As an example. with density x(3/e1aQ/e fV (x) _ e 6X ' x > 0. RT(u) has an exponential distribution with rate (S) and hence E [H(RT(u))I r(u) < oo] L Pe6'r (P/C . Proof 2 We use Theorem 2.b P/E dx /' P/ ' (p/  x)p/e 150/f I' (/3/E) (6P1'E.g.5 The analysis leading to Theorem 2./3 log(b + a)] = log ePa/f (a + a ) e which shows that Z is distributed as p/E . H(u) = P(Z r < u) = P(V > u + p/E) = (8(p + Eu)/E.f.V.V < x)]0 + f P(V > p/E ) + eby fv (p/E . . 13/E).13 /E) r (.2) follows by elementary algebra. and the c.
[129]. Q2/2E). that the analysis does not seem to carry over to general phasetype distributions. write y* for the solution of the Lundberg equation f3(B[ry *] . [282]. and recall Lundberg 's inequality . A r. of the form Ei° p"X" with the X„ i. or to nonlinear premium rules p(•).1) .. Logarithmic asymptotics For the classical risk model with constant premium rule p(x) . it is also used as basis for a diffusion approximation by these authors. Z is normal (p/E. 134 (the time scale there is discrete but the argument is easily adapted to the continuous case).d. Some of these references also go into a stochastic interest rate. however.Y*p* W*(u) < ery*u = 0.v. Paulsen & Gjessing [286] and Sundt & Teugels [356].g. Goldie & Griibel [167].2 is a special case of a perpetuity.i. Corollary 2. see e. The solution is in terms of Bessel functions for an Erlang(2) B and in terms of confluent hypergeometric functions for a H2 B (a mixture of two exponentials). Paulsen [281]. se e.3 is from Harrison [185].g.3. Delbaen & Haezendonck [104]. not even Erlang(3) or H3. Gerber [155]. 3 The local adjustment coefficient.8. It must be noted.p*. write Vi* (u) for the ruin probability etc. as in the proof of Proposition 2. Emanuel et at. [129] and Harrison [185]. Paulsen & Gjessing [286] found some remarkable explicit formulas for 0(u) beyond the exponential case in Corollary 1. Gerber [157] p. The formula (2.3) was derived by Emanuel et at. Further studies of the model with interest can be found in Boogaert & Crijns [71]. THE LOCAL ADJUSTMENT COEFFICIENT _ Q2a2 pa 4e E 201 I.4 is classical.. [357]. it follows that the ruin probability is Cu) H(u) H(0) 11 Notes and references Theorem 2. [283]. and since RT = 0 by the continuity of Brownian motion.e. for a martingale proof.
1. it holds that f3[s] T oo. i. as solution of the equation n(x.5) which implies inf. (3.e.1.3) When trying to extend these results to the model of this chapter where p(x) depends on x. x * oo. we will use the local adjustment coefficient 'y(x). B(x) > C(2)e(ao+f)x for all x. For the last asssertion . which in turn by Lundberg's inequality can be bounded by ery*(1E)" Hence limsup„.E). and that p(x) * oo. oo for all E > 0.ap(x). The steepness assumption and p(x) + oo ensure 'y(x) * So.. Proof of Theorem 3. When u > uo. c(.4) we assume existence of y(x) for all x.1 ). a first step is the following: Theorem 3 . 1) and for a given E > 0. Then lim sup u>oo u and e E''p(r) + 0.C*ef*". let p* be a in (3.w (u) J dt > c(3)eeu v 1 p(u+ t) .2) such that p(x) < c(. choose uo such that p( x) > p* when x > u0E.>o 7(x) > 0. 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.'y ( x)) = 0 where r. a) = f3(B[a] . e(1o+e)2 (x ) u > 00.1) .i)eex. log ?i(u) < < 00 JO . i. as will hold under the steepness assumption of Theorem 3.1 Assume that for some 0 < 5o < oo. Then we have the following lower bound for the time for the reserve to go from level u to level u + v without a claim: w(u + v) . the 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). (x. then log u (u) In the proof as well as in the remaining part of the section . If 60 s f 6o.e.. x>0 (3.202 CHAPTER VII.log '(u)/u < ry*(1 . RESERVEDEPENDENT PREMIUMS and the CramerLundberg approximation V.*(u) . choose c(. 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 . Letting first E * 0 and next ry * T 5o yields the first statement of the theorem. Let y* < So. (3. and (for simplicity) that inf p(x) > (3µs .
The first main result in this direction is the following version of Lundberg's inequality: Theorem 3 .7) CIO Remarks: 1.3 Assume that either (a) p(r) is a non decreasing function of r. Bucklew [81]). which essentially says that an overshoot r. Condition 3.3.6) The second main result to be derived states that the bound in Theorem 3.4)eE" Given such an arrival.2 is also an approximation under appropriate conditions. (3. 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. and in particular. 3) = (1 . Theorem 3 .g.ea°/(ecf1)). the result is not very informative if bo = oo. by cU2. ruin will occur if the claim is at least u + v. Theorem 3. If p(x) = pis constant . 3. I. For e > 0. 2 Assume that p(x) is a nondecreasing function of x and let I(u) = fo ry(x)dx. the limit is not u + oo but the slow Markov walk limit in large deviations theory (see e. (u) = O(u/e). 2.1 only presents a first step. . The rest of this section deals with tail estimates involving the local adjustment coefficient. THE LOCAL ADJUSTMENT COEFFICIENT 203 where c. Therefore the probability that a claim arrives ( before the reserve has reached level u + v is at least c(.0 are the same. one can then assume that e = 1 is small enough for Theorem 3.13 is a technical condition on the claim size distribution B.' (u) < eI(").v. the asymptotics u * oo and c .(u) > so.e. noting that in many cases the constant C is close to 1. u Obviously. then Rte) = CRtie for all t so that V). let 0e (u) be evaluated for the process only with 3 replaced by /0/e and U. or (b) Condition 3. The form of the result is superficially similar to the CramerLundberg approximation. Then limelog l/ie (u) = I(u).2). UJU > x cannot have a much heavier tail than the claim U itself..3 to be reasonably precise and use e` (u) as approximation to 0 (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. (3.13 below holds. However. Then .
bx} dx fo 00 1 + [exp {/(3w(x) .bx} dx = 1 + J0 dodx(x) exp {.bu}. J0 ^oo g(x ) dx f AX) lexp IOW (X ) bx + b u 1 exp low(x) . we show how to rewrite the explicit solution for ti(u) in Corollary 1. As typical in large deviations theory.(iw(x) . rather than eI(u)). say. However.1 + b f e.bx} dx . we get = 1+ J" AX) exp {(3w (x) . u . First.4 Consider again the exponential case B(x) = eax as in Corollary 1.(3/p(x). One would expect the behaviour in 2) to be important for the quantitative performance of the Lundberg inequality (3. the logaritmic form of (3.3. and r j 1 'Yo v(x)dx = bu  a J0 p(x)ldx = Integrating by parts. 3a Examples Before giving the proofs of Theorems 3.bx} dx 1+0.7) 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). it is formally needed only for Theorem 3.(x) dx.3.bx} dx oo exp low(x) bx dx 70 Ju r oo = b J exp low (x) .204 CHAPTER VII.6). 5.exp {/33w(u) .2. 3.8. RESERVEDEPENDENT PREMIUMS 4.8 in terms of I(u) when the claims are exponential: Example 3 . we consider some simple examples.bx}]o + b /' oo exp low (x) . Then y(x) = b .
.fory(x+u)dxdy ( 3. 0.9) yields e log .I ( v )dy fo +u) dxdy . u .7) follows. THE LOCAL ADJUSTMENT COEFFICIENT and hence 205 f°° eI(v )dy . It is well known that (see Theorem XI. applying the inequality 7(x + u) > 7(x) yields immediately the conclusion of Theorem 3. BE * 0. (3.2. Be = e log U000 e.2(X) = ev2(x) so that 7e(x) = 7(x)/e.10 or Karlin & Taylor [222] pp.3. and (3. (3. (X) = µ(x). in the definition of AE converges to 0. Choosing yo.fa 7(x+u)dx/Edy o The analogue of (3. note first that the appropriate slow Markov walk assumption amounts to u.e. Similarly. and (3.0. ry(x /b I u o e f0 °° e  e. In particular. For Theorem 3.5) is infx>o 7(x) > 0 which implies that f °O . > lime log e = 0 and AE * 0.BE.ev 0 O /E) J0 70 70 Yo This implies lim inf A.5 Assume that {Rt} is a diffusion on [0. The appropriate definition of the local adjustment coefficient 7(x) is then as the one 2p(x)la2(x) for the locally approximating Brownian motion.3 in the particularly simple case of diffusions: Example 3.2.10) where AE = e log 000 e. the integral is bounded by 1 eventually and hence lim sup AE < lim sup a log 1 = 0. 70 > 0 such that 7(x) < 7o for y < yo.9 ) 11000 eI(v)dy f000 e.. we get r 00 e. IE(u) = I(u)/e.1.8) 7(x)dxdy 11 We next give a direct derivations of Theorems 3.fo 7(x)dx/Edy f .f y(x)dxd y If 7(x) is increasing .3. 3. (u) = I(u) + AE . oo) with drift µ(x) and variance a2 (x) > 0 at x.I ( u) fool.1/8 . 191195) that 1P (U) = fu0 eI(v)dy = eI(u) follo e.fo 7(x) dx /E dy > av 'yo /Edy = E (1 .
6/p* so that u 1 I (U) = bu . I(u ) = G1(u) + .. G.7o C 15 I I.6 Assume that B is exponential with rate S.0) = 0.1 3. Further. Ignoring 1/5 in the formula there. the results are suggestive in their form and much more explicit than anything else in the literature.5 for risk processes with exponential claims is as follows: Example 3 .(u) oo.. Of course.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./3 1 AX dx.6) exactly as in Example 3.. Thus 7e(x) _7(x)/e and (3.e. . that the interchange of the slow Markov walk oo is not justified and in fact.10) holds if we redefine AE as AE = flog (j °° efo 7(x)dx/edy _ E/5 I and similarly for B. As in Example 3. RESERVEDEPENDENT PREMIUMS The analogue of Example 3.5.0/e.5) and 7* = 5 . (u) representing the first few terms in the asymptotic expansion of I(u) as u + oo. .Q/p*. G.206 CHAPTER VII. lim sup Af < lim sup c log(1 . 0. . _ . the slow Markov limit a * 0 and the limit u walk approximation deteriorates as x becomes large. the slow Markov walk assumption means 5E = b/c.4. however .. Then the solution of the Lundberg equation is y* = b .+1 (u) = o( 1). . + Gq(u) + o(G9(u))• Gi (u) It should be noted . ) Note that this expression shows up also in the explicit formula for lk(u) in the form given in Example 3.5. E+o e*O By (3.5. Nevertheless . . 7(x) is typically not explicit. so our approach is to determine standard functions Gl (u).0. I. G. we have 5 > 7o and get lim inf AE > lime log e . 0 Now (3.
ry* loge*+ g7loglogp*.c3 logu a= 1 J 0 a + bx 1/ ( c4ul 1/° a > 1 where c3 = c2 /b.1 =$ f cse8 Sn f e"B(x)dx = e8 Jo s eIB ( 1 .8). More precisely.:. u Example 3 . phasetype distributions (Example 1. the typical case is a = 1 which holds . Here B[s] is defined for all s and B[s] .11) that b[s] * co as s f S and hence 7* T S as p* + oo.C2p* C2 = (3clr( a))11'. 1. say 1 is the upper limit and B(x) .Y .12) with y > 1. B[s] = 1 + s exB(x)dx = 1 +c1SF(a) ('+o(')) (S .x)n1.g. 2. c4 = c2b 1/'/(1 . Hence (3. THE LOCAL ADJUSTMENT COEFFICIENT Example 3 .8 Assume next that B has bounded support. and hence (3.clxale5x 207 (3.y/s)dy sn 1 1 f ' evy'7ldy = cse8r(T7) as s T oc.3cse7*I7(77) . (3. 1) and 17 = k + 1 if B is the convolution of k uniforms on (0.11) with a > 0.1) leads to . if the phase generator is irreducible ( Proposition VIII. It follows from (3.s)C' f "o o as s T S.. . I(u) Pt.1/k). u(logu + r7loglogu). more generally. fu I(u) Su .1) leads to (S7T N Ocp a. x T 1. . 77 = 1 if B is degenerate at 1. This covers mixtures or convolutions of exponentials or. e.7 Assume that B(x) .cs(1 . in the phasetype case .1/a).c2 Su a dx ) Su a<1 Su .3. For example.ry*°p*.4) or gamma distributions. y = 2 if B is uniform on (0.
u .15) Proposition 3.Ul up to the first claim (here ru (•) denotes the solution of i = p (r) starting from ru(0) = u). RESERVEDEPENDENT PREMIUMS Example 3 .3 (B[s] . h].1 Cgs o"O 0 esxex2/2c7 dx = cgsec782/2 f .B[7o (u)] .9 As a case intermediate between (3.(t))dt. 1 = E.css 2%rc7eC782/2. of U1 + v . Hence for u<V. Then: (a) y(x) and 7o(x) are also nondecreasing functions of x.f.2 We first remark that the definition (3. 7 * .(T1)) > Ee7o(u)(ul+vr»(Ti)). h 10.c8 log .g. ec78)2/2c7 dx C7 .u is a nondecreasing function of u.4) is the formula h logEues ( Rhu) ..12). (b) 'y(x) <'Yo(x)• Proof That 7(x) is nondecreasing follows easily by inspection of (3.•. By convexity of the m . 1 0 3e. The assumption implies that ru(t) .4) of the local adjustment coefficient is not the only possible one: whereas the motivation for (3.e7o ( u)(ul+u r. of the increment in a small time interval [0.log p*. (3.13) We get b[s] . (3. x f oo .f.r^. This leads to an alternative local adjustment coefficient 7o(u) defined as solution of 1 = Ee''o(u)(vi+u .10 Assume that p(x) is a nondecreasing function of x.208 CHAPTER VII. g. (3. I (u) c8u log u 0 where c8 = 2/c7. this is only possible if 7o(v) 2 7o(u)• .Ote7o( u)(u.ru(TI)) .14) for the m .11) and (3.1) .4). .sp(u). one could also have considered the increment ru (T1) . 3b Proof of Theorem 3. assume that B(x) CO x2/2c7.r„(Ti).
u We prove Theorem 3.7o (u)p(u)• Since (3.11 Assume that p(x) is a nondecreasing function of x.1) .10(b): Theorem 3.17) from which the theorem follows by letting n + oo. the case of 7 then follows immediately by Proposition 3.u > tp(u). Separating after whether ruin occurs at the first claim or not.17) shown for n and let Fu(x) = P(U1 + u .u[70(u)] fo eyo(x)dx .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. (3. We shall show by induction that (' Y'(n) (u) < e fo 'Yo(x)dx (3. Hence „/. this is only possible if yo(u) > 7(u).4) considered as function of 7 is convex and 0 for y = 0.(n+l) (u) 1 .e70(u)(U1P(u)T1) 209 0 + 7o(u)p(u)' 0 <_ 00['Yo( u)] . The case n = 0 is clear since here To = 0 so that ik(°)(u) = 0.3. we obtain „I. Also. Assume (3.Fu(u ) + J ^(n)(u . fa 7o(y)dy < u7o(u) < xyo (u) for x > u. note that the assumption implies that ru(t) .2 in terms of 7o. it is easily seen that fu x7o(y)dy < xyo (u).(n+1) (u) efo Yo(x)dxI^"Q exyo( I u u)Fu(dx )+ J .16) Proof Define 411(n)(u) = P('r(u) < on) as the ruin probability after at most n claims (on = TI + • • • + Tn). Hence 1 = EeYo(u)(U1+uru(T1)) < E.ru(T1) < x).es'Yo(u)Fu(dx)} o0 e fo yo( x)dx j. THE LOCAL ADJUSTMENT COEFFICIENT For (b). Then (u) < efo Yo(x)dx.
11 be reasonably tight something like the slow Markov walk conditions in Theorem 3. 3. by €U=.n. pk n = uk_l..n) > k =1 II v ^k n. given downcrossing occurs. Let Ck.n.E (u/n). However. in . we have chosen to work with y(u) as the fundamental local adjustment coefficient. C*e where the first equality follows by an easy scaling argument and the approximation by (3.. (u) < I(u). 0 It follows from Proposition 3.n <Z auk}l.n AX). let Op*.3 The idea of the proof is to bound { R( f) } above and below in a small interval [x .E (u/ n) Y'E (un .. Also... Lemma 3. For these reasons. for either of Theorems 3. the value of {R(E)} at the time of downcrossing is < unl. yo(u) appears more difficult to evaluate than y(u). W O .: y(u). in accordance with the notation i/iE (u).3).n = sup p(x).E (u/n) Now as e .. 0.E (u) denote the ruin probability for the classical model with 0 replaced by . resp.2).10(b ) that the bound provided by Theorem 3.11 is sharper than the one given by Theorem 3. To this end.2.12 lim sup4^o f log O. and. we used also Proposition 3.210 CHAPTER VII.n) pn niE (u /n) n n_1 n. 0. 3c Proof of Theorem 3.n so that n.n u k}1.10(a) for some of the inequalities.nbe C*. RESERVEDEPENDENT PREMIUMS where the last identity immediately follows from (3.3).3 is required. y* evaluated for p* = Pk. op*.3/e and U.15). P k. and here it is easily seen that yo(u) . Y*u /E. (3. (u). x + x/n] by two classical risk processes with a constant p and appeal to the classical results (3. (un2.n. Proof For ruin to occur.n (starting from u/n) without that 2u/n is upcrossed before ruin.I.n = ku. {RtE)} (starting from u = un.n inf n uk1. define uk. the probability that ruin occurs in the CramerLundberg model with p* = pn.x/n.2.E ( u/n) ^•e.e (u) = v'.n) must first downcross unl. Further. ryk.n. The probability of this is at least n n.
y > 0 it holds that F(U>x +yIU>x) B(x + y) < F (V > y).E (u/n) OP +^p•.19) . /' (u/n) 'T nk. 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.nu/en(1 + where o(1) refers to the limit e . 0 with n and u fixed. THE LOCAL ADJUSTMENT COEFFICIENT particular. B(x) (3... Indeed .E (u/n) Op•. ne7k. so that Theorem 3. 40 Combining with the upper bound of Lemma 3. v.2 gives 7PE (u) < eIi"i/f = lim inf Clog 0E (u) > I (u).F (2u/n).7k. in obvious notation one has tC (x) = y(x)/e. 211 Clearly.E (urn) < \ *I.i.n <X<Uk. It follows that n log V'C (u) k =1 log Ypk. (3. 11 Theorem 3.n + 0(1). 3 now follows easily in case (a).n cE (2u/n) Ck ne7k. i. k=1 k=1 n u _ nE7 k. .18) (ii) the family of claim overshoot distributions is stochastically dominated by V.e.3. *p•. since ry' is an increasing function of p'. In case (b).n.nu /En) o(1)). we need the following condition: Condition 3.nk=1 limsupelogv).. (u) CIO < Letting n 4 oo and using a Riemann sum approximation completes the proof.! (u/n) n n m 7k.n . for all x .a( u)z.12 completes the proof.log Ck.n = sup ?'(x)..nu /fn( 1 Ck  e. uk_1. also ryk.13 There exists a r.
.R<) (u v). For E2.QEU 1 . we first note that the number of downcrossings of 2u/n starting from RoE) = 2u/n is bounded by a geometric r.1 n. RESERVEDEPENDENT PREMIUMS To complete the proof.EV) = EiI 1 . Ei + E2 < e71.E (2u/n . v ) = inf { t > 0 : R(c ) < v R) = u } . let v < u and define T(E) (u. (u/n . T() (u. u/n) < oo] E [OE (u/n . u /n) < oo] l = = < E [OE (u/n . u/n) < oo] .E (u/n .nu /EnE [e71.E (0) cf. Then Y'E (u) ^(E) (u.( . ) (u u /n)) . u/n) < oo) EV).^(E) (u.EV) = El + E2. Then the standard Lundberg inequality yields El < E?.E(E) (u.eV) • P (T(E) (u.2y 1 ' . Write EO. u/n)) .of:>2 in n(x). where El is the contribution from the event that the process does not reach level 2u/n before ruin and E2 is the rest. u/n)) I T(E) (u. N with EN < 1 = infx>2u/nA(x) = 0(1). u /en 0(i) _n so that E2 < e2ryl nu/En0(1). ..n < ery1.nu/En0(1) . v ) = v . u/n) < oo) ..V) = e71 nu/Eno(l) (using (3.^'' = E [ . (R. (u/n .EV) = e.v. infx>2u /n P(x) ..212 CHAPTER VII.5) and the standard formula for b(0). P (T(E) (u. V < u/En] + P(V > u/En) (u/En . (3.18) for the last equality). The probability of ruin in between two downcrossings is bounded by Epp .n V. T(E) (u.
7) for ruin probabilities in the presence of an upper barrier b appears in Cottrell et al. (u) 40 213 lim inf e log(Ei +E2) + logP (r(`) (u. Similarly. s). u Notes and references With the exception of Theorem 3. Comparing these references with the present work shows that in the slow Markov walk setup.J y(Rs)dR.21) to pass from u to 0. . one can in fact arrive at the optimal path by showing that the approximation for 0(u. the risk process itself is close to the solution of the differential equation r(x) _ r (x. u/n) < oo { 40 )I U nryl n+liminfelogP (T(')(u. THE LOCAL ADJUSTMENT COEFFICIENT Hence lim inf e log Ali. the approximation (3. the results are from Asmussen & Nielsen [39]. [89].1. where the key mathematical tool is the deep WentzellFreidlin theory of slow Markov walks (see e . whereas the most probable path leading to ruin is the solution of r(x) _ k (x.20) (with ic(x.g.13. they also discuss simulation based upon 'local exponential change of measure' for which the likelihood ratio is ( /'t /'t Ns Lt = exp S .T) = P „(info<t <T Rt < 0) via related large deviations techniques.t.3EU) (3. Djehiche [122] gives an approximation for tp(u. 0 and b T 00 are interchangeable in the setting of [89]. s) as in (3.7) then comes out (at least heuristically) by analytical manipulations with the action integral. Typically.4) and the prime meaning differentiation w. 0 ) (= p(x) . it might be possible to show that the limits e . Whereas the result of [122] is given in terms of an action integral which does not look very explicit. Bucklew [81]). 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 3.3. l o JJJ o .7(x)) (3.r. T) is maximized over T by taking T as the time for (3.) = exp .J r(Rs)p(R.=1 J An approximation similar to (3.u/n) < oo) CI  > u n n ryi n' i=1 Another Riemann sum approximation completes the proof.)ds + Y(R2.)Ui } .21) (the initial condition is r(0) = u in both cases).
3. We should like. For different types of applications of large deviations to ruin probabilities .. RESERVEDEPENDENT PREMIUMS the simplest being to require b[s] to be defined for all s > 0 (thus excluding . to point out as a maybe much more important fact that the present approach is far more elementary and selfcontained than that using large deviations theory. . the exponential distribution ). see XI. e.214 CHAPTER VII. however.g.
Chapter VIII Matrixanalytic methods 1 Definition and basic properties of phasetype distributions Phasetype distributions are the computational vehicle of much of modern applied probability. We often write p for the number of elements of E. This implies in particular that the intensity matrix for { it } can be written in blockpartitioned form as T 0 0 . 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<. More precisely. 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. A proper knowledge of phasetype distributions seems therefore a must for anyone working in an applied probability area like risk theory. Typically. and not in other cases.1) is 'Here as usual . that is. F (Jt = A eventually) = 1 for all i E E 1 and where all states i E E are transient. Note that since (1. A distribution B on (0. if a problem can be solved explicitly when the relevant distributions are exponentials. P. if v = (vi)iEE is a probability distribution. refers to the case Jo = i. 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. we write Pv for the case where Jo has distribution v so that Pv = KER viPi• 215 . on Eo = E U {A} where A is some extra state which is absorbing.
i. Thus the phasetype distributions with p = 1 is exactly the class of exponential distributions. j.0 = t11. the ith component ti gives the intensity in state i for leaving E and going to the absorbing state A. E = {i. We now say that B is of phasetype with representation (E. The initial vector a is written as a row vector. C is the lifetime sup It > 0 : Jt E E} of {Jt}. (1. t1 = /3. the rows sum to one which in matrix notation can be rewritten as t + Te = 0 where e is the column Evector with all components equal to one. MATRIXANALYTIC METHODS the intensity matrix of a nonterminating Markov process. an exponential distribution with rate parameter . i. B(t) = Fa(^ < t ).3.1 The phase diagram of a phasetype distribution with 3 phases. 0 2this means that tii < 0.e. Here are some important special cases: Example 1 . Then a = a1 = 1. T is a subintensity matrix2. k}. In particular. A convenient graphical representation is the phase diagram in terms of the entrance probabilities ai.e. the exit rates ti and the transition rates (intensities) tij: tj 3 aj ai i ti tk tjk FkJ ak Figure 1. tij > 0 for i 54 j and EjEE tij < 0 . and the phasetype distribution is the lifetime of a particle with constant failure rate /3. T) (or sometimes just (a. and we have t = Te.1 Suppose that p = 1 and write .216 CHAPTER VIII. Equivalently. a.T)) if B is the Padistribution of the absorption time C = inf{t > 0 : it = A}.2) The interpretation of the column vector t is as the exit rate vector. that is.
1)!e Since this corresponds to a convolution of p exponential densities with the same rate S. 0 0 0 T= t= 0 ••• S S 0 0 0 0 0 0 . 0 •.2 The Erlang distribution EP with p phases is defined Gamma distribution with integer parameter p and density bp XP1 6x (p. .. PHASETYPE DISTRIBUTIONS 217 Example 1... 0 S 6 Example 1.1.... a = (1 0 0 .. 6. so that the density is P E ai6ie6.. p}.3 The hyperexponential distribution HP with p parallel channels is defined as a mixture of p exponential distributions with rates 51. 0 0 0 0 S 6 . . the EP distribution may be represented by the phase diagram (p = 3) Figure 1. .. 00)) S s o . . 0 ••• 0 0 Sp1 0 0 t= 0 0 00 •... 0 SP 0 and the phase diagram is (p = 2) ... . .2 corresponding to E = {1.x i=1 Thus E _ Si 0 T 0 S2 0 0 ..
218 CHAPTER VIII. Proof Let P8 = (p ^) be the sstep EA x EA transition matrix for {Jt } and P8 the sstep E x Etransition matrix for {Jt} .3 .f is B (x) = 1 . 36) yields s d. The basic analytical properties of phase type distributions are given by the following result . B[s] = f0°O esxB (dx) is a(sI T)lt (d) the nth moment f0°O xnB(dx) is (.3 0 Example 1 . and is defined as the class of phasetype distributions with a phase diagram of the following form: 1 617 ti t2 2 b2. Theorem 1 . the backwards equation for {Jt} (e. Then: (a) the c.d.4 (COXIAN DISTRIBUTIONS) This class of distributions is popular in much of the applied literature. (b) the density is b(x ) = B'(x) = aeTxt. p:. see A. .1 tP1 1 Figure 1. the restriction of P8 to E.aeTxe. the Erlang distribution is a special case of a Coxian distribution. (c) the m. ds^ = ds' = ttlaj + tikpkj. 5 Let B be phasetype with representation (E. [APQ ] p.t2 yt bP.f.1)"n! aT"e. j E E. dp.4 For example.e. a. Recall that the matrixexponential eK is defined by the standard series expansion Eo K"/n! 3. MATRIXANALYTIC METHODS Figure 1.g. T).g. Then for i . i. E t ikp kj = kEE kEE 3For a number of additional important properties of matrixexponentials and discussion of computational aspects .
1 ) n +l n ! a (s I + T ) .s j# tii i (1.tii we go to A. (1) n+1n!aT .5) as hi(tii + s) = ti  t ij hj.tii is the rate of the exponential holding time of state i and hence (tii)/(tii .1. or w.f.jEE B'(x) _ cx Pxe = aeTxTe = aeTxt (since T and eTx commute).5) ki = 1 + tii L jj:Ai tii (1. i. = aPxe.T) 't = (. d8 P8 = TP8.s) is the m . tij / .p.T) 1t. the rule (A. For (c).f. Rewriting ( 1. i. of the initial sojourn in state i.5) Indeed . this means in vector notation that (T + sI)h = t.g.. Alternatively. Part (d) follows by differentiating the m. Then h tit ti + ti3 h j .B(x) = 1'a (( > x) = P.f.12) for integrating matrixexponentials yields B[s] = J esxaeTxt dx = a ( f°°e(81+T)dx ) t a(sI .n1t = (1)nn!aTn1Te (1)nn! aTne..6) . After that. ti/ .e. . hj . we i w. for n = 1 we may put ki = Ei( and get as in (1. and (b) then follows from 1: aipF.p.tii and have an additional time to absorption either go to state j which has m .g. Since 1 . the solution is P8 = eT8.g. j#i jEE tijhj + his = ti. h = (T + sI)1t.g. 1. B(n)[0] = _ Alternatively.n lt . d" dsn a (. PHASETYPE DISTRIBUTIONS 219 That is. and since b[s] = ah. we arrive once more at the stated expression for B[s]. and since obviously P° = I. (Jx E E) = this proves (a). in which case the time to absorption is 0 with m .f.tii tii . define hi = Eie8S.s I .
are idempotent. Example A3.7) 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."n! ( ( l 2 2 ) 17 9 0 \ 1 / 10 10 32 n! 35 6" +n!353 Similarly. another the case p = 2 where explicit diagonalization formulas are always available.h. 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. making the problem trivial. see the Appendix.220 CHAPTER VIII. Consider for example 3 9 a= (2 2). 0 Example 1. One obvious instance is the hyperexponential distribution. T= 2 111 so that 2 2 Then (cf.s. MATRIXANALYTIC METHODS which is solved as above to get k = aTle. there are some examples where it is appealing to write T on diagonal form. we get the density as 9 9 6 (1 1) 10 7 1 0 10 2 aeTyt = e x .6 Though typically the evaluation of matrixexponentials is most conveniently carried out on a computer.
where the initial vector a is substochastic.hall. hail = E=EE a.7 If B is phasetype with representation (v.1. 00 B[Q] = J0 f veTxteQx dx = (v ® I) ( f° eT x edx I (t I) (v (& I) ( (T ®Q)xdx f o" e o )( t ® I) _ (v ® I)(T ® Q)1(t ® I). This is the traditional choice in the literature.29) and Proposition A4. and in fact one also most often there allows a to have a component ao at A.f. or one just lets U be undefined on this additional set. 5 and serves at this stage to introduce Kronecker notation and calculus (see A. 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. 0 Sometimes it is relevant also to consider phasetype distributions. i. a random variable U having a defective phasetype distribution with representation (a.e 11BIJ = 1laDD < 1. < 1.e a mixture of a phasetype distribution with representation (a/llall. • The phasetype distribution B is zeromodified. .11aDD.T) with weight hall and an atom at zero with weight 1 . B[Q] of B is f3[Q] = J e'1zB(dx) _ (v (9 I)(T ® Q)1(t ® I).7) Proof According to (A.g.4. i. then the matrix m. There are two ways to interpret this: • The phasetype distribution B is defective. T) is then defined to be oo on a set of probability 1. (1.T).4b for definitions and basic rules): Proposition 1.
and we have eTx . See in particular the notes to Section 6. MATRIXANALYTIC METHODS la Asymptotic exponentiality Writing T on the Jordan canonical form. see his book [269] (a historical important intermediate step is Jensen [214]). (or Laplace transform) are often used where one would now work instead with phasetype distributions. where C.. we give a criterion for asymptotical exponentiality of a phasetype distribution B. Example A5. cf. Neuts. 2. B(x) . v.F.hve7x.8). Rolski. Then the tail B(x) is asymptotically exponential.Ce7'. but in many practical cases. Using B(x) = aeTxe . In Proposition A5.222 CHAPTER VIII. 77 > 0 and k = 0. the result follows (with C = (ah)(ve)).4c). x * oo.f. 1. the text is essentially identical to Section 2 of Asmussen [26]. here k = p1). (1. let . All material of the present section is standard. In older literature. it is easily seen that the asymptotic form of the tail of a general phasetype distribution has the form B(x) _ Cxkenx. .q be the eigenvalue of largest real part of T.g. h be the corresponding left and right eigenvectors normalized by vh = 1 and define C = ah • ve . 0 Of course.. one has k = 0. distributions with a rational m. let v. Lipsky [247]. the conditions of Proposition 1. the Erlang case). T). Notes and references The idea behind using phasetype distributions goes back to Erlang. 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.8) Proof By PerronFrobenius theory (A. Here is a sufficient condition: Proposition 1.f. but the relevant T is not irreducible.8 Let B be phasetype with representation (a.1 of the Appendix. Schmidt & Teugels [307] and Wolff [384]. assume that T is irreducible .8 are far from necessary ( a mixture of phasetype distributions with the respective T(') irreducible has obviously an asymptotically exponential tail. Schmidli. The Erlang distribution gives an example where k > 0 (in fact. not only in the tail but in the whole distribution. i is real and positive.g. No satisfying . but todays interest in the topic was largely initiated by M. cf. h can be chosen with strictly positive component. Other expositions of the basic theory of phasetype distributions can be found in [APQ]. O'Cinneide [276] gave a necessary and sufficient for a distribution B with a rational m.
U1<t < U1+U2. is Markov and has two types of jumps . with common distribution B and define4 U(A) = E# {n = 0. we refer to U as the renewal measure. Jt={Jt?ul}.d. but is in part repeated below. and U(A) is then the expected number of replacements (renewals) in A..1 of the Appendix.. the renewals form a Poisson process and we have u(x) = 0.2.. 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.f.. . the jumps of the j(k) and the it } k) to the next J( k+l) A jump jumps corresponding to a transition from one Jt 4Here the empty sum U1 +. JtJt1) Then { 0<t<U1 .r... If B is exponential with rate 0. For this reason. RENEWAL THEORY 223 algorithm for finding a phase representation of a distribution B (which is known to be phasetype and for which the m. Let U1... + U0 is 0 . if U is absolutely continuous on (0...i. . be i. . +UnEA} 00 = EEI(U1 +.1. however. we denote the density by u(x) and refer to u as the renewal density.. but nevertheless. (2...g. U2. A related important unsolved problem deals with minimal representations: given a phasetype distribution .1) Proof Let {Jtk)} be the governing phase process for Uk and define {Jt} by piecing the { J(k) } together.t. as the lifetimes of items (say electrical bulbs) which are replaced upon failure. Then the renewal density exists and is given by u(x) = ae(T+ta)xt.+UnEA).: U1 + . or the density is available ) is. The explicit calculation of the renewal density (or the renewal measure) is often thought of as infeasible for other distributions. n=O We may think of the U. known.T). oo) w. Lebesgue measure.1 Consider a renewal process with interarrivals which are phasetype with representation (cr. the problem has an algorithmically tractable solution if B is phasetype: Theorem 2.
IIafl < 1. Corollary 2.2 Consider a terminating renewal process with interarrivals which are defective phasetype with representation (a.T).. Then: (a) the excess life t(t) at time t is phasetype with representation ( vt. + Uit_1 where s. However.T) where vt = ae (T+ta)t . (b) £(t) has a limiting distribution as t * oo. u Returning to nonterminating renewal processes . see Fig.1) remains valid for that case.U1 U3 U2 U3 U4 Figure 2. and the jumps of the first type are governed by T. and let µB = aTle be the mean of B.1) follows by the law of total probability. Then the lifetime is zeromodified phase type with representation (a.3 Consider a renewal process with interarrivals which are phasetype with representation (a. that is.IIBII which is > 0 in the defective case. u The argument goes through without change if the renewal process is terminating. fi(t) U2 U1 . .224 CHAPTER VIII. Hence the intensity matrix is T + ta. MATRIXANALYTIC METHODS of the last type from i to j occurs at rate tiaj .1 Corollary 2. which is ti in state i. B is defective . since Uk = oo with probability 1 . is the first k with Uk = 00.T + ta).e.1. this is welldefined. i. the phasetype assumptions also yield the distribution of a further quantity of fundamental importance in later parts of this chapter . 2. Hence ( 2. define the excess life e(t) at time t as the time until the next renewal following t.e.T) where v = aT1 /µB. the density is veTxt = B(x)/µB. . This is defined as U1 + . T). Equivalently. and hence ( 2.. and the distribution of Jx is ae ( T+t«)x. which is phase type with representation (v. The renewal density at x is now just the rate of jumps of the second type. as the time of the last renewal. i. Proof Just note that { it } is a governing phase process for the lifetime. the lifetime of the renewal process.
According to Example A3.4 Consider a nonterminating renewal process with two phases.) ( t2 ) .2) v(T + ta) = 0. Al.T) where vt is the distribution of it which is obviously given by the expression in (a).2): aT1 e = AB = 1 µB µB a + aT'Tea aT1(T + ta) µB PB a + aea a + a µB µB =0.1. Next appeal to the standard fact from renewal theory that the limiting distribution of e(x) has density B(x)/µB.q2.2. The formulas involve the matrixexponential of the intensity matrix Q = T + to = ( tll + tlal t12 + t2al tlz + tlaz _ q1 ql t22 + t2a2 q2 q2 (say). hence e(t) is phasetype with representation (vt.e. T1 and eTx commute. (ii) First check the asserted identity for the density: since T. The renewal density is then aeQtt = (al a2) ( 7i 7"2. Here are two different arguments that this yields the asserted expression: (i) Just check that aT1/µB satisfies (2. The time of the next renewal after t is the time of the next jump of the second type. (2. u Example 2 . we get B(x) aeTxe aT1eTxTe µB µB PB = veTxt. = qz ql (x1 xz) = ql + qz ql + q ' and the nonzero eigenvalue A = ql . i. cf. we first compute the stationary distribution of Q. RENEWAL THEORY 225 Proof Consider again the process { Jt } in the proof of Theorem 2.e.6. the unique positive solution of ve = 1. Hence in (b) it is immediate that v exists and is the stationary limiting distribution of it.
6 Let B be hyperexponential.t2) . A = 25. . t1B 0 Example 2 .e2bt) 13 Example 2 .4 yields the renewal density as u(t) = 2 (1 . and Example 2. )t (51 .52a1.(biaz + aza. Then _ Q Hence 51 0 0 52 + 51 52 _ 5152 51a2 ) (al a2) 52a1 62a1 Slat + 52a1 51a2 51a2+52a1 A = 51a2 . Hence 7r = (1/2 1/2).52) 25152 51x2+5251 51a2+5251 Notes and references Renewal theory for phasetype distributions is treated in Neuts [268] and Kao [221].4 yields the renewal density as u(t) = 5152 e.226 CHAPTER VIII.t2) 1 + eat (a17r2 . The present treatment is somewhat more probabilistic.`t (al a2) + C 11 172 ir12 / \ t 2 ) r1 (7r1 7r2) ( t2 7rltl + J + eAt (al a2) ( 71(t2 . Then Q= 0 55 )+(1o)=( j ad ).5 Let B be Erlang(2).a27rl) (tl . and Example 2.tl) 7r2t2 + eat (a17r2 .a27r1) (t1 . MATRIXANALYTIC METHODS e.
use the phasetype representation of Bo. T). (b) V. i.f3aT1. and rewriting in matrix form yields the phase generator of {my} as T + ta+. Thus the total rate is tip + tia+. however. Since the results is so basic. 3. cf. 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 COMPOUND POISSON MODEL 227 3 The compound Poisson model 3a Phasetype claims Consider the compound Poisson (CramerLundberg) model in the notation of Section 1. a+j.1 on the next page. Considering the first. Proof The result follows immediately by combining the PollaczeckKhinchine formula by general results on phasetype distributions: for (a). Now just observe that the initial vector of {mx} is a+ and that the lifelength is M.) = F(ST(o) E •. Then: (a) G+ is defective phasetype with representation (a+. For (b). We asssume that B is phasetype with representation (a. The stars represent the ladder points ST+(k). r(u) the time of ruin with initial reserve u.3.(u) = a+e(T+tQ+)u Note in particular that p = IIG+II = a+e. marked by thin and thick lines on the figure. T + to+). Next. Here we have taken the terminating Markov process underlying B with two states.1 Assume that the claim size distribution B is phasetype with representation (a. and M is zeromodified phasetype with representation (a+. represent the maximum M as the lifetime of a terminating renewal process and use Corollary 2. 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.i. Corollary 3. we shall. T(0) < oo) the ladder height distribution and M = supt>o St. T).2. 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. with 0 denoting the Poisson intensity.e.p. the Markov processes representing ladder steps can be pieced together to one {my}. add a more selfcontained explanation of why of the phasetype structure is preserved. which occurs at rate ti. . Corollary 2. Then each claim (jump) corresponds to one (finite) sample path of the Markov process. T) where a+ is given by a+ = . {St} the claim surplus process. The essence is contained in Fig.3. B the claim size distribution. Within ladder steps. and if there is a subsequent ladder step starting in j whic occurs w. G+(.
