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
. . . . . .. . . . . . . . . . . . .. . 294 XI Miscellaneous topics 297 1 The ruin problem for Bernoulli random walk and Brownian motion. . . 336 A3 Matrixexponentials . . The twobarrier ruin problem . . . . . . . .. . 316 5 Principles for premium calculation . . 323 6 Reinsurance . . . . . . . . . 304 3 Large deviations . . . . . . . . . . . . . .CONTENTS vii IX Ruin probabilities in the presence of heavy tails 251 1 Subexponential distributions . . . . . 261 4 Models with dependent input . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . .. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . 264 5 Finitehorizon ruin probabilities . . . . . . . . . . . . . . . . . . . . . 271 6 Reservedependent premiums . .. . . . . . . . . . . . . . . . .. . . . . . . . . . . . . 331 A2 WienerHopf factorization . . .. . . . . .. 259 3 The renewal model . . . . . . . . . . . . . . . . .. . . . . . . . 287 4 Importance sampling for the finite horizon case . . . . . . . . .. . .. . .. . . . . 350 Bibliography Index 363 383 . . . .. . . . . . . 290 5 Regenerative simulation . . . .. . . . . .. . 297 2 Further applications of martingales . . . . . . . . . . . . . . . . . . . 292 6 Sensitivity analysis . . . 340 A4 Some linear algebra . 281 2 Simulation via the PollaczeckKhinchine formula . . . . . . . . . . . . . . . 279 X Simulation methodology 281 1 Generalities . . . . . . .. .. . . . . . . . . . . . . . . . . . .. 326 Appendix 331 Al Renewal theory . . . . . . . . . . 344 AS Complements on phasetype distributions . . . . . . . . . . . 251 2 The compound Poisson model . . . . . . . . 306 4 The distribution of the aggregate claims . . . . . . . . . . . 285 3 Importance sampling via Lundberg conjugation . . . .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . .. . . . . .. . . . . . .
This page is intentionally left blank .
and other projects absorbed my interest. and has been an active area of research from the days of Lundberg all the way up to today. Let me take this opportunity to thank above all my publisher World Scientific Publishing Co.Preface The most important to say about the history of this book is: it took too long time to write it! In 1991. the book is basically mathematical in its flavour. Thus. it would not be fair not to say that the practical relevance of the area has been questioned repeatedly. which has in particular removed one of the standard criticisms of the area. Apart from these remarks. It has obviously not been possible to cover all subareas. the idea was close to expand these to a short book on the subject. The course was never realized. Risk theory in general and ruin probablities in particular is traditionally considered as part of insurance mathematics. I was invited to give a course on ruin probabilities at the Laboratory of Insurance Mathematics. A similar thank goes to all colleagues who encouraged me to finish the project and continued to refer to the book by Asmussen which was to appear in a year which continued to be postponed. I have deliberately stayed away from discussing the practical relevance of the theory. As an excuse: many of these projects were related to the book. One reason for writing this book is a feeling that the area has in the recent years achieved a considerable mathematical maturity. In particular. However. and my belief was that this could be done rather quickly. it is not by intention. if the formulations occasionally give a different impression. Since I was to produce some handouts for the students anyway. this applies to longrange dependence which is intensely studied in the neighboring ix . University of Copenhagen. 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!). that it can only say something about very simple models and questions. and the result is now that the book is much more related to my own research than the initial outline. and the series editor Ole BarndorffNielsen for their patience. But the pace was much slower than expected.
VI.2 more properly).89.4a.2. see e.13. For a brief orientation. the standard stochastic control setting of diffusion models has been considered.45.maths . Finally. 111.g.lth. some basic discussion can be found in the books by Biihlmann [82] and Gerber [157].g. A book like this can be organized in many ways. Here is a suggestion on how to get started with the book. for the effects on tail probabilities. The main motivation comes from statistical data for network traffic (e.1.13 and IX. The book does not go into the broader aspects of the interface between insurance mathematics and mathematical finance. some papers not cited in the text but judged to be of interest are included in the Bibliography. an area which is becoming increasingly important.lth. IV. Asmussen. Hojgaard & Taksar [206].se Lund February 2000 Soren Asmussen . another by method.13 and XI. Resnick & Samorodnitsky [303] and references therein. for which I apologize to the reader and the authors of the many papers who ought to have been on the list.3.15.2. X. Chapters IIIVII introduce some of the main models and give a first derivation of some of their properties. IX. Hojgaard & Taksar [35] and Paulsen & Gjessing [284]. I regret that due to time constraints.5. The rest is up to your specific interests. IV. 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 has not been possible to incorporate more numerical examples than the few there are. VIII. I intend to keep a list of misprints and remarks posted on my web page. read Chapter I. IV. see also Schmidli [325] and the references in Asmussen & Taksar [52]. Another interesting area which is not covered is dynamic control. For a second reading.x PREFACE field of queueing theory. VII. The present book is in between these two possibilities. incorporate 11. [381]). In addition. the first part of 11.se/matstat / staff/asmus and I am therefore grateful to get relevant material sent by email to asmusfmaths . Good luck! I have tried to be fairly exhaustive in citing references close to the text. 111.6 (to understand the PollaczeckKhinchine formula in 111. More recently. e. Willinger et al. Concerning ruin probabilities. see in particular Michna [259]. http:// www. VII.14. It is obvious that such a system involves a number of inconsistencies and omissions.13.g. In the classical setting of CramerLundberg models. One is by model.
Parts of X.8 .3 are reprinted from Asmussen & Rubinstein [46] and parts of VIII. were produced by Lone Juul Hansen . 111 . Lund September 2001 Soren Asmussen Acknowledgements Many of the figures . Section VIII. Aarhus. supported by Center for Mathematical Physics and Stochastics (MaPhySto). 5. 1 is almost identical to Section 2 of Asmussen [26] and reprinted with permission of Blackwell Publishers. Schmidli & Schmidt [47] with the permission from Applied Probability Trust . 5 from Asmussen & Kliippelberg [36] with the permission from Elsevier Science . Parts of II.PREFACE xi The second printing differs from the first only by minor corrections. 3 is reprinted from Asmussen & Nielsen [39] and parts of IX. .1 and X. not least the more complicated ones.4 from Asmussen.6. IV.1 by Bjarne Hojgaard and the table in Example 111. Fig.5 from Asmussen [21] with permission from CRC Press. A number of other figures were supplied by Christian Geisler Asmussen . Section VII . More substantial remarks. Fig. many of which were pointed out by Hanspeter Schmidli . of which there are not many at this stage .6 by my 1999 simulation class in Lund. as well as some additional references continue to be at the web page.6 is reprinted from Asmussen & Schmidt [49] and parts of IX.2 by Rafal Kulik .
This page is intentionally left blank .
(1. it is frequently more convenient to work with the claim surplus process {St}t>0 defined by St = u . and give a very brief summary of some of the models. results and topics to be studied in the rest of the book. (1. They are the main topics of study of the present book.3) sup St. (1. For mathematical purposes. The probability O(u) of ultimate ruin is the probability that the reserve ever drops below zero.1) We also refer to t/) ( u) and 0(u. we introduce some general notation and terminology.4) O<t<oo O<t<T 1 . A risk reserve process { Rt}t>o. Letting T(u) = inf {t > 0 : Rt < 0} = inf It > 0 : St > u}. is a model for the time evolution of the reserves of an insurance company. as defined in broad terms .Rt. respectively. We denote throughout the initial reserve by u = Ro. MT = sup St. M = (1.2) (O<t<T Ro=ul.T) = P inf Rt < 0 I . T) as ruin probabilities with infinite horizon and finite horizon .i(u. t/i(u) = P (infRt < 0) = P (infR t < 0 t>0 t>0 The probability of ruin before time T is t.Chapter I Introduction 1 The risk process In this chapter .
and Nt = min {n > 0 : 0rn+1 > t} = max {n > 0: Un < t}• The size of the nth claim is denoted by Un. (1. That is. Thus. St = E Uk .2 CHAPTER I. and T1 is the time of the first claim. t] is finite. (1.pt. the number Nt of arrivals in [0. (1. 1. Figure 1.. respectively. say.T) = F (MT > u) = P(r(u) < T). the following setup will cover the vast majority of the book: • There are only finitely many claims in finite time intervals.i(u.6) Sofar we have not imposed any assumptions on the risk reserve process. Putting things together. However. INTRODUCTION be the time to ruin and the maxima with infinite and finite horizon.E Uk.1. the ruin probabilities can then alternatively be written as .7) k=1 k=1 The sample paths of {Rt} and {St} and the connection between the two processes are illustrated in Fig.. . the time of arrival of the nth claim is an = T1 + • • • + Tn.1 . we see that Nt Nt Rt = u + pt . per unit time.5) i.b(u) = P (r(u) < oo) = P(M > u). T3. We denote the interarrival times of claims by T2. • Premiums flow in at rate p.
on Fig. not discuss whether this actually corresponds to practice. however. If 77 > 0. We shall not deal with this case either. then M < oo a. Thus. We study this case in Ch. that the insurance company should try to ensure 77 > 0.1.8) The interpretation of p is as the average amount of claim per unit time. rl= pP P It is sometimes stated in the theoretical literature that the typical values of the safety loading 77 are relatively small. For the purpose of studying ruin probabilities this distinction is. we shall.20%. but as an approximation to the risk process rather than as a model of intrinsic merit. and in fact: Proposition 1.s. THE RISK PROCESS 3 Note that it is a matter of taste (or mathematical convenience) whether one allows {Rt} and/or {St} to continue its evolution after the time T(u) of ruin. allowing a countable infinity of jumps on Fig. of course. (1. • General Levy processes (defined as continuous time processes with stationary independent increments) where the jump component has infinite Levy measure. VII. however. 1.(. 1. If 77 < 0.8) holds.) V 0. say 10% . immaterial. It would appear obvious. for example.1 Assume that (1. one could well replace Rt by Rtnr(u) or RtA. then M = oo a. . • Brownian motion or more general diffusions.. The models we consider will typically have the property that there exists a constant p such that Nt a E Uk k=1 p. t * oo.s. We shall discuss Brownian motion somewhat in Chapter IV.1 the slope of {Rt} should depend also on the level). one may well argue that Brownian motion in itself could be a reasonable model. Some main examples of models not incorporated in the above setup are: • Models with a premium depending on the reserve (i. and hence . A further basic quantity is the safety loading (or the security loading) n defined as the relative amount by which the premium rate p exceeds p. However.b(u) = 1 for all u.e.1. and the basic ruin probabilities are derived in XI.1. a basic references is Gerber [127]. since any modeling involves some approximative assumptions.. and hence O(u) < 1 for all sufficiently large u. though many results are straightforward to generalize from the compound Poisson model.
in connection with risk processes in a Markovian or periodic environment (Chapter VI). . U2. This case is referred to as the mixed Poisson process.6EU (on the average.i. and independent of {Nt}. If U1.Q (say) and U1. The simplest example is 3(t) = V where V is a r . ..v.i. Then the connection between the ruin probabilities for the given risk process {Rt} and those ^(u). are i. 0 We shall only encounter a few instances of a Cox process.10) is a property which we will typically encounter.T) = i. this needs to be verified in each separate case.. _ St __ k =1 Uk pt a4. Here it is easy to see that p = . Proposition 1.11) . and independent of {(0(t).s. St In concrete models.8) that F N. However. 0(u..oo t 0 J (provided the limit exists). Nt)}. then this limit is > 0 which implies St a$ oo and hence M = oo a.10) Again. with the most notable special case being V having a Gamma distribution. where {Nt} is a Poisson process with rate .. INTRODUCTION Proof It follows from (1.4 CHAPTER I.10) hold with p constant. U2..3 Assume p 54 1 and define Rt = Rt1p. zP(u .8).d. and that .b(u) < 1 for all u when rl > 0.i(u.s. we obtain typically a somewhat stronger conclusion. Thus p may well be random for such processes. if {(3(t)} is nonergodic. it is not too difficult to show that p as defined by (1.d. . not all models considered in the literature have this feature: Example 1. corresponding to the Pdlya process.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.8) is given by ^t p = EU • lim it (3(s) ds t.s.Tp). If u oo. and here (1.2 (Cox PROCESSES) Here {Nt} is a Poisson process with random rate /3(t) (say) at time t. are i. then similarly limSt/t < 0.T) for {Rt} is given by V)(u) = t/i (u). If 77 < 0. M < oo a. namely. The simplest concrete example (to be studied in Chapter III) is the compound Poisson model. t t p  p' t ^ oo. (1. (1. (1. However. namely that M = oo a. rl > 0. k=1 (1. tb(u) = 1 for all u holds also when rl = 0.
U2. Grandell [171]. while the first mathematically substantial results were given in Lundberg [251] and Cramer [91]. often referred to as collective risk theory or just risk theory. Pentikainen & Pesonen [101]. Some main later textbooks are (in alphabetical order) Buhlmann [82]. and in fact p < 1 is the fundamental assumption of queueing theory ensuring steadystate behaviour (existence of a limiting stationary distribution). [134]. Heilmann [191].. which is feasible since in most cases the process { Rt } has a similar structure as {Rt} (for example. The Swedish school was pioneering not only in risk theory. Note that when p = 1. Cox processes are treated extensively in Grandell [171]. We roughly classify these into two groups .g. see e . and we do not get near to the topic anywhere in this book. An idea of the additional topics and problems one may incorporate under risk theory can be obtained from the survey paper [273] by Norberg. Segerdahl [334] and Philipson [289]. Some of the main general ideas were laid down by Lundberg [250].g. For mixed Poisson processes and Polya processes. the recent survey by Grandell [173] and references therein. In the even more general area of nonlife insurance mathematics. Embrechts et al. another important early Swedish work is Tacklind [373]. Daykin.. Gerber [159]) has a rather different flavour. lighttailed distributions (sometimes the term . CLAIM SIZE DISTRIBUTIONS 5 The proof is trivial. Note that life insurance (e. Sundt [354]. Hipp & Michel [198]. Mitteilungen der Verein der Schweizerischen Versicherungsmathematiker and the Scandinavian Actuarial Journal. De Vylder [110].2. Insurance: Mathematics and Economics. the claim arrivals are Poisson or renewal at the same time). in particular. was largely initiated in Sweden in the first half of the century. [101]. [330]. Since { Rt } has premium rate 1. but in probability and applied probability as a whole. Rolski. we shall be able to identify p with the traffic intensity of an associated queue. The term risk theory is often interpreted in a broader sense than as just to comprise the study of ruin probabilities. Schmidli. some main texts (typically incorporating some ruin theory but emphasizing the topic to a varying degree) are Bowers et al. Schmidt & Teugels [307] and Seal [326]. Straub [353]. the role of the result is to justify to take p = 1. Besides in standard journals in probability and applied probability. many results and methods in random walk theory originate from there and the area was ahead of related ones like queueing theory. in a number of models. Daykin et al. Notes and references The study of ruin probabilities. 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.. Taylor [364]. see also Chapter XI. the assumption > 0 is equivalent to p < 1. Gerber [157]. the research literature is often published in journals like Astin Bulletin . Some early surveys are given in Cramer [91]. Buhlmann [82]. [76]..
As in a number of other applied probability areas. where B(bo. 6 has density r(p)xPleax b(x) P and m.f. 2a Lighttailed distributions Example 2. the m.2 and /LB is the mean of B.1 (THE EXPONENTIAL DISTRIBUTION) Here the density is b(x) = beax (2. and can also be interpreted as the (constant) failure rate b(x)/B(x). INTRODUCTION 'Cramertype conditions' is used). B is heavytailed if b[s] = oo for all s > 0.2) = 0. i.1) The parameter 6 is referred to as the rate or the intensity. In contrast.x given X > x is again exponential with rate b (this is essentially equivalent to the failure rate being constant). The crucial feature is the lack of memory: if X is exponential with rate 6. a simple stopping time argument shows that this implies that the conditional distribution of the overshoot ST(u) . a fact which turns out to contain considerable information. regularly varying (see below) or even regularly varying with infinite variance. for the compound Poisson model with exponential claim sizes the ruin probability . s<8. 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'. the exponential distribution is by far the simplest to deal with in risk theory as well.g. if 1 °O AB Jbos x B(dx) > 0. Here lighttailed means that the tail B(x) = 1 . Example 2 . and heavytailed distributions. then the conditional distribution of X . but different more restrictive definitions are often used: subexponential. Equivalently.u at the time of ruin given r(u) is again exponential u with rate 8. On the more heuristical side.O(u) can be found in closed form.2 (THE GAMMA DISTRIBUTION) The gamma distribution with parameters p.g.f.e. P B[s]= (8Is ) .6 CHAPTER I. (2. B[s] is finite for some s > 0.B(x) satisfies B(x) = O(e8x) for some s > 0.3) . For example in the compound Poisson model.8. In particular.
Ruin probabilities for the general case has been studied.) VarX1 (EX )2 p is < 1 for p > 1. are i. x] so that B(x) = r` e.i.c. . CLAIM SIZE DISTRIBUTIONS 7 The mean EX is p/b and the variance Var X is p/b2. 0. 0 < ai < 1..2.c. P b(x) = r` aibiea.v. . In particular. B(x) = r(p) Asymptotically.. X2.. u . the squared coefficient of variation (s. JP 1 B(x) r(p ) XP ie ax In the sense of the theory of infinitely divisible distributions. p). p) = J tPletdt.v. An important property of the hyperexponential distribution is that its s. In particular.y i=1 where >i ai = 1. > 1 for p < 1 and = 1 for p = 1 (the exponential case).2) can be considered as the pth power of the exponential density (2.d. This special case is referred to as the Erlang distribution with p stages... then X v Xl + • • • + X.. if p is integer and X has the gamma distribution p. one has r(bx. where X1. . and exponential with rate d. u Example 2 .ate (b2 ): L• i=o In the present text. p) °° where r (x.1 Poisson events in [0. An appealing feature is its simple connection to the Poisson process: B(x) = P(Xi + • • • + XP > x) is the probability of at most p . by Grandell & Segerdahl [175] and Thorin [369]. 2.). i = 1. is > 1. or just the Erlang(p) distribution. we develop computationally tractable results mainly for the Erlang case (p = 1.1) (or the 1/pth root if p < 1). among others. p.3 (THE HYPEREXPONENTIAL DISTRIBUTION) This is defined as a finite mixture of exponential distributions. the Gamma density (2.. The exact form of the tail B(x) is given by the incomplete Gamma function r(x.
a. T) or sometimes the triple (E.f.f. q2 q3 (2.(2. However. This class of distributions is popular in older literature on both risk theory and queues. 1)' is the column vector with 1 at all entries. are b(x) = aeTxt. The parameters of a phasetype distribution is the set E of transient states.7) q1 b(x) = cjxieWWx + djxi cos(ajx)ea'x + > ejxi sin(bjx)e`ix . Important special cases are the exponential.8) j=1 j=1 j=1 where the parameters in (2..8 CHAPTER I. See XI. We give a more comprehensive treatment in VIII. We give some theory for matrixu exponential distribution in VIII.6. of which one is absorbing and the rest transient.g.8) are realvalued. Example 2 . it is notable from a practical point of view because of reinsurance: if excessofloss reinsurance has been arranged with retention level xo.. Equivalent characterizations are that the density b(x) has one of the forms q b(x) j=1 = cjxienbx. Example 2 . The density and c. but the current trend in applied probability is to restrict attention to the class of phasetype distributions. . INTRODUCTION Example 2 . B(x) > 0 for x < xo) is of course a trivial instance of a lighttailed distribution. equivalently.e. (or.1 and defer further details to u Chapter VIII. the restriction T of the intensity matrix of the Markov process to E and the row vector a = (ai)iEE of initial probabilities. B(x) = aeTxe where t = Te and e = (1 . resp.6. then the claim size which is relevant from the point of view of the insurance company itself is U A xo rather than U u (the excess (U . The couple (a.xo)+ is covered by the reinsurer). the Erlang and the hyperexponential distributions. a rational Laplace transform) if B[s] _ p(s)/q(s) with p(s) and q(s) being polynomials of finite degree.d.4 (PHASETYPE DISTRIBUTIONS) A phasetype distribution is the distribution of the absorption time in a Markov process with finitely many states.7) are possibly complexvalued but the parameters in (2. which is slightly smaller but more amenable to probabilistic reasoning.6 (DISTRIBUTIONS WITH BOUNDED SUPPORT) This example (i.5 (DISTRIBUTIONS WITH RATIONAL TRANSFORMS) A distribution B has a rational m. there exists a xo < oo such that B(x) = 0 for x > xo. T) is called the representation. This class of distributions plays a major role in this book as the one within computationally tractable exact forms of the ruin probability z/)(u) can be obtained.
CLAIM SIZE DISTRIBUTIONS 9 2b Heavytailed distributions Example 2. However. a)/A)a+1' x > a. All moments are finite. (2. 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). x < a. the mean u is eµ+a /2 and the second moment is e2µ+2o2. u Example 2 . (2. Here failure rates b(x) = b(x)/B(x) play an important role.12) Sometimes also a location parameter a > 0 and a scale parameter A > 0 is allowed.1.pl = 1 W (logx .1). There are various variants of the definition around.u l b(x) = d dx or J ax lor 1 exp Asymptotically.10) The loinormal distribution has moments of all orders. (2. b(x) = crx''le`xr. the tail is B (x ) 2 x.2. In particular.N(0. or equivalently as the distribution of a°U+µ where U .9 (THE PARETO DISTRIBUTION) Here the essence is that the tail B(x) decreases like a power of x.11) ex log logx 2r p 1 1 2 ( a ) f 1 (lox_P)2} (2.p a 1 (2.7 (THE WEIBULL DISTRIBUTION) This distribution originates from reliability theory. It follows that the density is 't (1ogX . the exponential distribution representing the simplest example since here b(x) is constant.N(p. p is defined as the distribution of ev where V .13) u . we obtain the Weibull distribution B(x) = eCx'. and then b(x) = 0. Example 2 .9) which is heavytailed when 0 < r < I. Writing c = d/r. a2). b(x) _ A(1 + (x a The pth moment is finite if and only if p < a . one being B(x) (1 + X)b(x) (1 + x)a+1' x > 0.8 (THE LOGNORMAL DISTRIBUTION) The lognormal distribution with parameters a2.