1 This derivation is a complete proof except for the identification of a+ with .228 CHAPTER VIII.M {mx} ST+(2)  S .QaT1.t t d kkt S...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 . This is in fact a simple consequence of the form of the excess distribution B0..1 . T = (3 . 0 Example 3.Q = 3 and b(x) = .2 Assume that . 3e3x + . MATRIXANALYTIC METHODS t .3. Figure 3. 7e7x 2 2 Thus b is hyperexponential (a mixture of exponential distributions) with a (2 2 ). see Corollary 2.
and Markovmodulated models. see Shin [340].6. THE RENEWAL MODEL This is the same matrix as is Example 1. the discussion around Fig.1): Proposition 4. The result carries over to B being matrixexponential. (a) G+ is of phasetype with representation (a+. We assume p = PB/µA < 1 and that B is phasetype with representation (a. this was obtained in Section 3. In the next sections. T) for some vector a+ = (a+. We shall derive phasetype representations of the ruin probabilities V) (u). we encounter similar expressions for the ruin probabilities in the renewal.^(u) = a+e( T+ta+)ue = 24eu + 1 e6u 35 35 0 Notes and references Corollary 3.1 can be found in Neuts [269] (in the setting of M/G/1 queues.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+. That is. cf. his derivation of +'(u) is different. but that such a simple and general solution exists does not appear to have been well known to the risk theoretic community. Fig. For further more or less explicit computations of ruin probabilities.2 are taken from Gerber [157]. The parameters of Example 3. 0(8) (u) (recall that z/i(u) refers to the zerodelayed case and iY(8) (u) to the stationary case). but there the vector a+ is not explicit but needs to be calculated (typically by an iteration). see Stanford & Stroinski [351] . the duality result given in Corollary 11. For the compound Poisson model. It is notable that the phasetype assumption does not seem to simplify the computation of finite horizon ruin probabilities substantially. 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 .j). and the argument for the renewal case starts in just the same way (cf. T). where a+ is the (defective) . 3. 4 The renewal model We consider the renewal model in the notation of Chapter V. see Section 6.T).6). For an attempt. if we define {mz} just as for the Poisson case (cf. 3.4. with A denoting the interarrival distribution and B the service time distribution.4.1 In the zerodelayed case.
Then {m.230 distribution of mo. the calculation of the first ladder height is simple in the stationary case: Proposition 4. Fig. MATRIXANALYTIC METHODS (b) The maximum claim surplus M is the lifetime of {mx}. 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).1) Proof We condition upon T1 = y and define {m. B0 is phasetype with representation (aT1/µa.5. Also. (c) {mx } is a (terminating) Markov process on E. In fact. We have now almost collected all pieces of the main result of this section: Theorem 4 . where a(8) = aT1/PA. The key difference from the Poisson case is that it is more difficult to evaluate a+. where B0 is the stationary excess life distribution corresponding to B. but with initial distribution a rather than a+.1. which for numerical purposes can be solved by iteration.3 a+ satisfies a+ = V(a+).T).1). 4.*'} is Markov with the same transition intensities as {mx}.4 Consider the renewal model with interarrival distribution A and the claim size distribution B being of phasetype with representation (a. Proof Obviously. G(') = pBo.6. (4. Since the conditional distribution of my given T1 = y is ae4y.T). But by Corollary 2. with intensity matrix Q given by Q = T + to+. the form in which we derive a+ for the renewal model is as the unique solution of a fixpoint problem a+ = cp(a+). the Palm distribution of the claim size is just B. where u w(a +) = aA[T + to+) = a J0 e(T+t+)1A(dy). it follows by integrating y out that the distribution a+ u of mo is given by the final expression in (4. obviously mo = m. Nevertheless.*} from {St+y . Hence by Theorem 11.Sy} in the same way as {mx} is defined from {St}.T)• Proposition 4. Then . CHAPTER VIII.3.
.0) is an increasing function of /3. a+) > 0 = a+o) implies a+) _ (a+) > W (a+)) = a+) . The second follows in a similar way by noting that only the first ladder step has a different distribution in the stationary case.1) and a(8) _ aT.0.2.. I {mx} .•.1 .3). Furthermore .. In particular .1 by noting that the distribution of mo is a+. THE RENEWAL MODEL 231 . a+l ) = cp (a+°)) . a+ can be computed by iteration of (4.3) (defined on the domain of subprobability vectors . by a+ = lim a +n) where a+°) . and that this is given by Proposition 4.2) where a+ satisfies (4.M. The term tf3 in cp(i3) represents feedback with rate vector t and feedback probability vector (3. thus .4. only with initial distribution a(*) for mo. a+2) = ^p (a+l)) . It remains to prove convergence of the iteration scheme (4. (4. i. Hence ^p(.2 ) follows from Proposition 4. i y ^ T1= y `•r Figure 4. .^(u) = a+e ( T+ta+)xe.1). the maximum claim surplus for the stationary case has a similar representation as in Proposition 4.1/pA. (4.1(b).^(8)(u) = a ( 8)e(T+ta +) xe.3) Proof The first expression in (4.e...
Then (4. (4.4) whenever EeR(S)U < oo. 7+ ]. Then e4'h = e82h and hence sh = Qh = (T + taA[Q])h = Th + A[s]tah. . F[s] being interpreted in the sense of the analytical continuation of the m.232 CHAPTER VIII. Assume the assertion shown for n .T1.. both quantities are just 0 . Fn ). To this end. let F be the distribution of U1 . (4.) = P(mTl = i.5) yields h = (sI . limn4oo a ) < a+.T)It. we use an argument similar to the proof of Proposition VI. However. with B[s]. Similarly. 0 = a+) < a+ yields a+) _ (a+0)) (a+) = a+ (n and by induction that a(n) < a+ for all n .ST.T)1t.2. s ¢ sp(T).5) Since s $ sp(T). the corresponding right eigenvector may be taken as (sI .1 arrivals (n arrivals are excluded because of the initial arrival at time T1 ). and hence we may assume that h has been normalized such that ahA[s] = 1. Then s is an eigenvalue of Q = T + ta+ if and only if 1 =.1. and let &+". which links together the phasetype setting and the classical complex plane approach to the renewal model (see further the notes). To prove the converse inequality. n) &+n) T a+. For n = 0. a+ ) exists .} can contain at most n . 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). 0 0 We next give an alternative algorithm. MATRIXANALYTIC METHODS and (by induction ) that { a+ n) } is an increasing sequence such that limn. Thus by (4.4).4) makes sense and provides an analytic continuation of F[•] as long as s ¢ sp(T). Then each subexcursion of {St+Tl . In that case. Thus .f. this implies that ahA[s] # 0.T)'t • A[s] (4.4. Let Fn = {T1 + • • • + Tn+1 > r+}be the event that {my} has at most n arrivals in [T1. Proof Suppose first Qh = sh. Obviously. Then F[s] = a(sI . the normalization is equivalent to F(s) = 1. Theorem 4.5 Let s be some complex number with k(s) > 0.P[s] = A[s]B[s]. so to complete the proof it suffices to show that &+ < a+) for all n.g.
. This immediately implies that Q has the form CD1 and the last assertion on the diagonal form . . we get at a(Q ... .p1i . As in Corollary 4.6.. W v M(d) in the notation of Chapter V).. This gives d roots 'y. Pd with corresponding eigenvectors hl.9) we have G+[s] = 1 which according to Theorem 1. in turn. the matrix Q in Theorem 2. Let d denote the number of phases. 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. pdhd. Further... Q has diagonal form d d Q = dpivi®hi = dpihivi.. 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.T)lt = sh. Pd in the domain ER(s) > 0 . 0).6) i=1 i=1 Proof Appealing to Theorem 4. Given T has been computed... Corollary 4. . hd. Hence with h = (sI T). and define hi = (piI .T)It.. THE RENEWAL MODEL 233 Suppose next F(s) = 1. Since R(s) > 0 and G _ is concentrated on (oo.T)lt + t = s(sI . (4. t(ry) > 0. and the solution is . and hence by the WienerHopf factorization identity (A.5. T) with a+ = a(QT)/at.. the classical algorithm starts by looking for roots in the complex plane of the equation f3[y]A[ry] = 1. Then G+ is phase. we have IG_ [s] I < 1 . D that with columns p1 hl.1 has the d distinct eigenvalues . . .. 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 ). letting vi be the left eigenvector of Q corresponding to pi and normalised by vihi = 1 ....T) = 1 ata+ = a+. Q = CD1 where C is the matrix with columns hl.type with representation (a+.5(c) means that a+(sI T)1t = 1.lt we get Qh = (T + to+)h = T(sI . yd satisfying R(ryi) > 0. . .' that the equation F(s) = 1 has d distinct roots p1. In older literature . hd.6 Suppose u < 0.. Notes and references Results like those of the present section have a long history.4.
Here phase. e. a pioneering paper in this direction is Tacklind [373]. and the distribution of an arrival claim is B.exponential form of the distribution was found by Sengupta [335] and the phasetype form by the author [18]. T('). Numerical examples appear in Asmussen & Rolski [43].g.) d (see. the background Markov process with p states is {Jt}. whereas the approach was introduced in queueing theory by Smith [350]. It turns out that subject to the phase. involving .234 then in transform terms CHAPTER VIII. see Dickson & Hipp [118]. [270] and Latouche & Ramaswami [241]. see Neuts [269]. is phasetype. That is .4.type assumptions are basic.contained derivation). The solutions are based upon iterations schemes like in Theorem 4.. Asmussen & O'Cinneide [ 41] for a short self.. which contains somewhat stronger results concerning the fixpoint problem and the iteration scheme. In risk theory. [119]. 5 Markovmodulated input We consider a risk process {St } in a Markovian environment in the notation of Chapter VI. The number of elements of El=> is denoted by q. The matrix.type assumption . The distribution of W comes out from the approach but in a rather complicated form . For further explicit computations of ruin probabilities in the phasetype renewal case . the ruin probability can be found in matrixexponential form just as for the renewal model. 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. an alternative approach (the matrixgeometric method ) has been developed largely by M. and appears already in some early work by Wallace [377]. E(t)). but the models solved are basically Markov chains and processes with countably many states ( for example queue length processes ).. similar discussion appears in Kemperman [227] and much of the queueing literature like Cohen [88]. the intensity matrix is A and the stationary row vector is ir . Neuts and his students.F. The arrival rate in background state i is a. We assume that each B. where R is an unknown matrix. with representation say (a(' ). starting around in 1975. In queueing theory. the fixpoint problems look like R=Ao+RAI+R2A2+ . The exposition here is based upon [18]. For surveys . MATRIXANALYTIC METHODS d F 1 + a J e°" ip(u) du = Ee°w = 11(t.
the analysis involves new features like an equivalence with first passage problems for Markovian fluids and the use of martingales (these ideas also apply to phasetype renewal models though we have not given the details). 5. (a) 0 0 ♦ o ° tl ♦ • 0 0 o } o o (b) 0 } ♦ • 0 o f o Figure 5.5. has states o. states . Section 5b then gives a representation along the lines of Theorem 4. The key unknown is the matrix K. The two environmental states are denoted o. the phase space E(°) for B. Diagonalization Consider a process {(It. 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. say with slope r(i) on intervals where It = i. and the one E(•) for B. Vt)}t>o such that {It} is a Markov process with a finite state space F and {Vt} has piecewiese linear paths. However. This calculation in a special case gives also the ruin probabilities for the Markovmodulated risk process with phasetype claims. The connection between the two models is a fluid representation of the Markovmodulated risk process given in Fig.4.1. MARKOVMODULATED INPUT 235 some parameters like the ones T or a+ for the renewal model which need to be determined by similar algorithms.2.Vt)} obtained by time reversing the I component.6. We start in Section 5a with an algorithm involving roots in a similar manner as Corollary 4.1. O. The stationary distribution is obtained by finding the maximum of the Vcomponent of the version of {(It. 5a Calculations via fluid models. p = ql = Q2 = 2. •. 5. for which the relevant fixpoint problem and iteration scheme has already been studied in VI.1 In Fig.
< oo for all s. 4.Vt)} is then obtained by changing the vertical jumps to segments with slope 1. 4}. t. corresponding to the partitioning + Epp).1))diag + sII = 0 (5. Let E denote the matrix .A 0 Or 1A/ _ t(i) 0 t(2) 0 0 0 0 0 t(3) 0 T1 0 0 0 .31a(l) (/3i)diag . a E E(i) } . j = 1. Recall that in the phasetype case. we have more martingales at our disposal. the probability in the Markovmodulated model of upcrossing level u in state i of {Jt} and phase a E Eli) is the same as the probability that the fluid model upcrosses level u in state (i. Eli) + Proposition 5. 4. o. whereas Ee8s' = oo for all t and all s > so where so < oo. •. consider the vector a satisfying (A + (13i(Bi[ s] .1(b) {(It . a) = 1. in the fluid model Eel'. i E E.92a(2) 0 0 T(2) 0 0 0 f33a(3) 0 0 T(3) with the four blocks denoted by Ei„ i. Bi[s] = a(i)(T(i) + sI)it('). MATRIXANALYTIC METHODS 4. r(i) _ 1. This implies that in the fluid context.1(a). Second. 5. a) of {It}.236 CHAPTER VIII. In the general formulation .(Ni)diag r(i. 2.1 A complex number s satisfies 'A+ (f3i(Bi[s] .1))diag ) a = sa and the eigenvector b = . Thus F = {o. If s is such a number. A claim in state i can then be represented by an E()valued Markov process as on Fig. First. F = E U { (i. resp.1) if and only if s is an eigenvalue of E. The intensity matrix for { It} is (taking p = 3 for simplicity) I A . of E into components indexed by E. F is the disjoint union of E and the Eli). '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. 5. The fluid model on Fig . a) : i E E. V.
MARKOVMODULATED INPUT 237 indexed by E. 0 .E22)1 E21a E21a . it follows that Ell E12 ( E 21 E22) (d) = s 1 d I .sI+ ((3ia(i)(T(i) . For the assertions on the eigenvectors.(sI . t(1) 0 0 then also 0 t(2) 0 . Then E21c+E22d = E21a . resp . where c.sI)1t)) iag I = 0 which is the same as (5.1).E21a + sd = sd. and let d = (sI . d = (sI E22)1E21a = E ai(sI .sI 0 0 0 T(3) .sI + E12 (sI .sI) (sI .T('))1t(i) .A . Noting that E11c + E12d = se by definition. d correspond to the partitioning of b into components Proof Using the wellknown determinant identity Ell E12 E21 E22 E22 I ' I Ell . E(1) + + E(P).E22)1 E21) a = 0.32a(2) (/3i)diag .sI 0 0 0 T(2) .sI.sI ()3i)diag .E22 .5. with Eii replaced by Eii .E22)1 E21a. c = a. it follows that if Qla(1) 0 0 .sI 0 0 t(3) 0 0 = 0.Nla(1) 0 0 T 1. Then (up to a constant) c = a. assume that a is chosen as asserted which means (Ell .A . iEE (a> of 0* 1 AI.E12E22 E21 I .
.. it follows by Proposition II. v.4 that {e"1b(v) is a martingale .2 Assume that E = Or 'Al has q = ql + + qp distinct eigenvalues si. We can take a = c = 1 and get d = (s + b)16 = 5/(3 = 1/p.. q..pi(u. Example 5 . v. j) pi( u . MATRIXANALYTIC METHODS Theorem 5. s2 are the negative eigenvalues of Al +01 A1 E _ A 2 b1 0 52 A2 +32 0 ..a)d^ ). sq with $2s. .v) = j)./' u = e' (esiuc ( 1) . v = 1. Proof Writing Or'Alb( v) = svb( v) as (AI .4 Assume that E has two states and that B1.v)=inf{t >0:Vtu orVt=.3 Consider the Poisson model with exponential claims with rate 5. . . c j. d("))1 e. Then . j.sv)b(v) = 0.. Example 5 . For u.O.. a )d(a + e8 °vpi (u .. j. Letting v ^ oo and using Rsv < 0 yields e8'u = Epi(u. j.v. define w(u. the result u follows. . a).v) = v) I. Then we get V)i (u) as sum of two exponential terms where the rates s1.5.v}. pi(u. a) and noting that i1 (u) = >I j. u) Iw(u. Thus 0(u) = esu/d = pe7 ° as u should be. I' i( V P2 (w (u) < oo..j.. a)).238 CHAPTER VIII. we first look for the negative eigenvalue s of E = I 0 I which is s = ry with yy = b . < 0 and let b(v) = I d(„)) be the right eigenvector corresponding to s.. To determine 0 (u).a Solving for the pi(u.j)c v . j.. a) = Pi (Vw(u.( u. Iw(u. j. B2 are both exponential with rates 51 i b2. w(u)=inf{t >O:Vtu}. w(u. v.v) = (j. v) yields C{V) = e8 ..Q.v) = Optional stopping at time w (u. a)).. e89uc(e)) (d(1) . a) = (j. . . v > 0.v) = = p i( u .upi(u. Here E has one state only.
(y. •) is phasetype with representation (E(i). (5.h.6 For i E E.b (u) = Pi(M > u) = 9(i)euue. is 0 /3 f R(i . In terms of K. MARKOVMODULATED INPUT 239 5b Computations via K Recall the definition of the matrix K from VI.xxej • a 00 oo el . Proof We must show that G+ (i.33(e = 0 a(j))(K ®T ( j))(ej (9 I). 8^')IT(j)) where e 3^') =.3j eye. 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. j.3) . i. the Pidistribution of M is phasetype with representation (E(1) + + E(P). 9(').2) the l. (') a T( However . j. U) where t(j) + t(j)O(j j = k uja.s.( 2.2. dx)Bj(y . j.y = to B k7 j # k In particular. 0 Theorem 5 .x) 00 f ° (') (j) eT (yy)edx .5 G+(i.k.5. according to VI. oo)) j)ye.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). For a transition from (j. Starting from Jo = i. if b* [0] = b1 +b20+b302 +. a) is obviously chosen according to e(`). that the density b(x) can be written as aeTxt for some row vector a. and lifelength M. which occurs at rate t(i). the ratio between two polynomials (for the form of the density. we have the additional possibility of a phase change from a to ry within the ladder step..k y. and it just remains to check that U has the asserted form. Then b*[0] is rational if and only b(x) is matrixexponential.240 CHAPTER VIII. equivalently. i. +bn0i1 0n +a10n1 +.. MATRIXANALYTIC METHODS Proof We decompose M in the familiar way as sum of ladder steps . a m. (6. An alternative characterization is that such a distribution is matrixexponential. oo) and b* [0] = f °O eBxb(x) dx the Laplace transform. 6 Matrixexponential distributions When deriving explicit or algorithmically tractable expressions for the ruin probability. Associated with each ladder step is a phase process. bn1 bn). Furthermore.. However. For j = k. t = (0 0 .. see Example 1. +aii10+anI then a matrixexponential representation is given by b(x) = aeTxt where a = (b1 b2 .e. we have sofar concentrated on a claim size distribution B of phasetype. u Notes and references Section 5a is based upon Asmussen [21] and Section 5b upon Asmussen [17]. which occurs at rate t^^7. 0 1)'...y) to occur when j # k.) which is rational. the initial value of (i. Piecing together these phase processes yields a terminating Markov process with state space EiEE E('). intensity matrix U.g. say. This yields the asserted form of uja.p. with phase space EU> whenever the corresponding arrival occurs in environmental state j (the ladder step is of type j).1 Let b(x) be an integrable function on [0. a) to (k. t) is the representation of the matrixexponential distribution/density): Proposition 6. and a new ladder step of type k must start in phase y.e.. the current ladder step of type j must terminate. Bk7 . . T. some square matrix T and some column vector t (the triple (a.f. Numerical illustrations are given in Asmussen & Rolski [43]. i..2) .2. 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 occurs w.
personal communication).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. (6.3) was suggested by Colm O'Cinneide. (6.3) 0 0 0 0 0 . Namely.an_3 an _ 4 .1 0 0 )3 = (111)..3) that we can take 0 1 0 0 a= (1 + 47r2 0 0). since 1 + 4ir2 03 + 302 + (3 + 47x2)0 + 1 + 47r2 it follows by (6.2 A remarkable feature of Proposition 6. of (6.2).2).3 A set of necessary and sufficient conditions for a distribution to be phasetype are given in O'Cinneide [276]. 0 1 an an1 an _2 .s. s = c/ 2 . u Remark 6. t). .h.47x2 3 1 0 . For a proof. a2 a1 Proof If b(x) = aeTxt.1. The converse follows from the last statement of the theorem. 0 0 0 0 1 0 0 . S.. T= 0 0 1 . cannot be phasetype.6.4) 0 0 1 c This representation is complex.. Writing b(x) = c(e( 2ni1 ) y/2 .. . matrixexponentiality implies a rational transform. see Asmussen & Bladt [29] (the representation (6. (6. Thus. bn of the denominator to be distinct and expand the r. 0 0 ..1) as E 1 c. .(6. but as follows from Proposition 6. One of his elementary criteria. namely to asssume the roots 6l. u giving b(x) = E 1 ciebiz/bY. shows that the distribution B with density b(x) = c(1 cos(21r x))ex.T)1 is so.1 0 ..47r2 3 . b(x) > 0 for x > 0. S = f c/2 0 21ri . 1 . s) is given by 27r i ./(0 + bi). t= 0 . where c = 1 + 1/47r 2. we can always obtain a real one (a. . Example 6 .e(tai1)x/2 + e'T) it follows that a matrixexponential representation ()3... T.. . then b*[0] = a(0I T)1t which is rational since each element of (01 . MATRIXEXPONENTIAL DISTRIBUTIONS 241 T = 0 1 0 0 0 .
7 + 155ex b(x) Then it is known from O'Cinneide [276] that b is phasetype when 6 > 0.4) 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)) . we shall only consider the compound Poisson model with arrival rate 0 and a matrixexponential claim size distribution B.1 shows that a matrixexponential representation can always be u obtained in dimension only 3 independently of J. Then (cf. T. and present two algorithms for calculating '(u) in that setting.3. We recall (see Section 3.4 This example shows why it is sometimes useful to work with matrixexponential distributions instead of phasetype distributions: for dimension reasons . As for the role of matrixexponential distributions in ruin probability calculations. and that the minimal number of phases in a phasetype representation increases to 0o as 5 . Consider the distribution with density = 15 ((2e2x .6) holds true also in the matrixexponential case. (6. recall that t = Te) that if B is phasetype and (a. But since 15(1 +6)02 + 1205 0 + 2255 + 105 b* [9] _ (7 + 155)03 + (1355 + 63)92 + (161 + 3455)9 + 2256 + 105 Proposition 6.5 (6. and can use this to invert by the method of Proposition 6. we have represented ti* [0] as ratio between polynomials (note that 0 must necessarily be a root of the numerator and cancels). t) a phasetype representation with a the initial vector. q are polynomials without common roots. (6. t) of b(x). 0.1 to get i (u) = f3esus.5) Thus.1)2 + 6). we take as starting point a representation of b* [0] as p( O)/q(9) where p. For the second algorithm. then: Proposition 6. MATRIXANALYTIC METHODS Example 6 . then 5(u) = a+e(T+t+)uTle where a+ = /3aT1. T the phase generator and t = Te.6) The remarkable fact is. Corollary 111. leading to matrix calculus in high dimensions when b is small. For the first. .6) in Section 3 seems to use the probabilistic interpretation of phasetype distribution in an essential way. T.242 CHAPTER VIII. that despite that the proof of (6. we use a representation (a.
1UB(B + BVA1UB).T)1 + (6I .T)1 (91.7) 9( cf.T).6).T)1 so that b* b** b** a+(9I .T)1t ( l .A .6.t.1 + b+ = b++ 1 . (6.T . Then in Laplace transform formulation .1t = f3a (0I T)1T1t . (91.6b* . we get b+ = 0aT1(9I T).to+)1 = (BI .'t.1 = A1 .T)1ta+(OI . but we shall give an algebraic proof. From the general matrix identity ([331] p.T)1 J0 00 b(x) dx = f aT1t. with A = 91T. . xb(x) dx = aT2t. 519) (A + UBV ).1BVA1.T)1T1t.to+)1T . MATRIXEXPONENTIAL DISTRIBUTIONS 243 Proof Write b* = a(9I .T .a+(9I . (6.T)1t.to+)1T . since (91T)1T . Now. U =.T)1 + 1 ib* (91.T)1T 2 = and 1 = AB IT2 + 82T .1 = ^(T1 + ( 91T)1). Presumably.T .T)1t)1a +(9I .1 + 82 (9I . b+ = a +(BI . b+ = a+(9I .b* (6.1t du = . we get (91.B=land V=a+.1t = b* . this can be verified by analytic continuation from the phasetype domain to the matrixexponential domain .5 ). the assertion is equivalent to a+(BI .
T)1)t = 8 (1 .1) is the same as the r.1. see the Notes to VII. 7.5 is similar to arguments used in [29] for formulas in renewal theory. see Asmussen & Bladt [29].3a (1 0 T 2 + 1 T 102 (9I + 02 1 T)1) t P + 7.82b*.244 CHAPTER VIII. cf. . /3aT1(0I .T)1T2t .1. which is selfexplanatory given Fig.3 is taken). T).T)1T. a. the ruin probability(u) was found in explicit form for the case of B being exponential.1t = /3a (9I . 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. 7 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate 0. of (6.8 a(T1 + (01. 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. 7a Computing O(u) via differential equations The representation we use is essentially the same as the ones used in Sections 3 and 4. A key tool is identifying poles and zeroes of transforms via WienerHopf factorization. The proof of Proposition 6. 3. VII. Lipsky [247] and Asmussen & O'Cinneide [41]. From this it is straightforward to check that b**/(b+ .1.s. 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. Much of the flavor of this classical approach and many examples are in Cohen [88]. MATRIXANALYTIC METHODS . to piece together the phases at downcrossing times of {Rt} (upcrossing times of {St}) to a Markov process {mx} with state space E. but the argument of [286] does not apply in any reasonable generality).7). 0 Notes and references As noted in the references to section 4. See Fig.1. In Corollary VII.b*). a key early paper is Cox [90] (from where the distribution in Example 6. (for some remarkable explicit formulas due to Paulsen & Gjessing [286].8.la.h.
Proof The first statement is clear by definition.1) where A(t) = v(u)P(0.I] I 80 { tq f Q(v) dvl t1 1 . we obtain V)(u) = P(m„ E E) = v(u)P(0. the definition of {m8} depends on the initial reserve u = Ro. In fact. By general results on timeinhomogeneous Markov processes. i. the A(t) and hence Vi(u) is available by solving differential equations: Proposition 7. RESERVEDEPENDENT PREMIUMS 245 Rt l0 u . Also.1 The difference from the case p(r) = p is that {m2}.1 A(0) = v(u) and A'(t) = A(t)(T + tv(u . though still Markov.7. Let P(tl.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 + •).e. O<.t2) be the matrix with ijth element P (mt2 =j I mtl = i). P(tl.tl < t2 < u. t + s) . Note that in general >iEE Vi (U) < 1. Figure 7. in contrast to Section 3.1z I. t2) = exp where Q(t) = ds [P(t. t) is the vector of state probabilities for mt. is no longer timehomogeneous. 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. Since v(u) = (vi(u))iEE is the (defective) initial probability vector for {m8}. Ai(t) = P(mt = i). 0 < t < u. Given the v(t) have been computed.
two things can happen: either the current jump continues from u + p(u)dt to u. we get vi(u) = aidt + (1 . Thus. In the first case. or it stops between level u + p(u)dt and u.246 CHAPTER VIII.t) for the second.2 For i E E.(u) p ( u) = . the probability of which is 1 . the probability that level u is downcrossed for the first time in phase i is ai.t)). the probability of downcrossing level u in phase i is 8ji(1 + p(u)dt • tii) + (1 . dt].Sj i)p(u)dt • tji = Sji + p(u)tji dt. vi. the probability that level u + p(u)dt is downcrossed for the first time in phase j is vj (u + p(u)dt). jEE . Proposition 7.Qdt) vi(u) + vi'(u)p(u)dt + p(u) dt E{tji+tjvi(u)}. 0 Thus. Given A'. given A.t and being followed by a downcrossing.(tai + vi(u) E vj(u)tjp (u)  Q + vj (u)tjip ( u). the interpretation of Q(t) as the intensity matrix of {my} at time t shows that Q(t) is made up of two terms: obviously.4) jEE jEE Proof Consider the event A that there are no arrivals in the interval [0. 0 < t < u. the probability of downcrossing level u in phase i for the first time is E vj (u + p(u)dt) (Sji + p( u)dt • tji + p(u)dt • tjvi(u)) jEE vi(u) + vi' (u)p(u)dt + p(u) dt E {tji + tjvi(u)} jEE Collecting terms. those corresponding to state changes in the underlying phase process and those corresponding to the present jump of {Rt} being terminated at level u .t). Given A. A'(t) = A(t)Q(t) = A(t)(T + tv(u . whereas in the second case the probability is p(u)dt • tjvi(u). Hence Q(t) _ T + tv(u . (7.3dt. {mx} has jumps of two types. Given this occurs. MATRIXANALYTIC METHODS However. The intensity of a jump from i to j is tij for jumps of the first type and tivj(u . from a computational point of view the remaining problem is to evaluate the v(t).
From Section 3.2.00 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.h. Since the processes Rt and Rt coincide under level B.^ 0. of (7. RESERVEDEPENDENT PREMIUMS 247 Subtracting v. say. we can first for a given v solve (7.) + 0 as v + oo. Thus. V .)P"(AnB. we have p(r) = p = vi (u) 0aTe. we face the difficulty that no boundary conditions is immediately available.) is the tail of a (defective) random variable so that P(Bv) + 0 as v 4 oo. supRt>v l t<7 I where o. When solving the differential equation in Proposition 7. denotes the time of downcrossing level u . Let p" (t). (v) is given by the r. (u) on both side and dividing by dt yields the asserted differential u equation.7. <oo.. Now since both P(A n Bv) 3 0 and P"(A n Bv) . Rt . starting from v"(v) = . P u which implies that v. Then P(B.0 as v + 00 we have P(A) P"(A) = P(AnBv)+P(AnBv) P"(AnB. To deal with this..4) backwards for {va (t)}v>t>o.s.5).) P"(AnBv) = P(AnB. and similarly P"(Bv) . F" etc. refer to the modified process. vi (U) = lim v= (u). then P(A n Bv) _ P"(A n BV'). consider a modification of the original process {Rt} by linearizing the process with some rate p.. (u) for any values of u and v such that u < v. . This yields v.3 For any fixed u > 0. Then pv(r) p(r) r < v p r>v ' and (no matter how p is chosen) we have: Lemma 7. after a certain level v.i7rT1/p. say..
say. p(r) P. Corollary 3.7) equals 01 (v . while the fourthorder RungeKutta method implemented in [30] gives 0(n5).1. assuming u > v for the moment. (u)}.1. 1/n.. as well p1(u). The trapezoidal rule used in [288] gives a precision of 0(n 3).RQ (defined for or < oo only) is defective phasetype with representation (v(u).zp1(u)/(1 . T).) are so. To evaluate p1(u). > 0} and the last term the contribu tion from {R. < 0}). 3u etc. the probability of ruin between a and the next upcrossing of v. We recall from Propositon VII. cf. such that Rt coincide with RI under level v and with Rt above level v. Recall that q(w) is the probability of upcrossing level v before ruin given the process starts at w < v. numerically implemented in Schock Petersen [288]) and the present one based upon differential equations require both discretization along a discrete grid 0. p2. the evaluation of Vi(u) requires q(u) = 1 .e. 0 < u < v. r<v r > v.1. The precision depends on the particular quadrature rule being employed.1a. However. MATRIXANALYTIC METHODS Next consider a sequence of solutions obtained from a sequence of initial values {v.7) f o (the integral is the contribution from {R. let v(u) = a+2ieiT +ta+>)(uv).10 that in addition to the O'(•).. for u > v the distribution of v .z51(v)). Let ii'( u) = a+'ie(T+ta +^)"e denote the ruin probability for R't where a+ = a+i) = laT1/pi.6) We may think of process Rt as pieced together of two standard risk processes RI and Rte with constant premiums p1.v v(u)eTat 1 1 . (7..x) dx f v(u)eT xt dx . v = u. i.248 CHAPTER VIII. where v = inf It > 0 : Rt < v}. 2/n.q(v dx +( ) ) = ( ) ( q( )) vueTva (7. typically the complexity in n is 0(n2) for integral equations but 0(n) for integral equations. The f iin in (7. Then v(u) is the initial distribution of the undershoot when downcrossing level v given that the process starts at u.9. which is available since the z/i'(. Therefore u pl(vvueTa t 1.V" M 0 .. 2u. The algorithm based upon numerical solution of a Volterra integral equation (Remark VII.. 7b Twostep premium rules We now assume the premium function to be constant in two levels as in VII. 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. where.
Since µB = 5/21. I.2.v(u ) eTV e J v(u)eTxtz/)l (v . Example 7.v(u)eTve).e.v(u)eTVe .1 from which we see that pl (u) = 1 + 1 249  1 .7.^1(v) 1 . 01(u) _ 24 u + 35 e6u 1 35 e 4(u) _ 35 .24e.x) dx} V 1 1(v) f V v(u) eTxt.8) The integral in (7. so we consider the nontrivial case example p2 = 4 and p1 = 1.u .01 (v . (7.(7 The arrival rate is (i = 3.4) can be written as (Y(u) ®a+)e(T+t°+>)°1 (T ® (T .and A2 = 3 .x) dx 1 ^(v) ( 1 .8) equals v v(u)eTxta+2) e(T+ta +))( vx)edx which using Kronecker calculus (see A.24ev .v) + (2^ + 3v2 ea'(u " .2V"2.21 = ? yields 0(u) = 1. RESERVEDEPENDENT PREMIUMS 1 . Then one gets X20 20 21 f 1ea1(u v) + 1 3 3 ^ A 2(u e ..e6v Let Al = 3 + 2V'2.e6u 35 .v) 1eai(u v) + 7 7 1 e\2(u v) 1 3 ^') eA2 (u.jl (t ®e) Thus.4 Let {Rt } be as in Example 3. From Example 3. B is hyperexponential corresponding to 3 0 3 a(2 2)' T= ( 0 7 t. all quantities involved in the computation of b(u) have been found in matrix form. p2 < 3.2.to+))11 {e{T®(Ttoy+ ))}„ .be the eigenvalues of T + to( 2 ).
Notes and references [30].v)esv + 7 4_ 2.1)' ?. ) e sv + ( 2v/2.+ it (3 4'I 1 ea2(uv e1\2(u") 7 + ( 32 +4.2 35e6v . 192esv + 8 P1 ./2) ea 1(u . pi (u) = p12(u)/p1 l(u) where p1i(u) p12(u) 35e6v . MATRIXANALYTIC METHODS From (7.250 CHAPTER VIII.1 Thus.1 V2 = 4e5"+6 35e6v .7) we see that we can write pi (u) = v(u)V2 where V2 depends only on v.24es" .24e5v . The analysis and the example are from Asmussen & Bladt .1. 21 3 In particular.24e5v ./2 ea1(u") .21(35e6v . and one gets 12e5" ..b(v) = 192esv +8 35e6v + 168esv + 7* Thus all terms involved in the formulae for the ruin probability have been exu plicitly derived.24es" .
we require that B is concentrated on (0. (b) the lognormal distribution (the distribution of eu where U . and instead we shall work within the class S of subexponential distributions .f.4. a2)) with density 1 e(logyFh) 2/2az . For further examples. 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. L(tx)/L(x) 4 1. For the definition .B(x).2b. Some main cases where this lighttail criterion are violated are (a) distributions with a regularly varying tail.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 all t > 0. oo ) and say then that B is subexponential (B E S) if 251 . For example. A rough distinction between light and heavy tails is that the m. B(x) = ex0 with 0<0<1.N(µ. B[s] = f e8x B(dx) is finite for some s > 0 in the lighttailed case and infinite for all s > 0 in the heavytailed case. x 2iror2 (c) the Weibull distribution with decreasing failure rate . B(x) = L(x)/x" where a > 0 and L(x) is slowly varying.g. see I. III.46 and at numerous later occasions require a light tail. x 4 oo. the exponential change of measure techniques discussed in II.
's X1.2. That is. In terms of r.p. oo). the r.1 Let B be any distribution on (0. Proof By the inclusionexclusion formula. then (with high probau bility) so are both of X1. the behaviour in the lighttailed case is illustrated in the following example: Example 1. X2) > x) ^' 2B(x). Thus .'s. X1 is w.v.2B(x). Thus the liminf in Proposition 1.2 If B E S. given X1 + X2 > x. B(x) Here B*2 is the convolution square. (b) liminf BB(() ) > 2. note first the following fact: Proposition 1. proving (a). X2) > x}. X2 with distribution B. X2) > x} C {X1 + X2 > x}. (1. B(x) ax. if X1 + X2 is large . X2 but none of them exceeds x. X2 > x) = 2B(x) . As contrast to Proposition 1. Then: (a) P(max(Xi. Since B is concentrated on (0. we have {max(Xi. To capture the intuition behind this definition. We later show: Proposition 1. .F(X1 > x. that is. one can check that x x where U is uniform on (0. 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. 1/2 'typical' (with distribution B) and w. P(Xi <yI Xi+X2>x) 1B(y).252 CHAPTER IX.1) then means P(X1 +X2 > x) 2P(Xi > x). In contrast. X2) > x) is P(X1 > x) + P(X2 > x) .B(x)2 . X2) > u x)/B(x) = 2.v. the distribution of independent r. HEAVY TAILS B*2\ 2. 1/2 it has the distribution of X1I X1 > x. then P(X1>xI X1+X2>x)* 2. x 3 00. Then X1 +X2 has an Erlang(2) distribution with density yeY so that B*2(x) xex. and thus the lim inf in (b) is at least lim inf P(max(Xi. in the subexponential case the only way X1 + X2 can get large is by one of the Xi becoming large.3 Consider the standard exponential distribution. That is.p.1(b) is oo. oo).v. P(max(Xi. 1).