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).17) where L (x) is slowly varying. In general. The simplest examples correspond to p small and integervalued. examples of distributions with regularly varying tails are the Pareto distribution (2. in particular.(1 + Zx + $ p = 3.12) (here L (x) * 1) and ( 2. A = 1 and X is standard exponential.'s of the form YX. Choudhury & Whitt [1] as the class of distributions of r.13). x + 00. u . another standard example is (log x)'). where Y is Pareto distributed with a = (p . Thus. the density is { 3 (1 . the loggamma distribution is a Pareto distribution.v. satisfies L(xt)/L(x) 4 1. (2.2). INTRODUCTION Example 2. x 4 oo (any L having a limit in (0. i.10 CHAPTER I. the loggamma distribution (with exponent 5) and a Pareto mixture of exponentials. The density is 8p(log x)pi b(x) .16) 11 Example 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) . u Example 2 .11 (PARETO MIXTURES OF EXPONENTIALS) This class was introduced by Abate.L( x ).x6+lr(p) (2. B(x) = O(xP).10 (THE LOGGAMMA DISTRIBUTION) The loggamma distribution with parameters p.(1 + 2x + 2x2)e2x) p = 2 (2. 6 is defined as the distribution of et' where V has the gamma density (2.14) The pth moment is finite if p < 5 and infinite if p > 5. For p = 1.15) x2 + 16x3 ) a3x/2) 3 (1 .e. oo) is slowly varying .1)/p. (2. { s () 1s+3s29s3log(1+2s I p=3.
4) or even to completely different applied probability areas like extreme value theory: if we are using a Gaussian process to predict extreme value behaviour. the claim size distribution represents of course only one aspect (though a major one). and so is the Weibull distribution with 0 < r < 1. From a practical point of view.18) B(x) It can be proved (see IX. We return to a closer study in IX. Namely. the subexponential class of distributions provide a convenient framework for studying large classes of heavyu tailed distributions.3. 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. We give some discussion on standard methods to distinguish between light and heavy tails in Section 4f. the knowledge of the claim size distribution will typically be based upon statistical data. one may argue that this difficulty is not resticted to ruin probability theory alone. The reason is in part mathematical since this model is the easiest to analyze.. 3 The arrival process For the purpose of modeling a risk process . this phenomenon represents one of the true controversies of the area. Similar discussion applies to the distribution of the accumulated claims (XI. THE ARRIVAL PROCESS 11 Example 2. but can never be sure whether this is also so for atypical levels for which far less detailed statistical information is available. U2. which each have a ( timehomogeneous) small rate of experiencing a . but the model also admits a natural interpretation : a large portfolio of insurance holders . it will be seen that we obtain completely different results depending on whether the claim size distribution is exponentially bounded or heavytailed. Thus.. 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. When studying ruin probabilities. and based upon such information it seems questionable to extrapolate to tail behaviour. though the proof of this is nontrivial..1. for example the lognormal distribution is subexponential (but not regularly varying).. However. Also. (2.1) that any distribution with a regularly varying tail is subexponential. 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.
I. with a common term {Nt} is a Markovmodulated Poisson process . Cox processes are. has some mathematically appealing random walk features . 5. which facilitate the analysis. Some of them have concentrated on the marginal distribution of NT (say T = one year ). gives rise to an arrival process which is very close to a Poisson process. T2. Mathematically. found the Poisson distribution to be inadequate and suggested various other univariate distributions as alternatives . getting away from the simple Poisson process seems a crucial step in making the model more realistic.6.d. too general and one neeed to specialize to more concrete assumptions . 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.e. The point of view we take here is Markov dependent random walks in continuous time (Markov additive processes ).8 (t) is a periodic function of t.(3. are i. where {/3 (t)}too is an arbitrary stochastic process . it is more questionable whether it provides a model with a similar intuitive content as the Poisson model.g. This model can be intuitively understood in some simple cases like { Jt} describing weather conditions in car insurance . not many detailed studies of the goodnessoffit of the Poisson model in insurance are available . Historically.. and also that the ruin problem may be hard to analyze . To the author 's knowledge . but with a general not necessarily exponential distribution ). so that . see 11. to be studied in Chapter V. A more appealing way to allow for inhomogeneity is by means of an intensity . The compound Poisson model is studied in detail in Chapters III. Nevertheless . An obvious example is 3(t) depending on the time of the year (the season). when Jt = i. INTRODUCTION claim .3(t) fluctuating over time..i. its basic feature is to allow more variation (bursty arrivals ) than inherent in the simple Poisson process. Another one is Cox processes. the negative binomial distribution. in particular to allow for certain inhomogeneities. The one we focus on (Chapter VI) is a Markovian environment : the environmental conditions are described by a finite Markov process {Jt }too. in just the same way as the Poisson process arises in telephone traffic (a large number of subscribers each calling with a small rate). the periodic and the Markov modulated models also have attractive features . in Chapter VII). This model . however. the first extension to be studied in detail was {Nt } to be renewal (the interarrival times T1 . 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 . In others. e. with the extension to premiums depending on the reserve. we study this case in VI . radioactive decay (a huge number of atoms each splitting with a tiny rate ) and many other applications. However ... In order to prove reasonably substantial and interesting results . such that 8(t) = . epidemics in life insurance etc. IV (and.12 CHAPTER I. This applies also to the case where the claim size distribution depends on the time of the year or .
R = p(R) in between jumps. this amounts to Vo having the stationary distribution of {Vt}).4. dam/storage processes. extreme value theory. methods or modeling ideas developed in one area often has relevance for the other one as well.v. The M/G/1 workload process { Vt } may also be seen as one of the simplest storage models.'s like V is available. others being branching processes. that quite often the emphasis is on computing expected values like EV. A general release rule p(x) means that {Vt} decreases according to the differential equation V = p(V) in between jumps. In the setting of (4. A SUMMARY OF MAIN RESULTS AND METHODS 13 the environment (VI. and which seems well motivated from a practical point of view as well. The study of the steady state is by far the most dominant topic of queueing and storage theory. Markovmodulated or periodic can be related to queues with similar characteristics. More generally. this gives only f0 O°i (u)du which is of limited . interacting particle systems. The ones which appear most related to risk theory are queueing theory and dam/storage processes. point processes and so on. (4.0 (u. It should be noted. however. A stochastic process {Vt } is said to be in the steady state if it is strictly stationary (in the Markov case. with Poisson arrivals and constant release rule p(x) = 1.1) permitting to translate freely between risk theory and the queueing/storage setting.1).1) where V is the limit in distribution of Vt as t + oo. and here (4. stochastic geometry. time series and Gaussian processes. and the limit t 4 oo is the steadystate limit. others are quite different. and a lot of information on steadystate r. queueing theory. reliability.T) = P(VT > u). genetics models. Some of these have a certain resemblance in flavour and methodology. stochastic differential equations.6) . Thus. 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 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 .1) holds as well provided the risk process has a premium rule depending on the reserve. In fact. Similarly. it is a recurrent theme of this book to stress this connection which is often neglected in the specialized literature on risk theory. 0(u) = P(V > u). it is desirable to have a set of formulas like (4. Mathematically. ruin probabilities for risk processes with an input process which is renewal.
• The compound Poisson model with a claim size distribution degenerate at one point..14 CHAPTER I. e .2). 3. p = 0/8 and y = 8 . as is typically the case.1 .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 ideal is to be able to come up with closed form solutions for the ruin probabilities 0(u). .1). see Boxma & Cohen [74] and Abate & Whitt [3].1 is a sample path relation should be stressed : in this way the approach also applies to models having supplementary r. 3.v. A prototype of the duality results in this book is Theorem 11. • The compound Poisson model with premium rate p(x) depending on the reserve and exponential claim size distribution B. The cases where this is possible are basically the following for the infinite horizon ruin probability 0(u): • The compound Poisson model with constant premium rate p = 1 and exponential claim size distribution B.p(y) y^ Jo p(x) can be written in closed form.8. The qualifier 'with just a few phases ' refers to the fact that the diagonalization has to be carried out numerically in higher dimensions.1. Thus .3.'s like the environmental process {Jt} in a Markovmodulated setting. The infinite horizon (steady state ) case is covered by letting T oo. see Corollary III. Here O(u) = peryu where 3 is the arrival intensity. INTRODUCTION intrinsic interest .g. the two areas.T).3. • The compound Poisson model with constant premium rate p = 1 and B being phasetype with a just few phases . The fact that Theorem H. the functions w x f d 1 exdx () . 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. Similarly. which gives a sample path version of (4. Here ?P(u) is explicit provided that . see Corollary VII. 4b Exact solutions Of course . which can be expanded into a sum of exponential terms by diagonalization (see. • The compound Poisson model with some rather special heavytailed claim size distributions. have to some extent a different flavour. Example VIII. Vi(u. B(x) = ebx.6.3. though overlapping. Here Vi(u) is given in terms of a matrixexponential function ( Corollary VIII.
Here are some of the main approaches: Laplace transform inversion Often. say the fast Fourier transform (FFT) as implemented in Grubel [179] for infinite horizon ruin probabilities for the renewal model.1) is the explicit form of the ruin probability when {Rt} is a diffusion with infinitesimal drift and variance µ(x). Grubel & Pitts [132] and Grubel & Hermesmeier [180] (see also the Bibliographical Notes in [307] p.Lef$er function. Abate & Whitt [2]. Ab(u).4.f f 2µ(y)/a2(y) dy} dx  (4. A notable fact ( see again XI.7. the second best alternative is a numerical procedure which allows to calculate the exact values of the ruin probabilities. [s. T) du dT 0 TO 00 in closed form than the ruin probabilities z/'(u).S(u) 1S(oo) f °D exp {.ff 2µ(y)/a2(y) dy} dx . see VIII. Also Brownian models or certain skip free random walks lead to explicit solutions (see XI .2) is the natural scale. a2 (x): Ip (u) = where S(u) = f °O exp {. Embrechts. O(u. f {eXp U LX 2. 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. 1). it is easier to find the Laplace transforms = f e8 . 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 . where Furrer [150] recently computed ii(u) as an infinite series involving the Mittag. (u. but are somewhat out of the mainstream of the area . relevant references are Grubel [179].b(u)du .u(y)/a2(y) dy} 4c Numerical methods Next to a closedform solution. T) themselves. However. For the finite horizon ruin probability 0(u. T). We don't discuss Laplace transform inversion much.1) are so complicated that they should rather be viewed as basis for numerical methods than as closedform solutions. T) can then be calculated numerically by some method for transform inversion. . 191). the formulas ( IV. esuTb( u. Given this can be done. • An astable Levy process with drift .
Ce"u. U is explicit in terms of the model parameters. and in particular the naive idea of conditioning upon process behaviour in [0.16 CHAPTER L INTRODUCTION Matrixanalytic methods This approach is relevant when the claim size distribution is of phasetype (or matrixexponential). Differential. whereas for the renewal arrival model and the Markovian environment model U has to be calculated numerically. 0(u) is then given in terms of a matrixexponential function euu (here U is some suitable matrix) which can be computed by diagonalization. and in quite a few cases (Chapter VIII). u * oo.4) 00['Y]1)'Y = 0.1) and y > 0 is the solution of the Lundberg equation (4. which can equivalently be written as f3 [7] = 1 +13 . most often it is more difficult to come up with reasonably simple equations than one may believe at a first sight. (4.3) in the compound Poisson model which is an integral equation of Volterra type.3) where C = (1 . dt] most often leads to equations involving both differential and integral terms.and integral equations The idea is here to express 'O(u) or '(u. T) as the solution to a differential. and carry out the solution by some standard numerical method. An example where this idea can be carried through by means of a suitable choice of supplementary variables is the case of statedependent premium p(x) and phasetype claims.3.g. it states that i/i(u) . For the compound Poisson model with p = 1 and claim size distribution B with moment generating function (m. as the solution of linear differential equations or by some series expansion (not necessarily the straightforward Eo U'u/n! one!).p)/(13B'[ry] . 4d Approximations The CramdrLundberg approximation This is one of the most celebrated result of risk theory (and probability theory as a whole).f. In the compound Poisson model with p = 1. 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.) B[s]. . One example where this is feasible is the renewal equation for tl'(u) (Corollary III.7. However. see VIII.or integral equation.
The CramerLundberg approximation is renowned not only for its mathematical beauty but also for being very precise. u > oo. Diffusion approximations Here the idea is simply to approximate the risk process by a Brownian motion (or a more general diffusion) by fitting the first and second moment. It has generalizations to the models with renewal arrivals. Approximations for O(u) as well as for 1(u. Large claims approximations In order for the CramerLundberg approximation to be valid. incorporating correction terms may change the picture dramatically. In particular. in some cases the results are even more complete than for light tails. a Markovian environment or periodically varying parameters. However. A SUMMARY OF MAIN RESULTS AND METHODS 17 It is rather standard to call ry the adjustment coefficient but a variety of other terms are also frequently encountered.2. the claim size distribution should have an exponentially decreasing tail B(x). In the case of heavytailed distributions.4. T). corrected diffusion approximations (see IV.6) 4e Bounds and inequalities The outstanding result in the area is Lundberg's inequality (u) < e"lu.6) are by far the best one can do in terms of finite horizon ruin probabilities '(u. However. for the compound Poisson model ^(u) p pu In fact . For example. See Chapter IX. In fact. and use the fact that first passage probabilities are more readily calculated for diffusions than for the risk process itself. Diffusion approximations are easy to calculate.7 and IV. some further possibilities are surveyed in 111 . in such cases the evaluation of C is more cumbersome. the exact solution is as easy to compute as the CramerLundberg approximation at least in the first two of these three models. T) for large u are available in most of the models we discuss. This list of approximations does by no means exhaust the topic. J B dx. but typically not very precise in their first naive implementation. often for all u > 0 and not just for large u. (4. . when the claim size distribution is of phasetype. other approaches are thus required.
as a general rule.d. it is a general principle that adding random variation to a model increases the risk. INTRODUCTION Compared to the CramerLundberg approximation (4.. to have smaller ruin probabilities than when B is nondegenerate with the same mean m.. it splits up into the estimation of the Poisson intensity (the estimator is /l3 = NT/T) and of the parameter(s) of the claim size distribution. . However. . which is a standard statistical problem since the claim sizes Ui. obtained say by observing the risk process in [0. of being somewhat easier to generalize beyond the compound Poisson setting. and to plot the empirical mean residual life 1 N . they have however to be estimated from data. UNT may be viewed as an interpolation in or smoothing of the histogram). In practice. given NT.) at various places and in various settings. 4f Statistical methods Any of the approaches and results above assume that the parameters of the model are completely known.8. 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. The standard suggestion is to observe that the mean residual life E[U .3). This procedure in itself is fairly straightforward. We return to various extensions and sharpenings of Lundberg's inequality (finite horizon versions.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. one expects a model with a deterministic claim size distribution B. UNT are i. it has the advantage of not involving approximations and also. can we trust the confidence intervals for the large values of u which are of interest? In the present author's opinion. However. . . more importantly.U(k)) i =k+ i . e. . say degenerate at m. the difficulty comes in when drawing inference about the ruin probabilities. in the compound Poisson model. For example. How do we produce a confidence interval? And. empirical evidence shows that the general principle holds in a broad variety of settings.k (U(`) . When comparing different risk models. fitting a parametric model to U1. T].18 CHAPTER I.i. one may question whether it is possible to distinguish between claim size distributions which are heavytailed or have an exponentially decaying tail. lower bounds etc. This is proved for the compound Poisson model in 111. though not too many precise mathematical results have been obtained. For example.. .g.
A main problem is that ruin is typically a rare event (i.. claims U1.5. 4g Simulation The development of modern computers have made simulation a popular experimental tool in all branches of applied probability and statistics. Truncation to a finite horizon has been used.. having small probability) and that therefore naive simulation is expensive or even infeasible in terms of computer time. Thus Proposition 4. and also discuss how to develop methods which are efficient in terms of producing a small variance for a fixed simulation budget. UN.3) and Section VI.. . but is not very satisfying. We look at a variety of such methods in Chapter X.. Klnppelberg & Mikosch [134]. formula VI. . in this book referred to as [APQ]).3) or Section 3 of Chapter VI are referred to as Proposition VI.(5.4.v's) which can be generated by simulation.d. See further Embrechts.2. The chapter number is specified only when it is not the current one.e. and in fact methods from that area can often be used in risk theory as well . Simulation may be used just to get some vague insight in the process under study: simulate one or several sample paths. good methods exist in a number of models and are based upon representing the ruin probability zb(u) as expected value of a r. to observe whether one or the other limiting behaviour is apparent in the tail.2. 5 Conventions Numbering and reference system The basic principles are just as in the author's earlier book Applied Probability and Queues (Wiley 1987. because it appears to require an infinitely long simulation. in all other chapters than VI where we just write . where U(1) < . For example. < U(N) are the order statistics based upon N i. (or a functional of the expectation of a set of r. reference [14].3 (or just VI. this is a straightforward way to estimate finite horizon ruin probabilities. formula (5. The problem is entirely analogous to estimating steadystate characteristics by simulation in queueing/storage theory. the more typical situation is to perform a Monte Carlo experiment to estimate probabilities (or expectations or distributions) which are not analytically available.i. However. CONVENTIONS 19 as function of U(k). The infinite horizon case presents a difficulty. Still. and look at them to see whether they exhibit the expected behaviour or some surprises come up.v. respectively. and of course the method is relevant in risk theory as well..3).
f. A different type of asymptotics: less precise.v.v.f.f. If.s.g. b[s] is defined always if Rs < 0 and sometimes in a larger strip (for example. (moment generating function) fm e82B(dx) of the distribution B. then for 1s < 5). moment generating function.t. cumulative distribution function P(X < x) c. for a probability distribution IIGII = 1.The same symbol B is used for a probability measure B(dx) = P(X E dx) and its c. independent identically distributed i.r. squared coefficient of variation. . if B(x) . B(dy).i. or a more precise one like eh .o. . i.g.d. oo).e. n! 27r nn+1/2en. Abbreviations c. h + 0. right hand side (of equation) r.Used in asymptotic relations to indicate that the ratio between two expressions is 1 in the limit. B[s] the m.2.g.29) refer to the Appendix. see under b[s] below.g.f.f. and for a defective probability distribution IIGII < 1. In particular.p. log E[s] where b[s] is the m.c. random variable s. E expectation.d.d. B(x) = P(X < x) = fx.s. EX2/(EX)2. The Laplace transform is b[s].3) or Section 3. formula (5.4.h. w.20 CHAPTER L INTRODUCTION Proposition 4. with probability Mathematical notation P probability. cumulant generating function.ceax. B is concentrated on [0. E.g. i. say a heuristic approxi1 + h + h2/2. (A. with respect to w. r. B(x) the tail 1 . infinitely often l. n i oo.B(x) = P(X > x) of B. mation.h. References like Proposition A. as for typical claim size distributions. left hand side (of equation) m.f. IIGII the total mass (variation ) of a (signed ) measure G .
of numbers. matrices have uppercase Roman or Greek letters like T. (xi)diag denotes the diagonal matrix with the xi on the diagonal (xi)row denotes the row vector with the xi as components (xi). only have finitely many jumps in each finite interval. a. 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 dimension is usually clear from the context and left unspecified in the notation). i. CONVENTIONS {6B the mean EX = f xB(dx) of B ABA' the nth moment EXn = f x"B(dx) of B.e. E[X. i. the value just before t.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).A] means E[XI(A)]. Usually. all stochastic processes considered in this book are assumed to have sample paths in this space.. R(s) the real part of a complex number s. Unless otherwise stated. I(A) the indicator function of the event A. intensity interpretation. Notation like f3i and 3(t) in Chapter VI has a similar .. N(it. 0 marks the end of a proof. often the term 'cadlag' (continues a droite avec limites a gauche) is used for the Dproperty. row vectors have lowercase Greek letters like a.5. a2) the normal distribution with mean p and variance oa2. .e. the processes we consider are piecewise continuous. 7r. 21 D [0. In the Frenchinspired literature. A. Usually. Thus. In particular: I is the identity matrix e is the column vector with all entries equal to 1 ei is the ith unit column vector. and column vectors have lowercase Roman letters like t. . xa. Xt_ the left limit limstt X8f i. though slightly more complicated.e. F o r a given set x1. oo) the space of Rvalued functions which are rightcontionuous and have left limits. the ith unit row vector is e'i. . an example or a remark. the ith entry is 1 and all other 0. Matrices and vectors are denoted by bold letters.
EL the probability measure and its corresponding expectation corresponding to the exponential change of measure given by Lundberg conjugation. or quantities with a similar time average interpretation.5. interpretation. cf. cf. ry The adjustment coefficient. FL. 111. VI. Notation like BE and B(t) in Chapter VI has a similar. e. p the net amount /3pB of claims per unit time.g.5. though slightly more complicated.1. 'q the safety loading . I. I.1. cf. J the rate parameter of B for the exponential case B(x) = eby.22 CHAPTER L INTRODUCTION B the claim size distribution. .
More precisely. 5 on random walks and Markov additive processes can be skipped until reading Chapter VI on the Markovian environment model. 23 . a parallel selfcontained treatment is given of the facts needed there. in most cases via likelihood ratio arguments. Due to the generality of the theory. Sections 4. The duality results in Section 3 (and. somewhat more advanced than in the rest of the book. 5) are. however. strictly speaking. All results are proved elsewhere . The likelihood ratio approach in Section 2 is basic for most of the models under study. however. in particular at a first reading of the book. 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. fundamental ( at least in the author' s opinion) and the probability involved is rather simple and intuitive. the relevance for the mainstream of exposition is the following: The martingale approach in Section 1 is essentially only used here. Sections 4. When encountered for the first time in connection with the compound Poisson model in Chapter III. not crucial for the rest of the book. used in Chapter VI on risk processes in a Markovian (or periodic) environment. the level of the exposition is. The general theory is. in part. The topic is. The reader should therefore observe that it is possible to skip most of the chapter.