The uniformity now follows from what has been shown for y = yo and the obvious inequality y E [0.y)/B(x) > 1. This follows since the probability of the overshoot to exceed y is B (x + y)/B(x ) which has limit 1.] Proof Consider first a fixed y. Let 0 < 5 < 1/2. yo] as X + 00.v.2) B(x) B(x ) B(x) Jo with n = 1 and splitting the integral into two corresponding to the intervals [0. SUBEXPONENTIAL DISTRIBUTIONS Here is the simplest example of subexponentiality: Proposition 1. [In terms of r. we get BZ(x)) > 1 + B(y) + B(B()y) (B(x) . 1 < B(x ) B( x) Y) < B( 0). y] and (y.y)/B(x) > 1 since y > 0. then the overshoot X .1(b) we get B*2(x)/B(x) * 2.4 Any B with a regularly varying tail is subexponential.S)x + B(Sx)2 < lim sup B(x) xaoo B(x) lim sup 2L((1 x^oo . Hence lim sup a+oo B*2(x) 2B((1 . a contradiction. then B(B(x)y) * 1 uniformly in y E [0.B*(n ) B(dz) (1. Proposition 1. 253 Proof Assume B(x) = L(x)/xa with L slowly varying and a > 0. x].B E S. If X1 + X2 > x.z B(x) . Finally lim inf B(x .6)x)/((1 .5)x)' + 0 _ 2 L(x)l xa (16) Letting S 10. we get limsupB*2(x)/B(x) < 2.B*n(x . we therefore get lim sup B*2(x)/B(x) > 1+B(y)+ 1 . or they both exceed Sx. then either one of the Xi exceeds (1 . and combining with Proposition u 1.5 If B E S.1.B(y)) . If lim sup B(x .'s: if X .S)x. B( 0 . We now turn to the mathematical theory of subexponential distributions.yo].xIX > x converges in distribution tooo.B(y) = 2. Using the identity B*(n+1)(x) = 1+ + 1)(x) 1+ 2 1 .
then for any n B*n(x)/B(x) * n.B*2 (x) B(x) (x .z ) (x ) = 1 + (^ B(x . This implies B(x) > c2e5x for all x.254 CHAPTER IX. and this immediately yields the desired conclusions.2. HEAVY TAILS Corollary 1. Proposition 1.7 If B E S. B*(n+1) (x I xy + Jxx y) W.5 that B(n) > e6B(n . Proof We use induction. choose y such that IB*n(x)/B(x) .B(x .z) B(dz) _y B(x) 111 Lx B . Then by (1.. x oo. B(x) \Jo _ B(x .5 and the induction hypothesis. Given e > 0. its intuitive content is the same as discussed in the case n = 2 above. O The following result is extremely important and is often taken as definition of the class S. we have by Proposition 1.nI < e for x > y. so assume the proposition has been shown for n. b[c] = oo for all e > 0.y) sup v>o B(v) B(x) which converges to 0 by Proposition 1. The case n = 2 is just the definition.6 If B E 8. then e"R(x) * oo. Proof For 0 < 5 < e. 0 Proof of Proposition 1.z) B(dz) 2B(x) o rv 2 0 2 using Proposition 1.z) B(dz) (n + O(e)) ^x JO B(x) (n + 0(0) I B (x) .5 and dominated convergence. The first integral is y B(x .2). 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 .1) for all large n so that B(n) > cle6n for all n.z) B(dz).z) B(x) Here the second integral can be bounded by B*n(y) B(x) .
Then Al * A2(x) = P(X1 + X2 > x). A2 be distributions on (0. SUBEXPONENTIAL DISTRIBUTIONS 255 Here the first term in {•} converges to 1 (by the definition of B E S) and the second to 0 since it is bounded by (B(x) .y)Ai(dy) v) f o . 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. Combining these estimates and letting a 4.9 Let A1. an+1 fX B*n( *n(x . choose T such that (B(x)B*2(x))/B(x) < 1 + b for x > T and let A = 1/B(T).y))/B(x).(al + a2)B(x).z) B(dz ) + sup < 1 + sup f x<T B ( x) x>T 0 B(x . X2 be independent r.0 completes the proof.2).ala2B(x)2 which can be neglected.i).'s such that Xi has distribution Ai.8 If B E S.1. Proof Define 5 > 0 by (1+5)2 = 1+e. it follows that it is necessary and sufficient for the assertion to be true that JX_VA (x .y)B(dy) = B(x)ov (1)• v (1. 0 Lemma 1.z) B(x . a2 with a1 + a2 > 0. then there exists a constant K = KE such that B*n(x) < K(1 + e)nB(x) for all n and x.z) B(dz) < 1 + A + an(1 + d) .v. an = supx>o B*n(x)/B(x).5 easily yields P(X1 + X2 > x.z) B(x) < 1 + A + an sup f x B(x .y)Ai(dy) = (x)o(1) (1. For any fixed v. e > 0. Proposition 1.X2 > xv) < A1(xv)A2(x v) . Since P(X1+X2 > x.ajB(x)Ai(v) = ajB( x)(1+o„(1)) (j = 3 . Xi <= v Ai (x . Proof Let X1.X1 > xv.4) . Then by (1.B(x . 0 Proposition 1.3) Using the necessity part in the case Al = A2 = B yields f xv B(x .z) B(dz) x . oo) such that Ai (x) _ aiB(x) for some B E S and some constants al. Then Al * A2 (x) .
Then B E S provided fo "O exA(x) b(x) dx < oo.(x) is decreasing for x > x0 with limit 0 at oo. HEAVY TAILS 'VV B(x .Bl (x) + B2 (x) when B1. whereas the two first yield B(x)(Ai(v) . Then L = L1 + L2 is slowly varying and B1 * B2(x) sim L(x)/x«. i = 1.y)B(dy). a1 = a2 = a yields A*2(x) . if q(x) aB(x) for some B E S and some constant a > 0. it is easy to see that if L1.11 Let B E S and let A be any distribution with a ligther tail. then A E S. of (1.10 The class S is closed under tailequivalence.5) Here approximately the last term is B(x)o„(1) by ( 1. We next give a classical sufficient (and close to necessary) condition for subexponentiality due to Pitman [290]. V (1. A(x) = o(B(x)).256 Now (1.Ai(x . u Corollary 1. B1 * B2 E S does not hold in full generality (but once B1 * B2 E S has been shown. Proof Taking Al = A2 = A.v)Ai(v) . L2 slowly varying. B1 * B2 (x) .Bl (x) + B2 (x) follows precisely as in the proof of Proposition 1. In the regularly varying case. u It is tempting to conjecture that S is closed under convolution. with a > 0 and L1. Then A * B E S and A * B(x) . Hence Corollary 1. then so is L = L1 + L2. L2 are slowly varying. That is. u Corollary 1.v)B(v) + _'U Aq(x . Recall that the failure rate A(x) of a distribution B with density b is A(x) = b(x)/B(x) Proposition 1. a2 = 1. f " By a change of variables.3) follows if CHAPTER LX. A2 = B so that a1 = 0.5) becomes x B(x .2A(x).h. .aiB(v)) = B(x)o„(1). it should hold that B1 * B2 E S and B1 * B2 (x) .13 Let B have density b and failure rate A(x) such that .s. However. B2 E S. the l.4).2aB(x) .12 Assume that Bi(x) = Li(x)lxa.9).2. That is.y)Ai(dy) = B(x)o„(1).B(x) Proof Take Al = A.
y) < yA(x . subexponentiality has alrady been proved in Corollary 1.12. and exa(x)b(x) = (3x01e(10)x9 is integrable. SUBEXPONENTIAL DISTRIBUTIONS 257 Proof We may assume that A(x) is everywhere decreasing (otherwise. B*2(x) . Thus.A(xy)A ( y). Thus. Then B(x) = eA(x). replace B by a tail equivalent distribution with a failure rate which is everywhere decreasing). Thus A(x) is everywhere decreasing.y) * 0.3 < 1. L(x) y° (a .y) < A (y) for y < x/2. By (1. we first quote Karamata's theorem (Bingham. f ' L(y) dy .(y) dy. the u lognormal distribution is subexponential.A(x .1)xcl1 .14 Consider the DFR Weibull case B(x) = ex0 with 0 <. 0 A(x) . we can use the same domination for the second integral but now the integrand has limit 0 . elementary but tedious calculations (which we omit) show that A(x) is ultimately decreasing.1. Since ) (x .y ) b(y)dy = B (x) o ox _ J = ox/2 eA( x)A(xy ).U) /or) v 2x This yields easily that ex.y) dy.A(y)a(y ) = ev'(y) b(y). Thus B*2(x )/ B(x) . The middle bound shows that it converges to b(y) for any fixed y since \ (x . Jo For y < x/2.. To illustrate how Proposition 1.1 B(x) eA( x)A(xv )A(y)A(y) dy f B(x . Thus by dominated convergence .(x . x .2). the first integral has limit 1 . Then b(x) = Ox0lexp. proving B E S.y) y\(y)• The rightmost bound shows that the integrand in the first integral is bounded by ey"(v). the DFR Weibull distriu bution is subexponential. Example 1.e009xv)2/2a2/(x 2irv2) logx ( ) 't ((logx . Goldie & Teugels [66]): Proposition 1..1 has limit 1 + 0.A(y)\(y) dy + fox/ 2 eA(x ). a(x) = ax01. Define A(x) = fo .`(x)b(x) is integrable.13 works in this setting. Example 1. Further.15 In the lognormal distribution. an integrable function by assumption.16 For L(x) slowly varying and a > 1. In the regularly varying case.
the monotonicity condition in Proposition 1.13 may present a problem in some cases so that the direct proof in Proposition 1. Thus exa(x)b(x) .y(x)B(x). Proof ( a): Using Karamata's theorem.1/A(x) and P(X ixil'Y (x) > y) * e'. then 7(x) x/(a . we get (1.258 From this we get CHAPTER IX.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 . let X W = X .17 If B has a density of the form b(x) = aL(x)/x°+1 with L(x) slowly varying and a > 1.1)] I X > x) L(x[1 + y/(a .)/Y(x) > y) (1 + y/(a .a/x.xjX > x. . More precisely. the overshoot properly normalized has a limit which is Pareto if B is regularly varying and exponential for distributions like the lognormal or Weibull. Then 7(x) .L(x)/x" and )t(x) .1)X(x)/x > y) = P(X > x[1 + y/(a .E(X . yo] .1) and P(X (. HEAVY TAILS Proposition 1. f O B(y) dy .ea b(x) is integrable. However.18 (a) If B has a density of the form b(x) = aL(x)/xa with L(x) slowly varying and a > 1.4 is necessary in full generality. Then: Proposition 1. 'y(x) = EXix>.1))a . then B(x) .1)]) xa L(x) (x[1 + y/(a . (c) Under the assumptions of either ( a) or (b).1)])a 1 1 . We conclude with a property of subexponential distributions which is often extremely important: under some mild smoothness assumptions. (1 + y/(a .1))^ ' (b) Assume that for any yo )t(x + y/A(x)) 1 A(x) uniformly for y E (0.6) EX(x) .
8) in (b) then follows from P (A(x)X (x) > y) = F(X > x + y/. with common distribution G E S and let K be an independent integervalued r.2 Let Y1. i.EK G(u). Theorem 2 . P The proof is based upon the following lemma (stated slightly more generally than needed at present).15. cf.A(x) I X > x) = exp {A(x) . Y2. 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.1 If Bo E S.f yl 0 0 = exp {y (1 + 0(1))} 0 fY A( x + u /A( x)) a(x) du } The property (1.(x).and lognormal distributions . d. St > u}. nG(u). Kliippelberg & Mikosch [134]..v.. We get p(yl+. r(u) = inf it > 0. then Vi(u) P Bo(u). Then P(Y1 + • • • + YK > u) . 2 The compound Poisson model Consider the compound Poisson model with arrival intensity /3 and claim size distribution B. with EzK < oo for some z > 1. u a oo. Examples 1.nn.A(x + y/A(x))} =a(x) a(x + x) dx = ex p ex P .7) is referred to as 1/A(x) being selfneglecting.. 1.+YK> u) = ^•P(K = n)G* n(u ) . . Let St = Ei ` Ui . It is trivially verified to hold for the Weibull.n0 1•P(K= n)•n = EK.t be the claim surplus at time t and M = sups>0 St. Bo(x) = f0 B(y) dy / µB. Notes and references A good general reference for subexponential distribution is Embrechts. be i. The remaining statement (1.2. We assume p = /3µB < 1 and are interested in the ruin probability V)(u) = P(M > u) = P(r(u) < oo).. THE COMPOUND POISSON MODEL 259 We omit the proof of (c) and that EX (x) . Recall that B0 denotes the stationary excess distribution. Lemma 2. . 0 G(u) L G(u) .1/.14.
µB(01 .p) and EzK < oo whenever pz < 1. (2..1) In particular . For some numerical studies..2. Bo is more heavytailed than B .p)p'. as well as examples where B ¢ S. a]. Bo E S. lognormal . then Bo(x)/B(x) + 00. see Abate. mathematically one must note that there exist (quite intricate) examples where B E S. Borovkov [73] and Pakes [280]. The tail of Bo is easily expressed in terms of the tail of B and the function y(x) in Proposition 1. _ B(x^sx Bo(x) µ8 I aoB(y )dy = (^) . The approximation in Theorem 2. u x+a Notes and references Theorem 2. u The condition Bo E S is for all practical purposes equivalent to B E S. However. Since EK = p/(1.x400 PBB(x) PB Leta+oo. Bo E S is immediate in the regularly varying case.18.µ J ) . In general: Proposition 2.1)xa1' vxe(109x11)2/202 2 +° /2 µB = eµ Bo(x) eµ+O2/2(log x)2 27r' = µB = F(1/0 ) Bo(x 1 ) .x^ ) B(x) _ f or ( lox . P(K = k) = (1. and for the lognormal and Weibull cases it can be verified using Pitman 's criterion (Proposition 1. r(1/Q) xlQexp B(x) = ex' From this . The problem is a very slow rate of convergence as u ^ oo. HEAVY TAILS u using dominated convergence with >2 P(K = n) Dz" as majorant. in our three main examples (regular variation .?(xµ 8 (x). Proof Since B(x + y)/B(x) * 1 uniformly in y E [0.1 is essentially due to von Bahr [56]. Proof of Theorem 2. Bo ¢ S. the result follows immediately from Lemma 2. See also Embrechts & Veraverbeeke [136].1 is notoriously not very accurate.2) M = Yl + • • • +YK where the Yt have distribution Bo and K is geometric with parameter p. x 4 00.1.260 CHAPTER IX.3 If B E S. The PollaczeckKhinchine formula states that (in the setup of Lemma 2. . Weibull) one has Bo(x ( B(x) . we have fx B(y)dy = a B0 (x) > lim inf lim inf x+oo B(x) . Note that in these examples .13).
. 3 The renewal model We consider the renewal model with claim size distribution B and interarrival distribution A as in Chapter V.1. E. Based upon ideas of Hogan [200].} Then ik(u) = F ( M > u) = P(i9 (u) < oo). Define further 0 = IIG+II = P(r9+ < oo). i.] The proof is based upon the observation that also in the renewal setting. one may have to go out to values of 1/'(u) which are unrealistically small before the fit is reasonable. This shows that even the approximation is asymptotically correct in the tail. Then l/i(u) 1 P P [Note that (b) in particular holds if B E S.+ E A. Snd) = Xl +. To Bo(u) u + 00. [279].1) this end . 195 there are numerical examples where tp(u) is of order 105 but Theorem 2.3.1 when u is small or moderately large.. Kalashnikov [219] and Asmussen & Binswanger [27]. Thus G+ is the ascending ladder height distribution (which is defective because of PB < PA). there is a representation of M similar to the PollaczeckKhinchine formula. G+ (A) = P(Sq+ E A. The main result is: Theorem 3 . in [219] p. Then K M=EY. .1 gives 1010. T+ < oo) where r+ = T1 + • • • + T.y)/B (x) > 1 uniformly on compact y internals.e. Let U= be the ith claim . t9(u) = inf {n : Snd> > u} .. T1 the ith interarrival time and Xi = U. 1 Assume that (a) the stationary excess distribution Bo of B is subexponential and that (b) B itself satisfies B(x . Asmussen & Binswanger [27] suggested an approximation which is substantially better than Theorem 2. i=1 . let t9+ = i9(0) be the first ascending ladder epoch of {Snd> }. THE RENEWAL MODEL 261 Choudhury & Whitt [1].9+ < oo) = P(S. Somewhat related work is in Omey & Willekens [278]. {n= 0.. (3. p = iB /µA < 1.g. We assume positive safety loading. also a second order term is introduced but unfortunately it does not present a great improvement.... M = sup s$ . + Xn. In [1].y + as usual denotes the first ascending ladder epoch of the continuous time claim surplus process {St}.Ti.
Lemma 3 . u a 00. U_ (dy) is close to Lebesgue measure on (. x > 0. P(K = k) = (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. x + oo. A.oo.1) holds for a general random walk satisfying the analogues of (a).d.d)) E A) denote the pre19+ occupation measure and let and U_ = Eo G'_" be the renewal measure corresponding to G_.1) is equivalent to P(M > u) " . x * oo.y_ E A) the descending ladder height distribution (IIG II = 1 because of PB < P A) and let PG_ be the mean of G_.1 IPG_ I / F(x .FI(u). As for the compound Poisson model. Proof Let R+(A) = E E'+ ' I(S. we will use the fact that the proof of (3. are independent of K and i. 0] normalized by IPG_ I so that we should have to G+(x) .3 G+ (x) . HEAVY TAILS where K is geometric with parameter 9. d+ < oo).9)9'' and Y1.262 CHAPTER IX. Lemma 3 . Let further 19_ _ inf {n > 0: S^d^ < 0} be the first descending ladder epoch. Proof By dominated convergence and (b).i.3) and we will prove it in this form (in the next Section.Ti).(. 0 The lemma implies that (3. Then 0 0 F( x .. with distribution G+/9 (the distribution of S.g+ > x. The heuristics is now that because of (b).. 0] to the integral is O(F(x)) = o(FI(x)). B(x) _ J O° B(B(x)y) A(dy) f 1 .y+ given r+ < oo).Y2. whereas for large y . the contribution from the interval (.B(x). oo) = F(S.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_. and hence FI(x) . (b) and does not rely on the structure Xi = Ui .FI(x) /IPG_I.2). A(dy) = 1. cf. Let F denote the distribution of the Xi and F1 the integrated tail. G_(A) = P(S. Write G+( x) = G+ ( x.PBBo(x)..2 F(x) .N. FI (x) _ fz ° F(y) dy.y) dy = 1 Pi (X) oo IPG_ I . (3.
n] < (1 + e)/1µc_ I for n > N. THE RENEWAL MODEL 263 We now make this precise.1. Similarly. > (1 .2) yields 00 F F I (u) P(M > u) _ E(1 .1)/F(n) < 1 + e for n > N (this is possible by (b) and Lemma 3. u Proof of Theorem 3. By Lemma 3.1. we can assume that the span is 1 and then the same conclusion holds since then U(n .3.3.(dy) fN FI ( x) + lim sup ZY00 N F(x . n] F1 ( n=N _1 1+e E F(x+n) 0 + limsup xr00 FI(x) FAG.=1 BIp G_ I (1. (3.y) U_ (dy) 00 FI (x) < lim sup F(x) U(N.1. In the lattice case. If G_ is nonlattice. choose N such that F(n .1 I .F[s] = (1 .O[s])(1 . n] is just the probability of a renewal at n. n] + 1/I µG_ I. then by Blackwell 's renewal theorem U_ (n .2).UG_ I x.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. and that U_(n .0)0k k I(u) A. and in the last that FI is asymptotically proportional to Bo E S.9)IpG_ I Differentiating the WienerHopf factorization identity (A.e) z lim inf G+(x)  FI (x) Ip G_ I Letting a 10.y) U. the proof is complete. Hence using dominated convergence precisely as for the compound Poisson model. F(Y= > x) FI(x)/(OIp _ 1). Given e.1.I n=N (1 E)2 r00 F(x + y) dy + e) lim sup .1. 0] x+00 FI(x) 00 + lim up 1 x) E F(x + n) U_ (n .9) 1 .G+[s]) . We then get lim sup G+(x) xro0 Fj(x) < lim sup X)00 o F(x .
a. Then P(M E (u . Therefore by Lemma 3.(u)+n .a. Mn < u}. S+9(u) .1 is due to Embrechts & Veraverbeke [136]. 4 Models with dependent input We now generalize one step further and consider risk processes with dependent interclaim times. FJ(u) UBBO(U) PBo(u) N = (10)Ipc_I JUA .So( u)_1 < a) < P (w(u) < oo)j/i(0) < 0(0) P(M E (u . P(M > u. we have P(M E (u .. u)). and {Su. Note that substantially sharper statements than Lemma 3.SS(u)}n=o.1)6+[0] . 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.So(u)) are available. Notes and references Theorem 3.(1 .0)ua_ .yiui_1. u)) > P(w(u) < oo)(i lp (0))• On the other hand. HEAVY TAILS µF = (1 .a..2. see Asmussen & Kliippelberg [36].IIG+II)µc_ = (1 . Proof Let w(u) = inf {n : Sid) E (u . S+q(u) .a) N P(M > u).264 and letting s = 0 yields CHAPTER IX. on the set {M > u.l. 10(0) But since P(M > u .Se(u)_1 < a) = o(Fj(u)).a.Sty(u)_I < a} we have w(u) < oo.u)) = o(P (M > u)) = o(FI(u)).4 For any a < oo. . In view of the `one large claim' heuristics it seems reasonable to expect that similar results as for the compound Poisson and renewal models should hold in great generality even when allowing for such dependence. allowing also for possible dependence between the arrival process and the claim sizes..4 on the joint distribution of (S. with roots in von Bahr [56] and Pakes [280]. Sty(u) .
M = sup St.F*(X) = P0(Si > x) .X1 is the generic cycle. The idea is now to observe that in the zerodelayed case.4 below. We give here one of them... 0o(u) etc. See Fig.Sxi}0<t<x2Xl . For further approaches.. see [47]. +1. The zerodelayed case corresponds to Xo = Xl = 0 and we write then F0. (viewed as random elements of the space of Dfunctions with finite lifelengths) are i.i.n n=0. Figure 4.. Theorem 4.d...X2 < .1 based upon a regenerative assumption.. (corresponding to the filled circles on Fig. = Sx. We let F* denote the Podistribution of Si. G(x) (4.Sxk}o<t<xk+1xk is the same for all k = 1. E0.. 4.. t>0 S.1. examples and counterexamples. 4.. {SX1+t . Thus the assumption .. MODELS WITH DEPENDENT INPUT 265 Various criteria for this to be true were recently given by Asmussen.1 except for the first one) is a random walk.. 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 <.. and apply it to the Markovmodulated model of Chapter VI. We return to this point in Example 4.1 where the filled circles symbolize a regeneration in the path..1.4. < 0 and EoX < oo where X = X2 ..1) is assumed. {Sn}n=o. M* = max S... Schmidli & Schmidt [47]. 4. 2..1) . such that {SXo+t  SXo}0<t< X 1Xo .. and the distribution of {Sxk+t . M.1 Note that no specific sample path structure of {St} (like in Fig.. assume pp. Define S. ..1 = max k=0.
The one we focus on is Fo (Mix) > x) ..2 Theorem 4. N N Xi=0 N Figure 4. Then '00 (u) = Fo(M > u) . Fo(Si > X).* i o<t<xn+1x. it suffices by (4. (4.3) applicable so that F(M* > u) 141 F*(u). the assumption means that Mix) and Sl are not too far away.3) where Mnx) = sup o<t<xn +1 X.2) Imposing suitable conditions on the behaviour of {St} within a cycle will then ensure that M and M* are sufficiently close to be tail equivalent.1 Assume that (4.S. jF11 F* (U). Since clearly M(x) > Sl .. 4..2) to show F(M* > u) > 1. u p 00. (4.3) hold. Proof Since M > M*. Sxn +t . (4.266 CHAPTER IX..1) and (4.2. HEAVY TAILS for some G such that both G E S and Go E S makes (3. See Fig.4) liminf u>oo F(M > u) .Sxn = sup Sxn+t .
Theorem 4. MODELS WITH DEPENDENT INPUT Define 79* (u) = inf {n = 1 .a. choose a such that Po(Si > x ) > (1 . Then by Lemma 3.Mn +1 >aV(u n=1 00 S. Given e > 0. E (u .(u) .S. Under suitable conditions .E) Po ( n max St u. .Sn+1Sn>aV(uSn*)) n=1 00 > (1E)EPo(Mn<u.2. To this end. Letting first u + oo and next e . )) > (1 .e)Po (M > u. M^xu)+l > a) .: S. /3(u) = inf{n=1. u))/P(M* = 0) = o(Po(M* > u)). x > a. u)} < P(M* E (u .e)Po (M > U). We shall use the estimate Po(M > u) Miu^+ 1 < a) = o(Po (M > u)) (4.6) 1 p pBo(u) u where B is the Palm distribution of claims and p .Sn 0<t<x„+j ( 1 ..1 can be rewritten as 00 (U) (4. Let a > 0 be fixed..1 = limti00 St/t.4)..a.Po (M* > u.. Po(M* > u) ..: Sn > u} .( u)1 > a) 00 1: Po(Mn<u. 0 yields (4. Mn+l > a V (u .+Mn+1>u} 267 (note that {M> u} = {3(u) < oo}). 2..4.5) which follows since Po (M > u. MW O(u)+1 < a) IN ( U n=1 A1. assume the path structure Nt St = EUit+Zt i=1 .e)Po (MMX> > x).(1 .4. S..
(iii) For some o field Y.3PB. and ENX Ui . Corollary 4. oX (see Proposition A1.268 with {Zt} continuous.} and satisfying Zt/t N. are Fmeasurable and NX Po J:U=>x i=1 (iv) Po sup Zt > x / (0:5t<x o(B(x)) Then (4. (ii) EozNX < oo for some z > 1. cf. The same is true for Sl.v. i=1 (4. the proof of Lemma 4. we get 00 (u) 1 IPF. independent of {> CHAPTER IX.4). since the tail of Zx is lighter than B(x) by (iv).8) x Write .Q = EoNx/EoX.7).I u J Po(Sl > x) dx 1 EoNxB(x) dx EoX(1 .X both have tails of sup Zt. and the rest is just rewriting of constants: since p = 1+tlim St = 1+ .6) holds with p = . Assume further that (i) both B and Bo are subexponential. Proof It is easily seen that the r. Mix) < > UE + i=1 o<t<x Thus Theorem 4. a4' 0. Then the Palm distribution of claims is B(x) = E N Eo 0 I( U1 < x) .2 Assume that {St} is regenerative and satisfies (4.1 is in force. and also for Mix) since Nx FNX U.6 below. HEAVY TAILS N` U. X and N.'s order EoNx • B(x).p) Ju P Bo(u) 1p 0 .
We consider the case where one or more of the claim size distributions Bi are heavytailed. (i) holds. We now return to the Markovmodulated risk model of Chapter VI with background Markov process {Jt} with p < oo states and stationary distribution 7r. Again . The number N.(NX). 3 The average arrival rate / and the Palm distribution B of the claim sizes are given by P P Q = ir i/i. Assume that B E S. consider the periodic model of VI. > 0.6 with arrival rate /3(t) at time t (periodic with period 1) and claims with distribution B (independent of the time at which they arrive). X2 = 1.. More precisely. Bo E S. . then (iv) holds since the distribution of supo<t<i Z(t) is the same as that of I Zl 1. of claims arriving in [0. we assume that B E S.. i. The regenerative assumption is satisfied if we take Xo = Xi = 0.5 Consider the Markovmodulated risk model with claim size distributions satisfying (4.t + EN'I Ui where {>N`1 Ui . 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 . < 1. we conclude just as in Example 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. Thus we conclude that (4.6) holds.4. in particular lighttailed.t} is standard compound Poisson and {Zt} an independent Brownian motion with mean zero and variance constant a2. 1) is Poisson with rate /3 = fo /3(s) ds so that (ii) holds. X3 = 2. In particular. Zt . X2 = 1.3 As a first quick application.e. Taking again Xo = Xi = 0.. (iii) is obvious. MODELS WITH DEPENDENT INPUT 269 Example 4 . .6) u holds.3 that (4. Bo E S.9). .4 Assume that St = Zt . Example 4 .. Then (4.. X3 = 2. Theorem 4. and taking F = o. 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. and for some constants ci < oo such that cl + • • • + c. The arrival rate is /3i and the claim size distribution Bi when Jt = i.0 (thus (iv) is trivial). .6) holds.
6 Let (N1. . .ciG(x). It follows by a slight extension of results from Section 1 that P P(Yo > x I Y) G( x) ci Ni. u Proof of Theorem 4. we can define the regenerations points as the times of returns to i. and that for some + cp distribution G on [0. .. i1 = E\ G(x) In the general case. For lighttailed distributions.F) = P(Yo > X+x I •^) G (x +x)>2ciNi i=1 .. as x a oo. Assume EzN1+"'+Np < oo for some z > 1 and all i.. NP ) and X are . ..}P. .v. i =1 P(Yo > x I ^ ) < CG(x)zN1+ +Np for some C = C(z) < oo.X i=1 j=1 where conditionally upon F the Xi.5. and the rest of the argument is then just as the proof of Corollary 4. oo) and define p Ni Yx = EEX'i . oo) such that G E S and some c1.^•) G(x) P ^ E ciNi = C. are independent with distribution Fi for Xij. .. If Jo = i.c'(x) where c = ciENi .. An easy conditioning argument then yields the result when Jo is u random. Markovmodulation typically decreases the adjustment coefficient y and thereby changes the order of magnitude of the ruin . HEAVY TAILS Lemma 4 . cp with cl + > 0 it holds that Fi(x) . The same dominated convergence argument completes the proof. i=1 Proof Consider first the case X = 0. i=1 P(Yx > x ^) < P(Y0 > x I. Then P P(Yx > x) .. .. Let {Fi}t=1 P be a family of distributions on [0. 2 . X > 0 a r."+Np . Thus dominated convergence yields ( P(Yo>x P(Yo>x .270 CHAPTER IX. NP ) be a random vector in {0.2.G( x ) > ciNi . and F a aalgebra such that (N1. P P P(YX and > x I.Fmeasurable. 1.F) < CG(x)zn'1+.
1. the discussion provides an alternative point of view to some results in Chapter IV. and the final reduction by Jelenkovic & Lazar [213].p)Bo(u). states that under mild additional conditions. 5 was first proved by Asmussen. > 0) matter for determining the order of magnitude of the ruin probabilities in the heavytailed case.. Theorem 5.4. Within the class of risk processes in a Markovian environment. Then O(u) .5 shows that basically only the tail dominant claim size distributions (those with c.3.T2.4.... r(u) is the time of ruin and as in IV.d.1 is from Asmussen.4. I T(u) < oo). this should be compared with the normal limit for the lighttailed case. ). We start by reviewing some general facts which are fundamental for the analysis. Theorem 4. for lighttailed distributions the value of the adjustment coefficient y is given by a delicate interaction between all B. Notes and references Theorem 4.7). The present approach via Theorem 4. IV. in particular Proposition 2. cf. this is applied for example to risk processes with Poisson cluster arrivals. cf.e. we let PN"N = P(.i.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.6) to hold in a situation where the interclaim times (T1. 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).pl(1 . FINITEHORIZON RUIN PROBABILITIES 271 probabilities for large u. That paper also contains further criteria for regenerative input (in particular also a treatment of the delayed case which we have omitted here).4. Essentially. Theorem 2. An improvement was given in Asmussen & Hojgaard [33]. i.5.2 and Example 4. 5 Finitehorizon ruin probabilities We consider the compound Poisson model with p = /3pB < 1 and the stationary excess distribution Bo subexponential.7. ) form a general stationary sequence and the U..T2. VI. For further studies of perturbations like in Corollary 4. i. as well as a condition for (4. It follows from Theorem 4. Floe Henriksen & Kliippelberg [31] by a lengthy argument which did not provide the constant in front of Bo(u) in final form. The main result of this section. As usual. Combined with the approximation for O(u). Schmidli & Schmidt [47]. cf. In contrast. and independent of (T1. see Schlegel [316]. m is a (orfinite) . this then easily yields approximations for the finite horizon ruin probabilities (Corollary 5.. 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.
j. x = 0+ and F = (0. to consider only the case Px(w(F`) = 0) 0. For the present purposes it suffices . (note that we allow x. Then (5. St is distributed as y . however . but the example of relevance for us is the following: Proposition 5. for states i.s. Let G denote the distribution of ENt U. 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} .s=j are the transition probabilities for {St}. r. The simplest example is a discrete time discrete state space chain. Proof Starting from Ro = x. and (5. HEAVY TAILS measure on E such that L for all measurable A C E and all t > 0. a familiar case is time reversion (here m is the stationary distribution). {St} and {Rt} are in classical duality w.>N` Ui.00).= y.z) dx G(dz) = ffh(y + z) k(y)dy G(dz). Say {St} is reflected Brownian motion on [0.1 A compound Poisson risk process {Rt} and its associated claim surplus process {St} are in classical duality w . . oo).h.s. resp.r. Rt is distributed as x + t . follows by the substitution y = x .h.). Thus.rij = mjsji where r13. the whole of R and not as usual impose the restrictions x > 0.z.t. {Rt}. in the terminology of general Markov process theory.y = Qx (.t. Lebesgue measure. y = 0). u For F C E.t + EI U. where we can take h. k as indicator functions. The equality of the l. y to vary in. . t. We let QS be the corresponding distribution and Qx.2) means ffh(a. and starting from So = y. m.272 CHAPTER IX. a main difficulty is to make sense to such excursions also when Px(w(F°) = 0) = 1. Then there is a Markov process {Rt} on E such that fE m(dx)h(x)Exk(Rt) = Lm(dy)k(y)Eyh(St) (5.2) with t = 1 means m.2) for all bounded measurable functions h. w(Fc) < oo ) 'In general Markov process theory.)k(x . Sw(F. say. to the r. k on E.
2 Qy.. Sw(F)1 = y) .3 The distribution of r(0) given r(0) < oo.13AB < 1] Proof of Theorem 5. io = x. [note that w(z) < oo a. That is. FINITEHORIZON RUIN PROBABILITIES 273 y E F (in discrete time. 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). The theorem states that the path in (b) has the same distribution as an excursion of {Rt} conditioned to start in y < 0 and to end in x = 0. We can then view Qy. Thus..1 for the case F = (oo.(0)_ = y < 0 is the same as the distribution of w(y) where w(z) = inf It > 0 : Rt = z}.y(2p21 . . and we let Qy y refer to the time reversed excursion ..1 The sample path in (a) is the excursion of {St} conditioned to start in x = 0 and to end in y > 0. .s. in = y. Sw(Fo)_ should be interpreted as Sw(F^)_1). . Sn+1 E Fc) nx. In particular: Corollary 5. this simply means the distribution of the path of {Rt} starting from y and stopped when 0 is hit. 5.SS(F.y as a measure on all strings of the form i0i1 ...= y) Theorem 5 . oo) = r(0) x= St y (a) Figure 5. But in the risk theory example (corresponding to which the sample paths are drawn). w(0.. Sn = in = y.5. . S.y = Qy Q. when p = . x = 0.itt) = P Px(w(Fc) < 00. in with i0. i1. The theorem is illustrated in Fig ..y() = P ({SW(F`)t} 0<t<w(F °) E So = x. 0]. Qx.). QR and QRy are defined similarly. the one in (b) is the time reversed path. in E F.2.. /^s x (S1 = Z1. z > 0. Qx y is the distribution of an excursion of {St} conditioned to start in x E F and terminate in y E F.
S.i„_iEF Similarly. (Fc)1 = y) 00 CHAPTER IX. = in = y.ik_1EF . ...... R .y(inin _ E SYj jEF` 00 Sxik _1 .274 note that Fx(w(Fc) < 00..... Silt' E SO k=1 i1. ... To show Q y x (i0 i 1 ..gilt' k=1 ii .J (i.. Si1y k=1 i1 . in)  Pt' (R1 = ii. in) = oo jEF^ Sxin1 .ik_1EF Sxin_1 . R Qy x(2p21 .. Sn+1 E Fc) n=1 i1. Si11 S 1 . i0) Q x. MY Thus Qx(ioii .. Si l io E mjSjx. ....in1 ........ Si1y 00 jEF° E E 5xik_ 1 .ii .. 21 . Rn = in = x. Rn = in = x.(F<)1 = Y) S S and Qx y( ipil ... Silt' E Sxik_1 .. in E F.ik1EF Similarly but easier Sxin_1 . 2n) = Qx.. note first that Pt' (R l = il. i0 = y... . Rn+1 E FC) TioilTili2.... 2p). in = x... Rn+1 E F`) F (w(Fc) < 00. ..in E F... .. 20 = y.... 2p) when 20. .. in with 20.rin_1in E Txj jEFC m21 s2120 m2252221 m in Ssn n1 mjSjx Mx m2p mil min1 jEF` 1 Sinin _ 1 . ..... t' y and Qy x are measures on all strings of the form ipi l .. S..TI( 2n2n _1 ..... in) = Qx. HEAVY TAILS E E Px (Si = 21i .. . in = x....
ST(o) > y. the case r (O) < oo.r.p. U T(O) = T (u) Y Figure 5.')distribution of Yu is Bo").UBBo(u)].2.'s are defined w. the P(u. the distribution w. To clarify the ideas we first consider the case where ruin occurs already in the first ladder segment .t. Y > y} .2. y > u. Y > u). 5. Let Y = Yl = Sr+( 1) be the value of the claim surplus process just after the first ladder epoch .t. Z) is described in Theorem 111. S. 1 w . Now the P(u. P(o) ). that is.')distribution of Z since P(Z>aIY>u) = 1 °° B(y) B(y + a) dy FLBBo(u) B (y) J°° (z) dy .B(a) +a PBBo(u) . FINITEHORIZON RUIN PROBABILITIES 275 5b The time to ruin Our approach to the study of the asymptotic distribution of the ruin time is to decompose the path of { St} in ladder segments . P(") = P(. Bo") is also the P(u.5. that is.2. That is.r. see Fig.2 The distribution of (Y. Z = Zl = ST+( 1)_ the value just before the first ladder epoch (these r.')density of Y is B(y)/[.v. We are interested in the conditional distribution of T(u) = T(0) given {T(0) < oo. 7(0) < oo. Z follows the excess distribution B(Y) given by B(Y) (x) _ B(y + x)/B(y).(o) > y} = {T(0) < oo. The formulation relevant for the present purposes states that Y has distribution Bo and that conditionally upon Y = y.
Since the conditional distribution of Z is known (viz.d. Zk be defined similarly as Y = Y1. conditionally upon r+ (n) < oo..: r+ (n) < oo. 1/(1 .e. cf. .. It is straightforward that under the conditions of Proposition 1.r+ (n . z ^ oo.. . That is . Fig.1.. r(u)/Z 4 1/(1 .p).p) then yields the final result T(u)/y(u) + W/(1 .. . . However..e.1) of the last ladder segment can be estimated by the same approach as we used above when n = 1. denote the ladder epochs and let Yk.. then by the subexponential property Yn must be large. Bo") ). (Y. Y1 + • • • + Yn > u} denote the number of ladder steps leading to ruin and P("'n) = P(• I r(u) < oo. we get the same asymptotics as when n = 1.. Then.T+(2). P(Z < a I Y > u) 3 0. 2. must be large and Z1.3 implies that the P("'1)distribution of T(u) = r(0) is that of w(Z)..3.p) in Pi"'')distribution. i. > u with high probability. and since its dominates the first n . 4 Assume that Bo E S and that (5. it therefore follows that T(u)/Z converges in Pi"'')probability to 1/(1 . Then Corollary 5. Recall the definition of the auxiliary function y(x) in Section 1. K(u) = n). . the random vectors (YI..18(c) Bo")(yY (u)) + P(W > y) ( 5. Z = ZI but relative to the kth ladder segment.. Zn_1 'typical' which implies that the first n1 ladder segment must be short and the last long. Then 7(u)/y(u) ^ W/(1 . Since w(z)/z a$. more precisely. in particular of Z.p) in F(u) distribution. . 5. In the proof. and YI. are i. HEAVY TAILS Let {w(z)}Z^. Now Bo E S implies that the Bo ")(a) + 0 for any fixed a.p). We now turn to the general case and will see that this conclusion also is true in P(")distribution: Theorem 5 . Z1).. The idea is now to observe that if K(u) = n.i. Hence Z.. We let K(u) = inf In = 1. Z/'y(u) * W in Pi "' ')distribution . this in principle determines the asymptotic behaviour of r(u). .3) holds. Zn). let r+(1) = T(0). i.. a slight rewriting may be more appealing. Yn_1 'typical'.o be defined by w(z) = inf It > 0: Rt = z} where {Rt} is is independent of {St}..3) where the distribution of W is Pareto with mean one in case ( a) and exponential with mean one in case (b). and distributed as (Y.276 CHAPTER IX. the duration T+ (n) . Z).
I A"(u ))II + 0.u) E •) . then IIP( I A'(u)) Taking A'(u) = {Y. Further. P(. Lemma 5... the condition on A'(u) A A"(u) follows from Bo being subexponential (Proposition 1. Y„1. Yn .n) (y1.5..2. A"(u) are events such that P(A'(u) AA"(u)) = o(F (A'(u)) (A = symmetrical difference of events).3 In the following.u) II 0. +Yn1<u. ... I A'(u)) = P(u. . > u}. .n). Proof We shall use the easily proved fact that if A'(u).Bo (ri1) ®B( . II ' II denotes the total variation norm between probability measures and ® product measure.. suitably adapted). .5 Ilp(u.Yl+ +Yf1>u}.. A"(u) _ {K(u)=n} = {Y1+ P(.Yn1iYn . P (Yj.u) E • I A'(u)) = Bo (n1) ®Bou) . FINITEHORIZON RUIN PROBABILITIES 277 16 Z3 Z1 r+(1) T+(1) T+(1) Figure 5.
Z11)..P) Bo(u) for n = 1. Zn)... the discussion just before the statement of Theorem 5. ..1 and Y„ .. P(u) since by Theorem 2. . Let {wl(z)}.. the density of Yn is B(y)/[IBBO(u)].P(Z' E •)II > 0 (here Y. k = 1. Notes and references Excursion theory for general Markov processes is a fairly abstract and advanced topic. ... be independent random vectors such that the conditional distribution of Zk given Y. n_1 < u. For Theorem 5. Y") u etc.P(Y' E •)II * 0.7 O (u. .6. The same calculation as given above when n = 1 shows then that the marginal distribution of Zn is Bou).. Zn) E •) . Proof Let (Y11.. n . Y'. Zn are independent... Z. the F'distribution of r(u) is the same as the P'distribution of w1(Zl) + • • • + wn(Zn).. y > u. and clearly Zi.. . Then according to Section 5a. copies of {w(z)}.6 IIPIu'n ) CHAPTER IX. ... 2..+y 1 p"F(Yn > u) P)Pn1 P/(1 . HEAVY TAILS ((Z1'. in our example Y = (Y1. Z' are arbitrary random vectors. and that Yk has marginal distribution B0 for k = 1.4). (Y. Proof of Theorem 5.' = y is BM.u has distribution Bout That is. n.d..1. then 11P(Z E •) . By Lemma 5. .1).p) < y)..t..P) > y) Corollary 5.4. The first step is to observe that K(u) has a proper limit distribution w. {wn(z)} be i. the marginal distribution of Zk is Bo for k < n...... see Fitzsimmons [144]). wk(Zk) has a proper limit distribution as u + oo for k < n. 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 •) .i. . in particular his Proposition (2..1 P PBo(u) • P(W/(1 ... .Bo (n1) ®Bo' 0.2. whereas wn(Zn) has the same limit behaviour as when n = 1 (cf. .).. + Y" > u) Flul (K (u ) = n) _ Cu) P"F(1'i +.y(u)T) .278 Lemma 5 . Thus F(u'n)(T(u) /7(u) > y) = F(u'n)((wl (Z1) + . . +wn(Z n))l7( u ) > 1y) ^' P(u'n)(wn (Zn)/7(u) > y) 4 NW/(1 .r. It therefore suffices to show that the P(u'")distribution of T(u) has the asserted limit. Y1 +. . Similarly (replace u by 0).
2.y) . the probability that is exceeds u is then B(u .B(u). that MQ becomes large as consequence of one big jump. Proof of Theorem 6. one expects the level y form which the big jump occurs to be 0(1). The form of the result then follows by noting that the process has mean time Ea to make this big jump and that it then occurs with intensity /3B(u).(3 u u J B(y) dy . p(Y) and the result follows. The rigorous proof is. We will show that the stationary density f (x) of {Vt} satisfies f (x) /B(x) r(x) We then get V.1./3Ea B(u). T) when T + oo with u fixed. Extensions to the Markovmodulated model of Chapter VI are in Asmussen & Hojgaard [33]. = supo<t<0. . max VB>0I Vo=0^ o<s<t J11JJJ Lemma 6 .. 6 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate /3.6. Corollary II. More precisely. Then 0 (u) Qf "O ^) dy.1 Assume that B is subexponential and that p(x) > 00. cf. nontrivial and we refer to Asmussen [22]. however. that fo p(x)1 dx < oo.e. The heuristic motivation is the usual in the heavytailed area.(u) = P(V > u) = f f (y) dy . and define the cycle as a = inf{t>0: Vt=0. Assume for simplicity that {Vt} regenerates in state 0 . the results only cover the regularly varying case. claim size distribution B. Theorem 6 . u (6. 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 . i.1) The key step in the proof is the following lemma on the cycle maximum of the associated storage process {Vt}. V. Then P(MT > u) . Asmussen & Teugels [53] studied approximations of i (u. and premium rate p(x) at level x of the reserve.2 Define M. 3. x > oo.
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.280 CHAPTER IX.q(u) Now just use that p(x) * oo implies q (x) + 0. u Notes and references The results are from Asmussen [22]. . 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). Then D(u) = f(u)p(u) and. It is also shown in that paper that typically. Hence f (u)r(u) = D(u) = Do(u) . by regenerative process theory.P(MT > u) $B(u) Ft µ(1 . D(u) = DQ(u)/µ. where also the (easier) case of p(x) having a finite limit is treated . there exist constants c(u) 4 0 such that the limiting distribution of r(u)/c(u) given r(u) < oo is exponential.
where a2 = Var(Z ).2) is an asymptotic 95% confidence interval . topics of direct relevance for the study of ruin probabilities are treated in more depth..i. la The crude Monte Carlo method Let Z be some random variable and assume that we want to evaluate z = EZ in a situation where z is not available analytically but Z can be simulated. and this is the form in which the result of the simulation experiment is commonly reported. The crude Monte Carlo ( CMC) method then amounts to simulating i.96s z f (1. 281 . vrN(z . We shall be brief concerning general aspects and refer to standard textbooks like Bratley. Rubinstein [310] or Rubinstein & Melamed [311] for more detail . Hence 1.d.z) 4 N(0. estimating z by the empirical mean (Z1 + • • + ZN)/N and the variance of Z by the empirical variance N s2 = E(Z{  N 2. Ripley [304]. replicates Zl.. z) 2 = Zit NE ii ii According to standard central limit theory . Fox & Schrage [77]. 4Z). . ZN.Chapter X Simulation methodology 1 Generalities This section gives a summary of some basic issues in simulation and Monte Carlo methods ..
Sections 24 deal with alternative representations of Vi(u) allowing to overcome this difficulty. generated at the same time as Z. We survey two methods which are used below to study ruin probabilities. SIMULATION METHODOLOGY In the setting of ruin probabilities. one can argue that unless Var(Z') is considerable smaller than Var(Z). variance reduction is hardly worthwhile. so that Z' is a candidate for a Monte Carlo estimator of z. we then have EZ = EZ = z. v. it is straightforward to use the CMC method to simulate the finite horizon ruin probability z = i. Typically variance reduction involves both some theoretical idea (in some cases also a mathematical calculation). Further. 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'.b(u. an added programming effort. writing Var(Z) = Var(E [Z I Y]) + E(Var[Z I Y]) . This is a classical area of the simulation literature. However. T): just simulate the risk process {Rt} up to time T (or T n 7(u)) and let Z be the indicator that ruin has occurred. there are others which are widely used in other areas and potentially useful also for ruin probabilities. conditional Monte Carlo and importance sampling. and in most cases this modest increase of N is totally unproblematic. Conditional Monte Carlo Let Z be a CMC estimator and Y some other r . The situation is more intricate for the infinite horizon ruin probability 0(u). and a longer CPU time to produce one replication. Therefore. 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. Say that Var(Z') = Var(Z)/2. We mention in particular ( regression adjusted) control variates and common random numbers. lb Variance reduction techniques The purpose of the techniques we study is to reduce the variance on a CMC estimator Z of z. typically by modifying Z to an alternative estimator Z' with EZ' = EZ = z and (hopefully) Var(Z') < Var(Z).282 CHAPTER X. and many sophisticated ideas have been developed. Z = I inf Rt < 0 (0<t<T = I('r(u) < T). Letting Z' = E[Z I Y].
Variance reduction may or may not be obtained: it depends on the choice of the alternative measure P. LN) from P and uses the estimator N zrs = N > L:Zj i=1 and the confidence interval zrs f 1.96 sis v^ N 2 1 where srs = N j(LiZi . . (1. L1).e.zrs) = 2 1 N 2 2 2 i=1 i=1 N > Lt Zi . L such that z = EZ = E[LZ].3) Thus. the argument cheats because we are simulating since z is not avaliable analytically. . Thus. This may also be difficult to assess .E [Z Z]2 = z2 . using the CMC method one generates (Z1. L = z/Z (the event {Z = 0} is not a concern because P(Z = 0) = 0). In order to achieve (1. one would try to choose P to make large values of Z more likely.. Then z Var(LZ) = E(LZ)2 . GENERALITIES 283 and ignoring the last term shows that Var(Z') < Var(Z) so that conditional Monte Carlo always leads to variance reduction. but tentatively. a crucial observation is that there is an optimal choice of P: define P by dP/dP = Z/EZ = Z/z. . Nevertheless.z2 = 0.. i.zrs. 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. Thus we cannot compute L = Z/z (further. To this end.[E(LZ)] = E Z2 Zz . the obvious possibility is to take F and P mutually equivalent and L = dP/dP as the likelihood ratio. it may often be impossible to describe P in such a way that it is straightforward to simulate from P). even if the optimal change of measure is not practical.3). However. (ZN. and the problem is to make an efficient choice. it appears that we have produced an estimator with variance zero. it gives a guidance: choose P such that dP/dP is as proportional to Z as possible.1.v.
96 2Z ( 1 . and further it is usually not practicable to simulate from P(•IA). let z(u) = P(A(u)). However.5 or even much smaller . the optimal P is the conditional distribution given A.. Two established efficiency criteria in rare events simulation are bounded relative error and logarithmic efficiency. assume that the rare event A = A(u) depends on a parameter u (say A = {r(u) < oo}).96oz /(zV) = 0. We shall focuse on importance sampling as a potential (though not the only) way to overcome this problem. N . say of the order 103 or less. A = {T(u) < T} or A = {r(u) < oo} and the rare events assumption amount to u being large. Z z V5 In other words . For each u. u + oo. SIMULATION METHODOLOGY 1c Rare events simulation The problem is to estimate z = P(A) when z is small . we may try to make P look as much like P(•IA) as possible.e.1. in terms of the halfwidth of the confidence interval. z I. as is the case of typical interest. We then .. if z is small.100 .e. large sample sizes are required. i. and let Z(u) be a Monte Carlo estimator of z(u).z) 1001.B = iP(AB) = P(BIA). assume that the A(u) are rare in the sense that z(u) * 0.284 CHAPTER X. To introduce these.z) 1 > 00.96 2 z2 z increases like z1 as z . 10 . 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.e. Z = I(A) and A is a rare event. but if the point estimate z is of the order 105. This leads to the equation 1. Again.1. it does not help telling whether z is of the magnitude 104.0. say 10%. The optimal change of measure ( as discussed above) is given by P(B) = E [ Z] i. the issue is not so much that the precision is good as that relative precision is bad: oZ z(1 . Thus. a confidence interval of width 10 4 may look small.z) which tends to zero as z ^ 0. I. However. The CMC method leads to a variance of oZ = z(1 . Another way to illustrate the problem is in terms of the sample size N needed to acquire a given relative precision . In ruin probability theory. An example where this works out nicely is given in Section 3.
2 Simulation via the PollaczeckKhinchine formula For the compound Poisson model.log Var(Z(u)) lim inf > 2 u+oo . 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. so that NE (u) may go to infinity. let Z +.log z(u) of (1. the PollaczeckKhinchine formula III. . Notes and references For surveys on rare events simulation. but as a CMC method .. let Z +. F(K = k) = (1 . 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. P(K = k) = (1 .1. it is appealing to combine with some variance reduction method . which gives a logarithmically efficient estimator .4) for any e > 0. where X1.0. with common density bo(x) = B(x)/µB and K is geometric with parameter p..(2..4). it is not efficient for large u . Therefore . Otherwise. 2. . where M = X1 + • • • + XK. 3.d. Generate X1. If M > u. and in practice. see Asmussen & Rubinstein [45] and Heidelberger [190].2.X1 + + XK. are i.p)pk. Var(Z(u)) hm sup U+00 z (u) 2E < oo (1.e. X2. this means that the sample size N = NE(u) required to obtain a given fixed relative precision (say a =10%) remains bounded. XK from the density bo(x).i. Let M . logarithmic efficiency is almost as good as bounded relative error. Generate K as geometric. The algorithm gives a solution to the infinite horizon problem . . i. We shall here present an algorithm developed by Asmussen & Binswanger [ 271..1) may be written as V) (u) = P(M > u). the mathematical definition puts certain restrictions on this growth rate. Thus. The term logarithmic comes from the equivalent form . where Z = I(M > u) may be generated as follows: 1. According to the above discussion. O (u) = z = EZ. This allows Var(Z(u)) to decrease slightly slower than z(u)2. However.p)pk.
Z(1)(u) is defined as 0)..L(x)/x`' with a > 0 and L(x) slowly varying. . Z(1) (u) has a smaller variance than Zl (x).X1 .2..Xl ..b(u) = P (Xl +•••+XK>u) = EF[Xl + .. XK1....286 CHAPTER X.1) V)(u) . asymptotically it presents no improvement : the variance is of the same order of magnitude F(x). and considering only the remaining ones. y < 0)... For the simulation. Then (cf. . form the order statistics X(1) < X(2) < . Xl > x... 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.. XK. conditional probability. note first that To check the formula for the P(X(n) > x I X(1). and the problem is to produce an estimator Z(u) with a variance going to zero not slower (in the logarithmic sense ) than Bo(u)2. Theorem IX... we thus generate K and X1i .S( K_1)) V X(K1)) / Bo(X(K 1)) where S(K_l) = X(1) + X(2) + • • • + X(K_1).p)Bo(x)..XK_1 and let Z( 1)(u) = Bo (Y) (if K = 0. and let Z(2)(u) = _ P (SK B0((u > u I X(l).X(K1)) .SK1)2. To see this. X1 + + XK_ 1 > x when X1 > x..XK_1). As a conditional Monte Carlo estimator .+XK > uIXl. assume in the following that Bo(x) .. .X(n_1)) Bo(X(„_l) V X) Bo(X(n1)) . A first obvious idea is to use conditional Monte Carlo: write i. So. The idea of [27] is to avoid this problem by discarding the largest X.p/(l .. However....X(2).. and that Bo(y) = 1.. K > 2] = P2p(Xl > x) = P2Bo(x) (here we used that by positivity of the X. SIMULATION METHODOLOGY when the claim size distribution B (and hence Bo) has a regularly varying tail.XK1] = EBo(uX1 .. Thus. < X(K) throw away the largest one X(K)..X(2).. we generate only X1.. .. compute Y = u . just note that EZ(1)(u ) 2 > E[Bo (x .
3. The algorithm is sofar the only efficient one which has been developed for the heavytailed case. the continuoustime process {St} is simulated by considering it at the discrete epochs {Qk} corresponding to claim arrivals. B. . X(n1)) Bo((x .. X (. BL instead of 0. it must be noted that a main restriction of both algorithms is that they are so intimately tied up with the compound Poisson model because the explicit form of the PollaczeckKhinchine formula is crucial (say.S(n_1) I X(1).. For practical purposes..u is the representation 0(u) = e7sr(u) overshoot (cf. Asmussen ._1) > P(X(n) > _ X X(1). . and that paper contains one more logarithmically efficient algorithm for the compound Poisson model using the Pollaczeck. that is. 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.Khinchine formula and importance sampling .1) . Notes and references The proof of Theorem 2. IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 287 We then get P(S" > x I X( 1).1 is elementary but lengty.S (n1)) V X (. X(2). Then the algorithm given by { Z (2) (u) } is logarithmically efficient. l)) BO(X(n1)) Theorem 2 . . l)) . the algorithm for generating Z = Z(u) is: 1.y. 3 Importance sampling via Lundberg conjugation We consider again the compound Poisson model and assume the conditions of Ce7". Thus. for the purpose of recording Z(u) = erysr(u). use the the CramerLundberg approximation so that z(u) = '(u) = e7"ELe7E(") where ^(u) = ST(") .. BL by I3L = /3B[y].. . and simulate from FL. in the renewal or Markov. X . . 111. . X(2). X(2). Binswanger and HOjgaard of [28] give a general survey of rare events simulation for heavy tailed distributions . and we refer to [27]. BL(dx) = e7sB(dx)/B[y]..modulated model P(r+ < oo) and G+ are not explicit ).5). 1 Assume that Bo (x) = L(x)/x° with L(x) slowly varying. using 13L. Compute y > 0 as solution of the Lundberg equation 0 = K(y) = )3(B[y] . However . and define )3L. X(n1)) P(X(TZ) + S(.
Ti. Generate T as being exponential with parameter . Xi = U. and we have: Theorem 3. SIMULATION METHODOLOGY 3. There are various intuitive reasons that this should be a good algorithm.i. A > AL as in Chapter V. Let Sf0 CHAPTER X. X2. the results of IV. + X. In detail . = X1 + . b different from . to deal with the infinite horizon problem .(u)) are asymptotically coincide on {r(u)} < oo. one must restrict attention to the case 4µB > 1. let Z F e_'s... We formulate this in a slightly more general random walk setting '. 4. and avoid simulating the known part e7". More precisely. be i. The proof is given below as a corollary to Theorem 3. and the change of measure F r FL corresponds to B > BL.2 The estimator (3.. B) is not logarithmically efficient when (/3. so that changing the measure to FL is close to the optimal scheme for importance sampling . BL). Let X1.1 The estimator Z(u) = e'rs* "u) (simulated from FL) has bounded relative error.F. Otherwise. Let S . namely ELery£("). and assume that µF < 0 and that F[y] = 1.1) (simulated with parameters ^3. Let FL (dx) = 'For the renewal model.S+U .3. let S. If S > u. Proof Just note that EZ(u)2 < e . We may expect a small variance since we have used our knowledge of the form of 0(u) to isolate what is really unknown.. return to 3.Q.l3 and U from B.. u It is tempting to ask whether choosing importance sampling parameters .T. . with distribution F.r(u) < oo) = 1. cf. . the discussion at the end of Section 1b. The answer is no. 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.2ryu _ z (u)2/C2.QL.d. M(u) = inf {n : S„ > u}. The algorithm generalizes easily to the renewal model . P'[y] < oo for some ry > 0. b) # (/3L. r(u) < oo) and FL (both measures restricted to..7 tell that P(. In fact: Theorem 3..288 2. It resolves the infinite horizon problem since FL(. BL could improve the variance of the estimator .
2ryELXi.i. More generally.2'X1 .. The importance sampling estimator is then Z( u) = e'rSM( ).+KM(u)}.P = FL. . where Kl og (X) (j) 2 ) = log dFL (Xi) ..3 The estimator (3... F(XM(u)).d..yu+elu u +oo etry' 1 > lim up C2e2.2) (simulated with distribution F of the X3 has bounded relative error when . EFZ(u)2 EFZ(u)2 lim sup z(u)2eeU = lim cop C2e2. + KM(u))} = exp {ELM(u)(E . Proof The first statement is proved exactly as Theorem 3 . K2.yu = G. Since ELM(u)/u + 1/ELXi. 1. are i.. For the second. By the chain rule for RadonNikodym derivatives. it thus follows that for 0 < e < e'/ELXi..2) dF Theorem 3. where e' = EL Iog dFL (Xi) > 0 by the information inequality.3. Jensen's inequality and Wald's identity yield EpZ(u)2 > exp {EL(K1 + . When F # FL. Here ELK..2ryELXi)} . (3. is not logarithmically efficient. = c'. IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 289 e7yF(dx).. let F be an importance sampling distribution equivalent to F and M(u) dF Z(u) _ I (Xi) . . Since K1. write W(F IF) _ F(XI). EFZ(u)2 = EeW2(FIF) = Ep [W2(FIFL)W2(FLIF)] = EL [W2 ( FIFL)w(FLIF)] = ELexp {Kl+.2 > 0.
In fact: Proposition 4. U".4.'(y). 4 Importance sampling for the finite horizon case The problem is to produce efficient simulation estimators for '0 (u. Consider compound Poisson risk process with intensities /3'.T' has a left exponential tail with rate /3' and U" .'(y) or y > 1/r. SIMULATION METHODOLOGY u Proof of Theorem 3. see e.3.T". U' . T".1 is from Lehtonen & Nyrhinen [244]. we write T = yu. the references in Asmussen & Rubinstein [45] and Heidelberger [190]. The easy case is y > 1/k'(y) where O(u. . Further discussion is in Lehtonen & Nyrhinen [245].B'=B".290 which completes the proof.4 indicate that we can expect a major difference according to whether y < 1/r. /3'eQ'YR'( x + y) dy = .i. claim size distributions B'.T' D U" . all that needs to be shown is that if U' . CHAPTER X. generic interarrival times T' . The extension to the Markovian environment model is straightforward and was suggested in Asmussen [ 16]. then the estimator Z(u) = e7Sr(°)I(r(u) < yu) (simulated with parameters /3L. optimality is discussed in a heavy traffic limit y 10 rather than when u + oo. Then according to Theorem 3.T". First by the memoryless distribution of the exponential distribution .1 If y > 1/ic'('y).3'eO'x f f P (U" . This immediately yields / = 3". yu) is close to zk(u). U' . B" and generic claim sizes U'. u Notes and references The importance sampling method was suggested by Siegmund [343] for discrete time random walks and further studied by Asmussen [ 13] in the setting of compound Poisson risk models . /3".2. The results of IV. T) with T < oo. The optimality result Theorem 3.T' = U" . with the present (shorter and more elementary) proof taken from Asmussen & Rubinstein [45]. BL) has bounded relative error. from 3' P(U'T'>x) ^ = ^ eQ'zB (z) dz. so that one would expect the change of measure F 4 FL to produce close to optimal results.3"eQ x 0 J eQ zB (z) dz x (x > 0) and /3' = /3". then /3' B' = B". As in IV. Next.T" > x) J /3"e0 yB (x + y) dy = . In [13].g. we conclude by differentiation that Bo(x)=B' (x)forallx > 0.T" has a left exponential tail with rate /3'.e. The queueing literature on related algorithms is extensive .
and in fact.yy> 2 . Remark 4 .(ay). the result follows as in the proof of Theorem 3.8 has a stronger conclusion than (4. T(u) < yu] e Hence by (4. IMPORTANCE SAMPLING FOR THE FINITE HORIZON CASE 291 Proof The assumption y > 1/n'(y) ensures that 1fi(u. Further . we have ic(ay ) > 0 and get Eay Z(u)2 = Eay [e  2aySr( u)+2r(u )r.4.1) which is all that is needed here can be showed much easier .(u) * 1 (Theorem IV.log Var(Z(u)) l im of . yu) is of order of magnitude a71.log 4')u) 4 u (Theorem IV.'(7). Bounding u ELZ(u)2 above by a7u. T( u) < yu] e2ryyuEay le 2ay^(u).1). Let Qy2 = . lim inf uoo 27yu . Bay) is logarithmically efficient. We recall that ay is defined as the solution of a'(a) = 1/y. Proof Since ryy > y. that ryy = ay . 7y (4. (4.3) (simulated with parameters /gay.1). yu) in the sense that .4.to g x ( u ) u u so that (1. We next consider the case y < 1/r.5) follows.3) and we have: Theorem 4.4. yu)/z.yu. T(u) < yu] . one would expect that the change of measure F Pay is in some sense optimal. .2 The estimator (4. yu) = eayu Eay Leay^(u)+r(u)K(ay). 3 Theorem IV.log Var(Z(u)) _ .yk(ay) determines the order of magnitude of z'(u.O(u.1) so that z(u) = zP(u. (4. The corresponding estimator is Z(u) = eavS' ( u)+T(u)K (ay)I(T( u) < yu).8).4.1. and that ryy > ry.2) Since the definition of ay is equivalent to Eay r(u) .
N(0. SIMULATION METHODOLOGY Vara„ (r(u))/u so that (T(u) . 0 rather than when u 3 oo.2) .1) where the identity for Vi(u) requires that Vt has a limit in distribution V.o . 0 Notes and references The results of the present section are new. Hence lira inf log ryyu + vyu 1/2 tc(ay) . zi(u) = INV. there exists a dual process { V t} such that i.1) may be useful.4. Z (u)2 above. > u) = E f I(VV > u) dt 0 (5.1): then by Proposition A1. yu . and (5.3: for many risk processes {Rt }.2). However.T) = P O<t<T inf Rt < 0 = P(VT > u).4).7y x(u) > hm inf u+Oo U . 5 Regenerative simulation Our starting point is the duality representations in 11. Then z(u) = Eay Z(u) > Eay avS'(u)+T( u)k(av 1 ).1) is used to study Voo by simulating {Rt} (for example.(av)Eav l e.yu)/(uyu1/2) . yu . '%(u) = P I info Rt < 0) = P(VV > u).ryyu +oy u1/2K'(av)Eo l v 1/2) where the last step follows by Stam's lemma (Proposition IV.a vt(u).1) (see Proposition IV.4. related discussion is given in a heavy traffic limit q J.u1/2 < r(u) < yu Le ] l = e. In most of the simulation literature (say in queueing applications).Qyu1/2 < T(u) C yu e.yu1/2 <1 T(u) < yu l r > e7vu +avul/ 2r.3.b(u. In Asmussen [13]. we believe that there are examples also in risk theory where (5.292 CHAPTER X. the algorithm in Section 3 produces simulation estimates for the tail P(W > u) of the GI/G/1 waiting time W).uaoo U That lim sup < follows similarly but easier as when estimating En.o.a yu +l/ur' (av)Ei`av reav^(u)+(T(++)(U) yu . (5. One main example is {Vt} being regenerative (see A. the object of interest is {Vt} rather than {Rt}..
is the cycle length. To derive confidence intervals . a standard transformation technique (sometimes called the delta method) yields 1 V 2 (h (Zi. . Then (Z1z1i Z2z2 ) 4 N2(0.. > u) (and more general expectations Eg(V.. Zl the LLN yields Z1 a$' Z(1) +. )). Z1 = (Zl1i +. EZ2'i = z2 = E Thus. which we survey below . consider first the case of independent cycles . Thus the method provides one answer on to how to avoid to simulate { Rt} for an infinitely long time period. Then Z(1). 02) (5. (u) ?2 = E fo I(Vt > u) dt = 0( u ) zl Ew as N > oo. Taking h(zl. REGENERATIVE SIMULATION 293 where w is the generic cycle for {Vt}.+Z(N) z 1. + Z1N>) . and Z2'>) where Zi'i = w.. For details .. Z2'> the time during the cycle where { Vt} exceeds u and zj = EZJ'). let E denote the 2 x 2 covariance matrix of Z('). Simulate a zerodelayed version of {V t } until a large number N of cycles have been completed. z2)) > N(O.5. For the ith cycle. the regenerative estimator z%(u) is consistent. Z(N) are i . 2. i (^(u) . The method of regenerative simulation. z2) z2/z1 yields Vh = (z2/z2 1/zl).h (zl. Vh = (8h/8z1 8h/ 8z2). Thus..t(u)) 4 N(0. Z2 a4* z2. Z2 = N (X21' + . . Therefore . letting J0 'o I (Vt > u) dt .. i. + Z2N)) . j = 1.3) . provides estimates for F ( V.. record Zi'i = (Z1')... Z2 ..E).d. oh) for h : R2 ^ R and Ch = VhEVh. EZ1'i = z1 = Ew..
() dx = E[SZ] f(X.g. 6 Sensitivity analysis We return to the problem of 111 . the expectation z = EZ of a single r. say risk processes with a complicated structure of the point process of claim arrivals and heavy tailed claims . However . Then z(() = f cp(x) f (x. () dx so that differentiation yields zS d( fco(x)f(x. see e. There is potential also for combining with some variance reduction method.Z) ^Z(=) .294 where 01 2 CHAPTER X. () dx f Ax) (dl d()f (x' () f ( z. asymptotic estimates were derived using the renewal equation for z /i(u). Z of the form Z = ^p(X) where X is a r . Here are the ideas of the two main appfoaches in today 's simulation literature: The score function ( SF) method .C)dx = f w(x) d( f ( x. v. We here consider simulation algorithms which have the potential of applying to substantially more complex situations. in some situations it may be the only one resolving the infinite horizon problem .g S12 (5. Rubinstein [310] and Rubinstein & Melamed [311]. consider an extremely simple example .2. Notes and references The literature on regenerative simulation is extensive.v.5) Z1 Z1 Z1 and the 95% confidence interval is z1 (u) ± 1. 9.2 E1 2 z1 z1 Z2 The natural estimator for E is the empirical covariance matrix N S = N 1 12 (ZW . to evaluate the sensitivity z/i( (u ) = (d/d() 0(u) where ( is some parameter governing the risk process . () depending on C. In 111. Let X have a density f (x. Before going into the complications of ruin probabilities .0 .9. The regenerative method is not likely to be efficient for large u but rather a brute force one.96s/v"N. SIMULATION METHODOLOGY 2 Eli = Z2 z1 + 2 E22 .z^ i=1 so a2 can be estimated by 2 2 = 72 S11+ 12 S22 . with distribution depending on a parameter (.
Then z(() = Ecp(h(U. For the SF method. The following example demonstrates how the SF method handles this situation. ()) is 0 w . Thus . ()) is 0 for C < Co and 1 for C > Co so that the sample path derivative cp'(h(U. /3o is M(u) Oe (3T: < oo) .() d( is the score function familiar from statistics . SZ is an unbiased Monte Carlo estimator of z(. I(r(u) . Infinitesimal perturbation analysis (IPA) uses sample path derivatives. C). () _ (eSx. if f (x. this is usually unproblematic and involves some application of dominated convergence . just take cp as an indicator function . say W(x) = I(x > xo) and assume that h(U. zc = E [d co(h(U. one can take h (U. with density f (x. r(u) = Tl + • • • +TM(u)). SENSITIVITY ANALYSIS where 295 S = (d/d()f (X. The likelihood ratio up to r(u) for two Poisson processes with rates /3. giving h( (U. = E [`d (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. for some Co = (o(U). ( where h( (u.v. ()). C)). however . cp' (h(U.1 Consider the sensitivity tka(u) w. C) f(X. /3 is 0. IPA will estimate zS by 0 which is obviously not correct. one. this phenomenon is particularly unpleasant since indicators occur widely in the CMC estimators . () where U is uniform(0. the Poisson rate /3 in the compound Poisson model. For IPA there are.t.1). A related difficulty occurs in situations involving the Poisson number Nt of claims: also here the sample path derivative w. 11 /3oeOoT. ()) d( hc(U.r. () is an unbiased Monte Carlo estimator of zS.log U/(. () = (8/8()h (u.t.r. p. () is increasing in C. cp(h(U. Thus.6. To see this. () = log U/(2. Then . For example . In the setting of ruin probabilities . nonpathological examples where sample path derivatives fail to produce estimators with the correct expectation. Example 6 . () = d log f (X. ()) h((U. () can be generated as h(U. Let M(u) be the number of claims up to the time r(u) of ruin (thus. () Thus. () = . So assume that a r.
SIMULATION METHODOLOGY Taking expectation.3 (u) is of the order of magnitude ue7u. However. whereas for the SF method we refer to Rubinstein & Shapiro [312].t. we get 1 M(u) 00(u) = E (_Ti)I(T(U)<) E [(M(u) . in part for different measures of risk than ruin probabilities.0(1) so that in fact the estimator Zf(u) has bounded relative error. different parameters.9 .3L.T(u)) e7uerVu) for ?P.t. In the setting of ruin probabilities. ) we have VarL(ZQ(u)) ZO(u)2 O(u2)e2 u2e2ryu yu . We recall (Proposition 111. 0 Notes and references A survey of IPA and references is given by Glasserman [161] (see also Suri [358] for a tutorial). Thus. 4) that V5. To resolve the infinite horizon problem .r. the estimation of z(ip(u) is subject to the same problem concerning relative precision as in rare events simulation . differentiating w. .r. We then arrive at the estimator ZZ(u) = (M(u) . for different models and for the sensitivities w. j3 and letting flo = 0. since ELZp(u)2 < (M(U) _T(u) \ 1 2 a2ryu = O(u2)e27u.3 (u) (to generate Zp (u). the risk process should be simulated with parameters .T(u)) I(T(u) < co) ] . Example 6. There have been much work on resolving the difficulties associated with IPA pointed out above. BL).296 CHAPTER X. a relevant reference is VazquezAbad [374]. change the measure to FL as when simulating tp(u).1 is from Asmussen & Rubinstein [46] who also work out a number of similar sensitivity estimators.
X2. are i.Chapter XI Miscellaneous topics 1 The ruin problem for Bernoulli random walk and Brownian motion. and {1. The twobarrier ruin problem The twobarrier ruin probability 0. in most cases ... a) = r(u)).. in the Bernoulli random walk example below. Consider first a Bernoulli random walk.+• • •+X..g. T+(a) = inf It > 0 : Rt > al.P(•r(u..d..1. Besides its intrinsic interest . That is. Y'a(U) = P(T (u) = r+(a)) = 1 ..1}valued . Oa(U ) can also be a useful vehicle for computing t/i(u) by letting a * oo. .. as e. 297 .(u) is defined as the probability of being ruined (starting from u) before the reserve reaches level a > u. }).i. either this makes no difference (P(R.. 'Note that in the definition of r(u ) differs from the rest of the book where we use r(u) = inf {t > 0 : Rt < 0} ( two sharp inequalities ). T(u. R„ = u+X. defined as Ro = u (with u E {0. where X1. a) = r(u) A T+(a).. with P(Xk = 1) = 9.(u) = 0 ) = 0) or it is trivial to translate from one setup to the other. wherel T(u) = inf {t > 0 : Rt < 0} .
C1_0\a. u + 1.a(u)).1.(1B)u oJ 0.0)/0..1) o If 0 = 1/ 2. The martingale is then {zuzXl+•••+X„ } = {zR° }.1) = (19)4/'0(a3)+9ba(a1). z and the solution is z = (1 . In a general random walk setting .a) Y. = (1 .a) = 0) + zap ( R. where a is any number such that Ee°X = F[a] <oo.2) Oa(a . and the other more advanced but applicable also in some other settings.. We choose a = ry where ry is the Lundberg exponent.e. u Proof 2.+Xn) F[ a]n n=0. one elementary but difficult to generalize to other models.y] = 1.2). i. 7/la(a . and insertion shows that ( 1. Wald's exponential martingale is defined as in 11. tba(2) _ (1 . then 'Oa(u) _ au a We give two proofs . (1.4) by ea(u+Xl+.1 For a Bernoulli random walk with 0 0 1/2. zu = EzRO = EzRT(u.1) is solution. Conditioning upon X1 yields immediately the recursion 'a(1) = 19+00a(2). MISCELLANEOUS TOPICS Proposition 1...o)T/la (1) + 8z/'u(3). = z°Va(u) + za(1  .. The Lundberg equation becomes 1=F[ry]=(19)+9z. Proof 1.o» = z°P (RT ( u.(u) I\ e = 1 oa ' ()i a = u. and in view of the discrete nature of a Bernoulli random walk we write z = e7.. By optional stopping.r(u..(4..o)'t/1a(a .298 CHAPTER XI. the solution of F[..
If p<0. If p = 0. . u Proposition 1. {Rt} is itself a martingale and just the same calculation as in the u proof of Proposition 1.1 If p = 0. TWO BARRIERS 299 and solving for 4/la(u) yields t/ia(u) = (za . } yields e7u = Ee7R° = e°Wa(u) + e7a(1 .u)/u. (1.2 For a Bernoulli random walk with 9 > 1/2.e7u)/(e7° .zu)/(za .ba(u) = e2µa .u) = et(a2 /2 +aµ) the Lundberg equation is rye/2'yp = 0 with solution y = 2p.1. (1.• a2µa e2µu . Applying optional stopping to the exponential martingale {e7R. However.1) for p # 0.5) .} is then itself a martingale and we get in a similar manner u = ER° = ER ra( u) = 0 • Y'a (u) + all  au Y'a( u)). then Vi(u) = 1. RANDOM WALK. then Proof Since 'Oa (U)  au a Eea(R°.1). i1(u) = e211 .1). Then for p 0 0.3 Let {Rt} be Brownian motion starting from u and with drift p and unit variance .0a(u)). 1h (u) = a el u \1 If 9 < 1/ 2.4 For a Brownian motion with drift u > 0. BROWNIAN MOTION.2) is trivial (z = 1).. pa( u) _ u Corollary 1. Proof Let a+ oo in (1. If 9 = 1/2.1 yields 't/la(u) = (a . Corollary 1. thenz1 (u)=1. and solving for 9/la(u) yields Z/)a(u) = (e 76 . {R.
It may then be easier to first compute the onebarrier ruin probability O(u): Proposition 1.3. say.vi(a) Proof By the upwards skipfree property. 5). 7/'(u) = 1).0(a) 0 < u < a. the paths are upwards skipfree but not downwards.5a). For most standard risk processes .a) < 0) + e7°P (R(u. letting a * oo yields the standard expression pe7u for .a) = a ) + e ' ° ( 1 . 1 . this immediately yields (1. and hence e7u = Ee7Ro E [e7R(.a ) < 0) + e 7aF ( R (u.e7a Again .+^a(u))^(a) If 7k(a) < 1. valid if p < 1 (otherwise . implying R(u.0 (u) (where u p =. 0. Here the undershoot under 0 is exponential with rate 5.7). Ic 5ry 'pa(u) Using y = 6 . 7O(u) = 7/la(u) + (1 . (1.7/la(u)). Here is one more case where this is feasible: Example 1.. However.5 Consider the compound Poisson model with exponential claims (with rate. CHAPTER XI.7) .4).a) = r+ (a)} and similarly for the boundary 0. we obtain 'Oa a7u .6 If the paths of {Rt} are upwards skipfree and 7//(a) < 1.a) = a on {r (u. a) I R(u a ) < 0] P (R(u . and thus one encounters the problem of controlling the undershoot under level 0. (u) _ O(u) .300 Proof Let a * oo in (1. VIII.a) = a) = 5 y = P (R (u. passing to even more general cases the method quickly becomes unfeasible (see.616).e7a (u) = 6 /0 . MISCELLANEOUS TOPICS u The reason that the calculations work out so smoothly for Bernoulli random walks and Brownian motion is the skipfree nature of the paths. . however.
8) Proof In terms of the claim surplus process { St} = {u . we have ili(u. + µ2T) } . Hence P(MT>u. (1. TWO BARRIERS 301 Note thas this argument has already been used in VII.1. We now return to Bernoulli random walk and Brownian motion to consider finite horizon ruin probabilities. of r(u) are ( U2 Pµ (T(u ) E dT) = 2^T 3/2 exp µu .1a for computing ruin probabilities for a twostep premium function. For µ # 0. = eµuTµ2/2Po (T( u) E dT) 2 eµuTµ2/2 u T3/2 ex p u 27r p 12 T .T) P(MT > u) where MT = maxo<t<T St. MT > U) = P(ST > u) + P(ST > u) (1.2 . 0(u.11) VIT ) Proof For p = 0.11 ) is the same as (1. P(MT > u) = P(ST > u) + P(ST < u. MT > u) = P (ST > u) + P (ST > u. 10) follows then by straightforward differentiation.µ%T (1.ST<u) = P(MT>u. in particular symmetric so that from time r(u) (where the level is level u) it is equally likely to go to levels < u and levels > u in time T .. Here {St } is Brownian motion with drift 0 (starting from 0). For the symmetric (drift 0) case these are easily computable by means of the reflection principle: Proposition 1.4) I = . (i).8 ).µ so that {St} is Brownian motion with drift µ .µ T I + e2µ"4) ( . = 1 .f. T(u) E dT. Corollary 1.r(u). ( 1. BROWNIAN MOTION. RANDOM WALK.7 For Brownian motion with drift 0.. and (1 .9) = 2P(ST > u).. and hence Pµ('r(u) E dT) = Eo [e µsr(.ST>U).. the density dPµ / dP0 of St is eµsttµ2/2.10) Pµ (T(u) < T) !. (1.. Then the density and c.d.Rt}.)_ _( u)µ2 /2.8 Let {Rt} be Brownian motion with drift . T ) = P(T(u) < T ) = 241.