{e'YS° }t>0 is a martingale. The more general Theorem 6.)AT = E [e7ST(°).1 Assume that (a) for some ry > 0. Our first result is a representation formula for O(u) obtained by using the martingale optional stopping theorem . and the ruin probabilities are ip(u) = P (T(u) < oo). claim size distribution B and p = . f1 . T(u) < oo] + 0 = eryuE [e7Vu). 1 Martingales We consider the claim surplus process {St} of a general risk process. Thus N. We get 1 = Ee7So = E e'Y S(. (b) St a$ oo on {T(u) = oo}.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/.. the second term converges to 0 by (b) and dominated convergence (e7ST < eryu on {r(u) > T}).(u).QµB < 1. T) = P(T(u) < T).2) As T > oo. As usual. and in the limit (1. T(u) > T] .2 Consider the compound Poisson model with Poisson arrival rate .0. T(u) < T] + E [eryST .u denote the overshoot. Proposition 1. Then e7u (u) = E[e74(u)j7(u) < oo] Proof We shall use optional stopping at time r(u)AT (we cannot use the stopping time T(u) directly because P(T(u) = oo) > 0 and also because the conditions of the optional stopping time theorem present a problem. (1. V) (u. Let e(u) = ST(u) . Example 1 . using r(u) A T invokes no problems because r(u) A T is bounded by T).5 can be skipped. SOME GENERAL TOOLS AND RESULTS The ladder height formula in Theorem 6.2) takes the form 1 = E [e'ys().24 CHAPTER II. StUit. the time to ruin r(u) is inf It > 0 : St > u}. however.
Under the conditions of Proposi Proof Just note that C(u) > 0. it follows that E [e7st+v I J] = e"rstE [e7(st+vSt) I Ft] = e7StEe"rs° = elst where .i. and p =. B(x) _ edx. From this it is readily seen (see III. Thus. the martingale property now follows just as in Example 1. The available information on this jump is that the distribution given r(u) = t and S.= e"(') where K(a) = .u + x is again just exponential with rate S. and thus by the memoryless property of the exponential distribution . u Corollary 1.Q and the U.Q. with common distribution B (and independent of {Nt}).1) .5 For the compound Poisson model with B exponential. of the normal distribution.1) shows that Eels.3 Assume that {Rt} is Brownian motion with variance constant o.3/6 < 1. O(u ) < e7".r" where y = S . Thus.1.1) . From this it is immediately seen that the solution to the Lundberg equation ic(y) = 0 is y = 2p/a2.. Example 1 . the conditional distribution of the overshoot e(u) = U . are i. Since {St} has stationary independent increments.6a for details) that typically a solution to the Lundberg equation K(y) = 0 exists.2.f.g. Thus 00 E [e'rt (") I T(u) < oo] = I e5e  dx = f 5edx .ap.a it is immediately seen that y = S .4 (LUNDBERG ' S INEQUALITY ) tion 1 . condition (a) of Proposition 1.Q(B[a] . the conditions of Proposition 1. and thus Ee7s° = 1.x. the ruin probability is O(u) = pe.1 are satisfied.(„)_ = x is that of a claim size U given U > u . and thus Ee'rs° = 1.a = a . By standard formulas for the m. and (b) follows from p < 1 and the law of large numbers (see Proposition III. Now at the time r(u) of ruin {St} upcrosses level u by making a jump .Ft = a(S" : v < t). Then {St } is Brownian motion with variance constant o2 and drift p < 0.a.1 is satisfied.2 and drift p > 0. Since {St} has stationary independent increments./3. Proof Since c(a) = /3 (B[a] .1. A simple calculation (see Proposition III. u Corollary 1. 1. Eeas° = e"(°) where n(a) = a2a2/2 .1.d. MARTINGALES 25 where {Nt} is a Poisson process with rate .2(c)).
But if a $ ^ .1) 'though not always: it is not difficult to construct a counterexample say in terms of transient Markov processes. A].Ft}too and the Borel afield F. on (DE.6 If {Rt} is Brownian motion with variance constant a2 and drift p > 0. Proof Just note that ^(u) = 0 by continuity of Brownian motion. I. the parameters of the two processes can be reconstructed from a single infinite path.6 N S = { lim Nt I t +00 t gJ are both in F.e. [172]. Two such processes may be represented by probability measures F. A somewhat similar u argument gives singularity when B $ B.26 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Corollary 1. F(S) = P(S) = 1. Delbaen & Haezendonck [103] and Schmidli [320].. and in analogy with the theory of measures on finite dimensional spaces one could study conditions for the RadonNikodym derivative dP/dP to exist. we look for a process {Lt} (the likelihood ratio process) such that P(A) = E[Lt. and is further exploited in his book [157]. Theorem 2. then z/'(u) = e7" where 'y = 21A/a2. Thus the sets S = I tlim +oot t =. P on (DE. 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. u Notes and references The first use of martingales in risk theory is due to Gerber [156]. Example 2 . oo).F). The interesting concept is therefore to look for absolute continuity only on finite time intervals (possibly random. then S and S are disjoint . . P correspond to the claim surplus process of two compound Poisson risk processes with Poisson rates /3. P are then singular (concentrated on two disjoint measurable sets). and by the law of large numbers for the Poisson process . Embrechts. F). which we equip with the natural filtration {. hence so is Nt = limfyo N2`i. and F. 0 and claim size distributions B.v. More recent references are Dassios & Embrechts [98]. (2. However. B. Grandell & Schmidli [131]. cf. A E Ft. The number Nt F) of jumps > e before time t is a (measurable) r. Grandell [171].3 below). as shown by the following example this setup is too restrictive: typically'.1 Let F.
3 Let {Lt}. . Proposition 2.2) Proof Assume first G C {T < T} for some fixed deterministic T < oo. Then Ft (A) = E[Lt. If r is a stopping time and G E PT. A].F). F) such that ELt = 1. if for some probability measure P and some {. 1 J (2. Hence E [_ . LIKELIHOOD RATIOS AND CHANGE OF MEASURE 27 (i.t. P) such that LLt = 1. Then P(A) = E[Lt. A E Ft . _. then { 1 P(G) = EG .F such that (2.2.1) holds. ELt = 1 follows by taking A = DE in (2. implies that E[LtI. G ] = E [LT . Hence the family {Pt} is t>o consistent and hence extendable to a probability measure F on (DE. P be as in Proposition 2. A]. Ft).2(i). A E F.r. Finally.t. that the restriction of P to (DE. ({. A E F8. we have E [ LTIFT]1 = LT on {T < T}.Tt) is absolutely continuous w.1) and nonnegativity by letting A = {Lt < 0}. then {Lt} is a nonnegative martingale w. This proves (i). .e.1) holds. the restriction of P to (DE. Lt < 0] can only be nonnegative if P(A) = 0. Then Lt > 0 and ELt = 1 ensure that Pt is a probability measure on (DE.Ft}.t. u The following likelihood ratio identity (typically with r being the time r(u) to ruin) is a fundamental tool throughout the book: Theorem 2 . Proof Under the assumptions of (i). By the martingale property. using the martingale property in the fourth step. A] = E[Lt. then there exists a unique probability measure Pon .Pt)) The following result gives the connection to martingales.2 Let {Ft}t>o be the natural filtration on DE.A] = EE[LtI(A)IF8] = EI(A)E[LtIFB] = EI(A)L8 = PS(A). Lets < t..Pt}adapted process {Lt}t>o (2. The truth of this for all A E Y. F the Borel o•field and P a given probability measure on (DE. under the assumptions of (ii) we have for A E rg and s < t that A E Ft as well and hence E[L8.Y) such that P(A) = Pt(A).F8] = L8 and the martingale property. G C {T < oo}. G l ] = E [_I(G)E[LTIFT] ] = E { _I(G)Lr ] = P(G). define P by Pt (A) = E[Lt. Conversely. (ii) Conversely.r. ({Ft} . (i) If {Lt}t> o is a nonnegative martingale w.r.
we assume for simplicity that {Xt} has Dpaths. is Markov w.3) to G of{r < T} we get 1111 F(Gn {r <T}) = E[ 1 .2) by noting that 1 = ersr(„) = e1'ue7Ou). First we ask when the Markov property is preserved. r(u) < oo] occuring there than with the (conditional) expectation E[e'r{(u ) Jr(u) < oo] in (1. one would typically have Xt = Rt.28 CHAPTER II. we need the concept of a multiplicative functional. is nonnegative and has Ey Lt = 1 for all x. both sides are increasing in T. 5) for processes with some randomwalklike structure. 1 Since everything is non negative. Xt = St. The problem is thus to investigate which characteristics of {Xt} and {Lt} ensure a given set of properties of the changed probability measure..1) is that it seems in general easier to deal with the (unconditional) expectation E[eryVu). of (2.r. In the context of ruin probabilities. first in the Markov case and next (Sections 4.1). each F..t.O(u) = eryuE[e'YC(u).t.Gn {r <T} . T(u) < oo]. we have F(G) = V )(u).Rt} the claim surplus process and {Jt} a process of supplementary variables possibly needed to make the process Markovian. Now just rewrite the r.1: Corollary 2. SOME GENERAL TOOLS AND RESULTS In the general case .r.h. t.4 Under condition (a) of Proposition 1.4) compared to (1. and letting T * oo.2) follows by monotone convergence. (2. To this end. {St} = {u . Rt) or Xt = (Jt.3 we obtain a likelihood ratio representation of the ruin probability V) (u) parallel to the martingale representation (1.Ft} . u From Theorem 2. Xt = (Jt. say. applying (2.1) in Proposition 1. .s. St). the natural filtration {. Consider a (timehomogeneous) Markov process {Xt} with state space E.4) Proof Letting G = {r(u) < oo}. where {Rt} is the risk reserve process. Lr(u) 11 The advantage of (2. and this problem will now be studied. (2. in continuous time (the discrete time case is parallel but slightly simpler). The crucial step is to obtain information on the process evolving according to F. For the definition. A change of measure is performed by finding a process {Lt} which is a martingale w.1.
The converse follows since the class of r. for all x.s. Proof Since both sides of (2.Y8f t < s. s. t] * [0. 0 .F8measurable r. nonnegative and Lt+8 = Lt•(Lso9t) (2. the Markov property can be written E.(A) = Ex [Lt. (2.Ft }.[Y.v. (2. since Ext [L8Y8] = E[(Y8 o et)(L8 o 8t)I.5 Let {Xt} be Markov w.Ft+8measurable.v.v.v.7) for any Ftmeasurable Zt and any . and then L. Vt+e. Zt. The precise meaning of this is the following: being . oo). o 9t = V.'s of the form Zt • (Y8 o 0t) comprises all r. Indeed.2.T9measurable Y8. (2.YB] for any Ftmeasurable r. which is the same as Ex[Zt(Y8 o Bt)] = E8[ZtEx.. (2. Similarly.v. where Ot is the shift operator. A]. Ex[Lt+8Zt(Y8 o et)] = Ex[LtZt(Y8 o 0t)(L8 o Bt)].. t. Lt has the form Lt = 'Pt ({x }0<u<t) for some mapping cot : DE[O. let {Lt} be a nonnegative martingale with Ex Lt = 1 for all x. Then the family {Px}xEE defines a timehomogeneous Markov process if and only if {Lt} is a multiplicative functional. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 29 on DE and define {Lt} to be a multiplicative functional if {Lt} is adapted to {. t and let Px be the probability measure given by t. the natural filtration {Ft} on DE. ({Xt+u} 0<u<8) Theorem 2.7).6) implies (2.Ft]. Y8.7).5) is equivalent to Ex[Lt+8Vt+8] = E8[Lt • (L8 o 91)Vt+8] for any .t.r.Pt+8measurable r.8) which is the same as (2.Ft] = Ex. since Zt • (Y8 o Ot ) is .5) Pxa.'s of the form fl' f.Ftmeasurable.5) are Tt+e measurable.6) for any . this in turn means Ex[Lt+8Zt(V8 oet)] = Ex[LtZtExt[L8Y8]]. which in turn is the same as Ex[Lt+8Zt • (V8 o Bt)] = Ex[Lt • (L8 o 91)Z1 • (Y8 o et)] (2. o 9tI. or.(Xtitl) with all t(i) < t + s. By definition of Px.
The corresponding claim sizes are Ul.. u Notes and references The results of the present section are essentially known in a very general Markov process formulation. .. The storage process {Vt }o<t<T is essentially defined by time reversion. Thus R = Ro + f p(R8) ds . SOME GENERAL TOOLS AND RESULTS to define a timehomogeneous Markov process.. . it xEE suffices to assume that {Lt} is a multiplicative functional with Ex Lt = 1 for all x. the premium rate is p(r) when the reserve is r (i.. Ro = u (say). T] in the following setup: The risk process {Rt}o<t<T has arrivals at epochs or.1) The initial condition is arbitrary. More precisely . Indeed. 3 Duality with other applied probability models In this section. In between jumps. see Dynkin [128] and Kunita [239].e. 0 < vl < . (using the Markov property in the second step) so that the martingale property is automatic. In between jumps. } E[Lt+B I.. aN where or* = T UN_k+l. .6 For {u .Ft] = LtE[L8 o 9t I. t] = LtExt L8 = Lt. the arrival epochs are Qi. CN..30 CHAPTER H. t. .5 can be found in Kuchler & Sorensen [240]. {Vt} . we shall establish a general connection between ruin probabilities and certain stochastic processes which occurs for example as models for queueing and storage. UN. The formulation has applications to virtually all the risk models studied in this book. and the time to ruin is 7(u) = inf {t > 0: Rt < 0}. (3. then Remark 2. and thus for the moment no parametric assumptions (on say the structure of the arrival process) are needed.. with a proof somewhat different from the present one. < aN < T... A more elementary version along the lines of Theorem 2. The result is a sample path relation.. and just after time or* {Vt} makes an upwards jump of size UU = UN _k+l.. We work on a finite time interval [0. reflection at zero and initiar condition Vo = 0. . R = p(R)).At where At = k: vk <t U.
.2) k: ok <t and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0.T) = P(VT > u). instead of (3.__. (3.x.11 4. Theorem 3. That is.1) we have Vt = At  f P(Vs)ds where A= U= AT .T) = inf Rt < 0 P (O<t<T P(r(u) < T) be the ruin probability. {Vt} remains at 0 until the next arrival).. The sample path relation between these two processes is illustrated in Fig.3) (u) Proof Let rt' denote the solution of R = p(R) subject to r0 = u..1 The events {T(u) < T} and {VT > u} coincide. In particular.._.. 3.____•_.__.1..1 Define r(u) = inf It > 0: Rt < 0} (r(u) = oo if Rt > 0 for all t < T) and let ii(u.3.. :._: 1} 0 011 =T01N ^N3 To 0 011 014 01N Figure 3. .. (3.___ . V)(u. Note that these definitions make {Rt} rightcontinuous (as standard) and {Vt} leftcontinuous.e. V = p(V)).. DUALITY WITH OTHER APPLIED PROBABILITY MODELS 31 decreases at rate p(r) when Vt = r (i. Then rt°) > rt°) for all t when u > v.AT_t.
The arrival epochs correspond to rainfalls. u A basic example is when {Rt} is the risk reserve process corresponding to claims arriving at Poisson rate . Theorem 3. if nothing else n = N). The results can be viewed as special cases of Siegmund duality. and in between rainfalls water is released at rate p(r) when Vt (the content) is r.i.1 with Ro = u = u2). Then Vo.d. = r(VT) . this represents a model for storage. see Siegmund [344]. and so on.T l . If VaN > 0. say of water in a dam though other interpretations like the amount of goods stored are also possible.and left continuity is immaterial because the probability of a Poisson arrival at any fixed time t is zero). Nevertheless.3 and being i. Then the time reversibility of the Poisson process ensures that {At } and {At } have the same distribution (for finitedimensional distributions. . Thus we may think of {Vt} as having compound Poisson input and being defined for all t < oo. Historically. We get: Corollary 3. Then the storage process {Vt} has a proper limit in distribution. the distinction between right.1 and its proof is from Asmussen & Schock Petersen [50]. 3. Hence if n satisfies VVN_n+1 = 0 (such a n exists. the connection between risk theory and other applied probability areas appears first to have been noted by Prabhu [293] in a queueing context. Suppose next VT < u (this situation corresponds to the broken path of {Rt} in Fig.1 with Ro = u = ul).32 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Suppose first VT > u (this situation corresponds to the solid path of {Rt} in Fig. with distribution B. Hence RQ„ > 0 for all n < N.U1 > roil . u Notes and references Some main reference on storage processes are Harrison & Resnick [187] and Brockwell. we have RQ„ < 0 so that indeed r(u) < T. and a general premium rule p(r) when the reserve is r. Then similarly VVN = r0. Corollary 3. Proof Let T ^ oo in (3.Ul < roil  Ul = RQ„ Va1V_1 < RQ2. Some further relevant more general references are Asmussen [21] and Asmussen & Sigman [51].2 from Harrison & Resnick [188].2 Consider the compound Poisson risk model with a general premium rule p(r). say V. Resnick & Tweedie [79].3). 3. we have r(u) > T. Historically. and then '0 (u) = P(V > u).U1 = Rol. we can repeat the argument and get VoN_1 > Ra2 and so on. one may feel that the interaction between the different areas has been surprisingly limited even up to today. if and only if O(u) < 1 for all u. and since ruin can only occur at the times of claims.
evolves as a random walk with increments Z1i Z2.. if Wo = 0 then (Z1+•••+Zn) WN = Zl+•••+ZN. Theorem 4. Z2 .1)). can be viewed as the reflected version of the random walk with increments Z1. with common distribution F (say). = Xo + Y1 + • • • + Y..2 The following assertions are equivalent: (a) 0(u) = P(r(u) < oo) < 1 for all u > 0. ZN = . N min (Y1 + + Yn). N min (Y1 + • • • + YNn) n=0.. In particular. ..d. (b) 1/i(u) = P(•r(u) < oo) > 0 as u * oo.. Xo = 0. W1.. Z2 = YN_1 i .. Z2..s. .1.. the Lindley process Wo..N (4...1 in terms of Lindley processes .Y1 according to Wo = 0... i.. Most often. WN be the Lindley process generated by Z1 = YN.YNn+1) n=0.. . where the Yi are i .d... .. and is reset to 0 once the r. W2. 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.. there is an analogue of Theorem 3..Yl min (YN . Let further N be fixed and let Wo..... Then the events {r(u) < N} and {WN > u} coincide. 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. For discrete time random walks . 0 Corollary 4.e..min n=0.2).1. 0). generated by Z1..2) satisfies the same recursion as in (4.1... {Wn}n=0. just verify that the r . .2) (for a rigorous proof.. Z2 = Y2. .. hits (oo. Z2.. 1} is often referred to as simple random walk or Bernoulli random walk). Here F is a general probability distribution on R (the special case of F being concentrated on {1.1) Thus {Wn}n=o. Proof By (4. R valued sequence Z1. WN = YN ..4... . of (4. has a proper limit W in distribution as n + oo.i.1. For a given i....1 Let r(u) = inf In: u + Y1 + • • • + Yn < 0}. ...1...1. .. n=0..h. (c) The Lindley process {WN} generated by Zl = Y1.N From this the result immediately follows...w. W1.
. either M = oo a... Px to Fn are equivalent (have the same null sets) so that the likelihood ratio Ln exists. or M < oo a.. In that case . a Markovian change of measure as in Theorem 2..1 is actually not necessary . Remark 4 . .. . One then assumes Yn to be a stationary sequence.. (e). W v m and P(W > u) = P (m > u) = 0(u). .5 does not necessarily lead to a random walk: if.l. the Y1. For a random walk. Proof Since (YN.1.. N. 176) but appears to be rather intractable. In general. there is a more general version of Corollary 4..1..ooa..s.2 and Theorem 4. the condition 00 F(YI+•••+ Yn<0)<00 n=1 is known to be necessary and sufficient ((APQ] p.... doubly u infinite (n'= 0.3 The i. ±2.. . YN). SOME GENERAL TOOLS AND RESULTS (d) m = inf. F has a strictly positive density and the Px corresponds to a Markov chain such that the density of X1 given Xo = x is also strictly positive.s. e. Combining these facts gives easily the equivalence of (a)(d). assumption on the Z1. Next consider change of measure via likelihood ratios. ZN or.i..) sup n=0.d.l...s. (e)Yi+•••+Yn 74 .1. . (Yi + • • • + Yn) > oo a.1.. w. 0 By the law of large numbers. The converse follows from general random walk theory since it is standard that lim sup (Y1 + • • + Yn) = oo when Y1 + • • • + Yn 74 oo. then the restrictions of Fx.2.s .1 have the same distribution for n = 0. equivalently. Y1) has the same distribution as (Y1. + Z. .=o..g.o. (d) #.. Clearly. . By Kolmogorov's 01 law.... The following result gives the necessary and sufficient condition for {Ln} to define a new random walk: . . YN in Theorem 4. Similarly... Thus the assertion of Theorem 4.the result is a sample path relation as is Theorem 3. the Lindley processes in Corollary 4. ±1.g.1 is equivalent to WN D MN = (Z1 + .N so that WN _P4 M = supra=0..) and defines Zn = Yn. . (Z1 + • • • + Zn) = m and P(W > u) = P(M > u) = i (u ). a sufficient condition for (e) is that EY is welldefined and > 0. .34 CHAPTER II.
(4. . Y ) f (Y)] = E[g(O.g. = 1 for all n and x.4.5 Consider a random walk and an a such that c(a) = log F[a] = log Ee° ' is finite.4) ({Ln} is the familiar Wald martingale . cf. Since L1 has the form g (Xo. Proof If (4. this means E(g(x.. Breiman [78] p. 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. In that case. the random walk property implies Ex f (Y1) = Eo f (Y1 ).g. Then the change of measure in Theorem 2. implying g (x. (a) = log F(a] is the c. the changed increment distribution is F(x ) = E[h(Y). Then the change of measure in Theorem 2.. and define Ln by (4. and so onforn =3. we get L2 = L1 (L1 o91 ) = h(Y1)g(X1. Y1). y ). We get: Corollary 4. h(Yn) (4. RANDOM WALKS INDISCRETE OR CONTINUOUS TIME 35 Proposition 4.3) 1Pxa..4. of F). In particular. where h (y) = g(0..5 corresponds to a new random walk if and only if Ln = h(Y1) . For n = 2.4 Let {Ln} be a multiplicative functional of a random walk with E_L.3) holds. Conversely.4).s...nrc(a )} (4.3) holds for n = 1.5 corresponds to a new random walk with changed increment distribution F(x) = e'(a) Jr e"'F(dy) . Y < x]. e. 100 ).. Y) f (Y)] for all f and x.s.f. Y) = h(Y ) a. The corresponding likelihood ratio is Ln = exp {a (Y1 + • • • + Yn) . u A particular important example is exponential change of measure (h(y) = e°y'(") where r. for some function h with Eh(Y) = 1.Y2) = h(1'i)h(I'a).
. {Xt} can be written as the independent sum of a pure drift {pt}. say by recording the reserve or claim surplus just before or just after claims (see Chapter V for some fundamental examples). The traditional formal definition is that {Xt} is Rvalued with the increments Xt(1)_t(o)..36 CHAPTER II. SOME GENERAL TOOLS AND RESULTS Discrete time random walks have classical applications in queueing theory via the Lindley process representation of the waiting time . the claim surplus process for the compound Poisson risk model .1). or imbedded into continuous time processes . which corresponds to the compound Poisson case: here jumps of {Mt} occur at rate 0 and have distribution B = v/0 (in particular . but we omit the details ). corresponds to a process with stationary independent increments and u = p. In continuous time. v2 = 0 and v = 3B). The simplest case is 3 = JhvMM < oo. this description needs some amendments.6) More precisely.7) for all e > 0. see Chapter V. Roughly.5) Note that the structure of such a process admits a complete description.e. Xt(2)_t(l). However. In risk theory. . (4. they arise as models for the reserve or claim surplus at a discrete sequence of instants.Xt)I. a Brownian component {Bt} (scaled by a variance constant) and a pure jump process {Mt}. a positive measure on R with the properties e J x2v(dx) < oo. the interpretation is that the rate of a jump of size x is v(dx) (if f of Ixlv (dx) = oo. given by a the increment distribution F(x) = P(Xn+l ..t(i . (4. we are . The appropriate generalization of random walks to continuous time is processes with stationary independent increments (Levy processes).Ft] = Eof (X. . An equivalent characterisation is {Xt} being Markov with state space R and E [f (Xt+e . e f x:IxJ>e} v(dx) < oo (4. say the beginning of each month or year . < t(n) and with Xt( i)_t(i_l) having distribution depending only on t(i) . however.. Xt (n)t(n1) being independent whenever t(O) < t(1 ) < . the tradition in the area is to use continuous time models. 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. In discrete time. with premium rate p.). {Xt} is a random walk.Xn < x). i. the pure jump process is given by its Levy measure v(dx). Xt =Xo+pt+oBt+Mt.
2). V E [0. A different interpretation is as the workload or virtual waiting time process in an M/G/1 queue.7 If {Xt} has stationary independent increments as in (4.v. 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).1. Corollary 4.10) . O(u.s. (ex . Now consider reflected versions of processes with stationary independent increments. VT + V for some r.] Processes with a more complicated path structure like Brownian motion or jump processes with unbounded Levy measure are not covered by Section 3. Furthermore. WT = XT .4. Proposition 4.6 In the compound Poisson risk model with constant premium rate p(r) . defined as a system with a single server working at a unit rate.min Xt (4. oo]. and b(u) = P(V > u). 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.3 and decreases linearly at rate 1 in between jumps. having Poisson arrivals with rate . Then the storage process {Vt} has constant release rate 1. then Ee'(xtxo) = Eoeaxt = etx(a). has upwards jumps governed by B at the epochs of a Poisson process with rate .6).e. and the reflected version is then defined by means of the abstract reflection operator as in (4. where c(a) = ap + a2a2/2 + f 00 provided the Levy measure of the jump part {Mt} satisfies f".1)v(dx) (4. Here workload refers to the fact that we can interpret Vt as the amount of time the server will have to work until the system is empty provided no new customers arrive. cf.8) O<t<T (assuming Wo = Xo = 0 for simplicity).Q and distribution B of the service times of the arriving customers. [The condition for V < oo a. 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. 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. T) = P(VT > u). Chapter III. jxJ v(dx) < oo. i. is easily seen to be f3pB < 1.
let e" (a ) = Eoeaxl.Xt)L8 o 0tIFt] = E [f (Xt+s .f. 8 Assume that {Xt} has stationary independent increments and that {Lt} is a nonnegative multiplicative functional of the form Lt = g(t. 5 has stationary independent increments as well.11) (eax . Xt +B .1. we show in the compound Poisson case ( IlvIl < oo) in Proposition III.Xt)g(s. e. By explicit calculation . v(dx) = e9xv (dx). Theorem 4 ..1 ) v(dx) . we use the characterization (4. then the changed parameters in the representation (4. Then the Markov process given by Theorem 2. Proof For the first statement . if Lt = e9(xt .g.38 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Proof By standard formulas for the normal distribution. In particular.10) is the LevyKhinchine representation of the c.5) and get E [f(Xt+B . of an infinitely divisible distribution (see.g. Chung [86]).xo)tk ( e). This is of course no coincidence since the distribution of Xl . 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) .Xt)I'Ftl = E [f(Xt+B . use the representation as limit of compound Poisson processes. t. X8) = Eof (X8)L8 = Eof (X8)• For the second. u Note that (4.Xo is necessarily infinitely divisible when {Xt} has stationary independent increments. Xt Xo) with E2Lt = 1 for all x.4 o) aµ + ((a + 0 ) 2  0 2 )o 2 /2+ r w J 00 (e (a + 9)x  a 9x )v(dx) 00 a(µ + O 2) + a2a2 / 2 + J (eax . Eea(µt + QBt) = et{aµ +a2OZ/2}.Xt)I Ftl = Eof (X8)g(s.6) are µ = µ + Oo2 .1 that E eaMt = exp fmoo In the general case . .1)eexv(dx). Q2 = v2. (4.
(5. a = 0. are the arrival times and U1. the structure of MAP's is completely understood when E is finite: 2and only there . corresponding to p = 1. U2.8. b with b(x) > 0 for all x such that b(x) > 0).2.0. abbreviated as MAP in this section2. the corresponding claim sizes . Thus (since µ = p = 1. Then we can write v(dx) _ /3eOxB(dx) = / (dx)..3 =. one reason is that in parts of the applied probability literature. . .5. Example 4 . b = a = 0) the changed process is the claim surplus process of another compound Poisson risk process with Poisson rate . 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 ..(3B(dx). Recalling that U1. MAP stands for the Markovian arrival process discussed below. it is then easily seen that Lt = H dB(Ui) i:o.(3 = . is defined as a bivariate Markov process {Xt} = {(Jt. <t whenever the RadonNikodym derivative dB/dB exists (e.11 For an example of a likelihood ratio not covered by Theorem 4. u 5 Markov additive processes A Markov additive processes.3B[B]. v(dx) _ ..l3 and claim u size distribution B. dB/dB = b/b when B.10 Let Xt be the claim surplus process of a compound Poisson risk process with Poisson rate . B have densities b.4).3 and claim size distribution B # B.g. MARKOV ADDITIVE PROCESSES 39 Remark 4. B(dx) = B[9] B(dx).0 in the following. then the martingale {eex(t)tk(e)} is the continuous u time analogue of the Wald martingale (4. and let the Px refer to the claim surplus process of another compound Poisson risk process with Poisson rate.o[f (S8)g(J8)]. we write Pi. Ei instead of P2. As for processes with stationary independent increments .9 If X0 = 0..Ft] = Ejt. Ei.1) For shorthand .St)g(Jt+s)I. where . Example 4 . 0.3 and claim size distribution B.. 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.
the converse requires a proof.[a) = (Ei[easl.9 EE = (iii&ij[a])i j EE .jEE• On an interval [t. the distribution of which has some distribution Bij. SOME GENERAL TOOLS AND RESULTS In discrete time. {Jt} is specified by its intensity matrix A = (Aij)i.1 For a MAP in discrete time and with E finite. v.jEE (here pij = Pi(J1 = j)) and the probability measures Hij(dx)=P(Y1 EdxlJo=i. 1 J1 ='^])iJEE = (Fij[a])i . If all Fij are concentrated on (0.6) depending on i. with the Y„ being interpreted as interarrival times. An alternative description is in terms of the transition matrix P = (piA.o(Ji = j. (That a process with this description is a MAP is obvious.. Y1 E dx) where Y„ = S„ . As an example. Proposition 5. vi(dx) in (4. {St} evolves like a process with stationary independent increments and the parameters pi. consider the matrix Ft [a] with ijth element least Ei .i. As a generalization of the m.) If E is infinite a MAP may be much more complicated. Fn[a] = F[a]n where P[a] = P .g. In continuous time (assuming Dpaths). Y2... which we omit and refer to Neveu [272] or cinlar [87]. a MAP is specified by the measurevalued matrix (kernel) F(dx) whose ijth element is the defective probability distribution Fij(dx) = Pi.40 CHAPTER H.Sr_1.f. Then a Markov additive process can be defined by letting t St = lim 1 I(IJB1 < e)ds E1o 2d o be the local time at 0 up to time t. let {Jt} be standard Brownian motion on the line. oo). In addition. a jump of {Jt} from i to j # i has probability qij of giving rise to a jump of {St} at the same time.J1=j)= Fij (dx) Pij In simulation language. a MAP is the same as a semiMarkov or Markov renewal process.. t+s) where Jt . Jn = j.it = A. . this means that the MAP can be simulated by first simulating the Markov chain {J„} and next the Y1.. by generating Yn according to Hij when J„_1 = i.
u Proposition 5. Jt = k] { 1 .1)) . Sn ) yields Ei[easn+ '. MARKOV ADDITIVE PROCESSES Proof Conditioning upon (Jn. up to o( h) terms. aSt h = (1 + Ajjh) Ei [east . Jt = k] { xk kEE j la] . Jt = j] is given by etK[a]. assume that the Markov chain/process {Jt} is ergodic. we infer that in the discrete time case the . pi. which in conjunction with Fo[a] = I implies Ft[a] = etK[a) according to the standard solution formula for systems of linear differential equations.5.1 )v(dx). Then the matrix Pt[a] with ijth element Ei [east. vi(dx). where K[a] = A+ (r.2 Let E be finite and consider a continuous time Markov additive process with parameters A.qkj + k?^j qkj Bkj [a] } = Ei [east. Jt = j] (1 + htc (j) (a)) j + Ak j qk j (Bk +h E Ei [east .(')(a)) diag + (). u In the following. J1 = A which in matrix formulation is the same as Fn+1 [a] = Fn[a]F[a]. vi(dx) (i E E). this means that F't+h [a] = Ft[a] II+h(rc(i)(a)) +hA+h(Aijgij(Bij[a]1)) I. 00 r(i) (a) = api + a2ot /2 + f (e° . Jt = j] Ejesh'^ + E Ak j hEi [ease . a= . kEE Jn = k]Ek[e"Y" . Bij (i. Then. 013 . qij. Proof Let {Stt) } be a process with stationary independent increments and pa rameters pi .ijgij(Bij[a] . j E E) and So = 0. In matrix formulation .4c). By PerronFrobenius theory (see A.1) } (recall that qjj = 0). \ diag Ft[a] = Ft[a]K. Jn+1 = A] = 41 Ei[ e 5„.
h(") may be chosen with strictly positive components. Then h(°) = e. 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. Eie"sth^a) = e'Pt[a]h( a) = e. of a random walk.h(a)vva)etw(a). h(") are only given up to a constants. Furthermore. cf.e=e°tk. Corollary 5. we are free to impose two normalizations. and appropriate generalizations of the Wald martingale (and the associated change of measure) can be defined in terms of . just note that [a]h(a) = eietK (a)h(a) = etK(a)h(a). Corollary 5.f. The corresponding left and right eigenvectors v("). as will be seen from the following results.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).4 Eie"sth(a) = h=a)et?("). Corollary 5. We also get an analogue of the Wald martingale for random walks: Proposition 5. Jeast.Eikjt = ttc'(0) + ki .tK(a)h(a) J jj it L o is a martingale. In particular. The function ic(a) plays in many respects the same role as the cumulant g.42 CHAPTER II.etx It then follows that E feast+^(t+v)K(a)h(a) I ^tl l . Since v("). and write k = k(°). Yrh(a ) = 1. .r. a. Jt = j] . Proposition 5.2) where 7r = v(°) is the stationary distribution.3 Ei [east. Proof For the first assertion. (5.4c).7.c(a) (and h(")).4. its derivatives are 'asymptotic cumulants'. u Let k(a) denote the derivative of h() w. cf.t. Proof By PerronFrobenius theory (see A. and we shall take V(a)h(a) = 1.5 EiSt = tK'(0) + ki .
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.5. u The argument is slightly heuristic (e.3) Let a = 0 and recall that h(°) = e so that 0=°) = h(o) = 1. For the second . ] = t2tc (0)2 + 2tK'(0)vk + ttc"(0) + O(1) . we differentiate (5.4. In the same way. t im v^"St = '(0) Proof The first assertion is immediate by dividing by tin Corollary 5. one obtains a generalization of Wald's identity EST = Er • ES.. tam E tSt a (0). E=ST = tc'(0)E7. Since it is easily seen by an asymptotic independence argument that E„ [Stkjt] u = trc'(0) E„kjt + O(1).e. . summing and letting a = 0 yields E„ [St + 2Stkj.. Remark 5 . subtraction yields Vary St = tic"(0) + O(1).g.") }) . Squaring in Corollary 5. 43 Ei [Steast h(a) + east k^a)1 = et"(a) (kia) + tic (a)hia)) .5 yields + W (a)k.Eikjr .4) .5. [E. t a oo. the distribution of Jo). MARKOV ADDITIVE PROCESSES Proof By differentiation in Proposition 5. 8 Also for E being infinite (possibly uncountable ). Ee"st typically grows asymptotically exponential with a rate ic(a) independent of the initial condition (i. More precisely.+ k. (5.St]2 = t2/c'(0 ) 2 + 2ttc (0)vk . Corollary 5.7 No matter the initial distribution v of Jo. for a random walk: Corollary 5.a) + ttc (a)2hia ) Multiplying by v=. there is typically a function h = h(") on E and a ^c(a) such that Ey a"st t" (") * h(x).6 For any stopping time T with finite mean.2ttc (0)Evkjt + 0(i). the existence of exponential moments is assumed ) but can be made rigorous by passing to characteristic functions. (5.
5 defines a new MAP. xEE . h(Jo) Lo is a Px martingale for each x E E. An example beyond the finite case occurs for periodic risk processes in VI.f (x) tyo t provided the limit exists. however.. Usually. let ha(i. From (5. 0) = n(a) h(i).5.5) is a martingale can be expressed via the infinitesimal generator G of {Xt } = { (Jt. G is defined as Gf (x) = lim Exf (Xt) .5) is a martingale . s) = ea8h(i). we take the martingale property as our basic condition below (though this is automatic in the finite case). however. Then {Lt } is a multiplicative functional. In view of this discussion . this leads to h(i) + tcha( i. Remark 5. We then want to determine h and x(a) such that Ejeasth (Jt) = etK(a)h(i). V.4 that { h(Jt) easttK(a) L o (5. First. Given a function h on E.6) We shall not exploit this approach systematically.9 The condition that (5.e. gha(i.for the present purposes. some extra conditions are imposed. u forsEE). see.. 1) (i.3b and Remark VI. For t small .6.1) one then ( at least heuristically) obtains lim Ex eaSv v a) K( v+oo nEx easttK(a)EJt east(vt)K(a) u[J = Ex easttk(a)h(Jt) It then follows as in the proof of Proposition 5. this is.s. St) } as follows. 0 Proposition 5.(9) {Lt}t>o = .6. St)} be a MAP and let 0 be such that h(Jt) OStt. where {Jt} is deterministic period motion on E = [0. inconvenient due to the unboundedness of ea8 so we shall not aim for complete rigour but interpret C in a broader sense. Jt = (s+t) mod 1 P8a.44 CHAPTER IL SOME GENERAL TOOLS AND RESULTS for all x E E. in particular that f is bounded. (5.10 Let {(Jt. 0) = h(i )( 1 + ttc(a)). and the family {f LEE given by Theorem 2.
5. Then the MAP in Proposition 5. . 1 + q(b .Qi < oo and Bi a probability measure. Ai = µi + 0Q.11 Consider the irreducible case with E finite. Bi [0] Remark 5.10 is given by P = eK(e) Oh e) F[e]Oh('). We omit the details.11 below in the finite case. if vi(dx) is compound Poisson. u Theorem 5. in the discrete time case.1) eft ea' f ij (dx) = Hij (dx) Hij [0] ..1) holds for the P. Bi(dx) = Bi(dx). Bi..tc(0)e = 0 .1) . 0<q<1. one can directly verify that (5. 0<b<oo.12 The expression for A means h(e) Aij = hie) Aij [1 + gij(Bij[0] i 0 j. ^i = of qij Bij [0] 1 + qij ( Bij [0] . 0 < qij < 1 and Bij [0] > 0.7(dx) Bij [0] Bij(dx) in the continuous time case . In the infinite case .c(0)e = tc(0)e . then also vi (dx) is compound Poisson with e Ox ^i = /3iBi[0].ic(0)e = ic(0)Oh e) h(e) . That 0 < qij < 1 follows from the inequality qb <1. Here Oh(e) is the diagonal matrix with the h=e) on the diagonal.(0)j. In particular. MARKOV ADDITIVE PROCESSES Proof That { Lt} is a multiplicative functional follows from L8 ogt = h(Jt+s) es(St+ . qij = r. vi (dx) = f3 Bi(dx) with . (5.St)sl(e) h(Jt) 45 The proof that we have a MAP is contained in the proof of Theorem 5. this gives a direct verification that A is an intensity matrix: the offdiagonal elements are nonnegative because Aij > 0.7) In particular. That the rows sum to 1 follows from Ae = Oh(e) K[O]h(B) . and by A = Oh(°) K [0]Oh(e ) vi(dx) = e"xvi (dx).
8). This shows that F. v. In matrix notation . (dx) of a process with stationary independent increments follows from Theorem 4.11. since Hij..46 CHAPTER II.t. are probability measures . Further Fib (dx ) = P=(YI E dx.e) Consider first the discrete time case .K [O])Oh(e) (0) l + ( A + (tc(') (a + 0) . in continuous time ( 5. Here the stated formula for P follows immediately by letting t = 1.r. Jt = A. this means that Ft[a] = etw ( e)Ohc) Ft[a + 9]oh (e) (5. Yi E dx.tc(0)I. this implies k[a] = A 1 ) (K[a + 0] . Hence the same is true for H=j and H. . (dx). a = 0 in (5. . is absolutely continuous w.. Letting a = 0 yields the stated expression for A.8) h(.tc(0)' )Ah() = Oh(o) K[a + 0]Oh() . Similarly.Bay [0]) That k(') (a + 0) .8. Jt = j] = hie) . Jl = j] :(Yi E dx.tc(') (8)/ d)ag h 7 Aiiii (Bii[a + 0] .tc (') (0) corresponds to the stated parameters µ.. F:j with a density proportional to eei . SOME GENERAL TOOLS AND RESULTS Proof of Theorem 5. Jl = j) = Ei[Lt. First note that the ijth element of Ft[a] is etK(e)Ej [e(a+B)st E:[east Jt = j] = Ej[Lteas' . H1. it follows that indeed the normalizing constant is H1 [0]. v= . Ji = j) h(e) eeyK(B)p h(8) h(e) eexK ( h=e) e)Fi.8) yields et'[a] = Ohie )et (K[a +e]K(e)I)Oh(°) By a general formula (A. Now we can write K[a] =A+A ) ( K[a + 0] .13) for matrixexponentials .
i.6. [261].6. Conditions for analogues of Corollary 5.. [262] in discrete time. is slightly less general than the present setting.7). h.e. there is. . however.1) = Aij4ij(Bij[a] . 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). however. the literature on the continuous time case tends more to deal with special cases. [226] and Miller [260].Bij[0]) = hjel)ijgijBij[0](Bij[a] . < x) = 11 (S.3 for an infinite E are given by Ney & Nummelin [266]. [225]. an extensive bibliography on aspects of the theory can be found in Asmussen [16].)Ajjgij(Bij[a+0] . 0]. 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]. Much of the pioneering was done in the sixties in papers like Keilson & Wishart [224]. has no mass on (oo. Write r+ = T(0) and define the associated ladder height ST+ and ladder height distribution by G+(x) = 11 (S. Though the literature on MAP's is extensive.+ < x. 6 The ladder height distribution We consider the claim surplus process {St } of a general risk process and the time 7. oo). For the Wald identity in Corollary 5. which. and is typically defective. 7+ < oo). see also Fuh & Lai [149] and Moustakides [264].1). Note that G+ is concentrated on (0.(u) = inf {t > 0 : St > u} to ruin in the particular case u = 0 . Notes and references The earliest paper on treatment of MAP's in the present spirit we know of is Nagaev [265].. THE LADDER HEIGHT DISTRIBUTION 47 Finally note that by (5. hardly a single comprehensive treatment.