12) is the same as ( 1. 226). T are integervalued and nonnegative.. such that the drift µ(x) and the variance a2(x) are continuous functions of x and that a2(x) > 0 . and in a similar spirit as in VII. The expression for F ( ST = v) is just a standard formula for the u binomial distribution. e. MISCELLANEOUS TOPICS which is the same as (1. Theorem 1. S(oo) = f c s(y)dy. Breiman [78] or Karlin & Taylor [222] p.T+2.T (1.T) = P(ST = u) + 2P (ST > u). is finite for all x > 0.10 Consider a diffusion process {Rt} on [0. as defined above as the probability of actually hitting 0. We finally consider a general diffusion {Rt} on [0. If this assumption fails... and (1. Thus.10) and that the value at 0 is 0. the behaviour at the boundary 0 is more complicated and it may happen.T2..9) goes through unchanged.10). whenever u. We assume that u(x) and a2 (x) are continuous with a2 (x) > 0 for x > 0.302 CHAPTER XI. is zero for all u > 0 but that nevertheless Rt ^4 0 (the problem leads into the complicated area of boundary classification of diffusions.T)dx. u Small modifications also apply to Bernoulli random walks: Proposition 1.g. oo) with drift µ(x) and variance a2 (x) at x. Here {2T( (v}TT)/2) v=T.3 we can define the local adjustment coefficient y(x) as the one 2µ(x)/a2(x) for the locally approximating Brownian motion.13) The following results gives a complete solution of the ruin problem for the diffusion subject to the assumption that S(x).9 For Bernoulli random walk with 9 = 1/2. The same argument as used for Corollary 1. see e.11) then follows by checking that the derivative of the r.g. close to x {Rt} behaves as Brownian motion with drift µ = u(x) and variance a2 = a2(x). Vi(u.12) P(ST = v) = 0 otherwise. S(x) = f x s(y)dy. that 0(u).s. as defined in (1. Proof The argument leading to ( 1. (1. is (1.13) with 0 as lower limit of integration. Let s(y) = ef0 ry(.9). but we omit the details. 0 0 (1. oo).h.8 also applies to the case 9 54 1/2.
ba. In view of (1.b(u) + L.e.16) yields 4b (u) = 1 . E„ q(Rdt) = q(u)+Lq(u)dt. then. O. 191195 for material related to Theorem 1.S(u)/S(a). [117].14) S(oo) < 00. . The obvious boundary conditions '0a. i. Letting a T oo and considering the cases S(oo) = oo. For generalizations of Proposition 1.10. Using s'/ s = 2p/a2.ba. A classical reference for further aspects of Bernoulli random walks is Feller [142].16). see Asmussen & Perry [42]. b = 0. (1 . we can ignore the possibility of ruin or hitting the upper barrier a before dt. where Lq(u) = 0'22u) q "(u) + p(u)q(u) is the differential operator associated with the diffusion.11 Let 0 < b < u < a and let t&0.b(Rdt).6 to Markovmodulated models .b('u) = Eu . 0 in (1.(u) < 1 for all u > 0 and ^ S^ Conversely. and we get Wo.13) is finite for all x > 0.17) Hence L.S(b) Proof Recall that under mild conditions on q.b(a) = 0 then yield the result. the function S(x) is .0(u) = 1 for all u > 0. Letting b J.b (Rdt) = Oa. RANDOM WALK. if (1. so that Y)n. If (1. 15) i. Assume further that S (x) as defined in (1. Further references on twobarrier ruin problems include Dickson & Gray [116].b = a+/3S.b('u) = Eu &0. 1'. A good introduction to diffusions is in Karlin & Taylor [222].b(b) = 1.e LVa. Notes and references All material of the present section is standard. then 0 < 2l. see in particular pp. elementary calculus shows that we can rewrite L as Lq(u) d 1a2 (u)s(u)d [ s (u) ? ] .b = 0 implies that VQ b/s is constant.b(u)dt. TWO BARRIERS 303 for x > 0. (1. BROWNIAN MOTION.S(u) (1. S(oo) < oo separately u completes the proof. 0 Proof of Theorem 1. If b < u < a. Lemma 1.10.16) S(a) .b(u) = S(a) .b(u) be the probability that {Rt} hits b before a starting from u. Then YIa.14) fails.1. Wa.
111 .1) (2. (2. correponding to piecewise linear paths or . but by duality. defined by the density 1/va(u)s(u) showing up in (1.4) I.5) A martingale proof of (2.(a) (2.2. which is motivated from the study of modern ATM (asynchronous transfer mode ) technology in telecommunications.. where C_ = B(x) _ B(x) sup 2no fy° e7(Y )B(dy)' f2e7(Y2)B(dy)' C+ i/i(u.3.5. IV.o•K(a) = Ee . with the drift and the variance depending on an underlying Markov process .(7) . yielding eau = Ee. (2. The emphasis is often on stationary distributions .4. y > . MISCELLANEOUS TOPICS referred to as the natural scale in the general theory of diffusions (in case of integrability problems at 0. one works instead with a lower limit 5 > 0 of integration in (1. and here are alternative martingale proofs of the rest . is currently an extremely active area of research. yu) '+/1(u) . 1 y < k (y). much of the literature dels with the pure drift case.)AT . variance 0. See Asmussen [20] and Rogers [305] for some recent treatments and references to the vast literature. Remark 11. 2 Further applications of martingales Consider the compound Poisson model with adjustment coefficient ry and the following versions of Lundberg 's inequality (see Theorems 111.1. yu) where W (ay) = y.6.t&(u.2) C_e7u < t(u) < C+e _7u.3) < e 7yu.aR.17). 7y = ay .(T(u)AT) r.6) .13)). Lo I.ytc (ay).304 CHAPTER XI.1 ) was given already in II. (2. They all use the fact that ( tx(a) l ( eaRt = eau + aSttx(a) < e7yu. Lo is a martingale (cf.9 ) and optional stopping applied to the stopping time r(u) A T.(. information on ruin probabilities can be obtained . Markovmodulated Brownian models .4. equivalently. (2. Another basic quantity is the speed measure M .5): _ z/'(u) < e 7u.aRo .
dr JO Zoo ) f e7'B(r + dy) B(r) Jo ^00 ^00 H(dt.7R. we have tc(ay) > 0 and we can bound (2. FURTHER APPLICATIONS OF MARTINGALES 305 (we cannot use the stopping time r(u) directly because P(r(u) = oo) > 0 and also because the conditions of the optional stopping time theorem present a problem). (B(y) .( u ) I T(U) < 00] .T)  V. Hence E [e7Rr (u) Jr(u) < ool ^00 H( dt. For (2. dr) 1 = 1 I0 /o C+ C+ From this the upper inequality follows. Equivalently. yu))• Letting T + oo yield e_ayu > eyur4ay)(0(u)  Notes and references See II..B(r))/B(r).6).1 . eyuk (ay) = e7yu e > eyu"(ay ) ij(u. Proof of ( 2. we have ic(ay ) < 0 and use the lower bound E [e7Rr („). dr) e 7( yr)B(dy) B(r) f oo o 0 r > H(dt. dr) denote the conditional distribution of (T(u). so that i/1(uL yu) < eayu .4): We take a = ay in (2.2.1.2): As noted in Proposition II. y > r. Let H(dt.3). RT(u)_) given r(u) < oo.4).)r(u)r. it follows easily from (2.yu))• b(u.E [e.(ay)I T(u) < yu] P(r(u) < yu) (using RT(u) < 0). (2. when Rt_ = r. u Proof of (2.T(u)K(ay) I yu < r(u) < T] F(yu < r(u) < T) > e.(u. .6) with = 'y that eyu . Rt has distribution B(r + dy)/B(r).6) below by 1 E Le7Rr(. and the proof of the lower inequality is similar.yuk (ay)(u&(u.yu) Y Similarly for (2. A claim leading to ruin at time t has c.3).d.1.f.
The classical result in the area is Cramer's theorem. gn with fn + 0 . .1). the parameter will be u rather than n).g.gn if nioo lim 109 fn = 1 log gn (later in this section. . large deviations results have usually a weaker form.the correct sharp asymptotics might as well have +. then P C S. however . For example.1) is an example of sharp asymptotics : .^ e nn 1 > x n 0o 2xn (3.1 We will go into some more detail concerning (3. cle . Thus .3na with a < 1. such that the cumulant generating function r..?n typically only give the dominant term in an asymptotic expression . which in the setting of (3. Cramer considered a random walk Sn = X1 + . Example 3.2) can be rewritten as F (Sn/n > x) 1g a'fin. in being capable of treating many models beyond simple random walks which are not easily treated by other models . we will write fn 1. logarithmic asymptotics is usually much easier to derive than sharp asymptotics but also less informative . nroo n n /// Note in particular that (3. gn 4 0. + X.1) amounts to the weaker statement lim 1 log P I Sn > x I = 17.means (as at other places in the book) that the ratio is one in the limit (here n * oo). Thus. og For sequences fn.2)... large deviations results been. MISCELLANEOUS TOPICS 3 Large deviations The area of large deviations is a set of asymptotic results on rare event probabilities and a set of methods to derive such results. The last decades have seen a boom in the area and a considerable body of applications in queueing theory. and that a considerable body of theory has been developed.. (3. and gave sharp asymptotics for probabilities of the form P (S. logarithmic asymptotics . its generality.nn or C2e.1) does not capture the \ in (3. e.(B) = log EeOX 1 is defined for sufficiently many 0. if x > EX1. not quite so much in insurance risk. v2 later./n E I) for intervals I C R. Accordingly. The advantage of the large deviations approach is. The limit result (3. (3. 1) but only the dominant exponential term . However . ri.306 CHAPTER XI.1) where we return to the values of 0.
(0)) e 307 (other names are the entropy. S rtn > x 1. replacing Sn in the exponent and ignoring the indicator yields the Chernoff bound P Sn > x 1 < e°n (3.96o /] > 0. P with mean nx and variance no.(e)i XI E dx]. exponential change of measure is a key tool in large deviations methods.9S„+n' ( 9).4 enn +1.r. Define .sseo f which in conjunction with (3.425. the LegendreFenchel transform or just the Legendre transform or the large deviations rate function). Since P nn > x) = E {e_8 ' ( 9).t. V > 0 e. 2 where o2 = o2(x) = rc"(0). we have P(nx < Sn < nx + 1.4) immediately yields (3. Most often. and hence for large n P(Sn/n > x) > E [e.r.3) is put equal to x. rc*(x) = sup(Ox . if we replace Sn by nx + o / V where V is N(0.4) n Next. LARGE DEVIATIONS Define rc* as the convex conjugate of rc. the sup in the definition of rc* can be evaluated by differentiation: rc*(x) = Ox . we get P(Sn/n > x) E [e9nx +nK(9)9" '. In fact. nx < Sn < nx + 1. which is a saddlepoint equation .1).3.rc(0) where 0 = 0(x) is the solution of x = rc'(0). i.960/) * 0. More precisely.the mean rc'(0) of the distribution of X1 exponentially tilted with 0. since Sn is asymptotically normal w.q = rc* (x). of P(X1 E dx) = E[e9X1K.tin f o') o e9o^y 1 1 ey2/2 dy 21r = etin 1 Bo 27rn . (3.e.2).
In the application of large deviations to ruin probabilities. Pn Sn1 ... For the proof.e < 8 < y + e.'s... and write Sn = X1 + • • • + Xn. We shall need: Lemma 3 . . Further main results in large deviations theory are the GartnerEllis theorem. The substitution by V needs.2 (GLYNN & WHITT [163]) Let X1. Ee9X n < oo for e < 0 < e. be a sequence of r.s. we shall concentrate on a result which give asymptotics under conditions similar to the GartnerEllis theorem: Theorem 3 .3.. see Jensen u [215] or [APQ] p.. Xn) and sn = x1 + • • • + xn (note that the r. 260 for details.h. 1]. Sanov's theorem which give rare events asymptotics for empirical distributions. we introduce a change of measure for X1. commonly denoted as is the saddlepoint approximation. (iii) #c (8) = limn. (iv) tc(ry) = 0 and r.dxn) = 05nKn(7)Fn(dx1.o log Ee9Sn /n. X2. (ii) lim supn. to be made rigorous. asymptotics for probabilities of the form P ({S[nti/n}o<t<l E r) for a suitable set r of functions on [0.. integrates to 1 by the definition of Icn). Xn given by Fn(dxl. that is.'(u) )Ng a"u. Assume that there exists 'y. there exists z E (0. MISCELLANEOUS TOPICS which is the same as (3. is differentiable at ry with 0 < K'(y) < 00.3 For each i > 0.p > 7 < zn.. Then i/. Mogulskii's theorem which gives path asymptotics. n Icn(0) exists and is finite for ry . . which is of similar spirit as the dicussion in VII.e < 8 < y + e. e > 0 such that (i) Kn (0) = log Ee°Sn is welldefined and finite for 'y .. which is a version of Cramer's theorem where independence is weakened to the existence of c(O) = limn.v..1). . and the WentzellFreidlin theory of slow Markov walks. r(u) = inf {n : Sn > u} and o(u) = P('r(u) < oo)...308 CHAPTER XI. . .. however.dxn) where Fn is the distribution of (X1i . We further write µ = tc'(ry).. 1) and no such that Sn ..../^ >7 < zn n for n n0.
+r. can be chosen strictly negative by taking p close enough to 1 and 0 close enough to 0.ne(µ limsup 1 log Pn (Sn/n > µ + 17) < ic(9 + ry) . can be chosen strictly negative by taking 9 small enough. ( U) P(S..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 +.77) follows by symmetry (note that the argument did not use µ > 0). 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 r .n m µ 1 + rl ..s. is of order . it is easy to see that the r.r (7) n = e. This establishes the first claim of the lemma .YS.. This proves the existence of z < 1 and no such that Pn (Sn/n > µ.+r7) < zn for n > no. The corresponding claim for Pn(Sn/n < µ . For Sn1i we have Fn(Sn 1/n > µ+r7) < ene(µ+ 1?)EneeS„1 = ene ( µ+n)EneeSneX„ eno(µ +n) Ee(e+7)Sn ex„ wn (7) < e. h.3. Clearly.µ?7 .ne(p+ 17).]1/q = e.91) + o(O ) as 0 J. > 1 +17] m(7).2.n e(µ +o)w"(7) [Eep(B +7)Sn]1 /p [EegoX. u Proof of Theorem 3.h. Since I EeqOX „ ] 1/q is bounded for large n by (ii).s.Bµ .W. 0. Then V. mµ Sm > u] km e7Sm+n. P n(Sn/n > {c+77) < e no(µ 309 +n)Enees n +n)elcn(B +7). S.y) .71 < e and jq9j < e. in particular the r.. h. log zl'(u)/u > 'y.s.> .m(7). S..n > u ) = [ Em [em Em 1e. We first show that lim inf„_. for Sn. LARGE DEVIATIONS Proof Let 0 < 9 < e where a is as in Theorem 3. The rest of the argument is as before.077 n^oo n and by Taylor expansion and (iv ).2. Let r7 > 0 be given and let m = m(77) = [u(1 + 77)/µ] + 1.
I2 = F(T(u) = n).I < µl1 1+77 I M 1_ 1+277 S.3.(Y). Sn > u] < eYu+Kn(7)pn(Sn > u) (3. 0 yields liminfu __... logO(u)/u > ry. 3. MISCELLANEOUS TOPICS (7). we get lum inf z/i(u) 1 +12r7 >_ ry + 77 Letting r7 J. For lim supu.n Yµ 1 + m + r ('Y) } U n \ 77 m µ µ7 1 < 1+ 77 ) Here E. 14 = = E Lu(16)/aJ+1 Lu(1+6)/µJ+l = n) and n(S) is chosen such that icn('y )/n < 6 A (. n=1 n=n(b)+1 00 Lu(1 +6) /µJ 13 F( T (u) = P(T(u) n). I > IL exp `S. (iv) and Lemma 3.. Pn \ > la+ 8 I < zn (3. and since Ic. P(T(u) = n) < P(Sn > u) = En [e7S.0 log i'(u )/u < 'y.n(ry)/u 4 0andm/u* (1 + r7)/µ..(•) goes to 1 by Lemma 3.log z) /2 and Sn Fn\ n >lb+S) <Zn. 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).. this is possible by (iii).YS +^c CHAPTER XI.7) so that n(b) I1 < e'Yu E en. n=1 . Obviously..+wn(7)..310 ]Em I e.6) for some z < 1 and all n > n(E).
S. 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 .10) 00 I4 < E F(Sn_1 < u. µ n=n(6)+1 \ 1u(16)/µ1 00 1 zn < e7u E Z n/2 < e(U xn/2 E n=n(6)+1 n=0 eYu = 1 . eryu en logz/2p n nt n. u .zl/z en6 [u(1 +6)/µJ 1u (1 +6) /µJ ekn(7) < e' 13 < C" E Yu l u(16)/lij+1 Lu(16)/µJ+l1 < e7U Finally. we get lim sup log u/00 O (U) < y + b(1 + b) U Letbl0. Sn1 C U. C 26u `p / +1 I e6u(1+6)/µ (3.' 1 + b) n e7u x 1 /2 1 n x n / 2x (3. LARGE DEVIATIONS Lu(16)/µJ 311 I2 < e"u n=n(6)+1 e'n(Y)P(Sn > u) < Lu(16)/µJ ^.11) [u(1+6)/µJ+1 1  Thus an upper bound for z/'(u) is n(6) e'Yu n=1 eKn (7) + 2 + (28U + 1) e6u(1+6)/µ Fi 1 zl /2 and using (i). > u) Lu(1+6) /µJ +l 00 )^n 'YSn+kn (7) .3.
I2.xl/2 to give the desired conclusion. we have rcn (a + 7) < 2n^c(7 + a) < 4narc' (7). 13 = P(T (u) E (u(1 b)l^ (7). (7 + a) < 2arc'(7)..4/3rc'(y) > 0.4. this is straightforward since the last inequality in (3. Corollary 3.b)/i(7). it holds for each b > 0 that 0(u) 1' g F(T(u) E (u(1 .11 ) can be sharpened to x 4 [u(1+6)/µJ /2 1 . the last steps of (3. For 14. cf. For 12. Letting c11 = maxn<n.2. u(1 + b)/i(7)) Proof Since V.(u) = I1+I2+I3+I4'^ ery( u). we get Lou] E exp {( 7 + a)u + Kn(a +7)} n=1 Il Lou] exp {(y + a)u} { 111 + exp {4narc'(7)} n=1 exp {('y + a)u} c1 exp {4/3uarc'(7)} = clewhere a1 = aw. ryue«iu . 2. Then for n large.3ui where .Q is so small that w = 1 .z 1/z For I1. we replace the bound P(Sn > u ) < 1 used in (3. 4 there is an aj > 0 and a cj < oo such that Ij < c3e. say n n1.4 Under the assumptions of Theorem 3. IV.8) by P(S. the typical time is u/rc'(7) just as for the compound Poisson model. u . we need to redefine n(b) as L.312 CHAPTER XI. > u) < e"' E eIsn = ectueKn (a+'Y)Kn(7) where 0 < a < e and a is so small that r. For I.7' a"ju. e'..u(1+b)/rc'(7)). it suffices to show that for j = 1.9) can then be sharpened to x LQuJ /2 I2 < e7u 1 . MISCELLANEOUS TOPICS The following corollary shows that given that ruin occurs.('+'Y).
2 is in force with y = 2p/wz.5 Assume the Xn form a stationary Gaussian sequence with mean p < 0.v. The reader not satisfied by this gap in the argument can easily construct a discrete time version of the models! The following formula (3.g. we shall give two continuous time examples and tacitly assume that this can be done.1.1. whether P ( sup St > u ltg a ^" 0<t<oo // (3. Obviously many of the most interesting examples have a continuous time scale. Let {Nt}t>0 be a possibly inhomogeneous Poisson process with arrival rate .12) k=0.3(s) at time s.14) is needed in both examples .e. 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.'(y) > 0. V(s) with m. Thus the total reward in the interval [0. i..2 shows that the discrete time structure is used in an essential way. Hence z z\ 2 z nrn(9) _ n Cn0p+BZn/ * . 11 Inspection of the proof of Theorem 3. t] is Rt = E V (Un) n: o„ <t . but nevertheless.. Xk+l) k=1 00 naoo n provided the sum converges absolutely.. and in fact.(O) = 9µ+02 for all 9 E R. and we conclude that Theorem 3 .3.. criteria are given in Duffield & O'Connell [124]. r. If {St}t> 0 is the claims surplus process. To verify these in concrete examples may well present considerable difficulties.. the key condition similar to (iii). (iv) becomes existence of a limit tc(9) of tct(9) _ log Ee8S° It and a y > 0 with a(y) = 0.LARGE DEVIATIONS 313 Example 3 .. An event occuring at time s is rewarded by a r.. Theorem 3.. Assuming that the further regularity conditions can be verified. 09(9). for the ruin probability z/'h(u) of any discrete skeleton {Skh}k=0.f. The problem is whether this is also the correct logarithmic asymptotics for the (larger) ruin probability O(u) of the whole process.2 then immediately yields the estimate log F( sup Skh > u) a7u (3.13) One would expect this to hold in considerable generality.
t. the CramerLundberg model implicitly assumes that the Poisson intensity /3 and the claim size distribution B (or at least its mean µB) are known. Of course . one would take p(t) = (1 + rt)At/ t. 7 Given the safety loading 77.'`1 U.6 We assume that claims arrive according to a homogeneous Poisson process with intensity 0 . We further assume that the processes {U1(s)}8>0 are i.14) (to see this . <t which is a shotnoise process.noise model is the same as the one for the Cramer Lundberg model where a claim is immediately settled by the amount Un.Lundberg model has the larger ruin probability. MISCELLANEOUS TOPICS are the event times. (9) < oo for 9 < 'y + C. 0 and since EeOUn(8) + Ee°U^ as s * oo. It is interesting and intuitively reasonable to note that the adjustment coefficient ry for the shot .1) . An apparent solution to this problem is to calculate the premium rate p = p(t) at time t based upon claims statistics . is At .15) . Kt (0) t (Ee9U"it8i J0 . = U„ ( t . We let ic (9) = 3(EeWU° . a differential equation in t).s). . the above discussion of discrete skeletons). this is not realistic . Since the remaining conditions of Theorem 3. where Ft = a(A8 : 0 < s < t). but that a claim is not settled immediately.0 and assume there are y. we have rct (9)/t 4 ic (9). Most obviously. i. if the nth claim arrives at time a. Then logEeOR° = J0 /3(s)(^8(9) .2 are trivial to verify. nondecreasing and with finite limits Un as s T oo ( thus.9t.g. n: o.314 where the an CHAPTER XI. Thus.It. leading to St = At(1+77) Joo t S8 ds.d.1) ds .. we conclude that Cu) log e7 u (cf.Q„) . (3. it contributes to St by the amount Un(t .. Of course. assuming a continuous premium inflow at unit rate.v.. e > 0 such that ic('y) = 0 and that r. the best estimator of /3µB based upon Ft. derive .9t = /3 J t (Ee8U° i8l . O'n +S] is a r . 0 Example 3 .. Example 3. we have S. At = .1) ds . then the payments from the company in [on. Thus by (3.14). Thus.1) ds rt (3. More precisely. e. Un represents the total payment for the nth claim). Un(s). If the nth claim arrives at time Qn = s. the Cramer.
i. equivalently. Thus.e.3. rewrite first rc as te(a) _ /3E 1 1 +(1+77)aUJ eau 1 .2 hold. we have Nt t N. standard exponential .b(u) IN a'Yu (cf.log Oi are i.14) that rt _ 13 Jo _ (a [1_( i+77)log]) ds_flt = t (a) (3. which yields eau f 1 t(1+n )audtl = E r Ee°Y = E [O(1+n)aueaul = E [eau J L Jo J L1+(l+r))aUJ . one has y > y' (3. To see this . and since the remaining conditions are trivial to verify. the solution of /3(Eelu . uniform (0.1) or .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 . the Vi = .19) with equality if and only if U is degenerate. LARGE DEVIATIONS With the Qi the arrival times. Indeed.17) K(a) f o 1 O (a[I + (1 + 77) log u]) du )3.18) Thus (iii) of Theorem 3. we conclude that t. (3.d.(1 + 17)0µB = 0./3.1) . i.i. again the above discussion of discrete skeletons) where y solves ic('y) = 0 It is interesting to compare the adjustment coefficient y with the one y* of the CramerLundberg model.(1 +i) f > i= 1 s ds = E Ui 1 .16) i=1 o i=1 Let ict (a) = log Eeast .(1 + r7) log t (3.d. It then follows from (3. typically the adaptive premium rule leads to a ruin probability which is asymptotically smaller than for the CramerLundberg model .20) (3. (3. Ui Nt / t 01i 315 St = Ui .
20) is due to Tatyana Turova. This implies n(y*) < 0.(1 + ri)y*x is convex with k(oo) = 00. MartinL6f [256]. For notational simplicity. see Nyrhinen [275] and Asmussen [25]. a* (s) are convex with tc'(0) < 0 . The main example is Nt being Poisson with rate fit. Further. This is a topic of practical importance in the insurance business for assessing the probability of a great loss in a period of length t. Dembo & Zeitouni [105] and Shwartz & Weiss [339].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.. this in turn yields y > y*. using that Ek(U) = 0 because of (3. the function k(x) = e7*x . 4 The distribution of the aggregate claims We study the distribution of the aggregate claims A = ^N' U. we are interested in estimating P(A > x) for large x. say one year. MISCELLANEOUS TOPICS Next. rc*' (0 ) < 0. Lehtonen & Nyrhinen [244].2.i.7. we then take t = 1 so that p. [245]. the study is motivated from the formulas in IV. y = y* can only occur if U .xo. Further. = P(N = n) = e(3an However. k(0) = 0. the proof of (3. x > x0. [257] and Nyrhinen [275]. and since tc(s). at time t. k'(0) < 0.316 CHAPTER XI.2 expressing the finite horizon ruin probabilities in terms of the distribution of A. Therefore e7'U _ k(U) E [1+(1+77)y*U] . .1 . For Example 3. are i. with common distribution B and independent of Nt. much of the analysis carries over to more general cases. and k(x) < 0. assuming that the U. Further applications of large deviations idea in risk theory occur in Djehiche [122]. In addition to Glynn & Whitt [163]. though we do not always spell this out. In particular. see also Nyrhinen [275] for Theorem 3.d. so there exists a unique zero xo = xo(r7) > 0 such that k(x) > 0. 11 Notes and references Some standard textbooks on large deviations are Bucklew [81].19). 0 < x < x0.
The exponential family generated by A is given by Pe(A E dx) = E [eeA K(9).1 Assume that lim8T8.9(Ax). Vare(A) = s. This shows that the Pedistribution of A has a similar compound Poisson form as the Fdistribution.4. For a given x.1).3e(bo[a] . e9x+K(°) P(A > x) B 2ir /3 B" [9] Proof Since EBA = x. A > x) eex+K(e ) ee AB°[ely 1 ev2/2 dy 0 2^ 00 9x+p(e) e ezez2/(2BZpB „[9)) dz 9 27r/3B" [9] fo eex+w ( e) oo z x)] ] 0 27r /3B" [9] o e 9 2 /3B" [9] J eex+w(B) dz .x)//3B"[9] is standard normal.[s])3/2 = 0. A > x)] = eex+K( e)E9 [e . Then Ee"A = e'(") where x(a) _ 0(B[a] . K'(0) _ ic'(9) = x. Proposition 4.2) implies that the limiting Pedistribution of (A . only with 0 replaced by a9 and B by B9. 818' where s' = sup{s : B[s] < oo}. A E dx] .e. In particular.3B"[9]. no(a) = logE9e'A = rc(a + 9) . B"[s] = oo. The analysis largely follows Example 3.1). we define the saddlepoint 9 = 9(x) by EBA = x. B"' [s] lim (B". THE DISTRIBUTION OF THE AGGREGATE CLAIMS 317 4a The saddlepoint approximation We impose the Poisson assumption (4. i."(0) = . (4.1) where )30 = .1.ic(9) = . Hence P(A > x) = E e [e9A+ ic(9). Then as x * oo.3B[9] and Be is the distribution given by eox B9(dx) = B [9] B(dx).
and (4.1).3) and related results u for the case of main interest .2 If B is subexponential and EzN < oo for some z > 1. For example. The (first order) Edgeworth expansion states that if the characteristic function g(u) = Ee"`}' of a r. B covers distributions with finite support or with a density not too far from ax° with a > 1. under the Poisson assumption (4. For example.(3µB)/(0µB^))1/2 has a limiting standard normal distribution as Q ^ oo. The present proof is somewhat heuristical in the CLT steps. i. either of the following is sufficient: A.Q{AB (4. In fact. 2). 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. the distribution of A is approximately normal . where q(x) is bounded away from 0 and oo and h (x) is convex on an interval of the form [xo.4) . Var(A) _ ^3p. it is quite questionable to use (4. bounded with b(x) . 1 . Y satisfies 9(u) ti eu2/2(1 + ibu3) (4. A covers the exponential distribution and phasetype distributions. large x.2i and that (A . [138]. some regularity of the density b(x) of B is required.318 CHAPTER XI.ycix °ie6x B.l3pB. Furthermore 00 b(x)Sdx < oo for some ( E (1. or. 4b The NP approximation In many cases . just the same dominated convergence argument as in the proof of Theorem 2.(D X .e.3) The result to be surveyed below improve upon this and related approximations by taking into account second order terms from the Edgeworth expansion.x') where x' = sup {x : b(x) > 0}.2) is often referred to as the Esscher approximation. MISCELLANEOUS TOPICS It should be noted that the heavytailed asymptotics is much more straightforward. For details. leading to P(A > x) :. For a rigorous proof.v. In particular. Remark 4 . it holds that EA = . Thus . see Embrechts et al. Jensen [215] and references therein. b is gammalike. b is logconcave.1 goes all the way back to Esscher [141]. Notes and references Proposition 4. more generally.EN B(x). b(x) = q(x)eh(z). then P(A > x) .
(4.5) may be negative and is not necessarily an increasing function of y for jyj large. Rather than with the tail probabilities F(A > x)..l = EY. s. in particular.2K3 + 4i 64 + .2 ^ \1 .equantile in the distribution of Y. where Kl . THE DISTRIBUTION OF THE AGGREGATE CLAIMS where b is a small parameter. so that 1(u) 3 exp { .y2)^P(y)• 319 Note as a further warning that the r. however.4.1).2 2 .h. the density of Y is 1 °° _ eiuy f(u) du 2x _.f.5 (y3 .e. the NP (normal power) approximation deals with the quantile al_E. of (4.s. one needs to show that 163. Heuristically. (4. K4 . zl_e be the 1 .c2i. the CLT for Y = Y6 is usually derived via expanding the ch. .. resp.EY)3. which is often denoted VaR (the Value at Risk).2X2 . .99.EA)/ Var(A) and let yl_E. If this holds .3!).6(1 .. In concrete examples . one expects the u3 term to dominate the terms of order u4. Let Y = (A . ylE should be close to zl_E (cf.. .6) . Thus if EY = 0.: EA + zl_E Var(A) . A particular case is a. then P(Y < y) 4(y) .i 3 K3 } Pt^ exp ..5) is obtained by noting that by Fourier inversion.. the standard normal distribution. as u2 u3 u4 9(u) = Ee'uY = exp {iuci .. are the cumulants . Remark 4. f °o 9(y) = 1 e'uye u2/2(1 + iSu3) du 27r _ cc(y) . K2 = Var (Y).. u5.3& (y). and so as a first approximation we obtain a1_E = EA + yle Var(A) . If the distribution of Y is close to N(0..5) follows by integration. are small. and from this (4.5). Var(Y) = 1 as above . K3 = E(Y . defined as the the solution of P(A < yle) = 1 .i 6 r 1 3 so that we should take b = ic3/6 in (4.
Another main reference is Daykin et at. the kth cumulant of A is /3PBk' and so s.k = /3µB^1 / (.1)! n ^eQ .EA)3 a1_E = EA + z1_E(Var (A))1/2 + 1 Var(A) Under the Poisson assumption (4. This leads to t( yl E) .E(/3PB^1 )1^2 + s(z1E .zlE)W(zlE) 1 . K5 . b = /3 for the Poisson distribution with rate /3 since Pn = Pn1 n! n (n .S(1 .7) as 1 (3) a1E = Qµa +z1 . k3 is small for large /3 but dominates 1c4. For example. .E .1)^ 2) µ'E Notes and references We have followed largely Sundt [354]. 4c Panjer 's recursion Consider A = constants a.1) E (A .. Using Y = (A ..E)A1 l E) 1 E 4)(yl E) ^' . In particular .1). and assume that there exist n ) Pn_i .zlE )w(zl _E) = which combined with S = EY3/6 leads to q^ 1 Y1 .zl E)^o(zl E) .E + (yl. let pn Pn = (a+ = P(N = n). this holds with a = 0. b such that EN 1 U%..E )Azl E) 4(z1E) + ( ylE .. MISCELLANEOUS TOPICS A correction term may be computed from (4. 21 .EA ) / Var(A). this yields the NP approximation 6(Z1 _E .5) by noting that the 4.yi. We can rewrite (4. as required ..5(1 .S(1 .320 CHAPTER XI.3ni /3 . Note.zi.6pBki) d/2. n = 1. that [101] distinguishes between the NP and Edgeworth approximations... [101].E = z1E + S(zi_E .1)EY3.zl E)V(zl_E) . however.(y) terms dominate the S(1 .y2)cp( y) term.6 (1 .
. The expression for fo is obvious.12). and calculating the gj*n recursively by 9*1 = 9j. fj = P(A = j). . E[a +bU=I >Ui =j l i=1 J (4. j = 1. then j (a + b!) 1ag k_1 3 gkfj. . (4. .1..4.13) but only O(j2) for Proposition 4. j = 0. Hence by (4. n. if go = 0.k . . THE DISTRIBUTION OF THE AGGREGATE CLAIMS 321 Proposition 4. (4.} and write gj = 2 . j = 1. 1. which would consist in noting that (in the case go = 0) fj = pn9jn n=1 (4.14) is independent of i = 1.9). (4. By symmetry. ..4 Assume that B is concentrated on {0..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 ..13) Namely.12) where g*n is the nth convolution power of g.10) f o = po..4 is that the algorithm is much faster than the naive method..4. 2.11) Remark 4. j1 g. 2. n = k=n1 9k(n1 )9j k • (4. u Proof of Proposition 4. the complexity (number of arithmetic operations required) is O(j3) for (4.14) is therefore a + b/n..4. .. fj = E (a+ b k =1 )9kfi_k . . Since the sum over i is na + b.. (4.. . . Then fo = >20 9onpn and fi = 1 E In particular. 2. the value of (4..5 The crux of Proposition 4.
322
00 J
CHAPTER XI. MISCELLANEOUS TOPICS
EE (a + bk I gkg3 _ k lieni n=ik=0 (a+bk l gkE g j'`kpn = E (a+b!)9kfi_k n=0 k=0 k=0 ^I 1 E(a+b. agofj+ k Jgkfjk, k=i /
and and (4.9) follows . (4.11) is a trivial special case.
u
If the distribution B of the Ui is nonlattice , it is natural to use a discrete approximation . To this end, let U(;+, U(h) be U; rounded upwards, resp. downwards , to the nearest multiple of h and let A}h) = EN U. An obvious modification of Proposition 4.4 applies to evaluate the distribution F(h) of A(h) letting f( ) = P(A() = jh) and
g(h) gkh+
= P (U(h2 = kh) = B((k + 1)h)  B(kh ), k = 0, 1, 2, ... , = P (U4;+ = kh) = B(kh)  B (( k  1)h) = gk  l,, k = 1, 2, ... .
Then the error on the tail probabilities (which can be taken arbitrarily small by choosing h small enough ) can be evaluated by
00 00
< P(A > x ) f (h) j=Lx/hl j=Lx/hl
Further examples ( and in fact the only ones , cf. Sundt & Jewell [355]) where (4.9) holds are the binomial distribution and the negative binomial (in particular, geometric ) distribution . The geometric case is of particular importance because of the following result which immediately follows from by combining Proposition 4.4 and the PollaczeckKhinchine representation: Corollary 4.6 Consider a compound Poisson risk process with Poisson rate 0 and claim size distribution B. Then for any h > 0, the ruin probability zb(u) satisfies 00 00
f^,h) Cu) < E ff,+, j=Lu/hJ j=Lu/hJ (4.15)
f! h)
5. PRINCIPLES FOR PREMIUM CALCULATION
where f^ +, f^ h) are given by the recursions
(h) 3 (h) (h)
323
fj,+ = P 9k fjk,+ ' I = 17 2, .. .
k=1 3 (h)
(h)
=
P
(h)
f9,  (h) gk,fAk, e 1  ago, k=1
j = 1+2,
starting from fo + = 1  p, f(h) = (1  p)/(1  pgoh) and using 07
g(kh) 1 (k+1)h
=
Bo((k + 1 ) h)  Bo(kh ) =  f
AB
kh
B(x) dx, k = 0, 1, 2, ... , k = 1,2 .....
gkh+
Bo(kh )  Bo((k  1 ) h) = 9kh)1 ,
Notes and references The literature on recursive algorithms related to Panjer's recursion is extensive, see e.g. Dickson [115] and references therein.
5 Principles for premium calculation
The standard setting for discussing premium calculation in the actuarial literature does not involve stochastic processes, but only a single risk X > 0. By this we mean that X is a r.v. representing the random payment to be made (possibly 0). A premium rule is then a [0, oo)valued function H of the distribution of X, often written H(X), such that H(X) is the premium to be paid, i.e. the amount for which the company is willing to insure the given risk. The standard premium rules discussed in the literature (not necessarily the same which are used in practice!) are the following: The net premium principle H(X) = EX (also called the equivalence principle). As follows from the fluctuation theory of r.v.'s with mean, this principle will lead to ruin if many independent risks are insured. This motivates the next principle, The expected value principle H(X) = (1 + 77)EX where 77 is a specified safety loading. For 77 = 0, we are back to the net premium principle. A criticism of the expected value principle is that it does not take into account the variability of X which leads to The variance principle H(X) = EX+77Var(X). A modification (motivated from EX and Var(X) not having the same dimension) is
324
CHAPTER XI. MISCELLANEOUS TOPICS
Var(X).
The standard deviation principle H(X) = EX +rl
The principle of zero utility. Here v(x) is a given utility function, assumed to be concave and increasing with (w.lo.g) v(O) = 0; v(x) represents the utility of a capital of size x . The zero utility principle then means v(0) = Ev (H(X)  X); (5.1)
a generalization v(u) = Ev (u + H(X)  X ) takes into account the initial reserve u of the company. By Jensen 's inequality, v(H(X)  EX) > Ev(H(X)  X) = 0 so that H(X) > EX. For v(x) = x, we have equality and are back to the net premium principle. There is also an approximate argument leading to the variance principle as follows. Assuming that the Taylor approximation
v(H(X)  X) ^ 0 +v'(0)(H (X)  X) + v 0 (H(X)  X)2 ,/2
is reasonable , taking expectations leads to the quadratic v"H(X )2 + H(X) (2v'  2v"EX) + v"EX2  2v'EX = 0 (with v', v" evaluated at 0) with solution
H(X)=EXv^±V( ^ )2Var(X).
Write
( vI ) 2 \
Var(X) v^  2v^Var(X)/ I  (
, Var(X) )2
If v"/v' is small, we can ignore the last term. Taking +f then yields H(X) ,:: EX 
2v'(0) VarX;
since v"(0) < 0 by concavity, this is approximately the variance principle. The most important special case of the principle of zero utility is The exponential principle which corresponds to v(x) = (1  e6x)/a for some a > 0. Here (5.1) is equivalent to 0 = 1  e0H(X)EeaX, and we get
H(X) = 1 log Ee 0X .
a
5. PRINCIPLES FOR PREMIUM CALCULATION
325
Since m.g.f.'s are logconcave, it follows that H,, (X) = H(X) is increasing as function of a. Further, limQyo Ha (X) = EX (the net premium princiHa (X) = b (the premium ple) and, provided b = ess supX < oo, lim,, H(X) = b is called the maximal loss principle but is clearly not principle very realistic). In view of this, a is called the risk aversion The percentile principle Here one chooses a (small ) number a, say 0.05 or 0.01, and determines H(X) by P(X < H(X)) = 1  a (assuming a continuous distribution for simplicity). Some standard criteria for evaluating the merits of premium rules are 1. 77 > 0, i .e. H(X) > EX. 2. H(X) < b when b (the ess sup above ) is finite 3. H(X + c) = H(X) + c for any constant c
4. H(X + Y) = H(X) + H(Y) when X, Y are independent
5. H(X) = H(H(XIY)). For example , if X = EN U= is a random sum with the U; independent of N, this yields
H
C^
U; I = H(H(U)N)
(where, of course, H(U) is a constant). Note that H(cX) = cH(X) is not on the list! Considering the examples above, the net premium principle and the exponential principle can be seen to the only ones satisfying all five properties. The expected value principle fails to satisy, e.g., 3), whereas (at least) 4) is violated for the variance principle, the standard deviation principle, and the zero utility principle (unless it is the exponential or net premium principle). For more detail, see e.g. Gerber [157] or Sundt [354]. Proposition 5.1 Consider the compound Poisson case and assume that the premium p is calculated using the exponential principle with time horizon h > 0. That is,
N,,
Ev I P  E U;
i =1
= 0 where
v(x) = 1(1  e°x
a
Then ry = a, i.e. the adjustment coefficient 'y coincides with the risk aversion a.