T+ > t)dt = E f 0T+I(St E A) dt. The main result of this section is Theorem 6.. which gives an explicit expression for G+ in a very general setting. For the proof of Theorem 6. it follows that for g > 0 measurable. i.1. see Fig. 6.1) The interpretation of R+(A ) is as the expected time {St} spends in the set A before T+. the ladder heights are i.2.B(x) denotes the tail of B. a fact which turns out to be extremely useful. Theorem 6 . 0 f T+ (6.48 CHAPTER K. In other cases like the Markovian environment model. the dependence structure seems too complicated to be useful). Also. g(y)R+(dy) = E f g(St)dt. at present we concentrate on the first ladder height.e.1 The term ladder height is motivated from the shape of the process {Mt} of relative maxima.d. there are only finitely many). the second ladder point is ST+(2) where r+(2) is the time of the next relative maximum after r+(1) = r+. and the maximum M is the total height of the ladder. SOME GENERAL TOOLS AND RESULTS M ST+(2) Sr. i. 6. has no mass on ( 0. 1 For the compound Poisson model with p = 01LB < 1. Recall that B(x) = 1 .00 ).i. Thus. we shall first consider the compound Poisson model in the notation of Example 1. the sum of all the ladder steps (if rl > 0. The first ladder step is precisely ST+. define the prer+occupation measure R+ by R+(A) = E f o "o I(St E A. o 00 (6.2) . by approximation with step functions .5 below. On Fig.ST+(1) and so on. they have a semiMarkov structure (but in complete generality. R+ is concentrated on (oo. In any case. Here bo(x) _ B(x)/µB. 0].1.e. To illustrate the ideas.1.(3B(x ) = pbo(x) on (0. the second ladder height (step) is ST+(2) . G+ is given by the defective density g + (x) =. = ST+(1) Figure 6.. oo). where basically only stationarity is assumed. In simple cases like the compound Poisson model.
ST<St.2 R+ is the restriction of Lebesgue measure to (00.ST_t.O<t<T) = P(STEA.2(a): T+ > t Figure 6. since the distribution of the Poisson process is invariant under time reversion. That is. 49 Proof Let T be fixed and define St = ST .ST<St. {St }o<t<T is constructed from {St}o<t<T by timereversion and hence. 0]. ST < ST_t. St S* t a Figure 6. THE LADDER HEIGHT DISTRIBUTION Lemma 6 .0<t<T) = F(ST E A. P(STEA.2(b): r+ < t Thus. has the same distribution as {St}o<t<T. . see Fig.2.6.O<t<T). 6. 0 < t < T) P(STEA.St<0. 0 < t < T.T+>T) = P(STEA.
cf.y) (here we used the fact that the probability of a jump at u t is zero in the second step.St). and (6.50 CHAPTER II.3 where the bold lines correspond to minimal values. Figure 6.3 Lemma 6 .T+ > t] dt 0 T+ _ /3E f g( St) dt = 0 f g(y) R+(dy) 0 00 where g(y) = B(A . U + St_ E A. s. and since the jump rate is /3. oo). this is just the Lebesgue measure of A. Fig. T+ > t] 0 _ /3 f E[B( A . SOME GENERAL TOOLS AND RESULTS Integrating w. oo). it follows that R+ (A) is the expected time when ST is in A and at a minimum at the same time .y)R+(dy) 00 Proof A jump of {St} at time t and of size U contributes to the event IS. we get G+ (A) = f 00 /3 dt E[B(A .t dT. for A C (0..2) in the last). But since St 4 oo a. G+(A) = Q f 0 B(A . The probability of this given { Su}u<t is B(A . That is.St _).. 6.r.3 G+ is the restriction of /3R+*B to (0.St _)I(r+ > t). E A} precisely when r+ > t. .
S8 )t> o = {St }t>o for all s > 0. Uk) (k = 1.* ) and the second the mark (the claim size Uk ). assuming basically stationarity in time and space.. we consider the claim surplus process {St }t>o of a risk reserve process in a very general setup.s. U k)} k=1 a is as a marked point process M *. Nt St =>Uk k=1 t where Nt = max{k = O.4). as a point process on [0. i. We call M * stationary if M* o B8 has the same distribution as M* for all s > 0. Uk) for those k for which ak .e. The sample path structure is assumed to be as for the compound Poisson case: {St*} is generated from interclaim times Tk and claim sizes Uk according to premium 1 per unit time. we define the arrival rate as E# { k : ak E [0 . In the stationary case.Q f r+(x .4) are (ak.. obviously. i. h]} /h (by stationarity.:T1 +•••+Tk <t}.M o 08 shifted by s is defined the obvious way. . The marked point process . this is equivalent to the risk process {St*} being stationary in the sense of (6. 4 (the points in the plane are (ak . cf.e.s > 0). 0 Generalizing the setup.6. oo). THE LADDER HEIGHT DISTRIBUTION 51 Proof of Theorem 6. . oo) x (0..) where ak = Ti + • • • + Tk .. Fig.. {St+8 .1 With r+(y) = I(y < 0) the density of R+.z)B(dz) _ f I(x < z)B(dz) _ f (x). 2. The traditional representation of the input sequence {(TT.3 yields g+ (x) = . Lemma 6. the first component representing time (the arrival time o. 6 .1. 6 . this does not depend on h). The first ladder epoch r is defined as inf It > 0 : St > 0} and the corresponding ladder height distribution is * G+ (A) = P(S** E A) = P(ST+ E A.T+ < oo). The points in the plane (marked by x on Fig.
Assume {Jt} irreducible so that a stationary distribution 7r = (1i)iGE exists. e. = 0 . Note also that (again by stationarity) the Palm distribution also represents the conditional distribution of M* o Ot given an arrival at time t.5) represents the conditional distribution of M* given vi = 0.4 Given a stationary marked point process M*.4 Consider a finite Markov additive process (cf. i 1 U2 Us 1_ 0 or Q2 $ U3 *1 L 0 7 X I 11 1 Figure 6. we define its Palm version M as a marked point process having the conditional distribution of M* given an arrival at time 0 .. where TI = 0. As above . 0.e.2. vi(dx) = .s. We represent M by the sequence (Tk. and let T = T2 denote the first proper interarrival time . Uk) k=1. k: vk E [0. o. See. i.QiBi(dx).. the r. h] and the sum approximately ^o(M*)I(ul < h). letting h J..52 CHAPTER II. where T is the first arrival time > 0 of M and h > 0 an arbitrary constant (in the literature. This more or less gives a proof that indeed (6. h.5) does not depend on h. Oh becomes the approximate probability F(ri < h) of an arrival in [0. most often one takes h = 1).. . Example 6 .g. The two fundamental formulas connecting M* and M are Eco(M) = aE E. of (6. SOME GENERAL TOOLS AND RESULTS M* U. Section 5) which has pure jump structure corresponding to pi = a = 0. h] Eco(M*) = 1 E f co(M o Bt)dt. Sigman [348] for these and further aspects of Palm theory. V(M* o eak ).
aij for (i. let U0 be a r. Assume that St * . If Jt_ = i.6 Under the assumptions of Theorem 6..O for i # j. 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 . .OEU0.s.O fo "o F(x)dx = . A stationary marked point process M* is obtained by assigning Jo distribution Tr. Jo) w. having the Palm distribution of the claim size and F (x) = F(Uo < x) its distribution . This follows by noting that iP*(0) = IIG+JI = J0 "o g+(x)dx = . let the arrivals and their marks be generated by {Jt} starting from Jo = j. Note in particular that the Palm distribution of the mark size (i.6iBi + Aijgij Bij j#i iEE iEE 0 Theorem 6 .6. j) and let the initial mark Ul have distribution Bi when i = j and Bij otherwise. First choose (Jo_. Jt = j is iri(3i /. we get a marked point process generated by Poisson arrivals at rate /3i and mark distribution Bi when Jt = i. It follows that we can describe the Palm version M as follows . the distribution of Ul) is the mixture B = E aii Bi + aij Bij J = j#i !i J. an arrival for M* occurs before time t + dt w.p. we note: Corollary 6. qij when {Jt} jumps from i to j and have mark distribution Bij.*(0) with initial reserve u = 0 is p = /3EU0. v.e. and that p = 0EU0 < 1./.oo a.O for i = j and iriAijgij/.= i. THE LADDER HEIGHT DISTRIBUTION 53 Interpreting jump times as arrival times and jump sizes as marks.p.p. 5 Consider a general stationary claim surplus process {St }t>o.OF(x). and by some additional arrivals which occur w. After that. Before giving the proof. Then the ladder height distribution G+ is given by the (defective) density g+(x) = . the ruin probability . the probability aij of Jt .5.
The last property is referred to as insensitivity in the applied probability literature.. the arrival times 0 < 0'1 < Q2 < . It follows that for A C (0.. in (0.s.Su_ <0.. moves down linearly at a unit rate in between jumps and starts from S0 = U. A standard argument for stationary processes ([78] p.Su<0..o. which makes an upwards jump at time . are point processes on (oo . The sample path relation between { Su } and { Su } amounts to S„ = St . The result is notable by giving an explicit expression for a ruin in great generality and by only depending on the parameters of the model through the arrival rate 0 and the average ( in the Palm sense) claim size EU0. the mark at time Qk is denoted by Uk. We then represent M by the mark (claim size ) Uo of the arrival at time 0. oo ) and the arrival times 0 > 0_1 > a_2 > .. . 0<u<t) = P(St EA.$St_ u. .o<u<t where a claim arrives at time t and has size Uo.Mt). Then clearly * G+ (A) = P(ST+ E A) = Consider a process { f p(t)f3dt.A. CHAPTER H.5.(left limit) when 0 < it < t and is illustrated on Fig . Proof of Theorem 6.St*_ u.. . and the kth preceding claim arrives at time t .1] here the r .Su< 0.. that M* and M have doubly infinite time (i.5. { Su}0<u<t is distributed as a process {Su} .St<Su.54 By (6. Let p(t) be the conditional probability that ST+ E A. (k = St}t>o 1.o. in (oo.0<u<tIAt) = P(St EA.Q_k and has size U_ k.0<u<t) = P(StEA.St <.5). 105) shows that one can assume w. has a very simple interpretation as the average amount of claims received per unit time . T+ = t given the event At that an arrival at t occurs . SOME GENERAL TOOLS AND RESULTS V` (0) = E E Uk k: ak E [0.e.0<u<t) = P(St EA.).. oo) x (0 . 6.l. h. oo)).g.0<u<tIAt) = P(St EA. oo) p(t) = P(St EA. Now conditionally upon At . 2. 0).
Uo]. cf.5.s. Mt)dt = i3EL(A) o"o . Thus. the support of L has right endpoint U0. G' (A) = 3 f P(St E A. In Fig. time instants corresponding to such minimal values have been marked with bold lines in the path of { St}. 2 therefore immediately shows that L(dy) is Lebesgue measure on (oo.5 where the boxes on the time axis correspond to time intervals where {St } is at a minimum belonging to A and split A into pieces corresponding to segments where {Su} is at a relative minimum. 6.6.5 where it = { St < Su. A sample path inspection just as in the proof of Lemma 6 . the left endpoint of the support is oo. Since So = U0. 0 < u < t } is the event that { Su } has a relative minimum at t . and we let L(dy) be the random measure L(A) = fo°° I(St E A. t a oo. Fig. 6. NIt)dt . THE LADDER HEIGHT DISTRIBUTION 55 { A Su}0<u<t U0 U0 \t tt u>0 N U_1 Figure 6. and since by assumption St * oo a..
5 is due to Schmidt & coworkers [48].6 is Bjork & Grandell [67]. A further relevant reference related to Corollary 6.1). [263] (a special case of the result appear in Proposition VI. SOME GENERAL TOOLS AND RESULTS = OE f 0 I(Uo>y)I (yEA)dy = Q f IP (Uo>y)dy A 0o a fA P(y) dy• 0 Notes and references Theorem 6. .2.56 CHAPTER II. [147].
i=1 i=1 An important omission of the discussion in this chapter is the numerical evaluation of the ruin probability. For finite horizon ruin probabilities . and assume that • { Nt}t>o is a Poisson process with rate j3. 3).4 below . . i.. • the claim sizes U1.6) and simulation methods ( Chapter X). Panjer's recursion ( Corollary XI. are i. say. {Rt} and the associated claims surplus process {St} are given by Nt Nt Rt = u+t EUi. A common view of the literature is to consider such processes as perturbed compound Poisson risk processes . with common distribution B. exact matrixexponential solutions under the assumption that B is phasetype (see further VIII. Thus .. see Chapter IV.d. 4. i.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. 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 . • the premium rate is p = 1.e. St = uRt = EUi t. being of the form Rt = Rt+Bt + Jt where {Rt } is a compound 57 . U2. Some possibilities are numerical Laplace transform inversion via Corollary 3. and independent of {Nt}.
of the claim surplus St .1) . 0 . {Bt} a Brownian motion and {Rt} a pure jump process.g.+Uk)P(Nt = k) k=O e8t k=O B[s]k .1) = t(p . m.. Dufresne & Gerber [126]. (b) Var St = t.t = t(p . Write pB^) = EUn' YB = Pali = EU.. and this immediately yields (a). and Schlegel [316]. Furrer [150].1). cumulants .f.'s etc.)3t (fit' k t} = etk(8) exp {st '3t + B[s]f Finally. 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). we get Ee8st = 00 e8t c` Ee8 (U1+.58 CHAPTER III. where K(k) (0) is the kth derivative of is at 0. See e. 1 Introduction For later reference. e .g. Schmidli [319]. We do not spell out in detail such generalizations. and that B(k)[0] = Pak).u . say stable Levy motion.t = E[Ntµs] . [324]. for (d) just note that the kth cumulant of St is tic(k) (0).1 is the expected claim surplus per unit time. THE COMPOUND POISSON MODEL Poisson risk process. For (c).6pBa).1 (a) ESt = t(13µ$ .Rt.t = E E [ U k k=1 k=1 Nt . A more formal proof goes as follows: Nt r Nt ESt = E > U k .s. Proof It was noted in Chapter I that p .1). (d) The kth cumulant of St is tf3p(k) for k > 2. P = PAB = 1/(1 + rl) Proposition 1. we shall start by giving the basic formulas for moments.t = fltpB . (c) Ee8St = et" (8) where c(s) = f3(B[s] .
II. S„+V > S„ .h < St < S(n+1)h + h.. In particular. Proof We first note that for u.2 (DRIFT AND OSCILLATION) St/ta3'p1 ast >oo.. however. For the proof. Indeed. St = oo. .Tk. The right hand inequality in (1.Tk are i. and the value is then precisely v. (d) If 17 = 0. St = oo. if t = nh + v with 0 < v < h.d. rather to view {St} directly as a random walk in continuous time.1 = . then Snh . we need the following lemma: Lemma 1. then St 4 co. v > 0.V. 2. In this way. cf. 1. INTRODUCTION 59 The linear way the index t enters in the formulas in Proposition 1.Sok_l = Uk . then St> SnhV>Snhh.S„ attains its minimal value when there are no arrivals in (u. which is often used in the literature for obtaining information about {St} and the ruin probabilities. lim supt. Sn+0 . (b) If 77 < 0. we get a discrete time random walk imbedded in the claim surplus process {St}. then St 00.3EU01 = 1µs where rt is the safety loading. (a) No matter the value of 77.3) is proved similarly.1. meaning that the increments are stationary and independent.. Obviously. we have Sok . Here is one immediate application: Proposition 1. u + v]. the Uk . where Tk is the time between the kth and the (k .4. The connections to random walks are in fact fundamental. then lien inft.1)th claim. For example.3 If nh < t < (n + 1)h.1 is the same as if {St} was a random walk indexed by t = 0. (c) If 77 > 0. and there are at least two ways to exploit this: Recalling that ok is the time of the kth claim. obviously 0(u) = F(maxk Sok > u). so that {Sok } is a random walk with mean EUET = EU.. The point of view in the present chapter is.i. We return to this approach in Chapter V.
at least once. u 307).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)...p. and (b). it is seen that upcrossing occurs at least twice.3. 169) stating that lim infra. p.. For any fixed h..5 The limiting distribution of St . and < 1 for all u when 77 > 0. 1 since St 4 oo) and repeating the argument... However. hence by induction i. Thus using Lemma 1. lim supn_. Snh/n a4' ESh = h(p . and hence by the strong law of large numbers.o.3. or by a general result on discrete skeletons ([APQ] p. is a discretetime random walk for any h > 0.1. 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 .60 CHAPTER III.1. If rl > 0.. 0 Snh = 00.2: Proposition 1. {Snh}n=o. then {St} upcrosses level 0 a. Notes and references All material of the present section is standard. and hence it folz lows from standard central limit theory and the expression Var(St) = tf3pB (Proposition 1. Proof The case of 17 < 0 is immediate since then M = oo by Proposition 1.4 The ruin probability 0(u) is 1 for all u when 77 < 0. (c) are immediate consequences of (a). h.s. h A similar argument for lim sup proves (a). .2.1(b)) that the assertion holds as t 4 oo through values of the form t = 0. Corollary 1. where the size of the portfolio at time t is M(t). Part (d) follows by a (slightly more intricate) general random walk result ([APQ]. The general case now follows either by another easy application of Lemma 1. it suffices to prove 4'(0) = F(M > 0) < 1.._. THE COMPOUND POISSON MODEL Proof of Proposition 1. Considering the next downcrossing (which occurs w. There is also a central limit version of Proposition 1.1). Remark 1 .. This contradicts u St400. if P(M > 0) = 1.t . 2h. {Snh}n=o. Assuming that each risk generates claims at Poisson intensity /3 and pays premium 1 per unit time.1.6 Often it is of interest to consider size fluctuations. Snh u = 00 (the lemma is not needed for (d)).. 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. is a discrete time random walk.2.
IIG +II)EG+ .1. The decomposition of M as a sum of ladder heights now yields: 00 Theorem 2 . The following results generalizes the fact that the conditional distribution of the deficit ST(o) just after ruin given that ruin occurs (i. Theorem 2. This follows simply by noting that the process repeats itself after reaching a relative maximum. and we further get information about the joint conditional distribution of the surplus and the deficit. p < 1. Fig. B(x)/aB. 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}.1) representing the distribution of M as a geometric compound. [APQ] Ch. we may view the ladder heights as a terminating renewal process and M becomes then the lifetime. The expression for g+ was proved in Theorem 11. IV. It is crucial to note that for the compound Poisson model. and we shall here exploit the decomposition of the maximum M as sum of ladder heights.e. 1e.2. that r(0) < oo) is Bo: taking y = 0 shows that the conditional distribution of (minus) the surplus ST(o). Note that the distribution B0 with density bo is familiar from renewal theory as the limiting stationary distribution of the overshoot (forwards recurrence time ). As a vehicle for computing tIi(u).1 provides a representation formula for 0(u). Thus . cf.. but we shall be able to extract substantial information from the formula. Here bo(x) _ Proof The probability that M is attained in precisely n ladder steps and does not exceed x is G+ (x)(1 . Summing over n.1) is not entirely satisfying because of the infinite sum of convolution powers. the formula for the distribution of M follows . d.34 or A. which we henceforth refer to as the PollaczeckKhinchine formula. equivalently. n=0 (2. where G+ is given n=0 by the defective density g+ (x) = 3B (x) = pbo(x) on (0. we can rewrite the PollaczeckKhinchine formula as 00 (u) = P (M > u) = (1 .. Note that this . 11. (2. the ladder heights are i. oo ). 1 The distribution of M is (1. Combined with i/i(u) = P ( M > u).1.just before ruin is again B0. nevertheless.6.6. We assume throughout rl > 0 or. i.P) E PnBon(u) . 0 Alternatively.IIG+II) (the parenthesis gives the probability that there are no further ladder steps after the nth ).
W are independent.1 is traditionally carried out for the imbedded discrete time random walk. ST(o) > y. (c) the marginal distribution of ST(o)_ is Bo . f +b (b) the joint distribution of (ST( o).just after ruin.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. [62]. (a) 11 (ST(o)_ > x. However. and the conditional distribution of ST(o)_ given ST(o)_ = z is Bo z) The proof is given in IV. 2 The joint distribution of (ST(o )_.5. ST(o)) is given by the following four equivalent statements: B(z) dz. cf. the form of G+ is surprisingly insensitive to the form of {St} and holds in a certain general marked point process setup.6. (d) the marginal distribution of ST(o)_ is B0. see for example [APQ]. 1) and W has distribution Fw given by dFyy/ dB(x) = x/µB. Theorem A1.2(a) is from Dufresne & Gerber [125]. Feller [143] or Wolff [384]. (1 .62 CHAPTER III. in this setting there is no decomposition of M as a sum of i. Asmussen & Schmidt [49]. see Schmidli [323] and references there. Theorem 2 . cf. V is uniform on (0. 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. Beekman [61]. where it requires slightly more calculation. 7r(0 ) < oo) = Q 3 Special cases of the PollaczeckKhinchine formula The model and notation is the same as in the preceding sections. Again. As shown in Theorem 11. Theorem 2. there is a general marked point process version.i.5. For the study of the joint distribution of the surplus ST(u)_ just before ruin and the deficit ST(„).d. ST(o )) given r (0) < oo is the same as the distribution of (VW. .V)W) where V. ladder heights so that the results do not appear not too useful for estimating 0(u) for u>0. and the conditional distribution of ST(o) given ST(o)_ = y is the overshoot distribution B(Y) given by Bov)(z) _ Bo (y + z )/Bo(y). cf. the PollaczeckKhinchine formula is often referred to as Beekman 's convolution formula. The proof of Theorem 11.6. We assume rt > 0 throughout. In the risk theory literature.
3b Exponential claims Corollary 3.e. use Laplace transforms.p) = S . and hence this overshoot has the same distribution as the claims themselves . I. Alternatively.p.1)1 00 ( 1 . however .. then. For a failure at x.2 If B is exponential with rate S. B0 is exponential with rate S and the result can now be proved from the Pollaczeck Khinchine formula by elementary calculations .0(u) = pe(aA)" Proof The distribution Bo of the ascending ladder height ( given that it is defined ) is the distribution of the overshoot of {St} at time r+ over level 0.1 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. 0 . But claims are exponential . Thus .p) E pn S n x n. hence without memory.1 e ax = n1 (n . the formula for P(O) holds in a more general setting.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. Bon is the Erlang distribution with n phases and thus the density of M at x > 0 is (1 .6.p. As shown in 11. 1 . Let r ( x) be the failure rate of M at x > 0. The result can.p)pSe a ( l v)x = p( S . Thus r(x) = S(1 . also be seen probabilistically without summing infinite series . a further relevant reference is Bjork & Grandell [67]. the current ladder step must terminate which occurs at rate S and there must be no further ones which occurs w.O)e(b0)x.3. SPECIAL CASES OF POLLACZECKKHINCHINE 3a The ruin probability when the initial reserve is zero 63 The case u = 0 is remarkable by giving a formula for V)(u) which depends only on the claim size distribution through its mean: Corollary 3. the result follows . Integrating from u to oo.
A variety of proofs are available .y)G+(dy ) = f U V(u . (3.1) For a heavytailed B. 0 Proof Write o (u) as P(M>u) = P(S.3)).y)/3B (y) dy.+ >u. then 24 1 V. u .y)f3 (y) dy.1.2). the survival probability Z(u) = 1 .64 CHAPTER III.+ <U.h. THE COMPOUND POISSON MODEL In VIII.y)G+(dy) For the last identity in (3. (b) use stopped martingales . Corollary 3. is ?7+ ( u).3).3 The ruin probability Vi(u) satisfies the defective renewal equation ik (u) = 6+ (u) + G+ * 0(u) = Q f B(y) dy + u 0 f u 0(u . (3. we use the PollaczeckKhinchine formula in Chapter IX to show that b(u) .g. The case of (3.p + G+ * Z(u) = 1 .3) Equivalently. Then the first term on the r. (u) 35eu + 35e6u. (Example VIII. 3c Some classical analytical results Recall the notation G+(u) = f^°° G+(dx). We mention in particular the following: (a) check that ip (u) = pe (60)u is solution of the renewal equation (3.T+ <oo)+P(M> u.T+ <oo).1 p pBo(u). II.4) can be derived by elementary algebra from (3. cf. 2 is one of the main classical early results in the area. E.+ = y yields P(M>u. we show that expression for /'(u) which are explicit (up to matrix exponentials) come out in a similar way also when B is phasetype.i(u) satisfies the defective renewal equation Z(u) = 1 .+ <u.3. and conditioning upon S.4) is similar (equivalently. T+ <00) (3.S.S.4) zu P(M > u .2) Notes and references Corollary 3. if 3 = 3 and B is a mixture of two exponential distributions with rates 3 and 7.p + f u Z(u .3.s. just insert the explicit form of G+. (3. u + oo.3) below. (3. and weights 1/2 for each.
8) Proof This can be shown.Ps s(.5) Proof We first find the m.6) 00 = (I .p)2 3(1 .p)s s /3 . for example.(3 . either of these sets of formulas are what many authors call the PollaczeckKhinchine formula.f.p)s . 0 Notes and references Corollary 3. eau B(u) du = f PB 3PB SPB 0 o (3. 206207)./3B[s] .p = (1 . Embrechts. SPECIAL CASES OF POLLACZECKKHINCHINE Corollary 3. Also (3. The approach there is to condition upon the first claim occuring at time t and having size x .3 is standard .g. 111112 or Feller [143].P)pB' (3. Of course. In view of (3.5).P)PB 2(1 . . g.1 Bo[s] = f oc.7).5 The first two moments of M are 2 EM . Some relevant references are Abate & Whitt [2].5 can be found in virtually any queueing book. it is not surprising that such arguments are more cumbersome since the ladder height representation is not used. We omit the details (see. 191).7) s +. Corollary 3.3 .7) and Corollary 3.. [APQ] pp. (3.3.4 The Laplace transform of the ruin probability is 65 fo Hence Ee8M 00 e8uiP(u)du .pBo[s] no (1 . In fact. see e.PPB2) EM2 = PPB) + QZPBl 2(1 . numerical inversion of the Laplace transform is one of the classical approaches for computing ruin probabilities.5). e.s .. Bo of B0 as m e8u B(du) = B[s] .3 .(3B[s] 1 ./3B[s] which is the same as (3.4) can be derived by elementary but tedious manipulations.)3B[s]) (3. which yields the survival probability as 00 f u }t Z(u) = f f3eRtdt 0 from which (3.Ee8M) f ao e8' ( u)du = a8uP (M > u)du = 0 o 1 ( 1+ (1 .g. by analytical manipulations (L'Hospital's rule) from (3. [APQ] pp.p) E p"Bo[s]" = 1 . Griibel [179] and Thorin & Wikstad [370] (see also the Bibliographical Notes in [307] p. Griibel & Pitts [132].s .
u)]k k! (1 L3) 1: e_O(ku) NIN (k (k . Z^ =eR(k.1). For n < u < n + 1. of (3.u) [N(k ./32(u . THE COMPOUND POISSON MODEL 3d Deterministic claims Corollary 3.h.u + 1 )]k = QZ(u) . eO('u) [)3(k .u) [p(k .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 < 1.1)! k=1 u1 .66 CHAPTER III.u)]k k! k0 The renewal equation (3. differentiation yields Z(u) _ /3Z(u) .9) follows for 0 < u < 1.u) a)Qea" + (1 .9). . Assume (3.s./3Z(u .9) shown for n .4) for Z( u) means f lhu Z(u) = 1.u)]k1 ku+1) [/3( k .u/p)]k ko k! Proof By replacing {St} by {Stu/p} if necessary.)3(1 .3+ 18+ J0 Z(uy). we may assume p = 1 so that the stated formula in terms of the survival probability Z(u) = 1 .1).6 If B is degenerate at p.Q) 3e.Q (k 1 k= n  [O(k .1 < u < n and let Z(u ) denote the r.Q) k=0 k! E e0( = /32(u) .3I( 0<y<1)dy Z(y)/3I(0<u.u)]k d 1 u) _ a) n ( du ( k! (1  . then p) 1: ep(k u/. 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.z/'(u) takes the form Z(u) L^J L.
0 Notes and references Corollary 3. 00 the standard definition of the exponential family {F9} generated by F is FB(dx) = e°xK(e)F(dx). but will now be repeated for the sake of selfcontainedness. F and c.(9). it follows that Z(u) = 2(u) for n<u<n+1. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 67 Since Z(n) = 2(n) by the induction hypothesis.1) .2) (Here 9 is any such number such that r. See also Iversen & Staalhagen [208] for a discussion of computational aspects and further references.g.2).f. (4.g. The answer is yes: inserting in (4.3) by t.r. we set up . B9(dx) = B[9] B(dx).1 that c(a) = /3(B[a] . co(a) = rc(a + 9) .f.4. and define rce by (4.f. Formalizing this for the purpose of studying the whole process {St}. say t = 1: recall from Proposition 1.d.g. in terms of the c. 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 (4.6 is identical to the formula for the M/D/1 waiting time distribution derived by Erlang [139]. and thus (4. 4 Change of measure via exponential families If X is a random variable with c.4) works as well.2) shows that the solution is Ox [O]0].rc(9) = .) The adaptation of this construction to stochastic processes with stationary independent increment as {St} has been carried out in 11.a.f. The question then naturally arises whether ie is the c. (4.Qe(Bo[a] .3B[9].(9) is welldefined.1) . K(a) = logEe'X = 109f 00 eaxF(dx) = logF[a].1) or equivalently. or equivalently BB[a] = B[^+ Repeating for t 54 1. of F9. (4.4.a. We could first tentatively consider the claim surplus X = St for a single t.4) .3B = .
Then FB denotes the probability measure governing the compound Poisson risk process with arrival intensity. n) for a given n. for G E FT.. Ti(a)/n.f. v(Xi.g. .9) Proof We first note that for any fixed t. .1) and multiply from 1 to n).1 Let P be the probability measure on D[0.8) follows by discrete exponential family theory.4). = exp {BST .S(k_1)Tln.10) . (4. . (4. in particular the expression (4.. G C {T < oo}. Then the Xk are i.r.5) for the density of n i.d.FT. (4.2 For any fixed T.7) now follows by taking Z = eBST+TK(e)I(G) u Theorem 4 . with T taking the role of n) is the analogue of the expression exp{8(x1 + • • • + xn) . But let Xk = SkT/n .t. The following result (Proposition 4.i. EeeBSt + tk(B) = 1. and dP(T) dP^T) That is.5) for the density. Then P(G) = Fo(G) = EB [exp {BST + TK(O)} . Proposition 4.68 CHAPTER III. THE COMPOUND POISSON MODEL Definition 4.(9)} (4. Let FT = o(St : t < T) denote the o•algebra spanned by the St. G]. (4. Z is measurable w. the PBT) are mutually equivalent on. and define 09.r.i. then EBZ = E [Ze9ST _T"(9)I . t < T.FTn) = Q(SkTIn : k = 0.Tic (0)} . and thus (4.3 and claim size distribution B. with common c.d. and PBT) the restriction of PB to FT.3 Let T be any stopping time and let G E FT.nr.8) By standard measure theory. G]. .2.7) Proof We must prove that if Z is FTmeasurable. (4. The identity (4. the corresponding expectation operator is E9. it suffices to consider the case where Z is measurable w. . Xn).. BB by (4.0e and claim size distribution Be. .6) F(G) = Po (G) = EB [exp {BST + Ttc(0)} . oo) governing a given compound Poisson risk process with arrival intensity.t.1. replications from Fe (replace x by xi in (4.
5.7) holds.(Y) = 13(B['Y] .ST) + (T .FT]] = EB [exp { BST + Trc(9)} I(G)] .1 It is seen that typically) a ry > 0 satisfying 0 = r. the typical shape of rc is as in Fig. Ee [exp { BST +Trc(9)} I(G) FT)] = 1. according to what has just been proved.9) holds with G replaced by GT.r)rc(9)}I . so that PG = EeE0 [exp { 9ST+Trc(9)}I(G)I FT)] = Ee [exp { BST + rrrc(O)} I(G)EB [ exp {9 (ST . LUNDBERG CONJUGATION 69 Now assume first that G C Jr < T} for some deterministic T.g. (a) rc (a) (b) KL(a) 'Y 'Y Figure 5. (0) + rc'(0)a = 0 + ES1 a = a (p . Then GT = G n Jr < T} satisfies GT E FT.1) . GT C_ Jr < T}. Thus by (4.10).r is deterministic.7 1 Some discussion further supporting this statement is given in the next section.1(a). Thus. subject to the basic assumption ij > 0 of a positive safety loading. and hence (4.. 77 Thus. Letting T t oo and using monotone convergence then shows u that (4.f. . c(a) is a convex function of a. (4. t = T . The behaviour at zero is given by the first order Taylor expansion c(a) r. Given FT. 5 Lundberg conjugation Being a c. Now consider a general G.9) holds for G as well. Then G E FT.1) _ 1 + a.5.
2 s As support for memory. Thus B[7] = 6/.2 is B(7) = 1 + ^.g. . Taking T = r(u). Example 5 .3.4) ELS1 = #L(0) cf.1(b).4) yields /3L = b and that BL is again exponential with rate bL =. (5. we write FL instead of F7. An established terminology is to call y the adjustment coefficient but there are various alternatives around.3.1(b).3.QL instead of /37 and so on in the following .2) 7 Figure 5. Thus. It is then readily seen that the nonzero solution of (5.a = i(a + 7). (5.70 CHAPTER III. the Lundberg exponent.1) . Fig. u It is a crucial fact that when governed by FL. Lundberg conjugation corresponds to interchanging the rates of the interarrival times and the claim sizes.1 Consider the case of exponential claims. b[s] = 5/(b . 5. Equation (5. and (4. Fig.3) cf. THE COMPOUND POISSON MODEL exists .2)) is 7 = 5/3. we further note that ( 5. Note that KL (a) = /L (BL [a] .s).1) (or (5. the claim surplus process has positive drift > 0. e. 5. G = {T(u) < oo} in Theorem 4.1) is precisely what is needed for one of the terms in the exponent . an equivalent version illustrated in Fig. (5. 5.1) is known as the Lundberg equation and plays a fundamental role in risk theory .
1p . (5.7) is the same as (5.G+L)(x)) dx ry^+L) J 00 f 0 (1 .u be the overshoot and noting that PL(T(u) < oo) = 1 by (5.+ E A} in Theorem 4. Letting e(u) = ST(u) .e7x)G+(dx).5. T(u) < oo] .t. T = T+.1e.Ce7u as u 4 oo.3. we can rewrite this as 0(u) = e"ELe7^(u). e(u) has a limit i.5).r.ascending ladder height distribution and µ+ its mean. which shows that G(L) (dx) = e7xG +(dx) = e7x /3 (x) dx.6 ). PL ) with density 1 . G = {S.8) . Then P(ST+ E A) = EL [exp { 7S?+} . To this end.1 (5.(oo) (in the sense of weak convergence w.6) Proof By renewal theory. (5. Since a7' is continuous and bounded.3 (THE CRAMERLUNDBERG APPROXIMATION) i'(u) . we therefore have ELe7t(u) + C where C ELe7 (00) = µ+) f e7(1 . see A . ST+ E A] .3 takes a particular simple form. Proof Just note that e(u) > 0 in (5. V) (u) = P(T(u) < oo) = EL [exp {ryS. take first 0 = ry. 0 Theorem 5 . where C .G+ L) (x) G+L) (x) IL(+) µ+L) L) where G+L) is the FL.7) 0 and all that is needed to check is that ( 5.4). LUNDBERG CONJUGATION 71 to vanish so that Theorem 4. (5.5) Theorem 5 . V)(u) < e7u.1.(u)} .2 (LUNDBERG'S INEQUALITY) For all u > 0.P Y j o' xeryxOB (x) dx /3k [Y] .
we get L where 00 (1 . Noting that SIG(L)II = 1 because of (5.1) (5. (5.1)) and 7µ+L) = 'y/3 [7] 7 1/0 = /3B ['y] . of course. From this it follows.12) Example 5 .3 (this was found already in Example 5.")G + (dx ) = 1  J0 00 3B(x) dx = 1p.10) VW = JI c* e° (x) dx = a (B[a] .7). this solves the problem of evaluating (5.8) yields +L) J0 xel'B ( x) dx (5. or equivalently of how close the safety loading 77 is to zero. but some tedious (though elementary) calculations remain to bring the expressions on a final form. u .e. THE COMPOUND POISSON MODEL In principle.11) so that I 7B ['Y](B[7]1) BI [7]Q VP (7) 72 7 (using (5.4 Consider first the exponential case b(x) = Seax.72 CHAPTER III.1 above) and that C = p.1 . Using (5. that 7 = S . 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. .1 = ^7 The accuracy of Lundberg's inequality in the exponential case thus depends on how close p is to one.4). Then 0(u) = pe(a_Q)u where p = /3/S.
5. LUNDBERG CONJUGATION Remark 5.5 Noting that PL  1 = ,3LIBL  1 = #ci (0 ) = k (ry) _ ,QB' ['Y]  1 ,
73
we can rewrite the CramerLundberg constant C in the nice symmetrical form G, _'(0)1  1  p K'(7) PL1
(5.13)
In Chapter IV, we shall need the following result which follows by a variant of the calculations in the proof of Theorem 5.3: 1  aB[ry  a]  1 Lemma 5 . 6 For a # ry, ELea^ (°°) = 7 aK'(7) 7  a Proof Replacing 7 by a in (5.7) and using ( 5.8), we obtain 1 (I 1  ^ e('ra) x,3 (x)dx) (L ) ELea^*) = a \\\ f
using integration by parts as in (3.6) in the last step . Inserting (5.12), the result follows. u
Notes and references The results of this section are classical, with Lundberg's inequality being given first in Lundberg [251] and the CramerLundberg approximation in Cramer [91]. Therefore, extensions and generalizations are main topics in the area of ruin probabilities, and in particular numerous such results can be found later in this book; in particular, see Sections IV.4, V.3, VI.3, VI.6.
The mathematical approach we have taken is less standard in risk theory (some of the classical ones can be found in the next subsection). The techniques are basically standard ones from sequential analysis, see for example Wald [376] and Siegmund [346].
5a Alternative proofs
For the sake of completeness, we shall here give some classical proofs, first one of Lundberg's inequality which is slightly longer but maybe also slightly more elementary:
74 CHAPTER III. THE COMPOUND POISSON MODEL
Alternative proof of Lundberg 's inequality Let X the value of {St} just after the first claim , F(x) = P(X < x). Then , since X is the independent difference U  T between an interarrival time T and a claim U, ,3+ry F'[7} = Ee7 ( UT) = Ee7U • Ee7T = B['Y] a = 1' where the last equality follows from c(ry) = 1. Let 0( n) (u) denote the probability of ruin after at most n claims. Conditioning upon the value x of X and considering the cases x > u and x < u separately yields
,0(n +1) (u) = F(u) +
Ju
0 (n) (u  x) F(dx).
We claim that this implies /,(n) (u) < e 7u, which completes the proof since Vi(u) = limniw 1/J(n) (u). Indeed , this is obvious for n = 0 since 00)(u) = 0. Assuming it proved for n, we get
„/, (n+1)(u) <
F(u) + e7u
00
Ju
e7(u=) F(dx)
00
<
f
e7x F(dx)
+ fu
e  7(u z) F(dx)
u
o0
= e 7uE[ 'Y] = e 7u.
Of further proofs of Lundberg's inequality, we mention in particular the martingale approach, see II.1. Next consider the CramerLundberg approximation. Here the most standard proof is via the renewal equation in Corollary 3.3 (however, as will be seen, the calculations needed to identify the constant C are precisely the same as above): Alternative proof of the CramerLundberg's approximation Recall from Corollary
3.3 that
(u) = )3
J OO B(x) dx + J U Vi(u  x)/3 (x) dx.
u 0
Multiplying by e7u and letting Z(u) = e7" O(u), we can rewrite this as
u Z(u) =
z(u) = e7u/
J
B(x)dx, F(dx) = e7x,QB(x)dx,
u
z(u)
f +
J
e7(ux ),Y' 1 • l•(u  x) • e7'/B(x) dx,
0
= z(u) +
J0 u Z(u  x)F(dx),
6. MORE ON THE ADJUSTMENT COEFFICIENT 75
i.e. Z = z+F*Z. Note that by (5.11) and the Lundberg equation, ry is precisely the correct exponent which will ensure that F is a proper distribution (IIFII = 1). It is then a matter of routine to verify the conditions of the key renewal theorem (Proposition A1.1) to conclude that Z (u) has the limit C = f z(x)dx/µF, so that it only remains to check that C reduces to the expression given above. However, µF is immediately seen to be the same as a+ calculated in (5.10), whereas
L
00
z(u) du =
f
J
/3e7udu "o
J "o B(x) dx = J "o B(x)dx J y,0eludu
u 0 0
B(x)^ (e7x  1) dx = ^' (B[7]  1)  As] [0 µs] = l y P^
using the Lundberg equation and the calculations in (5.11). Easy calculus now gives (5.6). u
6 Further topics related to the adjustment coefficient
6a On the existence of y
In order that the adjustment coefficient y exists, it is of course necessary that B is lighttailed in the sense of I.2a, i.e. that b[a] < oo for some a > 0. This excludes heavytailed distributions like the lognormal or Pareto, but may in many other cases not appear all that restrictive, and the following possibilities then occur: 1. B[a] < oo for all a < oo. 2. There exists a* < oo such that b[a] < oo for all a < a* and b[a] = 00 for all a > a*. 3. There exists a* < oo such that fl[a] < oo for all a < a* and b[a] = 00 for all a > a*. In particular , monotone convergence yields b[a] T oo as a T oo in case 1, and B[a] T oo as a f a* in case 2 (in exponential family theory , this is often referred to as the steep case). Thus the existence of y is automatic in cases 1 , 2; standard examples are distributions with finite support or tail satisfying B(x) = o(eax)
76 CHAPTER III. THE COMPOUND POISSON MODEL
for all a in case 1, and phasetype or Gamma distributions in case 2. Case 3 may be felt to be rather atypical, but some nonpathological examples exist, for example the inverse Gaussian distribution (see Example 9.7 below for details). In case 3, y exists provided B[a*] > 1+a*/,3 and not otherwise, that is, dependent on whether 0 is larger or smaller than the threshold value a*/(B[a*]  1). Notes and references Ruin probabilities in case 3 with y nonexistent are studied, e.g., by Borovkov [73] p. 132 and Embrechts & Veraverbeeke [136]. To the present authors mind, this is a somewhat special situation and therefore not treated in this book.
6b Bounds and approximations for 'y
Proposition 6.1 ry <
2(1  aps) 2µs
OMB PB)
Proof From U > 0 it follows that B[a] = Eea' > 1 + µsa + pB2)a2/2. Hence 1 = a(B[7]  1) > Q (YPB +72µs)/2) = 3µs + OYµa2) 2 (6.1) 7 'Y from which the results immediately follows. u
The upper bound in Proposition 6.1 is also an approximation for small safety loadings (heavy traffic, cf. Section 7c): Proposition 6.2 Let B be fixed but assume that 0 = ,3(77) varies with the safety loading such that 0 = 1 Then as 77 .0, µB(1 +rl) 2) Y = Y(77) 277 PB Further, the CramerLundberg constant satisfies C = C(r1)  1. Proof Since O(u) + 1 as r7 , 0, it follows from Lundberg's inequality that y * 0. Hence by Taylor expansion, the inequality in (6.1) is also an approximation so that OAY]  1) N Q (711s + 72µB2) /2) = p + 3,,,(2) B 'y 7 2 2(1  p) _ 271µB
QPB PB)
6. MORE ON THE ADJUSTMENT COEFFICIENT 77
That C 4 1 easily follows from y 4 0 and C = ELe7V°O) (in the limit, b(oo) is distributed as the overshoot corresponding to q = 0 ). For an alternative analytic proof, note that C  1P = rlµB 73B' [7]  1 B' [ry)  1/0 711µB µB +7µB2 )  µB(1 +77 ) 'l = 1. 277q
77
7PBIPB
 77
13 Obviously, the approximation (6.2) is easier to calculate than y itself. However, it needs to be used with caution say in Lundberg's inequality or the CramerLundberg approximation, in particular when u is large.