326
Proof The assumption means
CHAPTER XI. MISCELLANEOUS TOPICS
0 a (1  eareo (B[a11)
l
i.e. /3(B[a]  1)  ap = 0 which is the same as saying that a solves the Lundberg u equation. Notes and references The theory exposed is standard and can be found in many texts on insurance mathematics, e.g. Gerber [157], Heilman [191] and Sundt [354]. For an extensive treatment, see Goovaerts et al. [165].
6 Reinsurance
Reinsurance means that the company (the cedent) insures a part of the risk at another insurance company (the reinsurer). Again, we start by formulation the basic concepts within the framework of a single risk X _> 0. A reinsurance arrangement is then defined in terms of a function h(x) with the property h(x) < x. Here h(x) is the amount of the claim x to be paid by the reinsurer and x  h(x) by the the amount to be paid by the cedent. The function x  h(x) is referred to as the retention function. The most common examples are the following two: Proportional reinsurance h(x) = Ox for some 0 E (0, 1). Also called quota share reinsurance. Stoploss reinsurance h(x) = (x  b)+ for some b E (0, oo), referred to as the retention limit. Note that the retention function is x A b. Concerning terminology, note that in the actuarial literature the stoploss transform of F(x) = P(X < x) (or, equivalently, of X), is defined as the function
b * E(X  b)+ =
f
(s  b)F(dx) _ f
6 00
(x) dx.
An arrangement closely related to stoploss reinsurance is excessofloss reinsurance, see below.
Stoploss reinsurance and excessofloss reinsurance have a number of nice optimality properties. The first we prove is in terms of maximal utility: Proposition 6.1 Let X be a given risk, v a given concave nondecreasing utility function and h a given retention function. Let further b be determined by E(X b)+ = Eh(X). Then for any x,
Ev(x  {X  h(X)}) < Ev(x  X A b).
6. REINSURANCE
327
Remark 6 .2 Proposition 6.1 can be interpreted as follows. Assume that the cedent charges a premium P > EX for the risk X and is willing to pay P1 < P for reinsurance. If the reinsurer applies the expected value principle with safety loading q, this implies that the cedent is looking for retention functions with Eh(X) = P2 = P1/(1 + 77). The expected utility after settling the risk is thus
Ev(u + P  P1  {X  h(X)})
where u is the initial reserve . Letting x = u + P  P1, Proposition 6.1 shows that the stoploss rule h (X) = (X  b)+ with b chosen such that E(X  b)+ u = P2 maximizes the expected utility. For the proof of Proposition 6.1, we shall need the following lemma: Lemma 6 .3 (OHLIN'S LEMMA) Let X1, X2 be two risks with the same mean, such that Fj(x) < F2 (x), x < b, Fi(x) ? F2(x), x > b for some b where Fi(x) = P(Xi < x). Then Eg(X1) < g(X2) for any convex function g. Proof Let Yi=XiAb, Zi=Xivb.
Then
P(Yl < x) _ Fi(x) <_ F2 (x) = P(Y2 < x) x < b 1=P(Y2<x) x>b so that Y1 is larger than Y2 in the sense of stochastical ordering . Similarly, P(Zl < x) _ 0 = P(Z2 < x) x < b Fi(x) > F2(x) = P(Z2 < x) x > b
so that Z2 is larger than Zl in stochastical ordering. Since by convexity, v(x) = g(x)  g(b)  g'(b)(x  b) is nonincreasing on [0, b] and nondecreasing on [b, oo), it follows that Ev(Y1) < Ev(Y2), Ev(Zi) < Ev(Z2). Using v(Yi) + v(Zi) = v(Xi), it follows that
0 < Ev(X2)  Ev(Xi) = Eg(X2)  Eg(X1),
using EX1 = EX2 in the last step. u
Proof of Proposition 6.1. It is easily seen that the asssumptions of Ohlin' s lemma hold when X1 = X A b, X2 = X  h(X); in particular, the requirement EX1
328
CHAPTER XI. MISCELLANEOUS TOPICS
= EX2 is then equivalent to E(X  b)+ = Eh(X). Now just note that v is convex. u
We now turn to the case where the risk can be written as N
X = Ui
i=1
with the Ui independent; N may be random but should then be independent of the Ui. Typically, N could be the number of claims in a given period, say a year, and the Ui the corresponding claim sizes. A reinsurance arrangement of the form h(X) as above is called global; if instead h is applied to the individual claims so that the reinsurer pays the amount EN h(Ui), the arrangement is called local (more generally, one could consider EN hi(Ui) but we shall not discuss this). The following discussion will focus on maximizing the adjustment coefficient. For a global rule with retention function h* (x) and a given premium P* charged for X  h* (X), the cedents adjustment coefficient y* is determined by
1 = Eexp {ry*[X  h*(X)  P*]},
for a local rule corresponding to h(u) and premium P for X look instead for the ry solving
J _f
(6.2) N 1 h (Ui), we
[ X_P_^
1 = Eexp
[ Ei  h(Ui)] P [U
= Eexp{ry
h(Ui)]
l (6.3) This definition of the adjustment coefficients is motivated by considering ruin at a sequence of equally spaced time points, say consecutive years, such that N is the generic number of claims in a year and P, P* the total premiums charged in a year, and referring to the results of V.3a. The following result shows that if we compare only arrangements with P = P*, a global rule if preferable to a local one. Proposition 6.4 To any local rule with retention function h(u) and any
N
J}
P > E X  N h(Ui)
4 =1
(6.4)
there is a global rule with retention function h* (x) such that
N
Eh*(X) = Eh(U1)
i=1
and 'y* > ry where ry* is evaluated with P* = P in (6.3).
4). Assuming for simplicity that the Ui are i.h(u) and any P satisfying (6. appealing to (6.3). REINSURANCE Proof Define N 329 h* (x) = E > h(Ui) X = x .P > EexP{7[X . ' ii (6. that 01[ry] < 0[y] where 0[y] = Ee'r(U^') . ry* > 0 because of (6. u But since ry > 0. Proof As in the proof of Proposition 6. it suffices to show that Eexp {ry ii 'UiAb.b)+ = Eh(U) (and the same P) satisfies 71 > ry. Eexp 7 [E [Ui . Applying the inequality Ecp(Y ) > EW(E (YIX )) (with W convex ) to W(y ) = eryy.h * (X) .4). we get EX = EN • EU. as often local as global.5) reduce quite a lot. we get N 1 = Eexp ry E[Ui ii .h(Ui)] . (6.6 Assume the Ui are i. Then for any local retention function u .h(U) (as in the proof of Proposition 6.4).6) u where C[ry] = Ee'r(u4(u)).P]}.h(Ui)P JJJ l:='l {ry ] or.4.4) and u g(x) = e7x in Ohlin's lemma. then (6.P I = EC [7]N.b)+ with b determined by E(U .6.P } < 1 = Eexp E[Ui.h(Ui)] . y = Ei [Ui .h(Ui)] . i. Remark 6.6).b)+ is referred to as excessofloss reinsurance and plays a particular role: Proposition 6.5 Because of the independence assumptions . this implies 7* > 7.d. (6. however. X2 = U . expectations like those in (6. and so on. .i.d.5) holds trivially.P. This follows by taking Xl = U A b.h( UU) = EN • E[U .h(U)]. Local reinsurance with h(u) = (u .. N E X . the excess ofloss rule hl (u) = (u . The arrangement used in practice is.
Heilman [191] and Sundt [354]. .330 CHAPTER XI. Bowers et at. see also Sundt [354]. The original reference for Ohlin's lemma is Ohlin [277]. MISCELLANEOUS TOPICS Notes and references The theory exposed is standard and can be found in. See further Hesselager [194] and Dickson & Waters [120]. The present proof is from van Dawen [99].g.many texts on insurance mathematics. e. [76].
The mathematical representation is either the ordered set 0 < To < T1 < . then Stone 's decomposition holds : U = U. Y2.. t]) so that U(t + a) . when t is large. That is.. are independent and Y1.U(t) is the expected number of renewals in (t. .T„_1). = T„ . the renewal process is called zerodelayed. t] is denoted by Nt. Then Blackwell 's renewal theorem holds. all have the same distribution. Technically.. not concentrated on {h. t 00 (A.} for any h > 0. i..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. note in particular that U({0}) = 1.1) (here U(t) = U([0. of interarrival times and the time Yo = To of the first arrival (that is. stating that U(t+a)U (t) ^ a.r.t. If F satisfies the stronger condition of being spreadout (F*' is nonsingular w . Lebesgue measure for some n > 1). The number max k : Tk_j < t of renewals in [0... The associated renewal measure U is defined by U = u F*" where F*" is the nth convolution power of F. of epochs or the set Y1. Lebesgue measure dt normalized by the mean to of F. Y1...e. 2h. Y. U(A) is the expected number of renewals in A C R in a zerodelayed renewal process. some condition is needed: that F is nonlattice. . The point process is called a renewal process if Yo. denoted by F in the following and referred to as the interarrival distribution. Y2.. . If Yo = 0. The renewal theorem asserts that U(dt) is close to dt/µ... + U2 where U1 is a finite measure and U2(dt) = u(t)dt where 331 . . the distribution of Yo is called the delay distribution. t +a]).
(A. (A.5) 2This condition can be weakened considerably . µF (A. U Z(u .9.2). i.e. IV).3) Further.. (A. oo). then it suffices for (A. wee shall need the following less standard parallel to the key renewal theorem: Proposition A1. and F(dx) a known probability measure . and that F has a bounded density2. A weaker (and much easier to prove) statement than Blackwell's renewal theorem is the elementary renewal theorem. In 111.2) Z(u) = J0 u z(x)U(dx). z(x) = 0. Note in particular that F is spreadout if F has a density f.x)F(dx).2 Assume that Z solves the renewal equation (A.4) that z is Lebesgue integrable with limZ. Under weak regularity conditions (see [APQJ Ch.1 if F is nonlattice and z (u) is directly Riemann integrable (d. then Z(u) i f0 z(x)dx . Both result are valid for delayed renewal processes. that z(u) has a limit z(oo) (say) as u 4 oo. Then Z(u) 4 z(oo).i.332 APPENDIX u(t) has limit 1/µ as t 4 oo. the asymptotic behavior of Z(u) is given by the key renewal theorem: Proposition A1. in convolution notation Z = z + F * Z.a.2) has the unique solution Z = U * z. Equivalently. resp. z(u) a known function. (A.EN(t) .i". see [APQ] Ch. u u PF 4 00.4) If F is spread. IV). the statements being EN(t + a) .out. 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 .R. stating that U(t)/t > 1/p. but suffices for the present purposes .
multiply (A.5a. z(x) = e7xz(x)... or many queueing processes. . Assuming that y can be chosen such that f °° Ox F(dx) = 1. . However. • . Yk ). .. The simplest case is when {Xt} has i.x)u(x) dx = z(u( 1 . Here the relevant F does not have mass one (F is defective).i. A stochastic process {Xt}t>0 with a general state space E is called regenerative w. T1.(3.k+t }t>o is independent of To.d. The kth cycle is defined as {XTk+t}o<t<Yk . However.. Y1 . asymptotic properties can easily be obtained from the key renewal equation by an exponential transformation also when F(dx) does not integrate to one. Eo etc.2) by e7x to obtain Z = z +P * Z where Z(x) = e'Y'Z(x). To this end. {Tn} if for any k. this covers discrete Markov chains where we can take the Tn as the instants with Xt = i for some arbitrary but fixed state i. The distribution F of Y1. however.. refer to the zerodelayed case. of Yo. . Y2.. We let FO. equivalently.t. 1c Regenerative processes Let {T. 0 PF µF 11 In risk theory. i. this expression is to be interpreted as a random element of the space of all Evalued sequences with finite lifelengths.. is called the cycle length distribution and as before. This program has been carried out in III.} be a renewal process. Note.APPENDIX 333 Proof The condition on F implies that U(dx) has a bounded density u(x) with limit 1/µF as x * oo. . a basic reason that renewal theory is relevant is the renewal equation II. Z(u) U = 1 u 1 u f z(u .. The property of independent cycles is equivalent to the postTk process {XTk+t}t>0 being independent of To. where the Tn are the instants where a customer enters an empty system (then cycles = busy cycles). results from the case fo F(dx) = 1 can then be used to study Z and thereby Z.3) satisfied by the ruin probability for the compound Poisson model.e. Hence by dominated convergence. the postTk process {XT.t))u(ut) dt 0 0 J f z(oo) • 1 dt = z(OO). that the existence of y may fail for heavytailed F. A regenerative process converges in distribution under very mild conditions: .r. F(dx) = e7xF(dx). T1. Tk and {Xt }o<t<Tk • For example. Tk (or.. . cycles. we let µ denote its mean. and its distribution does not depend on k. that F is a probability measure. the present more general definition is needed to deal with say Harris recurrent Markov chains.
d..t : t < Tk}.. where the distribution of X.. i. fi (t) = inf {Tk .3..ZT }0<t<Y„+.i. µ 0 If F is spreadout.. We denote the limiting r. under the condition of Blackwell's renewal theorem.6) id Cumulative processes Let {Tn} be a renewal process with i. Y1) le Residual and past lifetime Consider a renewal process and define e ( t) as the residual lifetime of the renewal interval straddling t.0 be cumulative w.. [0.t. just the same proof as there carries over to show: Proposition A1.d. C). and we have: holds more generally that (rl(t). An example is Zt = fo f (X8) ds where {Xt} is regenerative w.tEU1/µ)/f has a limiting normal distribution with mean 0 and variance Var(Ui) + (!)2Var (Yi)_ 2EU1 Cov(U1.4 Let {Zt}t^.ZT Then: (a) If E sup I ZTo+t . oo).ZTOI < 00. C(t) and ij (t) both have a limiting stationary distribution F0 given by the density F (x)/p.3 Consider a regenerative process such that the cycle length distribution is nonlattice with p < oo. 0<t<Yi then Zt /t a$• EU1/µ. assume that p < 00 and define Un = ZT}1 . 2.+ X.e. {Tn}. cycles (we allow a different distribution of the first cycle)..v..i.t.. then (Zt ..t. Then {Zt}t^. (A. {i7(t)} are Markov with state spaces (0.0 is called cumulative w. and q(t) = sup It . is given by Eg(Xoo) = 1 E0 f Ylg (Xt)dt. then Xt .r. {Tn} if the processes {ZT +t . P(C ( t) < a) 4 0 for any a < oo) and ij (t) * oo.. Otherwise . for n = 1. but in fact. {Tn}. If p = oo. oo).334 APPENDIX Proposition A1. This is the case considered in [APQ] V. Then {e(t)}.oo (i..Tk : t < Tk} as the age.'s by e. then e (t) . (b) If in addition Var(Ul ) < oo. resp .r. in total variation.e. Then it (ii. r. e(t )) . Then Xt Di X. are i.r.
Yo > 0] + f Eo^ (t . ^) is given by the following four equivalent statements: (a) P (77 > x.d.6 Consider a renewal process with µ < oo. we used: Proposition A1.i.5 Under the condition of Blackwell's renewal theorem.U(x) (c < oo because it is easily seen that U(x + 1) .d.dy )z(y) < c ^ l z(k) Eoe(t 0 0 k=o where c = sup. if in addition EYo < oo. EC(t)/t + 0. Since the maximum Mn of n i.. 1) and W has distribution Fw given by dFw/dF(x) = x/pF. ^ > y) = 1 f +Y (z)dz.y) = f U(t . (b) the joint distribution of (ri.. the sum is o(t) so that Eo£(t)/t + 0 . l:) is the same as the distribution of (VW. Yl > t].v. and the conditional distribution of given 17 = y is the overshoot distribution R0(Y) given by FO(Y) (z) = Fo (y+z)/Fo(y). Since z ( k) < E[Yi . r.'s with finite mean satisfies Mn/n a$• 0 (BorelCantelli). (d) the marginal distribution of ^ is FO. Hence t t lt ) = f U(dy)z(t . the first statement follows.y)P(Yo E dy) . and the equivalence of (a) with (b)(d) is an easy exercise.t. Proof The number Nt of renewal before t satisfies Nt/t a4' p. use t E^(t)/t = E[Yo . Y1i Y2. 0 If Markov renewal theory By a Markov renewal process we understand a point process where the interarrival times Yo . but governed by a Markov chain {Jn} (we .t.(t). (1 V)W) where V. In the general case.4. . we can bound e(t) by M(t) = max {Yk : k < 2t/p}.i. Then fi(t)/t a4' 0 and. In IV.APPENDIX 335 Theorem A1. the joint distribution of (rl. are not i.U(x) < U( 1)). Y1 > t] 4 0. U(x + 1) . Then Eo^(t) satisfies a renewal equation with z(t) _ E[Y1 . W are independent. V is uniform on (0.^(t))} as a regenerative process. Hence for t large enough. = z is Foz) The proof of (a) is straightforward by viewing {(r.. (c) the marginal distribution of q is FO. For the second. assume first the renewal process is zerodelayed. and the conditional distribution of ri given l.
}..) and (Fij )i. Notes and references Renewal theory and regenerative processes are treated. oo). e.+ < x. These facts allow many definitions and results to be reduced to ordinary renewal. where the distribution of X. Alsmeyer [5] and Thorisson [372]. namely {Twk } where {Wk } is the sequence of instants w where Jo. T_=inf{n>0: Sn<0}. . = io for some arbitrary but fixed reference state io E E.d.336 APPENDIX assume here that /the state space E is// finite) in the sense that P(Y. Jn = i is the same as the P.. . Further: Proposition A1. For example.. Then Xt 4 Xo. Jn +1=j} where J = a(JO. We call r+ (T_) the strict ascending (weak descending) ladder epoch and G+ (G_) the corresponding ladder height distributions. . .r. Jn_1. . the Markov renewal process if for any n.jEE is a family of distributions on (0. A stochastic process {Xt}t>o is called semiregenerative w.. Sn = X1 + • • • + Xn the associated random walk. Let X1..i . G_(x) = P(ST_ < x.. IT.. Jo.and regenerative processes.T_ < oo). the conditional distribution of {XT„+t}t>o given Yo. < yIJ) = Fij( y) on {Jn= i. oo). A Markov renewal process {Tn} contains an imbedded renewal process. .t. A2 WienerHopf factorization Let F be a distribution which is not concentrated on (oo. Assume that uj = EjYo < oo for all j and that {J„} is irreducible with stationary distribution (v3)jEE. Y1. 0] or (0 . with common distribution F.. r+ < oo). The semiregenerative process is then regenerative w. be i. X2. . G+(x) = P(S... the semiregenerative process is called nonlattice if {T. J1 i . distribution ofjXt}t>o itself where Pi refers to the case Jo = i.t. Yn.g.. is given by Eg(X00) = 1 YO vjEj f g(Xt) dt µ jEE o where p = ujEEViAj.7 Consider a nonlattice semiregenerative process... in [APQ]. and define r+=inf{n>0: Sn>0}.. .} is nonlattice (it is easily seen that this definition does not depend on i). ..r. .
7). Proof Considering the restrictions of measures to (oc. . A C (0.=n w=m i Figure A. u . 0] and (0.1 .. >0.APPENDIX 337 Probabilistic WienerHopf theory deals with the relation between F. F(A .S.x)R+(dx).8) (e. 0). oo). 0]). n 0 R_(A) = E I(Sn E A). m<j<n}.7) (A.T_=n} = {S. In (A. we may rewrite (a) as G_ (A) = G+(A) = F(A) + (G+ * G_)(A). (e) R_ = U+. n=0 n=0 00 00 and the T+. A C (oo. Sr_ _1 is at its minimum . the renewal measures U+=>G+. U. .>0.and r_ preoccupation measures T+1 r_1 R+(A) = E E I(Sn E A).1 (a) F = G+ + G_ .x)R_ (dx). we consider the last such time (to make w unique) so that {w=m.S.. On {T_ > 2}. oo). 0<j<m. More rigorously. oo) (A. (A. define w as the time where the preT_ path S1..g.=EGn. F(A) is the contribution from the event {T_ = 1} = {X1 < 0}. G_.r. (d) R+ = U_. (c) G+(A) = f °.7) follows since G+(A) = 0 when A C (oo. S. A C (oo.G+ * G_: (b) G_ (A) = f °° F(A . A C (0. 0]. G+. F(A) + (G+ * G_)(A). n=0 The basic identities are the following: Theorem A2.
and reversing the order of summation yields P(T_ > 2.x)P(Sk < 0. ST_ E A) P(T+ = m. (b) follows from 00 G+ (A) _ E F(Sn E A.x)R+(dx). m it follows (see Fig.0<k<ri .3 8 APPENDIX Reversing the time points 0. A.m.1) that P(Sj Sn._ E A . SnEAIS. (A. and the proof of (A. m=1 f S mming over n = 2.3.. S._ E A) n1 f P(r_=nw=m Sm EduSrEA) m=1 n1 F(r+=mSr+Edu). Sn1 E dx) n=1  F(A . m < j <n.1).+ E du) E P(S.F(r_n_mSrEA_u). S.1. .XnEAx) 00 f 0 f 0 00 00 1: F(A .+ E du)P(S. r+ = n) n=1 n=1 0  C0 E fF(Sk< 0. ST_ E A .Sn_1Edx..du) (G+ * G)(A)• C llecting terms. ST+Edu). 0<j<m... A.8) is similar.. 0 < k < n.u) f0m m=1 n=m+1 00 J0 OO P(S. It follows that for n > 2 F (7.. ._ = n .>0. clearly (Sj Sm>0..= n. E du) = P(T_=nm. Sr_ E Adu) (s ee again Fig . . Aso. SmEdu) = P(T+=m..7) follows.
being concentrated at 0. . Summing over n yields R+ (A) = U_ (A). G_ are trivial. 11.SnEA) = P(Sn<Sk.6. the analogue of a random walk is a process with stationary independent increments (a Levy process.APPENDIX 339 and the proof of (c) is similar. consider a fixed n and let Xk = Xn_k+l. and G+.s. P(SnEA .G_[s]) (A.1). In continuous time.SnEA) = P(SnSn_ k. this holds always on the line its = 0. such developments motivate the approach in Chapter VI on the Markovian environment model.0<k<n. For example. then T+ = inf It > 0 : St = 0} is 0 a.P as a product H+H_ of functions with such properties. cf.Sn_k. G_ [s] are defined at the same time.O<k<n. oo)...f. if {St} is Brownian motion. 6+ [s]. In this generality of.O<k<n. see e. and similarly H_ (s) = 1 .SnEA) is the probability that n is a weak descending ladder point with Sn E A. H+ (s) = 1G+[s] is defined and bounded in the halfplane Is : ERs < 0} and nonzero in Is: Rs < 01 (because IIG+lI _< 1).G_ [s] is defined and bounded in the halfplane is : ERs > 01 and nonzero in Is : ERs > 0}. u Remark A2.4).g. Nevertheless.T+> n) = P(Sk < O. Then for A C (oo. Since G+ is concentrated on (0. a number of related identities can be derived. The classical analytical form of the WienerHopf problem is to write 1 . Sk = X1 + • • • + Xk = Sn . the survey [15] by the author and the extensive list of references there. E. u Notes and references In its above discrete time version.1.F[s] = (1 . there is no direct analogue of Theorem A2.SnEA) = P(Sn<Sk. However. there are direct analogues of Theorem A2. the derivation of the form of G+ for the compound Poisson model (Theorem 11.2 In terms of m. Again.'s. we can rewrite (a) as 1 .0<k<n. which is basic for the PollaczeckKhinchine formula. and the proof of (e) is similar.g. Another main extension of the theory deals with Markov dependence. see for example Bingham [65]. and using timereversion as in (d) to obtain the explicit form of R+ (Lebesgue measure). WienerHopf theory is only used at a few places in this book.9) whenever F[s].0+[s])(1 . is based upon representing G+ as in (b).g. For (d). The present proof of Theorem A2. it serves as model and motivation for a number of results and arguments in continuous time.1(a) is from Kennedy [228].1. In discrete time. 0]. and sometimes in a larger strip.
whereas there is no similar single established a proach in the case of matrix exponentials. Thus. 1. hen the elements of Q"/n! do not decrease very rapidly to zero and may contribute a nonnegligible amount to eQ even when n is quite large and very any terms of the series may be needed (one may even experience floating point overflow when computing Qn).10) d dteAt = AeAt = eAtA (A.11) A f eAtdt = eA. three of the c rrently most widely used ones: xample A3.13) henever A is a diagonal matrix with all diagonal elements nonzero. Eo Kn/n! converges rapidly and can be evaluated without p oblems.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. _I 0 (A. To circumvent this. Some fundamental properties are the following: sp(eA) = {e' : A E sp(A)} (A.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. however . It is seen from Theorem VIII.5 that when handling phase type distributi ons. Here it is standard to compute matrixinverses by GaussJordan el imination with full pivoting . ere A is the eigenvalue of largest absolute value. one needs to compute matrix inverses Q1 and matrix exponentials eQt ( r just eQ ). JAI = max {Jjt : µ E sp(A)} and sp(A) is the set of all eigenvalues of A (the spectrum). and eQ can then be computed as the mth power (by squaring if = 2). write eQ = (eK)m where = Q/m for some suitable integer m (this is the scaling step). 0 .12) eA'AO = Ale AA (A. if m is s fficiently large. Here are.
14) holds is therefore that the tstep transition matrix for {fft} is eQt = E ent (.e.. vp be the corresponding left . . However .]t)n (A. The probabilistic reason that (A. what is needed is quite often only Zt = TreQt (or eQth) with it (h) a given row (column) vector. The approach is in particular convenient if one wants eQt for many different u values of t. Zo = a (Z = QZ. Let vi.. One then can reduce to p linear differential equations by noting that k = ZQ. Here is a further method which appears quite appealing at a first sight: Example A3 . i.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}. p different eigenvalues Aj i .3 (DIFFERENTIAL EQUATIONS) Letting Kt = eQt.3 i (A. letting P = I + Q/i and truncating the series in the identity = e17t 00 Pn(. construction of {Xt} by realizing the jump times as a thinning of a Poisson process {Nt } with constant intensity 77. To this end..e. 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..APPENDIX 341 Example A3. condition upon the number n of Poisson events in [Olt])  Example A3.2 (UNIFORMIZATION) Formally. the procedure consists in choosing some suitable i > 0. . Ap. assume that Q is the intensity matrix for {Xt} and choose q with rt > max J%J = max qii• 1. 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 . In practice.7t) n=0 n! u °O n Pn (to see this. Zo = h). some jumps are dummy in the sense that no state transition occurs )..15) Then it is easily checked that P is a transition matrix . 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 = q22.4 (DIAGONALIZATION) Assume that Q has diagonal form.
16) (A. vi.. Complex calculus : Typically. however.18) Namely. say Al. i=1 i=1 Thus. Example A3. (A. and we may adapt some normalization convention ensuring vihi = 1. In view of this phenomenon alone care should be taken when using diagonalization as a general tool for computing matrixexponentials. The phenomenon occurs not least when the dimension p is large.. hp..g H1.17) eQt = E e\`thivi = E ea:thi ® vi.. Qhi = vihi.5 If Q= ( 411 ( q21 q12 q22 is 2 x 2.18) contains terms which almost cancel and the loss of digits may be disasterous. we can take H as the matrix with columns hl. of largest real part is often real (say. we have an explicit formula for eQt once the A j. Then vihj = 0.. this last step is equivalent to finding a matrix H such that H1QH is a diagonal matrix.. Nevertheless. under the conditions of the PerronFrobenius theorem). two serious drawbacks of this approach: u Numerical instability : If the A5 are too close. Then P P Q = > Aihivi = E Aihi (9 vi. i= 1 i=1 P P (A.342 APPENDIX (row) eigenvectors and hl. some cases remain where diagonalization may still be appealing. say A = (Ai)diag. There are.. and we need to have access to software permitting calculations with complex numbers or to perform the cumbersome translation into real and imaginary parts. v5Q = Aivi. i # j. hi have been computed. not all ai are real. and hence A2 is so because of A2 = tr(Q). Everything is nice and explicit here: 411+q2+D' )12_g11+q2^^ where (411422z + 4412421. the eigenvalue. and writing eQt as eQt = He°tH1 = H (e\it)di. (A. D = ) 2 2 . and vihi ¢ 0.. hp the corresponding right (column) eigenvectors.
6 A particular important case arises when Q = q1 qi ) q2 q2 J is an intensity matrix.APPENDIX 343 Write 7r (= v1) for the left eigenvector corresponding to a1 and k (= hl) for the right eigenvector. Then 7r = (ir1 7r2 ) = a (q21 Al . u Example A3.21) Here the first term is the stationary limit and the second term thus describes the rate of convergence to stationarity. b are any constants ensuring//Irk = 1. k  C k2 ) =b ( A1 q 1 Q11 / where a .19) Example A3 . i. However.k1).Q2i and after some trivial calculus one gets eQt = 7r 1 112 + eat 7r1 7r2 / (7fl 7r2) = ( 7r2 1r2 7r1 IF. replacing ai by A2. The other eigenvalue is A = A2 = q1 . where (A. Of course. 1) . l ab (g12g21 + (A1  411) 2) = 1. Then Al = 0 and the corresponding left and right eigenvectors are the stationary probability distribution 7r and e.20) ir = q2 ql qi +q 2 9l +q2 (A.q. it is easier to note that 7rh2 = 0 and v2k = 1 implies v2 = (k2 . h2 = Thus. 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.7 Let 3 9 2 14 7 11 2 2 .e.
APPENDIX x1 3/2 .5 . A+AA+ = A+. but only that dimensions match .11/2 .11/2 + 5 1..344 Then D= 2+ 11)' 7 T4 2 =52.satisfying AAA = A. e_6u A4 Some linear algebra 4a Generalized inverses A generalized inverse of a matrix A is defined as any matrix A.. 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_. Generalized inverses play an important role in statistics. (A+A)' = A+A. (A.6.22) Note that in this generality it is not assumed that A is necessarily square.23) . and a generalized inverse may not unique. (AA+)' = AA+. A2 = 3/2 . (A. 2 2 1=ab(142+(1+2)2 ) = tab. They are most often constructed by imposing some additional properties . for example AA+A = A.
Here is a typical result on the role of such matrices in applied probability: Proposition A4. ( Q . and define D = (A .. (I . 0 01 In applied probability.1 Let A be an irreducible intensity matrix with stationary row vector it.e ® 7r)1. and exists and is unique (see for example Rao [300]). Then for some b > 0. lt o eAx dx = te7r + D(eAt .1 goes under the name fundamental matrix of the Markov chain). most often either an intensity matrix Q or a matrix of the form IP where P is a transition matrix..I) (A.eir ). one then works with Q = (Q .P). E. and can define /ail 0 0 0 0 0 0 A+ = C A' 0 0 0 C' .P + e7r)1 (here ( I . one is also faced with singular matrices . if A is a possibly singular covariance matrix (nonnegative definite).25) . Am+1 = . (A. . are ordered such that Al > 0.. ..eir )1. _ A.ew..1Q = Q(Q . then there exists an orthogonal matrix C such that A = CDC' where 0 0 D = AP Here we can assume that the A .24) = te7r . = 0 where m < p is the rank of A.= (I . Assume that a unique stationary distribution w exists .g.e.eir)1 = I .D + O(ebt). These matrices are not generalized inverses but act roughly as inverses except that 7r and e play a particular role ..APPENDIX 345 A matrix A+ satisfying (A.23) is called the MoorePenrose inverse of A. Rather than with generalized inverses .g. Am > 0.P + e7r ).
I.I) (A.26) 2 = 2 e7r + tD . Interpreting 7r.27) Proof Let A(t). h as 1 x m and k x 1 matrices.s..346 t APPENDIX 2 xe Ax dx = eir + t(D + e7r) + D(eAt . Note that h ® it has rank 1. the r. . 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. (A.I) . in block notation i2h A®B= ( a11B a21 B a12B a22 B Example A4. and in fact any rank 1 matrix can be written on this form. see below. the rows are proportional to it.h.e. of (A.91a(2) . h ® it reduces to hit in standard matrix notation.2e7r . it follows that h ® it is the k x m matrix with ijth element hi7rj .I)}.eir)eAt = eAt = A'(t). (A.24). For example.s. B'(t) = e7r + DAeAt = eir + (I . Then A(O) _ B(O) = 0. .DZ(ent . resp. and the columns to h. the formulas involving O(e6t) follow by PerronFrobenius theory.J {xe^r + D(e .26) follows by integration by parts: t f t /' xeAx dx = [x {xe7r + D(eAx . o Finally. 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.D + D2 + O(ebt).h.2 Let it be a row vector with m components and h a column vector with k components.3 Let 2 A= 4 3 Vf' N7 5 )' B= ( 8 ). ()®(6 f 6/ 7f 8^ 7 8 )=! ^)( 6 7 8 )=(6^ 7^ 8^) \ u Example A4. respectively. B(t) denote the l.I)} dx. Equivalently.
APPENDIX 347 Then A®B = 2 f 20.31) Indeed. it follows that e® ® e B An _ 0o oo oo Bn 7 I F n! = ` k! (I .3v'6. 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. and v1B1h1 • v2B2h2 = v1B1h1 ® v2B2h2 = ( v1(&v2 )( B1(&B2 )( h1(&h2 ) .3V8. (AED B)1 = (A®I+I(9 B)l is the sum of all products of t factors. (A B)' = eA®B e! L 1=0 0 . Using (A.3vV/72f 20.29). such a factor is Ak (&B 1k according to (A. A2 = v2 are row vectors and C1 = h1.5v/.4vf. C2 = h2 are column vectors. then v1B1h1 and v2B2h2 are real numbers.5v'8 5vf9 11 A fundamental formula is (A1B1C1) ®(A2B2C2) = (A1 (9 A2)(B1 (9 B2)(C1®C2). if Al = vi. (A. each of which is A ® I or I ® B.k)! ( n0 n=0 t=0 k=0 J _ ® Ak ®Blk r ^.29) If A and B are both square (k1 = ml and k2 = m2). and the number of such factors is precisely given by the relevant binomial coefficient.3f 4v/.(A. if A ® I occurs k times.4 eA® B = eA ®eB.31).30) eA+B = eAeB function generalizes to Kronecker notation (note that in contrast typically only holds when A and B commute): Proposition A4.A9.50 6 7 6 4f 4.28) In particular. then the Kronecker sum is defined by A(1) ®A(2) = A(1) ®Ik2 + k ®A(2).
{ On the other hand. Then 2 0 ire At h • ve Bt kdt = (^®v)(A®B)1(e A®Ba . resp . Yt(2 ) }. and Q = Q(1) ® Q (2) = Q(1) ® I + I ® Q(2) (A. Ps 1) = exp {sQ ( 1) } > p(2 ) = exp {sQ(2) } can therefore be rewritten as Taking s = 1 for simplicity .4 can easily be obtained by probabilistic be the sstep transition reasoning along the same lines . represents ces Q( 1). Thus . P(t) Yt(2) }. X ) }. A special case of Proposition A4. P(2). where transition matrix of the bivariate Markov chain {X n1). n2 n1 ) {X(2) } are independent Markov chains with transition matrices P(1). in the definition (A.33) .6 Suppose that A and of B. Let P8f P(Sl).5 Many of the concepts and results in Kronecker calculus have p(2) is the intuitive illustrations in probabilistic terms. first term on the r .348 APPENDIX Remark A4. p = P(1) ® {X }. k any column vectors.I)(h ® k). h. Q(2). Yt(2) where independent Markov processes with intensity matri{y(2) } are {Y(1) }. From what has been said about matrices of {Yt( 1).32). independent Markov chains.32) is the intensity matrix of the bivariate continuous Markov process {Yt(1). the {Yt(2) } transitions in the {Yt(1) } component and the second transitions in the component . we have P8 = Pal) ® p(2). { 1't(1) }.3 < 0 Lemma A4 . P8 = exp {sQ} = exp {s (Q(1) ®Q(2)) } . the same time. {Yt(1). v whenever a is an eigenvalue of A and 0 is an eigenvalue be any row vectors and h. and the form of the bivariate intensity matrix reflects the fact that Yt(2) } cannot change state in both components at due to independence .s. 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 +. Let further it. (A.
(b) if in addition A is aperiodic. . then IN < Ao for all A E sp(A). h can be chosen with 3By this.12). j = 1. = j and atk_li. h such that vh = 1. E (0. the integrand can be written as ( 7r (9 v)( eAt ® eBt )(h ®k ) = ( 7r ®v)(eA (DBt)(h (& k). i. in such that io = i. Now note that the eigenvalues of A ® B are of the form a +. il. [APQ] X. . Then: (a) The spectral radius Ao = max{JAI : A E sp(A)} is itself a strictly positive and simple eigenvalue of A. . h = e and v = 7r (the stationary row vector). 4c The PerronFrobenius theorem Let A be a p x pmatrix with nonnegative elements. and appeal to (A. n.34) Note that for a transition matrix. We call A irreducible if the pattern of zero and nonzero elements is the same as for an irreducible transition matrix.. . so that by asssumption A ® B is u invertible.. . > 0 for k = 1. see e. A is called aperiodic if the pattern of zero and nonzero elements is the same as for an aperiodic transition matrix. Similarly.29)..7 Let A be a p x pmatrix with nonnegative elements. we have AO = 1. (A. and the corresponding left and right eigenvectors v. . .APPENDIX 349 Proof According to (A. we mean that the pattern of nonzero offdiagonal elements is the same as for an irreducible intensity matrix.The PerronFrobenius theorem has an analogue for matrices B with properties similar to intensity matrices: Corollary A4. ao). which can be found in a great number of books.. Then the eigenvalue Ao with largest real part is simple and real.1 and references there (to which we add Berman & Plemmons [63]): Theorem A4. .. Here is the PerronFrobenius theorem.. .. and if we normalize v. then An = Aohv+O(µ") = Aoh®v+O(µ") for some u. p there should exist io. h can be chosen with strictly positive elements.8 Let B be an irreducible3 p x pmatrix with nonnegative offdiagonal elements. f o r each i.g. and the corresponding left and right eigenvectors v.3 whenever a is an eigenvalue of A and 3 is an eigenvalue of B. That is.
The next result gives a condition for asymptotical exponentiality. not only in the tail but in the whole distribution. Corollary A4. if we normalize v. one can consider A = 77I + B where rl > 0 is so large that all diagonal elements of A are strictly positive (then A is irreducible and aperiodic). we have A0 = 0. let t = (ti)iEE # 0 have nonnegative entries and define T(°) = aQ . relate the eigenvalues of B to those of B via (A.2). h such that vh = 1. Then for any (3. To this end. Bi° (x) + at*x Proof Let { 4 } be the phase process associated with B(a) and (°) its lifelength. Note that for an intensity matrix.e.(3. the analogy of this procedure with unformization. it was shown that under mild conditions the tail of a phasetype distribution B is asymptotical exponential. let {Yti°i } be a Markov process with initial distribution a and intensity .(ti)ding.35) for some p E (oo. the condition is that t is small compared to Q. Proposition A5.350 APPENDIX strictly positive elements. The content is that B is approximately exponential if the exit rates ti are small compared to the feedback intensities tij (i # j). A5 Complements on phasetype distributions 5a Asymptotic exponentiality In Proposition VIII.. For example. Furthermore.(ti)diag where Q = T + (ti)diag is a proper intensity matrix (Qe = 0). I.n t AL n=0 n! (cf.8 is most often not stated explicitly in textbooks. the phasetype distribution B(a) with representation (. Example A3.1. but is an easy consequence of the PerronFrobenius theorem.1 Let Q be a proper irreducible intensity matrix with stationary distribution a. then eBt = ea0thv + O(eµt) = eA0th ® v + O(et t) (A. note that we can write the phase generator T as Q . T(°)) is asymptotically exponential with parameter t* _ r EiEE aiti as a 4 oo. 10) and use the formula me at e Bt = e 00 Antn = e . h = e and v = 7r (the stationary row vector).8. Ao).