6c A refinement of Lundberg 's inequality
The following result gives a sharpening of Lundberg 's inequality (because obviously C+ < 1) as well as a supplementary lower bound:
Theorem 6 .3 C_eryu < ,)(u) < C+ eryu where
= B(x) = C_ x>o f °° e7( Yx)B(dy )' C+
B(x) xuo f 0 e'r( vx)B(dy)
Proof Let H(dt, dx ) be the PLdistribution of the time r(u) of ruin and the reserve u  S7(„)_ just before ruin . Given r(u) = t, u  ST (u) = x, a claim occurs at time t and has distribution BL(dy)/BL(x), y > x. Hence ELe7£(u) 0
J
°o
H(dt, dx)
fX
eY(Y x) 00 f°° B(dy) x
BL dy
BL(x)
o
f
f H(dt, dx)
L ^ H(dt, dx) f e7B( x)B(dy) Jo oc, < C+
J0 0 o" H(dt, dx) = C. o" J
The upper bound then follows from ik(u) = e7uELeVu), and the proof of the u lower bound is similar.
78 CHAPTER III. THE COMPOUND POISSON MODEL
Example 6.4 If B(x) = eax, then an explicit calculation shows easily that B(x) _ e6X fz ° e7(Yx)B(dy) f x' e(6,6)(Yx)8esydy = 5 = P. Hence C_ = C+ = p so that the bounds in Theorem 6.3 collapse and yield the exact expression pey" for O(u). u The following concluding example illustrates a variety of the topics discussed above (though from a general point of view the calculations are deceivingly simple: typically, 7 and other quantities will have to be calculated numerically). Example 6.5 Assume as for (3.1) that /3 = 3 and b(x) = 2 .3e3x + 2 .7e7x, and recall that the ruin probability is 24 5su 5eu + 3e *(u) = 3 Since the dominant term is 24/35 • e", it follows immediately that 7 = 1 and C = 24/35 = 0.686 (also, bounding aS" by a" confirms Lundberg's inequality). For a direct verification, note that the Lundberg equation is
7 = /3(B['Y]1)
= 3\
2.337
+2.7771
which after some elementary algebra leads to the cubic equation 273  1472 + 127 = 0 with roots 0, 1, 6. Thus indeed 7 = 1 (6 is not in the domain of convergence of B[7] and therefore excluded). Further, 1P = B [7] 181B = 13 2.3+2.71 = 1 3 1 7 I 7'
_ 17
2 (3 a )2 + 2 (7  a)2 «=7=1 2 1p _ 7 _ 24
36 '
3.171 35* 36 For Theorem 6.3, note that the function QB[Y]1 f°°{L 3e_3x+
u
• 7e7x 1 dx
J
3 + 3e4u
f 0c, ex .
I 2 . 3e3x + 2 . 7e7x l dx
l J
9/2 + 7/2e4u
7. VARIOUS APPROXIMATIONS FOR THE RUIN PROBABILITY 79
attains its minimum C_ = 2/3 = 0.667 for u = oo and its maximum C+ = 3/4 = 0.750 for u = 0, so that 0.667 < C < 0.750 in accordance with C = 0.686.
Notes and references Theorem 6.3 is from Taylor [360]. Closely related results are given in a queueing setting in Kingman [231], Ross [308] and Rossberg & Siegel [309]. Some further references on variants and extensions of Lundberg's inequality are Kaas & Govaaerts [217], Willmot [382], Dickson [114] and Kalashnikov [218], [220], all of which also go into aspects of the heavytailed case.
7 Various approximations for the ruin probabil
ity
7a The BeekmanBowers approximation
The idea is to write i (u) as F(M > u), fit a gamma distribution with parameters A, 6 to the distribution of M by matching the two first moments and use the approximation
0(u)
f
u
Sa
r(A)
xa  leax dx.
According to Corollary 3.5, this means that A, 8 are given by A/S = a1, 2A/52 = a2 (2) PIB3) ^ZP(B)2 __ PPB a2 al 2(1  P)PB 3(1  P)µ8 + 2(1  p)2' i.e. S = 2a1 /a2, A = 2a2 1/a2.
Notes and references The approximation was introduced by Beekman [60], with the present version suggested by Bowers in the discussion of [60].
7b De Vylder's approximation
Given a risk process with parameters ,(3, B, p = 1, the idea is to approximate the ruin probability with the one for a different process with exponential claims, say with rate parameter S, arrival intensity a and premium rate p. In order to make the processes look so much as possible alike, we make the first three cumulants match, which according to Proposition 1.1 means p=AUB1=P1,
2N
(2) 6^= =OP
,
/3,4)
.
3 and Corollary 3.(3)2 P PB 2µB 2µB Letting /3* = /3/P.b(u) = p*e. 7c The heavy traffic approximation The term heavy traffic comes from queueing theory. p* _ . THE COMPOUND POISSON MODEL These three equations have solutions 9/3µB2)3 30µa2)2 3µa2) (3) P+ (3) ' 0 .P .7) that Ee$(Amex /j)M _ 1p _ 1p Eo [s (0max 1 . Grandell [171] pp. Though of course it is based upon purely empirical grounds. cf.Ps(/3max .3 )1 } _ 1p 1 .80 CHAPTER III.1 As /3 f Nmax.p + p { 1 1p ti 1 .PBo [s (/3max . the premiums exceed only slightly the expected claims. numerical evidence (e. [174]) shows that it may produce surprisingly good results.3* /S. 1924.g./3)M converges in distribution to the 2a exponential distribution with rate S = B' Proof Note first that 1 . Proposition 7./3)] 1 .p = (/3max 0)µB.)3 )PBo µB  . Mathematically.s(/3max .1.8µBo  Ss' u where 6 = µB/µBo = 2µa/µB 2) . Notes and references The approximation (7. heavy traffic conditions mean that the safety loading q is positive but small. or equivalently that /3 is only slightly smaller than /3max = 1/µ8.(bA*)". we shall represent this situation with a limit where /3 T fl but B is fixed. the approximating risk process has ruin probability z. we have according to the PollaczeckKhinchine formula in the form (3.p . Proposition 1. That is.1./3)PBo PB . Letting Bo be the stationary excess life distribution. and hence the ruin probability approximation is b(u) e(bAln)u.2) was suggested by De Vylder [109]. (/3max . but has an obvious interpretation also in risk theory: on the average.
[APQ] Ch.B AB ) 6()3max _'3) = However . and hence 2µ2B 1 . light traffic conditions mean that the safety loading rl is positive and large . then P(u) 4 e6„ Proof Write z'(u) as P((/3max . VARIOUS APPROXIMATIONS FOR THE RUIN PROBABILITY 81 Corollary 7. the premiums are much larger than the expected claims .g. as well as it is needed for the interpolation approximation to be studied in the next subsection. we shall represent this situation with a limit where 3 10 but B is fixed./3)u * v.2./3)u).3) is reasonable for g being say 1020% and u being small or moderate. These results suggest the approximation Vi(u) e6(0_. It is worth noting that this is essentially the same as the approximation (2) z/i(u) Ce. We return to heavy traffic from a different point of view (diffusion approximations) in Chapter IV and give further references there . VIII). Of course.ryu . In the setting of risk theory.4) suggested by the Cramer Lundberg approximation and Proposition 6.l3)M > (/3max . u * oo in such a way that (3max . light traffic is of some interest as a complement to heavy traffic .p. Mathematically.0)u. 7d The light traffic approximation As for heavy traffic . This follows since rl = 1/p . Notes and references Heavy traffic limit theory for queues goes back to Kingman [230].Q T /3max. in risk theory heavy traffic is most often argued to be the typical case rather than light traffic . The present situation of Poisson arrivals is somewhat more elementary to deal with than the renewal case (see e . Numerical evidence shows that the fit of (7.2 If . the first results of heavy traffic type seem to be due to Hadwiger [184]. obviously Corollary 7.p _ 2rl11B PB p. That is . the term light traffic comes from queueing theory. However . while the approximation may be far off for large u. 2 provides the better mathematical foundation. or equivalently that 0 is small compared to µB .ze a2unµB laB (7.1 1 . but has an obvious interpretation also in risk theory: on the average . .7.
( 3 J O B dx.u. u Note that heuristically the light traffic approximation in Proposition 7.= 1 aJ 1 a 0+ 1 = = p. 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)+.3 is the same which comes out by saying that basically ruin can only occur at the F(U .(3 10.Q limIP ( u) + Q lim z/'(u) Amax &0 amax ATAm. see Daley & Rolski [96]. 0 u Notes and references Light traffic limit theory for queues was initiated by Bloomfield & Cox [69]. z/' (u) convergence P(U . THE COMPOUND POISSON MODEL Proposition 7. i.T > u) = J o" B(x + u)/3eax dx . En'=2 • • • = O(/32) so that only the first terms matters. the Poisson case is much easier than the renewal case. The crude idea of interpolating between light and heavy traffic leads to 0 (u) C1 . Omax max m.e.T > u). 0(u) /3 J B(x)dx = /3E[U . U > u] = /3iE(U .5) u Proof According to the PollaczeckKhinchine formula. For a more comprehensive treatment. Again. 10 ( u The alternative expressions in (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.5) follow by integration by parts. Asmussen [19] and references there. (7.82 CHAPTER III.3 As . by monotone time T of the first claim . Indeed. . ao n=1 00 n=1 (u) P) anllBBon(U) onPaBon(u) • Asymptotically. and hence 00 (U) /3pBBo (u) = 0 / B(x)dx. Sigman [347]. [97]. cf. Light traffic does not appear to have been studied in risk theory.
O(E)(u) 1 (1 . The adaptation to risk theory is new. however. . Substituting v = u(. 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.6) is . we see that the following limits HT) (u'). Al .3n. ^ LT Q maxQ m"^ Qlo V LT) ( CHT(v) (say). with rate 1/µB = /3max. to get nondegenerate limits . 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). B(2). "/Qmex Cu) CLT(u ( /3max 0) + O16 CHT( U(Qmaz . 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.x . the idea of interpolating between light and heavy traffic is due to Burman & Smith [83 ]. 8 Comparing the risks of different claim size distributions Given two claim size distributions B(1).8. (U). Instead. [84]./3)) . even if the safety loading is not very small. where further references can be found . ^IE) exist: 1 (B) HT QmsxQ hm J e e6" 2µE/µE2)'" = e(1 6)" =  Q1Qm.3). z/i(E) (u) = pe(QmaxQ)u.VHT) ( ax QmQ ) h (B) ( .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. available. .Wmax f(x ) dx + pee6mQ. ) M. no empirical study of the fit of (7.O0 M. that is. Let OLT) (u) denote the light traffic approximation given by Proposition 7. (7. Thus .3 and use similar notation for %(B) (u) = (u). COMPARISONS OF CLAIM SIZE DISTRIBUTIONS 83 which is clearly useless . _(E) (u). Another main queueing paper is Whitt [380]. Notes and references In the queueing setting .
B(') <i. one can interpret f x°° B(y) dy as the net stoploss premium in a stoploss or excessofloss reinsurance arrangement with retention limit x. XI. Finally. B(' <. B(2)) if f fdB(1) < f fdB(2) for any convex function f. Here convex ordering is useful: Proposition 8.6. Taking probabilities. whereas (consider x2) B(2) has the larger variance. this implies St T(l)(u) > r(2)(u) for all u so that 17(I) (U) < oo} C_ {T(2)(u) < oo}.1 If B(') <d B(2). an equivalent characterization is f f dB(') < f f dB (2) for any nondecreasing convex function f.1 is quite weak. Recall that B(') is said to be stochastically smaller than B(2) (in symbols. cf. for more detail and background on which we refer to Stoyan [352] or Shaked & Shantikumar [337]. U(2) distribution B(2) and U(1) < U(2) a. this ordering measures difference in variability. we shall need various ordering properties of distributions. B(2)) in the increasing convex order if f BM (y) dy < f 00 Bi2i (y) dy x x for all x. then i.2 If B(') <j. we have the convex ordering. THE COMPOUND POISSON MODEL To this end. B(2) and PB(1) = µB(2). B(') <d B(2)) if B(1)(x) < B(2)(x) for all x. In terms of the time to ruin.84 CHAPTER III. we can assume that 1) < St 2l for all t. A weaker concept is increasing convex ordering: B(1) is said to be smaller than B(2) (in symbols. Proof According to the above characterization of stochastical ordering. Bill is said to be convexly smaller than B(2) (in symbols.s. or the existence of random variables U(l).' 1)(u) < V)(2) (U) for all u. Proposition 8.ill(u) < V)(2) (U) for all u. u Of course. Rather than measuring difference in size. In the literature on risk theory. Proposition 8. . the proof is complete. then . most often the term stoploss ordering is used instead of increasing convex ordering because for a given distribution B. In particular (consider the convex functions x and x) the definition implies that B(1) and B(2) must have the same mean. equivalent characterizations are f f dB(') < f f dB (2) for any nondecreasing function f. and a particular deficit is that we cannot compare the risks of claim size distributions with the same mean: if BM <d B(2) and µB«) = /IB(2). then Bill = B(2). U(2) such that U(l) has distribution B(').
(D) (u) < O(B) (U ) for all u. A first attempt would of course be to identify 'variation' with variance.u) = (1 _ P) E /3npnBo( 1):n(u) n=1 00 < (1.5 If '0(1)(u) < p(2) (U) for all u and a. Bo1) <_d Bo2) which implies the same order relation for all convolution powers.6 Fix /3 at 1/1.6 below is that (in a rough formulation) increased variation in B increases the risk (assuming that we fix the mean). with fixed mean. The problem is to specify what 'variation' means. we have Bol) (x) f ' B(1) (y) dy < ' f' B(2) (y) dy = Bo2) (x)• µ 85 I. it is seen that asymptotically in heavy traffic larger claim size variance leads to larger ruin probabilities. The heavy traffic approximation (7.. A partial converse to Proposition 8.e. Example 8. say to p.2 is the following: Proposition 8. . larger variance is paramount to larger second moment.1.4 Let D refer to the distribution degenerate at 'LB .3 provides another instance of this.1 and µB at 1 so that the safety loading 11 is 10%. Corollary 8. then B(1) <. Proof If f is convex. we have by Jensen 's inequality that E f (U) > f ( EU). and here is one more result of the same flavor: Corollary 8. A general picture that emerges from these results and numerical studies like in Example 8. then /'(')(u) < 0(2)(u) for all u..p ) E /3"µ"Bo2)* n(u) _ V(2) (u) n=1 = Corollary 8.(1) (. Then V.8. from which the result immediately follows. This u implies that D <. B. u We finally give a numerical example illustrating how differences in the claim size distribution B may lead to very different ruin probabilities even if we fix the mean p = PB. B(2). Hence by the PollaczeckKhinchine formula . Proof Consider the light traffic approximation in Proposition 7. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS Proof Since the means are equal.3 If B(1) <. and consider the following claim size distributions: B1: the standard exponential distribution with density ay. B(2).4) certainly supports this view: noting that.
the behaviour of which is governed by a parameter 9. We return to ordering of ruin probabilities in a special problem in VI. A standard example from queueing theory is .. B3: the Erlang distribution with density 4xe2x. 1%.000.e'\1x + 0. Let ua denote the a fractile of the ruin function.4.9A2e'2r where A. van Heerwarden [189]. 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). One then obtains the following table: U005 U0.1358. [166]. we have 0r3 = 2 < or2 = 1 < 02 = 10 < 04 = 00 so that in this sense B4 is the most variable.e. B3 the comparison is as expected from the intutition concerning the variability of these distributions.0' U0. which appears to be smaller than the range of interest in insurance risk (certainly not in queueing applications!).4142.01%. Pellerey [287] and (for the convex ordering) Makowski [ 252]. and consider a = 5%.001 u0.1%. 0.01%.86 CHAPTER III. 11 Notes and references Further relevant references are Goovaerts et al. in comparison to B2 the effect on the ua does not show before a = 0. 9 Sensitivity estimates In a broad setting. = 0. For B1i B2. i. Note to make the figures comparable. B.lA. with the hyperexponential distribution being more variable than the exponential distribution and the Erlang distribution less. all distributions have mean 1. B4: the Pareto distribution with density 3/(1 + 2x)5/2. 1/)(u. and this is presumably a consequence of a heavier tail rather than larger variance. A2 = 3. THE COMPOUND POISSON MODEL B2: the hyperexponential distribution with density 0. Kluppelberg [234].) = a. sensitivity analysis (or pertubation analysis) deals with the calculation of the derivative (the gradient in higher dimensions) of a performance measure s(O) of a stochastic or deterministic system. 0. In terms of variances o2. However.
9. while /3 = j3 is an estimate. t].2 Consider a risk process { Rt} with a general premium rate p. a0 as ao 80 19P . Assume for example that 8 is known. a2/t).3. increasing in u. the distribution of %3 0 is approximatively normal N(0„ Q/t). Then the arrival rate /3(P) for { R(P) } is )31p. the premium rate p and the claim size distribution B. Proof This is an easy time transformation argument in a similar way as in Proposition 1. and hence the effect of changing p from 1 to 1 + Ap corresponds to changing /3 to /3/(1 + Op) /3(1 . obtained say in the natural way as the empirical arrival rate Nt/t in [0. and hence a _ e(60)u + u e(60)u = ( i + which is of the order of magnitude uV.Ap).. i. Thus. SENSITIVITY ESTIMATES 87 a queueing network. Example 9.19P a/ . Thus at p = 1. where the partial derivatives are evaluated at p = 1. and s(9) the expected sojourn time of a customer in the network. For example. say estimated from data. the standard deviation on the normalized estimate ^/1' (the relative error ) is approximatively . s(9) is of course the ruin probability t' = Vi(u) (with u fixed) and 0 a set of parameters determining the arrival rate 0.1. it follows that ' is approximatively normal N(0. u Proposition 9. Similar conclusions will be found below. where Q2 = fl ( l2 1113 / _ Ou2v)2. Let R(P) = Rtli.1 Consider the case of claims which are exponential with rate 8 (the premium rate is one).01/2u. In the present setting. a/3 0 . Then a p ao = 00 Qa/. Then ib = Pe(613)u. Then if t is large . if = a e(6A)u. or we may be interested in aV)/0/3 as a measure of the uncertainty on '0 if 0 is only approximatively known. In particular . with 0 the vector of service rates at different nodes and routing probabilities.e. we may be interested in a'/ap for assesing the effects of a small change in the premium.(u) for large u.
In the case of the claim size distribution B. namely that of a twoparameter exponential family of the form Bo. However .()^ 1 . it suffices to fix the premium at p = 1 and consider only the effects of changing . 3) ( 9 .2) (see Remark 9. and write yp = 8y/8/3 and so on .()wC(e. Similar notation for partial derivatived are used below.^)] 1(/3+y)we(9+'y.r.3) follows by straightforward algebra. () = log(1 + y//3). we cannot expect in general to find explicit expressions like in Example 9.0 = t/'(u) and the CramerLundberg constant C. (.3. mathematically a proof is needed basically to show that two limits (u * oo and the differentiation as limit of finite differences) are interchangeable. we can rewrite the Lundberg equation as w(9+ y. this intuition is indeed correct.3. x > 0 (9. u Now consider the ruin probability 0 = 0 (u) itself. but we shall concentrate on a special structure covering a number of important cases.1 or Proposition 9.t. Viei '0(.w(O. 9.uypCe7u urypO.3 70 = 'Ye = = 7 /3(1 we(e +'y. Proposition 9. various parametric families of claim size distributions could be considered. (3+'y)PC (0+7. e. ^) .(/3 + y)we(9 + 7. ()} p(dx) . (9.()YC = 1 +y/ /3 \ Q2 From this (9. for the ruin probabilities .w(6.g.()(0 +'0) ' (9 . Consider first the adjustment coefficient y as function of 3. 4) (9 . 5) (Q+'Y)[we(0+7.88 CHAPTER III. Of course. Differentiating w. /3 yields w e(e + Y. THE COMPOUND POISSON MODEL As a consequence. Consider first the case of 8/8/3: .Owe (9.((dx ) = exp {Ox + (t(x) . so that heuristically we obtain '00 50ryu = Coe"u . and the proofs of (9.4).5) are similar. () Proof According to (9.6 below for some discussion of this assumption). (9.6) As will be seen below. The most intuitive approach is to rely on the accuracy of the CramerLundberg approximation .10) below. but must look for approximations for the sensitivities 0.3 or/and B.