By the law of large numbers for Markov processes . a' = a .9. {t Y( a) } v>0 . Then a(a'V)/a (aV) a' 1. we get dx F (Idx = j) = (1 + qij t )Sij + qij dt. it states that the state. We can assume that Jta) = Yt(°). and this easily yields a(x)/x a' 1/t*.x (1 .)_ = Y(a) = 1'aS(a) = Ya(av)^ it follows that Pi ((. and write Yt = Yt(1). fo tY dv/t a$' t*. has a limit distribution: Proposition A5. from which the phase process is terminated . J^O)_ = j) Pi (v(aaV) > x.a' + oo (e. Hence O ((a) aa. 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. in fact . and that Yt(a) = Yat for all t. dx/ti] or not. Since JJ(.YQ(av) = j) Pi ( ci(a'V) > x.Yj(av) = j f . a'/a + 1.aE where 0 < e < 1). Hence we can represent ( (a) as ((a) = inf { t > O : f tY( )dv=V } ^l = inf { t > O : t adv = V } l jat inf{t > 0: tydv =aV} = JJJ a J J where o (x) = inf {t >0: fo tY dv = x}. Then {Ix} is a Markov process with to = Yo. J(()) _ = i) + at•x t tt' . We can think of ( ( a) as the first event in an inhomogeneous Poisson process ( Cox process ) with intensity process matrix aQ .2 Pi (c(a) > x.1. Conditioning upon whether { Yt} changes state in [0. = YQ(x). Proof Assume first ti > 0 for all i and let I.(a) > x . In addition to the asymptotic exponentiality.APPENDIX 351 ((1) etc. prove a somewhat more general result which was used in the proof of Proposition VI. We shall . t < (a).jEE.bij) Hence the intensity matrix of { Ix} is (qij/ti)i. a .g. v/ t. Let further V be exponential with intensity V and independent of everything else.
a) if B is the lifelength of a terminating Markov chain (in discrete time) on E which has transition matrix P = (p. (c) the nth moment k 1 k"bkis 1)"n!aP"p.. say bk = 0.. 5b Discrete phasetype distributions The theory of discrete phasetype distributions is a close parallel of the continuous case..3 As the exponential distribution is the simplest continuous phasetype distribution.. (b) the generating function b[z] _ E' . > 0}. u Notes and references Propositions A5. so is the geometric distribution.p)k1 p.zP)'p.352 rr Ia(a'V) Ei I ( > x) P APPENDIX L at (Yo (aV) . = 0 for one or more i. so we shall be brief. Penev & Turbin [238]. Then: (a) The point probabilities are bk = aPklp.g. an easy modification of the argument yields finally the result for the case where t.1 and A5.Pe. P. these results are in the spirit of rare events theory for regenerative processes (e.+ at*x • a't' L ` at t* t* J Reducing the state space of {Ix } to {i E E : t. k>1. and thus the parameter p of the geometric distribution u can be identified with the exit probability vector p.4 Any discrete distribution B with finite support. Keilson [223]. zkbk is za(I . Gnedenko & Kovalenko [164] and Glasserman & Kou [162]). Et II I a(a^V) > x) at' .. A distribution B on {1. . See also Korolyuk. k = 1. . Indeed.j) and initial distribution a. ' pk 0 k>1 11 Theorem A5. the simplest discrete phasetype distribution: here E has only one element. let E and Pkj j=k1.2 do not appear to be in the literature. However. a = b = (bk)k=1.} is said to be discrete phasetype with representation (E.. 2.. .5 Let B be discrete phasetype with representation (P. K}. is discrete phasetype. with point probabilities bk = (1 .x k > K. Example A5.... Then P is substochastic and the vector of exit probabilities is p = e . 2. Example A5.. 1 k=1 1 0 otherwise. a).
The discrete counterpart is the negative binomial distribution with point probabilities bk k1) (1 k = r. (E(2).2 The form of these results is easily recognized if one considers two independent phase processes { Jt 1) }. r . { Jt 2) } with lifetimes U1 .. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2). T= ( 0 T(2) ) (A.7 (THE NEGATIVE BINOMIAL DISTRIBUTION) The most trivial special case of Example A5. .6. resp..{ 0.T(2)). and hence the negative binomial distribution is discrete phaseu type.6 is the Erlang distribution Er which is the convolution of r exponential distributions.T(1)). A.a(1). A reduced phase diagram (omitting transitions within the two blocks) is am E(1) t(1) a(2) (2) t(2) Figure A. initial distribution a and phase generator T. Then {Jt} has lifetime U1 + U2 . U2. a. Jt t > U1 + U2.APPENDIX 353 5c Closure properties Example A5.a(2). 11 Example A5. and piece the processes together by it = 41) 0<t<U1 U1 < t < U1 + U2 2U.r + 1.. _ i E E(1) T(1) t(1)a(2) i E E(2) . a' . resp. as is seen by minor modifications of Example A5. B2 be phasetype with representations (E(1). and a=1).36) in blockpartitioned notation (where we could also write a as (a (1) 0)).6 (CONVOLUTIONS) Let B1. Then the convolution B = B1 * B2 is phasetype with representation (E.1 This corresponds to a convolution of r geometric distributions with the same parameter p..
resp. A reduced phase diagram is 0a(1) E(1) A . and o'i Oa.E) where a(°) = fAa(a)v(da).T.p)pn1B*n. Example A5. Then it is trivial to see that B(") is u phasetype with representation (a("). Equivalently.354 APPENDIX Example A5. P(N = n) = (1 .'). are i.0)a(2))). p at each termination. this means that a = (Oa(1) (1 .a(1).a(2). B2 be phasetype with representations (E(1).p)pn1. i E E(2) 0 T(2) =IT (in blockpartitioned notation. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2).d.10 (GEOMETRIC COMPOUNDS) Let B be phasetype with representation (E. a mixture of more than two phasetype distributions is seen to be phasetype. a reduced phase diagram is f a E t Figure A. Thus.O)B2 (0 < 0 < 1) is phasetype with representation (E.37) (1) (1 .. U2. one obvious interpretation of the claim u size distribution B to be a mixture is several types of claims.T(1)). then C is the distribution of Ul + • • • + UN. In risk theory. i E E(1) T 0 I (A. Then the mixture B = 9B1 + (1 . Example A5. with common distribution and N is independent of the Uk and geometrically distributed with parameter p..p.T(2)). Let B(") be the corresponding phasetype distribution.0)a(2) E(2) Figure A. we need to restart the phase process for B w.3 In exactly the same way. and consider B(") = fA B(a) v(da) where v is a probability measure on A. T) and C = EO°_1(1 . a.4 .8 (FINITE MIXTURES) Let B1.. if U1. (E(2). a. To obtain a phase process for C.0)ai2).i.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.
. U2 be random variables with distributions B1. T). Proposition VIII.d.T) if U is phasetype with representation (E. For U1 A U2. j E F}. be the point probabilities of a discrete phasetype distribution with representation (E.7. cf. a. then U1 +• is phasetype with representation (E. 12 (PHASETYPE COMPOUNDS ) Let fl.APPENDIX 355 and C is phasetype with representation (E. Note that this was exactly the structure of the lifetime of a terminating renewal u process. . { Jt2) } be independent with lifetimes U1. let the initial vector be a ® v and u let the phase generator be I ® T + P ® (ta).g. Example A5 . resp.2. i. Example A5.TWWW). B2 of phasetype with representations (E(').. Thus the representation is (E(1) x E(2). (E(2). of F. cf. it follows by mixing (Example A5. but the same T. P). are i. Indeed. then C is the distribution of U1 + • • • + UN. T) and C = F. Minor modifications of the argument show that 1. a(2). To obtain a phase representation for C ..aeTx. f2. let B be a continuous phasetype distribution with representation (F.2. then U1 + • • + UN is zeromodified phasetype with representation (a. we then let the governing phase process be {Jt} _ {(411 Jt2))} 2) interpreting exit of either of {4 M }. a. with common distribution B and N is independent of the Uk with P(N = n) = f. T(1) ® T(2)). Example A5 . let {Jtl)}. say v.aF[T]. U2.. T + pta). T + ta. if {Jt} is a phase process for U. then Jy has distribution aeTx.T) where F[T] = J0 "o eTx F(dx) u is the matrix m. v.. Then the minimum U1 A U2 and the maximum U1 V U2 are again phasetype. a.. E). { 4 } as exit of {Jt}. T(2) ).9) that (U .°_1 f„ B*?l. To see this.T + pta). resp.1. 13 (MINIMA AND MAXIMA ) Let U1.. v. X independent of U.X)+ is zeromodified phasetype with representation (E. . let the phase space be E x F = {i j : i E E.. if B is defective and N + 1 is the first n with U„ = oo. U2.x)+.f.a(1).°.v. +UN 2. It is zeromodified phasetype with representation (E. say with distribution F.11 (OVERSHOOTS) The overshoot of U over x is defined as the distribution of (U . If U1 has a different initial vector. Equivalently. a(1) ® a(2 ). if U1. Corollary VIII. If we replace x by a r.
.n = I:pi(n)Er v ( __ ) n) ) a= 1 .. the fact that any distribution B can be approximated arbitrarily close by a distribution with finite support. Example A5. Then we must find phasetype distributions Bn with B. see Neuts [269] (where the proof. oo). Hence it is immediate that Bn 4 B. oo) can be approximated 'arbitrarily close' by a phasetype distribution B: Theorem A5. The general case now follows easily from this.(bk)'. r # oo.14 To a given distribution B on (0. the initial vector is (a(1) (& a (2) 0 0).356 APPENDIX For U1 V U2... cf.xq(n)(n)}. there is a sequence {B. we can assume that ID.(Sn) with Sn = n/b.} of phasetype distributions such that Bn 3 B as n + oo.B(bk) I < 1/n for n > k. Now we can find first a sequence {Dm} of distributions with finite support such that D. q(n) q(n) pi(n)a . By the diagonal argument (subsequent thinnings).(n) = D...(bk) + B(bk) for all k..8. however. 5d Phasetype approximation A fundamental property of phasetype distributions is denseness .2) } to go on (on E(2)) when { i 1) } exits. and let Bn be the Erlang distribution E. relies more on matrix algebra than the probabilistic interpretation exploited here). i= 1 C. Let the support of Dn be {xl(n). and the phase generator is T(1) ®T(2) T(1) ®t(2) t(1) ® T(2) 0 T(1) 0 0 0 T(2) Notes and references The results of the present section are standard . The mean of B„ is n/Sn = b and the variance is n/Sn = b2/n. Thus the state space is E(1 ) x E(2) U E(1) U E( 2). That is. any distribution B on (0. Then from above. Proof Assume first that B is a onepoint distribution.(bk) + B(bk) for all k as n * oo. and vice versa. Here are the details at two somewhat different levels of abstraction: (diagonal argument . with weight pi(n) for xi(n). we need to allow { Jt. elementary) Let {bk} be any dense sequence of continuity points for B(x). say degenerate at b.. and the closedness of the class of phasetype distributions under the formation of finite mixtures.
D(bk)I < n. k < n. For a general Bo. for some a < oo. . if information on Bo is given in terms of observations (i. i.t. u 2 (abstract topological ) The essence of the argument above is that the closure (w. and that cp is known to be continuous. oo) and any fl. say on the claim size distribution B in risk theory. Since PET is closed under the continuous operation of formation of finite mixtures. f2. then it is immediate that WI(B) = p2(B) for all distributions B on [0. k < n. Let E be the class of functions f : [0. and we can take Bn = Cr(n). E E. Hence G C PET and L = PIT.n( b k ) .n.. Then ICr( n ). u Theorem A5. one would use the B given by some statistical fitting procedure (see below).. It should be noted. Corollary A5.. we can then approximate Bo by a phasetype B. that this procedure should be used with care if ^p(B) is the ruin probability O(u) and u is large. But To is the class G of all distributions on [0. x 4 oo. 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. the topology for weak convergence) PET of the class PET of phasetype distributions contains all onepoint distributions. If Cpl (B) and ^02(B) are weakly continuous..APPENDIX 357 Hence we can choose r(n) in such a way that ICr( n). however. compute W(B) and use this quantity as an approximation to cp(B0).B(bk )I < ..(x)Bf.. there is a sequence {Bn} of phase type distributions such that Bn Di B as n 4 oo and f ' f. oo) * [0.i. the class CO of all discrete distributions. PIT contains all finite mixtures of onepoint distributions. 2. oo) approximation Assume that we can compute a functional W(B) when B is phasetype.n (bk) .e.d. replications).( dx) * f r f{(x)B(dx). oo) such that f (x) = O(e«x).15 To a given distribution B on (0 . oo)..r.14 is fundamental and can motivate phasetype assumptions.. i = 1. In particular.
n B=az..16 To a given distribution B on (0 . i = 1. then cc f (x)Bn ( dx) = (?!c ) e'= . Bn=En z f f (x)Bn(dx) fof (x)B(dx) = ° (A. and hence we may choose r(n) such that L 9l) f (x)Cr(n). .38) We first show that for each f E E.39) Indeed. \\ 0 Corollary A5. i = 1. 2.. for each i. and the case of a general f then follows from the definition of the class E and a uniform integrability argument.n(dx) < 1+... . .... we may assume that in the proof of Theorem A5. . TO (A.f (z) = f = 1 1 1 1n/ o .2 .. there is a sequence {Bn} of phase type distributions such that Bn Di B as n + oo and all moments converge. Now returning to the proof of (A.  APPENDIX B implies that 00 o o 00 n. i=1.(dx) > J fi(x)B(dx).oo J fi(x)B.. liminf B.. oo). if f (x ) = e°x. i = 1.f ' f (x)B(dx).358 Proof By Fatou' s lemma. f° xtBn(dx ) * f °° x`B( dx).14 Dn has been chosen such that 00 1 °° f fi(x)D n(dx ) < 1++ '  o \ n o f fi(x)B(dx). n..n(dx) + f 0 fi(x)Dn(dx).. n..39). . By (A. f00 fi(x)Cr.38 ). and hence it is sufficient to show that we can obtain limsup n4oo fi(x)Bn(dx) < Jo 0 f fi( x)B(dx ).f (x)B(dx).
If ei > 0. We shall formulate the problem in the slightly broader setting of fitting a phasetype distribution B to a given set of data (1i . . .l3µb < 1. . However. the problem thus arises of how to fit a phasetype distribution B to a given set of data (1.. lim inf > is proved similarly. from a more conceptual .14 is classical. then Bn['Y + ei] * B[y + ei] > 1 + 7 Q implies that 'yn < ry + ei for all sufficiently large n . lim sup ryn < 7. Proof Let fi(x) = el'r+E. The adjustment coefficient is a fundamental quantity. I. . but are certainly not unexpected. and therefore the following result is highly relevant as support for phasetype assumptions in risk theory: Corollary A5. . . oo) with B[y +e] < oo for some e > y = 7(B.APPENDIX 359 In compound Poisson risk processes with arrival intensity /3 and claim size distribution B satisfying . e ) and ei J. the remaining results may be slightly stronger than those given in the literature. For practical purposes. This is motivated in part from the fact that a number of nonphasetype distributions like the lognormal. the adjustment coefficient 'y = 7(B. (N or a given distribution Bo. 0 as i * oo.. O We state without proof the following result: Corollary A5. the loggamma or the Weibull have been argued to provide adequate descriptions of claim size distributions. /3) = ry for all n.16.. there is a sequence {B./3) is defined as the unique solution > 0 of B[y] = l+y/j3./3). Notes and references Theorem A5. . 5e Phasetype fitting As has been mentioned a number of times already. one can obtain 7(Bn. The present section is a survey of some of the available approaches and software for inplementing this.3).> y for some sequence {ei} with ei E (0.17 To a given /3 > 0 and a given distribution B on (0. there is substantial advantage in assuming the claim sizes to be phasetype when one wants to compute ruin probabilities.} of phasetype distributions such that Bfz + B as n * oo and Yn 4 ry where ryn = y(Bn. (N.e.18 In the setting of Corollary 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.
d. e . and this is what matters when using phasetype distributions as computational vehicle in say renewal theory. The constraints were the exact fit of the two first moments and the objective function to be minimized involved the deviation of the empirical and fitted c. A number of approaches restrict the phase type distribution to a suitable class of mixtures of Erlang distributions . (N is the empirical distribution Be. Of course. we do not not want to perform matrix calculus in hundreds or thousands dimensions). Johnson & Taaffe considered a mixture of two Erlangs (with different rates ) and matched (when possible ) the first three moments . cf.. A method developed by Bobbio and coworkers (see e. giving mass 1 /N to each S=.g. risk theory. Schmickler (the MEDA package. [216] ). In a series of papers (e. The likelihood function is maximized by a local linearization method allowing to use linear programming techniques. one could argue that the results of the preceding section concerning phasetype approximation contains a solution to our problem : given Bo (or Be). reliability or queueing theory. the number of phases required for a good fit will typically be much larger.} of phasetype distribution such that Bo. g. [70]) restrict attention to acyclic phase type distributions . three for a mixture of two Erlangs ).d. The earliest such reference is Bux & Herzog [85] who assumed that the Erlang distributions have the same rate parameter.f. the L1 distance between the c .360 APPENDIX point of view the two sets of problems are hardly different : an equivalent representation of a set of data (1 . [202]. Asmussen & Nerman [38] implemented maximum likelihood in the full class of phasetype distributions via the EM algorithm . . and in practice this sets a limitation to the usefulness (the curse of dimensionality . B„ The problem is that the constructions of {B„} are not economical : the number of phases grows rapidly.g. a program package written in C for the SUN workstation or the PC is available as shareware. . It seems therefore a key issue to develop methods allowing for a more general phase diagram... [317] ) has considered an extension of this setup. 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. The characteristics of all of these methods is that even the number of parameters may be low (e. we have constructed a sequence { B. and as fitted distribution we may take B. for some suitable large n.g. defined by the absence of loops in the phase diagram .. at a a number of selected points .g .'s). and we next describe two such approaches which also have the feature of being based upon the traditional statistical tool of like maximum likelihood. The observation is that the statistical problem would be straightforward if the whole ( EAvalued) phase process { Jtk)} o<t<( k associated with each observa .f. and used a nonlinear programming approach .
(N ) (^ 54 k )+ and similarly for the cn+1) The crux is the computation of the conditional expectations.APPENDIX 361 tion Sk was available.... .. . = j) f k=1 k =1 tE[0. E. Thus.(N) = E Ea(n).. In fact. the methods of [70] and [38] appear to produce almost identical results.T(n) (Ti ^^ 1. it is easy to see that N (k Ea(n).. N Ti = I(J= i) dt. since this is parameterdependent.x)t(n) 1 and this and similar expressions are then computed by numerical solution of a set of differential equations. one is lead to an iterative scheme.(k] (Ti is the total time spent in state i and Nii is the total number of jumps from i to j).. e.. (n+1) _ Ea (n).T(n) k=1 I (Jti) dt o \f a(n)eT(n )(kt(n) N f:i a(n)eT(n)xei . In practice. (N) tJk Ea ( n). eieT(n)((k. . then the estimators would be of simple occurenceexposure type.g..g. . Nii = = . . The general idea of the EM algorithm ([106]) is to replace such unobserved quantities by the conditional expectation given the observations. jEEA.T (n)(TiI(1.T(n) (Nik IC1. EN where ai = N 1 I ((k) = i) tii=i iEE. it seems open whether the restriction to the acyclic case is a severe loss of generality.
This page is intentionally left blank .
Probab. Appl. 219229. 253267. Asmussen ( 1982 ) Conditioned limit theorems relating a random walk to its associate. 363 . Sparre Andersen (1957) On the collective theory of risk in the case of contagion between the claims . [12] S. [10] K. Asmussen (1989a) Aspects of matrix WienerHopf factorisation in applied probability.). Preprint. New York. In: Limit Theorems and Related Problems (A. Asmussen (1987) Applied Probability and Queues. Aktuar Tidskr. J. Dover. B. Scand. Skand. Queueing Systems 10. Arfwedson ( 1955) Research in collective risk theory. Asmussen (1984) Approximations for the probability of ruin within finite time . Stegun ( 1972 ) Handbook of Mathematical Functions (10th ed. Proc. 213229. Teubner. Adv. 311338. Asmussen (1985) Conjugate processes and the simulation of ruin problems. 57. [11) S. John Wiley & Sons. [3] J. The case of equal risk sums. Abate & W. Anantharam (1988 ) How large delays build up in a GI/GI11 queue. [8] G. [4] M. 20. 1985. Skand. 37. AT&T. [9] G.Bibliography [1] J. [13] S. New York. 345368. [15] S. [7] E. 587. Optimizations Software. 14. 38. Appl. with applications to risk reserve processes and the GI /G/1 queue. Arndt (1984) On the distribution of the supremum of a random walk on a Markov chain . Chichester New York. 3157 . 143170. Transactions XVth International Congress of Actuaries. Alsmeyer (1991) Erneuerungstheorie. Whitt (1994) Waitingtime tail probabilities in queues with longtail service time distributions . Stuttgart. ibid.). ed. [14] S. Abramowitz & I. G. pp. 1984 . [5] G. Abate. Aktuar Tidskr. II. Act. 53100. Whitt ( 1992) The Fourierseries method for inverting transforms of probability distributions . 101116. Queueing Systems 5 . Whitt (1998) Explicit M/G/1 waitingtime distributions for a class of longtail service time distributions . [2] J.G. Queueing Systems 16. The Mathematical Scientist 14. [6] V. Arfwedson ( 1954) Research in collective risk theory. 191223. New York. Choudhury & W. Borokov. Stoch.A. Abate & W.L.
Proc. Act. stationary distributions and first passage probabilities. Extremes 1 . [26] S. Schmidt (1995) Does Markovmodulation increase the risk? ASTIN Bull. 555574. Proc. Asmussen (1999) On the ruin problem for some adapted premium rules. Scand. Floe Henriksen & C. CRC Press. Asmussen (1998a) Subexponential asymptotics for stochastic processes: extremal behaviour.M. Asmussen.). Asmussen (1992a) Phasetype representations in random walk and queueing problems. Asmussen & B. 96107. Appl. eds. 20. Appl. J. Methods V Problems (J. Advances in Queueing: Models. . Hojgaard (1996) Finite horizon ruin probabilities for Markovmodulated risk processes with heavy tails. Stoch. Ann. 354374. Marcel Dekker. eds. Asmussen (1995) Stationary distributions via first passage times. J. Boca Raton. Chakravarty. [27] S. [34] S. Scand. 1999 . Bladt (1996) Renewal theory and queueing algorithms for matrixexponential distributions. Andronov. Ann. 772789.). 25. [21] S. Statist. [31] S. Probab. Riga Aviation University. SIAM Review 40.). Stochastic Models 11. Asmussen (1998b) Extreme value theory for queues via cycle maxima. 2943. Asmussen (1991) Ladder heights and the Markovmodulated M/G/1 queue. 79102. Probab. Binswanger & B. 189201. [23] S. J. Kalashnikov & A. Asmussen (1992c) Stationary distributions for fluid flow models and Markovmodulated reflected Brownian motion. J. [33] S. Act. [17] S. Scand. [28] S. Probabilistic Analysis of Rare Events (V. 297318.364 BIBLIOGRAPHY [16] S. 303322. K. [22] S. Appl. Florida. Asmussen & B. T. Asmussen (1989b) Risk theory in a Markovian environment. 1989 . 315. [24] S. 37. Asmussen (2000) Matrixanalytic models and their analysis. [20] S. Bladt (1996) Phasetype distributions and risk processes with premiums dependent on the current reserve. Kliippelberg (1994) Large claims approximations for risk processes in a Markovian environment. Hojgaard (1999) Approximations for finite horizon ruin prob abilities in the renewal model. Random Processes 2. 193226. [30] S. Asmussen. Frey. 69100. Dshalalow ed. Rolski & V.K. 27. Binswanger (1997) Simulation of ruin probabilities for subexponential claims. Stoch. 313341. Appl. Asmussen & M. 1996. 8. Th. 2. [19] S. Hojgaard (2000) Rare events simulation for heavytailed distributions. Asmussen (1992b) Light traffic equivalence in single server queues. Ann. [25] S. Probab. Astin Bulletin 27. Asmussen (1998c) A probabilistic look at the WienerHopf equation. 137168. Asmussen & M. 4966. [29] S. Asmussen & K. Act. L. Scand. Bernoulli 6. Asmussen. Alfa & S. [32] S. New York. 313326. [18] S. 2149. 54. A.S. 1936. MatrixAnalytic Methods in Stochastic Models (A. 106119.
105119. Res. Insurance: Mathematics and Economics 20. Appl. J. 421458. 19. Johnson.). Asmussen & T. Adv. J. [36] S. Appl. 58. 23. Schock Petersen (1989) Ruin probabilities expressed in terms of storage processes. Rolski (1994) Risk theory in a periodic environment: Lundberg's inequality and the CramerLundberg approximation. Schmidli & V. Schmidt (1993) The ascending ladder height distribution for a class of dependent random walks. Asmussen & R.BIBLIOGRAPHY 365 [35] S. Appl. [48] S. Asmussen.M. [46] S. Boca Raton. 429466.). Taksar (1997) Controlled diffusion models for optimal dividend payout. Asmussen & C. Statist. Asmussen & V. Scand. Sc. Stochastic Models 8. 410433. 365372. 120. Hojgaard & M. Schmidt (1999) Tail approximations for nonstandard risk and queueing processes with subexponential tails. 419441.Y. Nielsen (1995) Ruin probabilities via local adjustment coefficients . Stoch. Probab. Asmussen & D. with applications to insurance risk Stoch. Management Science 45. Opns. Asmussen & T. Koole (1993). 115. Marked point processes as limits of Markovian arrival streams. Sigman (1996) Monotone stochastic recursions and their duals. Proc. Res. [45] S. . 31. [42] S. [50] S. O. [52] S. 103125. Appl.A. Dshalalow ed. Asmussen & C. Perry (1992) On cycle maxima. Asmussen & S. Prob 32 . CRC Press. 422447. Probab. [40] S. O'Cinneide (2000/01) On the tail of the waiting time in a Markovmodulated M/G/1 queue. Asmussen & H. Th. Insurance: Mathematics and Economics 10. [37] S. [41] S. Asmussen & K. Oper. Rubinstein (1999) Sensitivity analysis of insurance risk models. H. 20. Rubinstein (1995) Steadystate rare events simulation in queueing models and its complexity properties. Asmussen . [38] S. Eng. [51] S. Asmussen & M. Rolski (1991) Computational methods in risk theory: a matrixalgorithmic approach. 736755. [43] S. 30. (to appear). Asmussen & R. Florida. Statistica Neerlandica 47. Finance and Stochastics 4. Taksar (2000) Optimal risk control and dividend distribution policies . Example of excessoffloss reinsurance for an insurance corporation. first passage problems and extreme value theory for queues. [39] S. Schmidt (1995) Ladder height distributions with marks. Probab. Proc. 913916. Supplementary Volume (Kotz. [49] S. 64. 259274. Fitting phasetype distributions via the EM algorithm. J. 299324. O'Cinneide (2000/2001) Matrixexponential distributions [Distributions with a rational Laplace transform] Encyclopedia of Statistical Sciences. Appl. Asmussen. Adv. Wiley. [47] S.A. Nerman & M.Y. [44] S. 11251141. Read eds. Asmussen & V. Kliippelberg (1996) Large deviations results for subexponential tails. Appl. Methods 8 Problems (J. 10. Inf. Math. Probab. Advances in Queueing: Models. Asmussen & G. Asmussen & C. Olsson (1996). B.
[61] J. Wiley.L. 1995. Act. Clarendon Press. von Bahr (1974) Ruin probabilities expressed in terms of ladder height distributions. Beekman (1974) Two Stochastic Processes. [71] P. Teugels (1987) Regular Variation. 33. 1988 . [64] P. Springer . von Bahr (1975) Asymptotic ruin probabilities when exponential moments do not exist. Adv. Appl. Actuaries 21. Act. Bjork & J. Appl. [67] T. 1974. 4148. Teugels ( 1996) Convergence rates for M /G/1 queues and ruin problems with heavy tails . Wiley. 77111. Boogaert & V. SpringerVerlag. Bingham. Act.366 BIBLIOGRAPHY [53] S. [59] O. Cambridge University Press. BarndorffNielsen (1978) Information and Exponential Families in Statistical Theory. Appl. 610. [69] P. Berlin. [68] T. Ney ( 1972) Branching Processes . Berman & R. Probab. 705766.T.H. Boogaert & A.A. Schmidli (1995) Saddlepoint approximations for the probability of ruin in finite time. Bingham (1975) Fluctuation theory in continuous time. Soc. 1985 . Probab. J. Halsted Press. Insurance: Mathematics and Economics 9. [63] A. Crijns (1987) Upper bounds on ruin probabilities in case of negative loadings and positive interest rates. J.B .R.M. Bobbio & M. Bjork & J. Goldie & J. Act. Insurance: Mathematics and Economics 6. Scand. [65] N. Beekman (1969) A ruin function approximation. [70] A.J. 7. Oxford. 190204. New York. 832840. [57] C. [54] K.L. [73] A. Cambridge. Scand. C. 1975. Scand. Athreya & P. Borovkov (1976) Asymptotic Methods in Queueing Theory. 129134. Stochastic Models 10. J. de Waegenaere (1990) Simulation of ruin probabilities. [60] J.H. J. BarndorffNielsen & H. Insurance: Mathematics and Economics 4. [56] B. J. Plemmons (1994) Nonnegative Matrices in the Mathematical Sciences. Act. [58] O. 9599. Probab. New York. 661667. 169186. J. J. Grandell (1988) Exponential inequalities for ruin probabilities in the Cox case.H. Baker (1977) The Numerical Solution of Integral Equations. Trans.A. . Telek (1994) A bencmark for PH estimation algorithms: results for acyclic PH. New York. SIAM. 148156. [62] J. [55] B. 11811190. Scand. 9. Beekman (1985) A series for infinite time ruin probabilities. Grandell (1985) An insensitivity property of the ruin probability. Scand. Bloomfield & D. Asmussen & J. [66] N. Billingsley (1968) Convergence of Probability Measures. 275279. [72] P. 221232. Cox (1972) A low traffic approximation for queues.
J. Bucklew ( 1990) Large Deviation Techniques in Decision. Boxma & J. 749763. Stockholm.A. Aut. Reading. Boxma & J. Itasca. Illinois. Stockholm. Skandia Jubilee Volume. Springer. IEEE J. 93121. Wiley. Resnick & R. Chung (1974) A Course in Probability Theory (2nd ed.C. Cohen (1998) The M/G/1 queue with heavytailed service time distribution. Malgouyres (1983) Large deviations and rare events in the study of stochastic algorithms . Cottrell. 34. 127130. J.C. AddisonWesley.W. [88] J. NorthHolland. Reiser eds. [76] N. 392433. [91] H.L. [82] H. Computer Performance (K. 14.I. 105119. [93] K. 313319. J. Tech. Res. Area Commun.W. [79] P. New York. Proc. Gerber. Philos.L. Bratley. Nesbitt (1986) Actuarial Mathematics . The Society of Actuaries . D.J. 907920.M. 16. Cramer (1930) On the Mathematical Theory of Risk. 2338. Fort & G.Y.). Jr. [85] W. 3539. [87] E. Burman & D . [92] H. Broeck . S. [84] D.) NorthHolland. [80] F. Hickman. Insurance : Mathematics and Economics 9.R. Sel. Tweedie (1982) Storage processes with general release rule and additive inputs . Smith ( 1983) Asymptotic analysis of a queueing system with bursty traffic. The Jubilee volume of Forsakringsbolaget Skandia. Smith ( 1986) An asymptotic analysis of a queueing system with Markovmodulated arrivals . Cramer (1955) Collective risk theory. Breiman ( 1968) Probability. Simulation and Estimation. Cox (1955) Use of complex probabilities in the theory of stochastic processes. 177204. Bell. Bux & U. Amsterdam. Bowers . Cambr. Burman & D. 24. IEEE Trans. Insurance : Mathematics and Economics 5. [89] M. verve. Veraverbeke (1990).R. H. Geb.. Oper. II. Appl. [77] P. Control AC28. .L. Croux & N. Fox & L. Biihlmann ( 1970) Mathematical Methods in Risk Theory. Herzog (1977) The phase concept : approximations of measured data and performance analysis. Wahrscheinlichkeitsth. SpringerVerlag. [81] J. Cohen (1982) The Single Server Queue (2nd ed. [90] D. Probab. Queueing Systems 33. De Vylder (1986) Ordering of risks and ruin probabilities.W. Jones & C. Syst. Adv. Nonparametric estimators for the probability of ruin. [78] L. Cohen (1999) Heavytraffic analysis for the GI/G/1 queue with heavytailed service time distributions . Goovaerts & F. 14331453. J. Brockwell . [75] O.L.). Qinlar (1972) Markov additive processes.A.BIBLIOGRAPHY 367 [74] O. Academic Press.U. [86] K. Amsterdam. Z. 62.Y . Soc. New York San Francisco London. 51. [83] D.R. Schrage (1987) A Guide to Simulation. Berlin. B. New York. M. Chandy & M.
Insurance: Mathematics and Economics 7. J. Stockholm. Act. Martingales and insurance risk. Chapman & Hall. [95] M. Haezendonck (1987) Classical risk theory in an economic environment. 7083 (1969). Jones and Bartlett. Editions de 1'Universite de Bruxelles. [112] F. [108] F. English version published in Skand. Roy. 17. Oper. T. [106] A. Davidson (1946) On the ruin problem of collective risk theory under the assumption of a variable safety loading (in Swedish). J. [104] F. De Vylder (1977) A new proof of a known result in risk theory. Statist. De Vylder (1978) A practical solution to the problem of ultimate ruin probability. [99] R. Dempster.P. Actuarial J. Rubin (1977) Maximum likelihood from incomplete data via the EM algorithm. Zeitouni (1992) Large Deviations Techniques and Applications. 88101. Deheuvels & J. Insurance: Mathematics and Economics 10. Goovaerts (1984) Bounds for classical ruin probabilities. [105] A. 121131.D. Steinebach (1990) On some alternative estimators of the adjustment coefficient in risk theory. Delbaen & J. [103] F. Laird & D. 3. Csorgo & J. Embrechts (1989). Math. Forsiikringsmatematiska Studier Tillagnade Filip Lundberg. De Vylder & M. Springer. Csorgo & J. Soc. 201206. 135159. Daykin. van Dawen (1986) Ein einfacher Beweis fur Ohlins Lemma. Rolski (1991) Light traffic approximations in queues. Rolski (1984) A light traffic approximation for a single server queue. [102] P. [96] D. 5771.. [111] F. 624628. Stochastic Models 5. Appl. De Vylder (1996) Advanced Risk Theory. Scand. De Vylder & M. 1990 . Aktuar. 9 . Daley & T. Appl. Insurance: Mathematics and Economics 4. [97] D. Math. Goovaerts (1988) Recursive calculation of finitetime ruin probabilities. J.. Boston. [110] F. Teugels ( 1990) Empirical Laplace transform and approximation of compound distributions . Insurance: Mathematics and Economics 6. 16. 583602. Comp. Dembo & O. New York. Dassios & P. Steinebach ( 1991 ) On the estimation of the adjustment coefficient in risk theory via intermediate order statistics . Pesonen (1994) Practical Risk Theory for Actuaries.L. Res. . 27. [107] L. 181217. Devroye (1986) Nonuniform Random Variate Generation. 227229. 435436. [100] A. Pentikainen & E. Math.B. Res. 3750. Oper. J.M. 114119. N. Haezendonck (1985) Inversed martingales in risk theory. Blotter der deutschen Gesellschaft fur Versicherungsmathematik XVII. Probab. Scand. [109] F. 85116. [98] A. 22. [101] C. Daley & T. Delbaen & J. Suppl. Tidskr.368 BIBLIOGRAPHY [94] M. Insurance: Mathematics and Economics 3.
C. Dickson & C. 5159. [129] D . [125] F. Dynkin (1965) Markov Processes I. [117] D. [123] E. J. Ruin estimates for large claims .C. 1993 . 6180.C. Shiu (1991) Risk theory with the Gamma process. J. Proc. 118. Act. 177192. 229232. Springer . Scand. Hipp (1998) Ruin probabilities for Erlang (2) risk processes.C.S. and the amount of claim causing ruin. 2000 .M. Res.R. [127] F.C. [124] N. Scand. Dufresne. 251262. Insurance : Mathematics and Economics 7. [130] P. Insurance: Mathematics and Economics 25. Soc. Gerber & E.J.BIBLIOGRAPHY 369 [113] D. Dickson & C. J. Math. J. Hipp (1999) Ruin problems for phasetype(2) risk processes. with applications. Gray (1984) Approximations to the ruin probability in the presence of an upper absorbing barrier. van Doorn & G. 191207. Dickson & J. Scand. Act.M.M.R. British Actuarial J. 1993 . Scand.J. Dickson (1992) On the distribution of the surplus prior to ruin . [119] D.M. 193199. 1984 . Duffield & N. Grandell & H. Insurance: Mathematics and Economics 19.M. 1984 . Dickson & H.C. J. [114] D. Harrison & A. Philos.C. Gerber (1988) The surpluses immediately before and at ruin.B. Act. 1. 107124. [115] D. Taylor ( 1975) A diffusion approximation for the ruin probability with compounding assets. 4259. Insurance : Mathematics and Economics 10. Gerber (1991) Risk theory for the compound Poisson process that is perturbed by a diffusion. H.M. 131138.W.M. Gray (1984) Exact solutions for ruin probability in the presence of an upper absorbing barrier. [131] P.G.C. [128] E. Dickson & J. J.C. Act. 1741. [121] D.U. 1975 . [118] D. Regterschot (1988) Conditional PASTA. Act. Astin Bulletin 21. 147167.C. Letters 7. Camb. Insurance : Mathematics and Economics 11.M. Dufresne & H. Act. Scand. Oper.R. [122] B. Dickson (1995) A review of Panjer's recursion formula and its applications. Emanuel . Scand. Dickson & H.R. J. 1994 .U. J.M. O'Connell (1995) Large deviations and overflow probabilities for the general singleserver queue. 3745. [120] D. 269274. 174186. 4962. [126] F. Insurance: Mathematics and Economics 7. Act. Dufresne & H. 105115. Berlin Gottingen Heidelberg. Djehiche (1993) A large deviation estimate for ruin probabilities.U. Waters (1999) Ruin probabilities wih compounding. 363374. Waters (1996 ) Reinsurance and ruin . J. Schmidli (1993) Finitetime Lundberg inequalities in the Cox case .K. Insurance : Mathematics and Economics 22. Embrechts ( 1988 ). Dickson (1994) An upper bound for the probability of ultimate ruin. Embrechts . [116] D. Scand. .M.
Proschan & D. Villasenor (1988) Ruin estimates for large claims. 181190. New York. Springer. Jensen. Nyt Tidsskrift for Matematik B20. Embrechts.L. Fuh & T. C.M.L. Erlang 2. Statistica Neerlandica 47.D. R. Astin Bulletin 16. 1998 . 695712.J.D. 38. J. Embrechts. Th. 29. Appl. [139] A. 14661474.A. Wiley. Probab. Embrechts & N. 5572. Maejima & J. Statist. Mikosch (1991). Schmidt (1982) Queues and Point Processes. Insurance: Mathematics and Economics 7.D. Reprinted 1948 as 'The theory of probabilities and telephone conversations' in The Life and work of A. Teugels (1985). Insurance: Mathematics and Economics 10.W. 404422. U. [146] E. Appl. 269274. Adv. [144] P. Fuh (1997) Corrected ruin probabilities for ruin probabilities in Markov random walks. [145] P. New York. Fitzsimmons (1987). Probab. J. A bootstrap procedure for estimating the adjustment coefficient. Danish Academy Tech. [133] P. Approximations for compound Poisson and Polya processes. J. . Esscher (1932) On the probability function in the collective theory of risk. Sci. Math. Frees (1986) Nonparametric estimation of the probability of ruin. Rel. Scand. 5975. Wiley.K. [143] W. F. 616624. Appl. 29. Schmidli (1994) Ruin estimation for a general insurance risk model. [147] M. Embrechts & H. 17. 3339. 623637. 159178. [150] H. Probab. 566580. Probab. 175195. Veraverbeke (1982) Estimates for the probability of ruin with special emphasis on the possibility of large claims . Tidsskr. Trans. Skand. Griibel & S. Probab. 26.370 BIBLIOGRAPHY [132] P.K. 5974. [149] C. Arndt & V. Embrechts. [142] W. Erlang (1909) Sandsynlighedsregning og telefonsamtaler . Lai (1998) Wald's equations. [140] J.). [134] P. [137] P. Heidelberg. Frenz & V.W. Insurance : Mathematics and Economics 1. Fields 75. [135] P. M. 131137. Kliippelberg & T. Appl. Act. 8190. Esary. Probab. 35. Embrechts & T. [138] P. [148] C. [141] F. Appl. On the excursions of Markov processes in classical duality. Feller (1971) An Introduction to Probability Theory and its Applications II (2nd ed. Akt. Walkup (1967) Association of random variables. Mikosch (1997) Modelling Extremal Events for Finance and Insurance. Konig. D. Adv. first passage times and moments of ladder variables in Markov random walks. Ann. with applications. Schmidt (1992) An insensitivity property of ladder height distributions. J. Franken. Furrer (1998) Risk processes perturbed by astable Levy motion. Adv. Pitts (1993) Some applications of the fast Fourier transform in insurance mathematics .). Feller (1966) An Introduction to Probability Theory and its Applications I (3nd ed. [136] P. Embrechts & J. Wiley.
Basel. Glynn & W.E. 205210. Schweiz. 1523. Act. Glasserman & S.BIBLIOGRAPHY 371 [151] H. J. Goldie & R. 7778.U. [166] M. 97114. Kou (1995) Limits of first passage times to rare sets in regenerative processes. Math. 28. Vers. Dordrecht. Gani.V. S. Z. Galambos & J. de Vylder & J. Kovalenko (1989) Introduction to Queueing Theory (2nd ed. London. Probab. Gerber (1973) Martingales in risk theory. A. 73. Insurance: Mathematics and Economics 20.M. In Studies in Applied Probability (J. [157] H. 205216. Griibel (1996) Perpetuities with thin tails. Unpublished.J. 1981 . Act. 6370.). Schmidli (1994) Exponential inequalities for ruin probabilities of risk processes perturbed by diffusion. Appl. Aktuarietidskrift 1970 .N. Weron (1997) Stable Levy motion approximation in collective risk theory.U. Boston. [169] J. F. Gnedenko & I. [162] P. [155] H.U. 105115. Mitt. Kluwer.. Goovaerts. Furrer (1996) A note on the convergence of the infinitetime ruin probabilities when the weak approximation is astable Levy motion.. Kaas. Appl. Grandell (1977) A class of approximations of ruin probabilities. Bauwelinckx (1990) Effective Actuarial Methods. 1977. [154] H. [168] J. [152] H. 424445. .). Ann. 5. Amsterdam. 131156.U. J. Mitt. contained in the authors PhD thesis. Springer. J. Adv.U. Ver. [156] H. Gerber (1970) An extension of the renewal equation and its application in the collective theory of risk. Furrer & H. R.W. 463480. Whitt (1994) Logarithmic asymptotics for steadystate tail probabilities in a singleserver queue. 3752. Goovaerts. [165] M. University of Pennsylvania. Life Insurance Mathematics.J. Gerber (1988) Mathematical fun with ruin theory. NorthHolland. Suppl. Appl. 1978 . Skand. [153] H.71.U. Haezendonck (1984) Insurance Premiums. Insurance : Mathematics and Economics 7. Probab. J. Furrer. 2336.G. [159] H. NorthHolland. ETH Zurich. [158] H. Birkhauser . Gerber (1979) An Introduction to Mathematical Risk Theory. Scand. [161] P. Math. eds. Huebner Foundation Monographs. [163] P.U. Gerber (1981) On the probability of ruin in the presence of a linear dividend barrier. 31A. van Heerwarden & T. Michna & A. [164] B. Ver. Scand. Insurance: Mathematics and Economics 15.S. Vers. Amsterdam. Gerber (1971) Der Einfluss von Zins auf die Ruinwahrscheinlichkeit. [167] C. [160] H. Scand. Probab. Schweiz. Glasserman (1991) Gradient Estimation via Perturbation Analysis. Grandell (1978) A remark on 'A class of approximations of ruin probabilities'. Gerber (1986). Act.
Grigelionis (1996) Lundbergtype stochastic processes. Insurance: Mathematics and Economics 26. Skand. Preprint. [180] R. 14. Korolyuk (1969) On the joint distribution of a process with stationary increments and its maximum . Informatica 2. Grigelionis ( 1993) Twosided Lundberg inequalities in a Markovian environment.372 BIBLIOGRAPHY [170] J. Appl.). 8. Silvestrov (2000) Cram€rLundberg approximations for nonlinearly perturbed risk processes . van Heerwarden (1991) Ordering of Risks: Theory and Actuarial Applications. [175] J. [171] J. Appl. Segerdahl (1971) A comparison of some approximations of ruin probabilities. Harrison & S. Harrison (1977) Ruin problems with compounding assets . Stoch. Gut (1988 ) Stopped Random Walks . Math. 332. Ibragimov. Theory and Applications. Grnbel (1991) G/G/1 via FFT. [181] D. Harrison & S. Archiv v. Resnick (1976 ) The stationary distribution and first exit probabilities of a storage process with general release rule . . Grandell (1997) Mixed Poisson Processes. [178] B. 355365. [176] B. Resnick (1977) The recurrence classification of risk and storage processes . Gusak & V. Winter School on Stochastic Analysis and Appl . Appl. 14. 1. [188] J. [182] A. 54. Aktuar Tidskr.I. Lemoine (1977) Limit theorems for periodic queues. [177] B. Astin Bulletin 29. Griibel & R. M. 3041. Grandell ( 1992) Finite time ruin probabilities and martingales . Statistical Algorithm 265. Probability Theory and Mathematical Statistics (I. [184] H . SpringerVerlag. Math. 5 . Res. J. Hermesmeier (1999) Computation of compound distributions I: aliasing errors and exponential tilting.131135. 6779. Matem. 197214. M. Grandell & C. 400409. [173] J. Grandell (1990) Aspects of Risk Theory. Appl. [174] J. 7590. Opns . Stoch. [187] J. Chapman & Hall. E. [189] A. Gyllenberg & D. Probab. Proc. SpringerVerlag. Berlin.A. New York. 143158. Applied Statistics 40. Grigelionis ( 1992) On Lundberg inequalities in a Markovian environment. M. ed. M. Proc .O. Grandell (1979) Empirical bounds for ruin probabilities. mathematische Wirtschaft . 3 .. AkademieVerlag. Harrison & A.V. [172] J. Rink. Proc. 167176. 33.J. 347358. Res. Tinbergen Institute Research Series 20. [183] M. 5766.and Sozialforschung 6. Th. Hadwiger (1940) Uber die Wahrscheinlichkeit des Ruins bei einer grossen Zahl von Geschiiften . Grandell (1999) Simple approximations of ruin functions. [186] J. Opns . Probab. [179] R.S. Amsterdam. 566576. [185] J.I.243255. KTH. Liet.
Ann. Iglehart (1969) Diffusion approximations in collective risk theory.. [191] W. Proc. [197] C. New York. 8095. [198] C. Hesselager (1990) Some results on optimal reinsurance in terms of the adjustment coefficient. Probab.a program for fitting phasetype distributions. 1998 . [202] 0. Ann. [193] U. Hermann (1965) Bin Approximationssatz fiir Verteilungen stationarer zufalliger Punktfolgen. Insurance: Mathematics and Economics 5.L. 8996. J. [192] U. Astin Bulletin 19.L. 4385. [200] M. Chalmers University of Technology and the University of Goteborg. Appl. Taksar (2000) Optimal proportional reinsurance policies for diffusion models. [208] V. Act. Hipp (1989b) Estimators and bootstrap confidence intervals for ruin probabilities.A. Haggstrom. [206] B. Insurance: Mathematics and Economics 5. No. 377381. 123151. 285292. [201] L. 259268. 12981310. Scand. 166180. Probab. Hoglund (1990) An asymptotic expression for the probability of ruin within finite time. Asmussen & 0. 305313. 1990 . J.B. J. Wiley. Statist. Heilmann (1987) Grundbegriffe der Risikotheorie. Nachrichten 30. Karlsruhe. Act. 378389. Hill (1975) A simple general approach to inference about the tail of a distribution. Klugman (1984) Loss Distributions.M. Iversen & L. [199] R. [194] 0. [203] T. Ann. Hoglund (1974) Central limit theorems and statistical inference for Markov chains. Karlsruhe. Foruth Prague Symp. Math. Huskova. . [204] T. 285293. on Asymptotic Statistics (P. Appl. Probab. [207] D. 25. Hogan (1986) Comment on corrected diffusion approximations in certain random walk problems. Hipp (1989a) Efficient estimators for ruin probabilities. Staalhagen (1999) Waiting time distributions in M/D/1 queueing systems. Nerman (1992) EMPHT . 19. eds. Electronic Letters 35. ACM TOMACS 6. Herkenrath (1986) On the estimation of the adjustment coefficient in risk theory by means of stochastic approximation procedures. Z. S. Michel (1990) Risikotheorie: Stochastische Modelle and Statistische Methoden. Wahrscheinlichkeitsth. Geb. Horvath & E. Willekens (1986) Estimates for the probability of ruin starting with a large initial reserve.). 11631174.V. 29. 5770. verve. Hojgaard & M. 6. [195] B. J. Verlag Versicherungswirtschaft e. The ruin problem for finite Markov chains. Mathematical Statistics.BIBLIOGRAPHY 373 [190] P.R. Versicherungswirtschaft. 3. Prob. [196] C. Studies in Statistical Quality Control and Reliability 1992 :4. Mandl & M. 23. Hoglund (1991). 18. Heidelberger (1995) Fast simulation of rare events in queueing and reliability models. [205] T. Hipp & R.V. Scand. Hogg & S.
[211] J. Taylor (1981) A Second Course in Stochastic Processes. Chicago. Act.L. [227] J. Kennedy (1994) Understanding the WienerHopf factorization for the simple random walk. [216] M. ibid. 283306. 31.M. Jagerman (1991) Analytical and numerical solution of Volterra integral equations with applications to queues.M. Soc. Taaffe (1989/90) Matching moments to phase distributions. 239256. 187193. 6. 35. J. J.P. Technometrics 30. Karlin & H. Munksgaard. 61. [220] V.G. [222] S. 711743. Goovaerts (1986) General bound on ruin probabilities. [228] J.R. ibid.C. 1996. Appl.J. Addenda to for processes defined on a finite Markov chain. Wishart (1964) Boundary problems for additive processes defined on a finite Markov chain. Proc.M. Jensen (1995) Saddle Point Approximations. [219] V. 3. [221] E. Soc. New York. 37. Cambridge Philos. Kemperman (1961) The Passage Problem for a Markov Chain.H. 15. University of Chicago Press. Jensen (1953) A Distribution Model Applicable to Economics. Kao (1988) Computing the phasetype renewal and related functions. Princeton. Statist. 118. Appl. N. Academic Press. 866870.G. Kaas & M. 11. J. Kalashnikov (1996) Twosided bounds of ruin probabilities. Probab. Keilson (1966) A limit theorem for passage times in ergodic regenerative processes.J. [210] D. Proc. Kluwer. 8793. Clarendon Press. 561563. Keilson & D. Amer. Ann. 60. Lazar (1998) Subexponential asymptotics of a Markovmodulated random walk with a queueing application.A. Reinhard (1985) Probabilities de ruine pour une classe de modeles de risque semiMarkoviens. 165167. Cambridge Philos. Insurance: Mathematics and Economics 5. Stochastic Models 1. Copenhagen. Wishart (1964) A central limit theorem for processes defined on a finite Markov chain. Oxford. Astin Bull. Scand. 325247. 173190. 63. . 116129. Stochastic Models 5.374 BIBLIOGRAPHY [209] D. [224] J. Cambridge Philos. [217] R. Jelenkovic & A.G. [213] P. 547567. Soc. [214] A. J. Johnson & M.M. [212] J. N. [223] J. NEC Research Institute. Kalashnikov (1997) Geometric Sums: Bounds for Rare Event with Applications. Jagerman (1985) Certain Volterra integral equations arising in queueing. Keilson & D. [225] J. Janssen (1980) Some transient results on the M/SM/1 special semiMarkov model in risk and queueing theories.B. Math. Kalashnikov (1999) Bounds for ruin probabilities in the presence of large claims and their comparison. Keilson & D. 259281. Wishart (1964). Proc. 123133. [215] J. Astin Bull. Act. 4151. Janssen & J. [218] V. 6.M. Probab. [226] J. Manuscript.
Proc. Kliippelberg (1988) Subexponential distributions and integrated tails. Scand. Ramaswami (1999) Introduction to MatrixAnalytic Methods in Stochastic Modelling. Macmillan.S. Kingman (1962) On queues in heavy traffic. 359361. Insurance: Mathematics and Economics 8. Appl. 277289. [231] J. [230] J.. Lemoine (1981) On queues with periodic Poisson input. Mikosch (1995) Modelling delay in claim settlement. [244] T. 390397. Scand. 1995 . [241] G. J. 4477. 858874.J. [246] D. J. Proc. 383392. I. Act.F. Mikosch (1995) Explosive Poisson shot noise with applications to risk retention. Act. [236] C. 132141. Lucantoni (1991) New results on the single server queue with a batch Markovian arrival process. [234] C. Sorensen (1997) Exponential Families of Stochastic Processes. [233] C. 4958.C. Lindley (1952) The theory of queues with a single server. Klnppelberg (1993) Asymptotic ordering of risks and ruin probabilities. [240] U. [237] C. 279285. 133135 (in Russian). 146. Philos. Penev & A. Roy. Klnppelberg & T. Adv. 24. Soc. J.BIBLIOGRAPHY 375 [229] J. Kliippelberg & U. Kingman (1961) A convexity property of positive matrices . Stadtmiiller (1998) Ruin probabilities in the presence of heavytails and interest rates . [248] D. Kuchler & M. [232] C. Philos. J. Oxford 12. J. Soc.a Linear Algebraic Approach. [238] V. Camb. 1998 . Soc. Appl. 18. 60 . Statist. Appl. [239] H. Probab. J. Lehtonen & H. 6075. 25. SIAM. Klnppelberg & T. SpringerVerlag. Actuarial J. J. SpringerVerlag.C. (235] C. Nyrhinen (1992a) Simulating levelcrossing probabilities by importance sampling. 48. Lipsky (1992) Queueing Theory . Kunita (1976) Absolute continuity of Markov processes. Turbin (1973) An asymptotic expansion for the absorption time of a Markov chain distribution. Probab.P.C. Quart. Cambr. [247] L. Klnppelberg (1989) Estimation of ruin probabilities by means of hazard rates. Quart.J. Appl. 125147. Nyrhinen (1992b) On asymptotically efficient simulation of ruin probabilities in a Markovian environment. Kingman (1964) A martingale inequality in the theory of queues. Cybernetica 4.F. 889900. Probab.F. [242] A. Latouche & V. Prob.F. . 283284. Stochastic Models 7. [243] A. Lemoine (1989) Waiting time and workload in queues with periodic Poisson input. 259264. 154168. New York. Math. [245] T. Lecture Notes in Mathematics 511. Lehtonen & H. Korolyuk. Bernoulli 1 . Scand. B24. 26. Seminaire de Probabilties X . Insurance : Mathematics and Economics 12.
Eigenvalue properties and limit theorems . 36. [261] H. Probab. [253] V. 32. [266] P. [263] M . a useful concept in risk theory. [252] A. 124147. 834841. Malinovskii ( 1994) Corrected normal approximation for the probability of ruin within finite time.S. Probab.K. Scand. Schmidt ( 1993) On ladder height distributions of general risk processes . Moustakides ( 1999) Extension of Wald 's first lemma to Markov processes. J. Meier ( 1984). Michna ( 1998) Selfsimilar processes in collective risk theory. Lundberg (1926) Forsdkringsteknisk Riskutjdmning. [250] F. Scand. Ney & E. 4859.F. Statist. 15.M. . Ann. Insurance : Mathematics and Economics 5. MeierHellstern & M . Probab. 561592. K. J.D. Hoist eds.V. 22. Makowski ( 1994) On an elementary characterization of the increasing convex order . Probab. 2939. Math. Lundberg (1903) I Approximerad Framstdllning av Sannolikhetsfunktionen. [265] S. Miller (1962 ) A matrix factorization problem in the theory of random variables defined on a finite Markov chain . Cambridge Philos. II Aterforsdkring av Kollektivrisker . Ann. Gut & J. 1994. [258] K . Miyazawa & V. 223235. [254] V . 268285. 1986 . Miller ( 1962 ) Absorption probabilities for sums of random variables defined on a finite Markov chain . Miller ( 1961 ) A convexity property in the theory of random variables defined on a finite Markov chain. Europ. Act. 1261 (0?)1270. Stoch. Proc.V. 378406.376 BIBLIOGRAPHY [249] D. Almqvist & Wiksell. Probability and Mathematical Statistics (A. J.D. Appl.K.). Mammitzsch ( 1986 ) A note on the adjustment coefficient in ruin theory. J. 1996 . Math. Probab . [257] A. Scand. Appl. MartinL3f (1986) Entropy. [260] H. Adv. MartinL6f (1983) Entropy estimates for ruin probabilities . Stockholm. with an application . [251] F. [264] G. Th. Soc. Opns.S. 31. Res. Act. 161174. Ann. [259] Z . Soc. 58 . [262] H . 11 . Act. J. Lucantoni . F. J. A fitting algorithm for Markovmodulated Poisson processes having two arrival rates . Malinovskii ( 1996) Approximation and upper bounds on probabilities of large deviations of ruin within finite time. 3.D. 763776. 2. 58. Appl. Uppsala. 286298. Cambridge Philos. Proc. 429448. 29 . Anal. [255] V. Probab. Appl. [256] A. Appl. Nagaev (1957) Some limit theorems for stationary Markov chains. J. 676705. Nummelin (1987) Markov additive processes I. 370377. . Englunds Boktryckeri AB. Appl. 147149. Neuts ( 1990) A single server queue with server vacations and a class of nonrenewal arrival processes .
[282] J .a survey. Proc. Norberg ( 1990) Risk theory and its statistics environment . IEICE Trans.F.F. 36. Appl. London. [274] H . Appl.K. Proc. Neuts (1977) A versatile Markovian point process. Marcel Dekker. Paulsen ( 1998) Sharp conditions for certain ruin in a risk process with stochastic return on investments. 316. [281] J. Stochastic Models 3 . 555564. Statistics 21. Omey & E. 71. Ohlin (1969) On a class of measures for dispersion with application to optimal insurance.F. [284] J. Stoch. Willekens (1987) Second order behaviour of distributions subordinate to a distribution with finite mean . Proc. Probab. Neuts ( 1978) Renewal processes of phase type. 135148. Astin Bulletin 5 . 215223. 273299. Neveu ( 1961 ) Une generalisation des processus a accroisances independantes. Paulsen (1993) Risk theory in a stochastic economic environment. [271] M. Probab. [276] C. Insurance: Mathematics and Economics 22. Stoch. [283] J. Stoch. Neuts ( 1981 ) MatrixGeometric Solutions in Stochastic Models. [270] M . Johns Hopkins University Press. E75B . Hamburg. 311342. Models 6. [273] R. 339353. [269] M .BIBLIOGRAPHY 377 [267] M . Gjessing (1997b) Present value distributions with applications to ruin theory and stochastic equations. Appl. Willekens (1986) Second order behaviour of the tail of a subordinated probability distributions. Probab. 10081026. Probab.F. Stoch. Baltimore . Nyrhinen (1999) Large deviations for the time of ruin. 733746. 249266. Adv. [280] A. [268] M . Appl. Omey & E. Stoch. Pakes (1975) On the tail of waiting time distributions. Insurance: Mathematics and Economics 20. . 327361. 46. 30. Paulsen & H. New York. J. 21. [278] E. 764779. Appl. O'Cinneide (1990) Characterization of phasetype distributions. 123144. Sem. Appl. 12. 12551265. Paulsen & H. Appl. Neuts (1992) Models based on the Markovian arrival process. 16. [279] E. Gjessing (1997a) Optimal choice of dividend barriers for a risk process with stochastic return on investments. Abh. Paulsen (1998) Ruin theory with compounding assets . Proc. Commun. 75.A. Appl.F. [275] H.G. 445454. Neuts ( 1989) Structured Stochastic Matrices of the M/G/1 Type and their Applications. J. [272] J . J. Naval Research Logistics Quarterly 25. [277] J.K. 157. Nyrhinen ( 1998 ) Rough descriptions of ruin for a general class of surplus processes. [285] J. Math.
Prabhu (1980) Stochastic Storage Processes. Math.U. Opns. [290] E. 390413. [305] C. 46. Ann. [288] S. Adv. New York. Ramaswami (1980) The N/G/1 queue and its detailed analysis. [287] F. 2343. 337347. 147159. Soc. [292] S. Oper. Math. [294] N.A. 61. [302] J. Act. Astin Bulletin XIV. Appl. 2330. Embrechts (1996) Confidence bounds for the adjustment coefficient. Gjessing (1997c) Ruin theory with stochastic return on investments. Austr. Rolski (1987) Approximation of periodic queues. Pitman (1980) Subexponential distribution functions. Probab. Queues. Wiley. [298] V. Math. [301] G. Probab.. 215246.K. On the ruin problem of collective risk theory. 29A. 117133. 4568. 28. Ramsay (1984) The asymptotic ruin problem when the healthy and sick periods form an alternating renewal process. Probab. Appl. Insurance Risk. de Smit (1986) The queue M/G/1 with Markovmodulated arrivals and services.378 BIBLIOGRAPHY [286] J.L.U. 32. [300] C. 11. and Dams. [291] S.G. Appl. Res. . 139143. Reinhard (1984) On a class of semiMarkov risk models obtained as classical risk models in a markovian environment. Inst. Aktuar. Insurance: Mathematics and Economics 16. Prabhu & Zhu (1989) Markovmodulated queueing systems. 691707. Schock Petersen (1989) Calculation of ruin probabilities when the premium depends on the current reserve. Berlin. Statist. Probab . Pellerey (1995) On the preservation of some orderings of risks under convolution. Adv. Prabhu (1961). J. 965985. Skand. Prabhu (1965) Queues and Inventories. 4. R. 757764. [303] S.M. Philipson (1968) A review of the collective theory of risk. Pyke (1959) The supremum and infimum of the Poisson process. 235243. Math.H. [306] T. 29.U.J. Tidskr. 19. Insurance: Mathematics and Economics 3. Rogers (1994) Fluid models in queueing theory and WienerHopf factorisation of Markov chains. 568576. Ann. Wiley. Appl. Pitts. 465483. 537555. Queueing Systems 5. Res. Paulsen & H. [289] C. Springer. Stat. 820827.K.G.J. Resnick & G. Griibel & P. Heidelberg. [299] C. J. [293] N. Pitts (1994) Nonparametric estimation of compound distributions with applications in insurance. 222261.M. [304] B. 1989 .U. Probab. Math. J. [297] R.M. [296] N. Ann. Statist. Adv. New York. Appl. Ripley (1987) Stochastic Simulation .R. 12. Scand. Ann. Regterschot & J. 30. Wiley. Samorodnitsky (1997) Performance degradation in a single server exponential queueing model with long range dependence. [295] N. Rao (1965 ) Linear Statistical Inference and Its Applications.
135149. 145165. Math. V. Appi. [313] M. [308] S. Ross (1974) Bounds on the delay distribution in GI/G/1 queues. Probab. [310] R. Wiley. Rolski. Operationsforsch. 4068. 795829. 7. Ann. Appl. Appl. Insurance: Mathematics and Economics 22. 5. 1997. Wiley. Stochastic Models 6. 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. Comp. [309] H. 417421. [322] H. Melamed (1998) Modern Simulation and Modelling. 93104. Schmidli. [312] R. Schlegel ( 1998) Ruin probabilities in perturbed risk models. 262286. Scand. Science Culture Technology Publishing . Wiley. Astin Bulletin 29. Adv. Wiley. J. Singapore.J. Rubinstein & A. Stochastic Models 10 .Y. Data Anal. 121133. [324] H . . [315] T. Schmidli ( 1997b) An extension to the renewal theorem and an application in risk theory. 5. Proc. 21. Probab. 2001 . 431447 [316] S. Rubinstein (1981) Simulation and the Monte Carlo Method. [323] H. Rudemo (1973) Point processes generated by transitions of a Markov chain. 365388. Schmidt & J. J.L. Insurance: Mathematics and Economics 16. Act. Shapiro (1993) Discrete Event Systems: Sensitivity Analysis and Stochastic Optimization via the Score Function Method. [317] L. Schmidli ( 1995 ) CramerLundberg approximations for ruin probabilities of risk processes perturbed by a diffusion . 227244. Statist. [321] H.Y. 687699. J. Statist. Seal ( 1969) The Stochastic Theory of a Risk Business. Schmidli ( 1996) Martingales and insurance risk. 11. Ryden (1994) Parameter estimation for Markov modulated Poisson processes.M. [314] T. H.Y. Teugels (1999) Stochastic Processes for Insurance and Finance. [318] H. Wiley. 131156. [325] H . Schmidli ( 1997a) Estimation of the Lundberg coefficient for a Markov modulated risk model .BIBLIOGRAPHY 379 [307] T. Schmidli ( 2001 ) Optimal proportional reinsurance policies in a dynamic setting . Scand. 4857. Th. [311] R. 155188. Schmidli ( 1999b) Perturbed risk processes: a review . Stochastic Models 10. Schmidli ( 1999a) On the distribution of the surplus prior and at ruin. Rubinstein & B. Lecture Notes of the 8th Summer School on Probability and Mathematical Statistics ( Varna). [320] H. Probab . [319] H. Stoch. Schmidli ( 1994) Diffusion approximations for a risk process with the possibility of borrowing and interest. Act. [326] H. 5. Schmickler ( 1992 ) MEDA : mixed Erlang distributions as phasetype representations of empirical distribution functions . Ryden (1996) An EM algorithm for estimation in Markovmodulated Poisson processes.
Versich. 243249. Weiss (1995) Large Deviations for Performance Analysis. Siegmund (1975) The time until ruin in collective risk theory. [339] A. 701719.380 BIBLIOGRAPHY [327] H. Grenander). Adv. Act. Appl. Shaked & J. Appl. [346) D . Sengupta ( 1989) Markov processes whose steadystate distribution is matrixgeometric with an application to the GI/PH/1 queue. Segerdahl (1955) When does ruin occur in the collective theory of risk? Skand. . Scand. Stockholm. Mitteil. Math. [337] M. Shtatland ( 1966) On the distribution of the maximum of a process with independent increments.F. 4. Probab .W. Adv. 11. Ann. New York. 191197. Versich. Queueing Systems 11. Shantikumar (1993) Stochastic Orders and Their Applications. [331] G. [330] H. Risk teory and the single server queue . Appl. Sengupta (1990) The semiMarkov queue: theory and applications. 673684. 121139. Seal (1974) The numerical calculation of U(w. 483487. [328] H. In Probability and statistics . Mitt. Seal (1972) Numerical calculcation of the probability of ruin in the Poisson/Exponential case. Seal ( 1972). J. 1987. . Siegmund (1976) The equivalence of absorbing and reflecting barrier problems for stochastically monotone Markov processes . t). Math. 21.W. 75. 157166. 914924. 72. Sigman (1992) Light traffic for workload in queues. Seber (1984) Multivariate Observations. Act. Segerdahl (1959) A survey of results in the collective theory of risk. Siegmund ( 1985) Sequential Analysis . [341] E. Chapman and Hall. 171178. [340] E. Mitt. Shin (1989) Ruin probability by operational calculus. 1955. [329] H. Math. Sigman ( 1994 ) Stationary Marked Point Processes: An Intuitive Approach. Chapman & Hall. Aktuar Tidsskr.L. 77100. 429442. [333] C: O. [334] C: O. [343] D. 383413. Shwartz & A. pp. Verein Schweiz. Aktuar Tidsskr.A. [335] B . Probab. Statist. Shin (1987) Convolution of uniform distributions and ruin probability. Seal (1978) Survival Probabilities. Insurance: Mathematics and Economics 8. Probab. [344] D. 4.S.L. Siegmund (1976) Importance sampling in the Monte Carlo study of sequential tests. Wiley. [345] D. Siegmund (1979) Corrected diffusion approximations in certain random walk problems. 61. 2236. Th. t). Verein Schweiz. SpringerVerlag. 279299. Academic Press. [338] E. Stochastic Models 6. J. 4383. Probab. Almqvist & Wiksell. L.S. [336] B. Verein Schweiz Versich. [332] C : O. 159180. the probability of nonruin in an interval (0.the Harald Cramer volume (ed. [347] K. Skand. Wiley. 1974.S. [348] K . Segerdahl ( 1942) Uber einige Risikotheoretische Fagestellungen .G. Ann. Scand. 72. 11. [342] D.L.
[358] R. 118.C. Thorin (1977). Thorin (1986) Ruin probabilities when the claim amounts are gamma distributed.. 1976. New York. 1978. J. Probab. Taylor (1978) Representation and explicit calculation of finitetime ruin probabilities.).J. Act. . Astin Bulletin 24. 149162. [355] B . Insurance: Mathematics and Economics 19. John Wiley & Sons. 163175. Thorin (1982) Probabilities of ruin. Scand. Actuarial J. 21. [352] D. Slud & C. Straub (1988) NonLife Insurance Mathematics. Smith ( 1953) Distribution of queueing times. 1982 . Insurance: Mathematics and Economics 16. [364] G. J. 8199. [365] J. 8594. Taylor (1979) Probability of ruin with variable premium rate.BIBLIOGRAPHY 381 [349] E. Act. Act.L. [357] B.). [367] O. Unpublished manuscript. Ruin probabilities prepared for numerical calculations. Stanford & K. Daley ed. [359] L. Sundt & J. J.J.L. [351] D . NorthHolland. 65102. [368] O. Appl. Scand. 235254. Astin Bulletin 12.A. J. 725741. Teugels ( 1995 ) Ruin estimates under interest force .C. Cambridge Philos. Stoyan (1983) Comparison Methods for Queues and Other Stochastic Models (D. [350] W. Adv. New York. Jewell ( 1981 ) Further results on recursive evaluation of compound distributions . 1980 . Taylor (1986) Claims reserving in NonLife Insurance. 197208. Taylor (1976 ) Use of differential and integral inequalities to bound ruin and queueing probabilities . 2739. Act. Verlag Versicherungswirtschaft e. Scand.L. [360] G. [362] G. Suri ( 1989) Perturbation analysis : the state of the art and research issues explained via the GI/G/1 queue. 114137.C. Taylor (1979) Probability of ruin under inflationary conditions or under experience ratings . [366] O.C. Act. Sundt ( 1993) An Introduction to NonLife Insurance Mathematics (3rd ed. Hoesman ( 1989) Moderate. [356] B . J. Scand.L. 65102. 722. Astin Bulletin 16 . Proceeedings of the IEEE 77. Takhcs (1967) Combinatorial Methods in the Theory of Stochastic Processes. 1974. [363] G. [354] B.C. Karlsruhe. Springer. Soc 49. Stroinski ( 1994) Recursive method for computing finitetime ruin probabilities for phasedistributed claim sizes.S. Insurance: Mathematics and Economics 1.V.and largedeviations probabilities in actuarial risk theory. Sundt & W. [353] E. Wiley. Sundt & J. [369] O. Scand. 5776. Teugels (1997) The adjustment coefficient in ruin estimates under interest force . [361] G. Proc. Teugels (1982) Estimation of ruin probabilities . Scand.. Thorin (1974) On the asymptotic behaviour of the ruin probability when the epochs of claims form a renewal process. 449461.
Wolff (1990) Stochastic Modeling and the Theory of Queues. Van Wouve. Unpublished Ph. Tacklind (1942) Sur le risque dans les jeux inequitables. Probab. Tidskr. [371] 0. [376] A. [372] H. Wilson (1995) Selfsimilarity in highspeed traffic: analysis and modeling of ethernet traffic measurements. 31.E. M. [381] W.R. Wikstad (1977) Calculation of ruin probabilities when the claim distribution is lognormal. Dordrecht Boston Lancaster. Haezendonck eds. Sherman & D. Astin Bulletin VII. Wald (1947) Sequential Analysis. J. Astin Bulletin IX. Willinger. Act. 231246. 142. Veraverbeke (1993) Asymptotic estimates for the probability of ruin in a Poisson model with diffusion. Goovaerts (1983) The influence of reinsurance limits on infinite time ruin probabilities. Res. 4963. M. Thorisson (2000) Coupling. 1983 . Wikstad (1977) Numerical evaluation of ruin probabilities for a finite period. Waters (1983) Probability of ruin for a risk process with claims cost inflation. Aktuar. Scand. [384] R. Appl. De Vylder & M. PrenticeHall. 148164. Insurance: Mathematics and Economics 15. [382] G. VazquezAbad (1999) RPA pathwise derivatives of ruin probabilities. D. 743756.). Thorin & N. 6785. R. [377] V. J. F. Submitted. Whitt (1989) An interpolation approximation for the mean workload in a GI/G/1 queue. 5762. [375] N. 37. 1942.W. [383] G. Wiley. Taqqu. Willmot (1994) Refinements and distributional generalizations of Lundberg's inequality. De Vylder. In: Premium Calculation in Insurance (F. SpringerVerlag. [379] M. Wallace (1969) The solution of quasi birth and death processes arising from multiple access computer systems. Oper. Goovaerts. [373] S. Skand. Thorin & N. Statistical Science 10. Stationarity and Regeneration. [374] F. . Reidel. [380] W. J. thesis.E. Insurance : Mathematics and Economics 13. 936952.382 BIBLIOGRAPHY [370] 0. 137153. Willmot & X. Lin (1994) Lundberg bounds on the tails of compound distributions. [378] H. University of Michigan.
218 Cox process 4.and sum 221. 271274.308 CramerLundberg model: see compound Poisson model cumulative process 334 dams: see storage process differential equation 16.287292. 17.150. 80 81. 91. 5. 170173.318320 change of measure 2630. 341.272. 79. 162164.Index adjustment coefficient 17. 141144. 8283 hyperexponential distribution 7. 283.100. 248 WienerHopf 144 interest rate 190.242. 119. 3334. 226. 323 Coxian distribution 147. 17.292293 Edgeworth expansion 113.200201.160167. 239. 117127 corrected 121127 duality 1314. 39. 2526. 117128. 301 Kronecker product.249250 integral equation 16 Lindley 143 renewal 64. 316323 Bessel function 102. 111117.328330. 7879.359 aggregate claims 103106. 15.226. 360 excursion 155156.135. 205. 89. 217.121129. 278 gamma distribution 67.285292.203. 9496.4447.249.178184. 3436.137141.6779. 318319 Erlang distribution 7. 3839. 12 CramerLundberg approximation 1617. 189. 7475. 97. 37. 7079. 308. 97129.182. 196201 inverse Gaussian distribution 76.251280 heavy traffic 76.301 central limit theorem 60 . 3032. 1112. 2425.9899.228229.293294. 1415. 180182. 245248.259261.281. 9396. 86. 1819. 4851. 201 Brownian motion 3 . 207 heavytailed distribution 6.269. 5796.314316. 40.86.346349 383 .307312 compound Poisson model 4. 332333 Volterra 192194. 122. 302303 diffusion approximation 17. 138139. 361 diffusion 3. 9293. 201214. 14. 135.299. 110113. 7179.217.185187. 227229.203.
113114.134135. 15.287.234.349 350 perturbation 172173. 71. 134. 106108.227230. 185187 GI/G/1 141144 M/D/1 6667 equation 16. 154. 38.215250.218221.287291 INDEX matrix equation . 133. 44. 245 M/G/1 13. 141144. 203204. 234 matrixexponential distribution 240244 matrixexponentials 14.123. 35. 86 periodicity 12. 44. 37.261264.108.180. 149.259261. 267269 Panjer's recursion 320323 Pareto distribution 910. 41. 162. 5758.298299.201. 25.161. 96.315 inequality 1718. 138. 100.178182. 38. 3639. 171.234240. 2730. 251. 59. 52 53. 16. 7576.269271. 9899.304 process 2830.336339 . 42.161164. 35. nonlinear 155.275278. 134135. 144.336339 Laplace transform 15.152160. 6162. 133. 112113. 145187.340350 multiplicative functional 2830. 7179. 260 Lundberg conjugation 6979 . 230.240244. 157.285287 queue 14 . 227228. 99.384 ladder heights 4756.146148. 213214. 14. 229 M/M/1 101 Markovmodulated 185187 periodic 187 martingale 2426. see also sensitivity analysis phasetype distribution 8. 304305 random walk 3336.238. 3947. 175 light traffic 8183 Lindley integral equation 143 process 3334. 179 NP approximation 318320 Palm distribution 5253.161.288290. 25. 306316 Levy process 3. 176185 nonhomogeneous 60 PollaczeckKhinchine formula 6167.297299. 4446. 269 PerronFrobenius theory 4142.174. 16. 271274. 108 life insurance 5. 3947.148. 39.350361 Poisson process Markovmodulated 12 periodic 12. 80. 142 likelihood ratio : see change of measure lognormal distribution 9.139141. 108109. 6970.302. 203 Markov additive process 12.160161.339 large deviations 129.348 terminating 215216. 178 modulation 12. 295. 32. 257. 176185. 137139. 132133. 65. 261264.128129.
292294. 168172 storage process 13. 281296 stable process 15. 123. 5455. 7475. 152. 333334 regular variation 10. 261264 reservedependent premiums 14.336339 workload 13. 307308. 8386.273274. 141144. 233. 120 statistics x. 260 reinsurance 8.INDEX 385 waiting time 141. 335336 sensitivity analysis 8693. 240.262263. 260 WienerHopf theory 144. 244.279280 Rouche roots 158. 326330 Weibull distribution 9. 229234.154157. 162. 222. 131144. 177 timereversion 14. 12. 251. 332333 model 12. see also matrixexponential distribution regenerative process 264 268.359361 stochastic control x stochastic ordering 18. 280. 60. 172173. 1819. 331336 equation 64. 9693. 186187 renewal process 131. 256258. 327 . 89. 317318 semiMarkov 147. 253. 251. 191192. 174. 3032. 189214. 107. 147. 233234. 238 saddlepoint method 115117. 87. 186187 virtual: see workload rational Laplace transform 8. 37. 294296 shotnoise process 314 simulation 19. 160.186. 31.244250. 4950. 257. 213. 146. 251280 time change 4. 279280 subexponential distribution 11. 223226. 338 utility 324. 11.314.
com 2779 he 9 "789810ll22293211 . Special features of the book are the emphasis on change of measure techniques. extensions of the classical compound Poisson model to allow f o r reservedependent premiums. the ^W A l \ i l ' ''' CramerLundberg approximation. exact solutions.. It is a comprehensive treatment of the known results on ruin probabilities. Markovmodulation or periodicity.." Short Book Reviews ISBN 9810222939 mi u inn i nun I I I I I I i in u www.g. worldscientific.T [Ail i The book is a comprehensive treatment of  I i I \ classical and modern ruin probability theory. P'i yfliother approximations (e. y finite horizon ruin probabilities.Vol. Some i (l I JL I J r of the topics are Lundberg's inequality. I 1! Ruin Probabilities . for heavytailed claim size distributions). "This book is a must for anybody working in applied probability.Advanced Series on Statistical Science & Applied Probability . 2 A I 11 JjVb l' i  i Yj . phasetype distributions as a computational vehicle and the connection to other applied probability areas like queueing theory.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue reading from where you left off, or restart the preview.