2 of the Appendix ).C).g.w(9. and alsoo zl(u) + 0 because of B['y ] < oo.we(9 .QB(x) dx. Z(u)/u a C//3PF where PF is the mean of F.11) Ee.4 As u oo.9) (9. w((9 + a. But from the proof of Theorem 5. By dominated convergence. the proof is complete. Hence by a variant of the key renewal theorem (Proposition A1. ()} .12)). u 0 Proceeding in a similar way as in the proof of the CramerLundberg approximation based upon (9.3 (see in particular (5. 11 For the following.8) (Section 5). () . 0(u) = /3 Ju"O B(x) dx + f 0 0(u . we note the formulas Ee.3) for z/'(u).(e"U = = wS(O.10) (9.x). F(dx) = e'yy/3B(x)dx. Z= zl + z2 where zl (u) = e7u J m B(x)dx.([a] = exp {w(9 + a.w(9.x). Be. PF = (1 . it holds that 89 a ue ryu a/3 Q(1 P) 7C2 Proof We shall use the renewal equation (3. ()] exp {w (O + y.4t (U)e°`U = which are wellknown and easy to show (see e.8). O} (9. z2(U) = e7" J u b(u . z2(u) _ 1 ^ e'ri`i7i( u .x) F(dx ) f u J C F(dx) = C as u 4 oo.p)/C'y.9. SENSITIVITY ESTIMATES Proposition 9.x)B(x)dx. we get p(u) = J "O B(x) dx + J U O(u . (9. ()} . BarndorffNielsen [58]). Combining these estimates .w(O. () . Further write de = [we (9 +'y. u 0 Then Z = z + F * Z and F is a proper probability distribution . () . we multiply by e7" and let Z(u) = elt" cp(u). () . () exp {w(9 + a.St (U) Ee.x)B(x) dx + J U W(u .3(x) dx.8) Letting cp = e0/e/3 and differentiating (9.
this implies Z = z + F * Z.lB(x) dx = e7uzl(u) + e7°zz(u) + V(u T where zl (u) = . ()} 1z(dy) = f [t(y) . Then as u > oo.wc (O. 8^ ue7u.2) holds. ^) . )}B(dy)• Letting cp it thus follows from (9.w( (0.1) B(dy) 'f '[t(y) . ()]B(dy) dx. ^)} [wc (0 + 7.6e7u f "o f[t(y) . By dominated convergence and (9. THE COMPOUND POISSON MODEL [we(e+7.6C do 89 1p 8( 1p Proof By straightforward differentiation.e7x/3 f 00 [t(y) .x).9)(9.90 CHAPTER III. 2 z 07P N ue7u (3C de .x)f3 f ^[t(y) .QB(x)dx. 0 x Multiplying by e7" and letting Z(u) = e"uV(u).5 Assume that (9. C)] (1 + 7 ) Proposition 9. ())B(dy) dx. 8 8() 8( (9.8) that cp(u) . u Z2(U) = e7° f u ^/i(u .we (0. ()]e7vB(dy) 'fCd 7 c . F(dx) = e7x.w( (0. C) .w (9.w(0.w(e. 01 (i+) do = +'Y.11). ()]B(dy) dx x 0 0C T ON O . oo z2 (u) f C .wc(9.12) f exp {O y + (t(y) . ()](e7v .wc(O. z = zl + z2.
.a log S = log r(c) .1edz = 1 exp {Sx + a log x .15) (9. w(e.) log(9) = %F(a) logs where %1 = F'/]F is the Digamma function.12) takes the form y) alp a . . by inserting in the above formulas.17) (9.pa+1 . Z(u) /3C 91 o c'o e11(t (y) . we (9..rye) S 5rya. () = C/9 = a/S.(log r(a) a log S)} • r(a) 1. and also zj (u) 4 0 because of f Hence.C log(9).12) follows.QS 1 .1 . ue_Yu 'C2d( 8a 8( 1 p . () = log r(a) .a/35a&y' ' (9./35' a/i'y + aryl 625ry.16) (9. Example 9. ())B(dy) < oo.Y)a+1 ' (9. and the proof of the first one u is similar.6 Consider the gamma density b (x ) = Sa xa. a /(S . We get w( (0. that C = a.9. t(x) = logx.Sry a/32 + a/37 + /37 .w((9.18) (05 + 57 _'3_y . U 7µF from which the second assertion of (9.14) de = d( 7!3 76 = 7e = log ( \ ( \5a_ / \SSry ) 72 . ( 9.3ary tog('Finally.. () ='I'(t. < = a. SENSITIVITY ESTIMATES as u 3 oo. 9 = S.ry) 5a1 cry (5 . It follows after some elementary calculus that p = a)3/5 and.yu/3C2do u86 89 1p' az/) = 8z/.2) holds with p(dx) = xldx. Here (9.13) (9.
([Y] . Be. Straightforward but tedious calculations . ()} = exp {c (C .w(9."62 .3. w(e. C) = B = Yc = de = do = .22.9) (() .2a) } Thus the condition B[a*] > 1 + a* /.1 = eXP {c(C . () = Cc .2) with µ(dx) = 2x3zrdx. which we omit in part .2 log (0.l3 of Section 6a needed for the existence of ry becomes e^Q > 1+62 / 2. t(x) _ .7 Consider the inverse Gaussian density ( b(x) Zx37 exp This has the form (9.CZ try)} 1 C C2 try .1 16 +ry c C22ry 2( = + 70 We (e. C = .2 . for a < a* = z (. C) . further yield . THE COMPOUND POISSON MODEL Example 9.21og 2. 9 = .log c = 2 In particular.3Ee.92 CHAPTER III.S[a] = exp {w (9 + a.
3C2de 1p' z a = c .12) takes the form a = a 93 ar.2) is motivated as follows. let NT 16T = ^T . if 1 PT = /3TNT(U1+. However.7 and references there. thus. Thus. by the LLN both F (NT = 0) and F (PT > 1) converge to 0 as T . 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). in which case we can just let t(x) = 0. 10 Estimation of the adjustment coefficient We consider a nonparametric setup where /3.10.1) . In general. Thus. (9. kT (a) = /T (BT [a] . [379] consider a special problem related to reinsurance. the exponent is either Ox.2 of the parameters. the results presented here are new. and we estimate y by means of the empirical solution ryT to the Lundberg equation. to our knowledge. However .+UNT) > 1. Van Wouve et al. and hence explicit or asymptotic estimates are in general not possible.cue_7u)3C P Remark 9. Comparatively less work seems to have been done in risk theory.. ae t 1lEY u S _ . in u which case the extension just described applies. we can just fix k . Finally if k = 1. the models there (e. the main tool is simulation.. the exponent of the density in an exponential family has the form 01 tl (x) + • • • + 9ktk (x). B are assumed to be completely unknown.8 The specific form of (9. Note that if NT = 0. Notes and references The general area of sensitivity analysis (gradient estimation) is currently receiving considerable interest in queueing theory.a. then ryT < 0.. That it is no restriction to assume k < 2 follows since if k > 2. Also.oo. queueing networks) are typically much more complicated than the one considered here. . sj=1 and let YT be defined by IKT('ryT) = 0. then BT and hence ryT is undefined. or Ct(x).g. for which we refer to X. To this end. BT [a]= NT ^` e"U. we have assumed k = 2 and ti (x) = x. ESTIMATION OF THE ADJUSTMENT COEFFICIENT Finally.
it is easy to see that we can write \ V1 1 l _ . a2 where a2 = /3r.: N ()3.T y .3/ T). If furthermore B[27] < oo.1) . B [7]2 (10. (10. .B[7]2 n Hence ( 10.'Y . we need a lemma.. Lemma 10 .(27)/K'(7)2.a BT[7] I B[7] I + .v.94 CHAPTER III. 7T a4' 7.7 + (.B[7]2 }) ( T 0 .i3)(B[7] 1) + (3(BT[7]  .: N 0. since NT /T . then (10. Hence KT(7) = (F' + (OT a(B[7l 0))((BT [7] .Q and Anscombe 's theorem.B[7]2 V2 . 1) r.1) 'YT . N ( n[7].3T .3) Proof Since Var(eryU) = we have B[7]. 16T where V1.2) follows from NT/T a4' .If .B[7]) + B [7] .)vl+ N CO. V2 are independent N (0.B[7]) 0+ Iv/o(b[y]. For the proof.'s.1)2 + E[27] . THE COMPOUND POISSON MODEL Theorem 10 . B[27] . vfoVFB[2y].b[Yp'V21 T { (E[7] . More generally.2) rT(7) N N (0.1) . B[2'Y]  /3T ) ./^ B[27] .1 As T 4 oo.2 As T * oo.
1 can be used to obtain error bounds on the ruin probabilities when the parameters .e. Theorem 10. and the truth of this for all e > 0 implies ryT at 'y. 6"Y (10.(ry . NT i =1 n'(a) for all a so that for all sufficiently large T K7 . Let 0 < E < ry.(ry + e) and hence KT(7 . °7IT) .KT(7) kT(7) K'(7) . lcT(a ) 4 /c(a).e. NT BT [a] Hence r. 7T E (y ..'T(a) = 1 E Uie°U' a$' EUe "u = B'[a].10. Combining ( 10. Proof of Theorem 10. where ryT is some point between ryT and ry. I.c'(7) N (0' T (2(7) / N (0.e) < 0 < r. it follows that 7T7 KT(7T) .E ) < 4T(7T) < (7 +0' which implies 'T(ry4) a$' r.3). first note that e7TU N (e7U u2e27Uo'2/T) 7 .4) + E).'(y).E) < 0 < kT(7 + E) for all sufficiently large T .1 By the law of large numbers. OT a 95 u 4 /3. Then r. To this end .4) and Lemma 10. Now write KT(7T)  kT(7) = 4T(7T)( 7T 7).2. By the law of large numbers. ESTIMATION OF THE ADJUSTMENT COEFFICIENT which is the same as (10. y + E) eventually. 0 are estimated from data . If ryT E (7  we have KT(7 .Q. BT[a] 3 B[a].E) < 4T(7T) < 4T(7 + E).
Wn). satisfies b(. Letting Wo = 0.g. Hipp [196].. ft. Csorgo & Teugels [95].T VIT where r7ry. 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. Mammitzsch [253] and Pitts. Embrechts & Mikosch [133]. i . [197]. various alternatives have been developed.e. Griibel & Embrechts [292].C1e"a ( see e..g. Further work on estimation of y with different methods can be found in Csorgo & Steinebach [94]. Frees [146]. A major restriction of the approach is the condition B[2ry] < oo which may be quite restrictive. V. Asmussen [23]) can then be used to produce an estimate of ry. if B is exponential with rate 8 so that ry = 8 .f. Herkenrath [192]. t]}. 6 < 2. with a tail of the form P(Y > y) .d.info< „< t S.96 CHAPTER III. > 0 for some t E [Wn_ 1.Q. For this reason .0) < 5.e.. Deheuvels & Steinebach [102]..1 : Vt = 0. it means 2 (8 . This approach in fact applies also for many models more general than the compound Poisson one. Notes and references Theorem 10.Wn) are i .3 or equivalently p > 1/2 or 11 < 100%. .. U2. i.i.) = a (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.ueryuU ".5%).. = 1.. Vt = St ..1 is from Grandell [170]. wn = inf{t > W. the nth busy cycle is then [Wn1.T = 3TKT ( 21T)IKT (^T)2 is the empirical estimate of vy and fc. For example .96 if a = 2. and the known fact that the Y„ = max Vt tE[W„1.
B[•] and mean AB. Unless otherwise stated. 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 where p = 13µB. exists. defined as solution of c(ry) = 0 where ic(s) _ /3(B[s] . the premium rate is 1. it is assumed that i > 0 and that the adjustment coefficient (Lundberg exponent) y.1) . Further let 'Yo be the unique point in (0.1 (the role of ryy will be explained in Section 4b).Chapter IV The probability of ruin within finite time This chapter is concerned with the finite time ruin probabilities 0(u.g. The safety loading is q = 1/p . 0.f. 'y) where c(a) attains it minimum value. In particular. generalizations to other models are either discussed in the Notes and References or in relevant chapters. The notation is essentially as in Chapter III. 97 .s. See Fig. the Poisson intensity is 0 and the claim size distribution is B with m.
(4. 2 that E [T(u)k. EL refer to the exponentially tilted process 3 with arrival intensity S and exponential claims with rate / (thus .) = e7u ELe'Y^(u) ELT(U)k = e'Yu b ELT(u)k = O(u)ELT(u)k.u is the overshoot. PROBABILITY OF RUIN IN FINITE TIME Figure 0. PL = 6/0 = 1/p > 1). Var[T(u) I T(u) < 00] = VarL T( U) . 1 Exponential claims Proposition 1.. 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.r.(U) < 00] = ELT(u)ke'YS.t.1) (1. 7. using that the overshoot l. 1 FL. (u) is exponential with rate 0 w. .2) Proof Let as in Example 111. In particular. E[T(u) I T(u) < 00 ] = ELT (U). we have for k = 1.9). the time of ruin is T(u) and ^(u) = ST(t&) .(.98 CHAPTER IV. FL and independent of T(u). By the likelihood identity III.5 .1 In the compound Poisson model with exponential claims with rate S and safety loading 77 > 0.1 The claims surplus is {St}.
B = a. where = eBu I 1 .h. .1)) ELST(u) ELT(u) (PL . which leads to the quadratic 02 + (/3 .1) .1)T(u))2 = UL where = s. the 1.6 + a)0 .1//32 (6/)3 1)2 26(/3u + 1)/(6 .s."(ry) = 26//32. Since Sr (u) and (PL . Let 0 > yo be determined by ^c(0 ) = a.V/ is as asserted.12 Thus the l.2 In the compound Poisson model with exponential claims with rate 6 and safety loading rl > 0.(PL . is V1rLSr( u) +VarL ((PL . EXPONENTIAL CLAIMS For (1 .2).s.1.2) is aLELT( u) . 0 Proposition 1.3) B = 0(a) = + (6/3a)2+4a6 2 and hence that the value of ic(yo) Proof It is readily checked that yo = 6 .0) . Wald's second moment identity yields 2 EL (Sr(u) .6a = 0 with solution 0 (the . of (1.h.I (1. u + ELe(u) _ PL . T(u) < oo] fora > r.1)ELT(u). This means that /3(6/(6 .1)T(u) are independent with QL the same mean . 1).(yo) = 2V ./3) .1)T(u)) = VarLe(u) + (PL .1 /3u + 1 u + 1 //3 = 6/3 6/01 For (1. we have by Wald's identity that (note that ELSt = t(pL . the Laplace transform of the time to ruin is given by Eea7( u) = E [eaT (u)./3 .h.s.1)2VarLT(u) + 2 Ca 1I VarLT(u).1 (6)3)2 which is the same as the r.6.
3) we have E [e«T(u ).T+ Ti a t U T I 1 a i F.. 1.v. Ti. . More precisely.'s with rate 5.4) The interpretation of this that T(u) can be written as the independent sum of T(0) plus a r. are the ladder heights which form a terminating sequence of exponential r.. T(u) < oo] = EB [exp {aT(u ) . But by the fundamental likelihood ratio identity ( Theorem 111..1 . M(u) T(u) = T + E Tk k=1 where T = T(0) is the length of the first ladder segment . Cf. (1. Using 5 = 6 .. St Ti F. PROBABILITY OF RUIN IN FINITE TIME sign of the square root is + because 0 > 0).. and M(u)+1 is the index of the ladder segment corresponding to T(u).4..0.3 that we can write EeaT( u) = eeuEe 017(o).Y1 Y2 Figure 1. Y(u) belonging to a convolution semigroup ..9ST(u) +T(u)!c(0)} . are the lengths of of the ladder segments 2.3. .. the result follows. T(u) < oo] = e. Fig.100 CHAPTER IV. T2 . Y2.1 where Y1.v.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. Note that it follows from Proposition 1.
T) = P(VT > u) where {Vt } is the workload process in an initially empty M/M/1 queue with arrival rate 0 and service rate S = 1. .T + • • • + UN. Corollary 11.1. including the customer being currently served).. Then V(u.ST).T) 1 I fl(O)h(0) fdO where (1.4. T) to be evaluated by numerical integration: Proposition 1.0. U2. 2..e. Note that the case 6 # 1 is easily reduced to the case S = 1 via the formula V.6(u) = Vfl/j l(Su.v.6) fl(9) f2(0) = = fexp {2iTcos9(1+/3)T+u(/cos91) cos (uisin9) . .d.6.T are conditionally i.cos (u/.i.i (u.3 sin0 + 29) f3(0) = 1+/32/cos9.3 Assume that claims are exponential with rate b = 1.1 )!. 1).T. If QT = N > 0. Proof We use the formula .T the service times of the customers awaiting service . UN. let (cf. Hence 00 F(VT > u ) P(QT = N)P(EN > u) N=1 00 N1 k F(QT = N) eu N=1 k=1 °O u k! k Ee k=0 1t P(QT ... i.I ex cos B cos j O dB fo " .1(u.k + 1). UN. where U1.1. [4]) 00 (x/2)2n+3 Ij (x) OnI(n+j)! . the following formula is convenient by allowing t. the conditional distribution of VT given QT = N is that of EN where the r.T. . . T.. EN has an Erlang distribution with parameters (N.. Since U1 . and exponential with rate S = 1. For j = 0. then VT = U1. .T.. Let {Qt} be the queue length process of the queue (number in system.T is the residual service time of the customer being currently served and U2 . EXPONENTIAL CLAIMS 101 For numerical purposes . density xN lex/(N . cf.
44).1 R [. PROBABILITY OF RUIN IN FINITE TIME denote the modified Bessel function of order j.38).(31/2eie . 912.1)] L _112 /(k+1)/2 [.8 ) yields F(QT > k + 1) . 8789) 00 E aj j= 00 = 1.cos((k + 2)9)] d9.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 .(31/2 cos (( k + 2)9) .ie .31 /2eie L 1)] 1 I/31/2eie .cos (( k + 1)0)] f3(9) Hence the integral expression in (1.k + 1) = 1 k +1 + bj j=00 j=00 00 j=kk+1 j=k1 By Euler 's formulas.)3k+1 = e(1+0)T e201/2Tcos 7r 0 e )3(k +l)/2 [31/ 2 cos ( kO) .13(k +l)/2ei(k +1)9 R E .i(k +1)e R [/3( klal/2e:0 (01 /2 e . let I _ j (x) = Ij (x).112 l 1( k +1)/2 [ 31/ 2 cos(kO) . k k2 + $k+1 E bj 00 t j . Then (see Prabhu [294] pp.102 CHAPTER IV. similar formulas are in [APQ] pp. in particular equations (1. f3(0) .)3k +1 tj g'(QT >. 00 E '3j/2 cos(je) j=k+1 00 _ j=k+1 ^j/ zeij = .3(k +1)/ 2ei(k + l)6 (.1 00 ok+lR 00 j=k1 +1)/2e . (1. and define tj = e(1+R)Taj/2Ij(2vT T).
we are concerned with describing the distribution of the ruin time T(0) in the case where the initial reserve is u = 0. . and the next one (often called Seal's formula but originating from Prabhu [293]) shows how to reduce the case u 54 0 to this. it follows as in (1. however. or. T). E Fk. t) = P . t )). Related formulas are in Takacs [359]. equivalently. from the accumulated claim distribution N. We allow a general claim size distribution B and recall that we have the explicit formula z/i(0) _ P(7(0) Goo) = p. 2 The ruin probability with no initial reserve In this section .7) that _ [^ au ak+l (30 k L.. The rest of the proof is easy algebra. k! k=O k0 i/z Co Uk ate" o'/z e . oo (u)31/2e^e)k = )3k z cos(k9) = R k. The first formula.e = e' COS a cos(uf31/2 sin 0). there are several misprints in the formula there. the numerical examples in [12] are correct).2. Seal [327] gives a different numerical integration fomula for 1 .0(u.3 was given in Asmussen [12] (as pointed out by BarndorffNielsen & Schmidli [59]. T) in terms of F(. going back to Cramer. We first prove two classical formulas which are remarkable by showing that the ruin probabilities can be reconstructed from the distributions of the St.T) which. Ui < x I / (note that P(St < x ) = F(x + t. k=0 103 Cu) A further application of Euler's formulas yields cc k =0 k 'ese)k __ U #kJ2 cos((k + 2)9) = R eNO ^` (u^1 L k= = eup i/z L OI = =ateU161/2 e '0+2iO COS a cos(u(31/2 sin 9 + 20). THE RUIN PROBABILITY WITH NO INITIAL RESERVE Since P(QOO > k + 1) = flk+1. is numerically unstable for large T. expresses V)(0. however. F(x. u Notes and references Proposition 1.
T)dx.b (0.T)) 1 fT P(M(v. PROBABILITY OF RUIN IN FINITE TIME Theorem 2 . .i.T))dv. 1 1 . f T lStv)} 0<t<T by a 'cyclic translation'.1 In formulas.T) T F(x. Proof For any v E [0. See Fig.S„ 0 <t<Tv STS„+St_T+v Tv<t<T as the event that IS. ") } is at a minimum at time t. 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.t)= {Stv) < SM.T)) does not {Stv)} depend on v. 2.T))dv E^T I(M(v.0<w<t} St+v . T) = P(Tr(0) > T) = P(M(0.1. T T o where the second equality follows from II. co ).T]. resp.(. we define a new claim surplus process St StM NJ Figure 2. T]. [v. v].(0.104 CHAPTER IV. Then 1 .3) with A = (0. meaning that we interchange the two segments of the arrival process of {St}o<t<_T corresponding to the intervals [0.(6.
then M(v. 0<t <Tv}n{ST<ST Sv+St T+v.t)dt. T) = M(0.v<t<T} = {ST<StSv.. 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)) dv f T I(M(0. letting w = inf It > 0 : St_ = mino<w<T Sw}. t). T) as {ST<St+ vS. v).Sv.T) and Sv < 0 on M(0. Obviously. cf.T) = F(u+T. v). Fig 2. w) for some small E. v) = M(0. v<t<T}n{ST<STSv+St. 0<t<v} = {ST < St .2. there exist v such that M(v. in which case there is a last time o where St downcrosses level u. T) occurs. T)). THE RUIN PROBABILITY WITH NO INITIAL RESERVE 105 Now consider the evaluation of fo I(M(v. or it occurs. this integral is 0 if STv) .ST on M (0. Proof The event {ST < u} = { Ei T Ui < u + T j can occur in two ways: either ruin does not occur in [0.T) occurs or not as long as ST < 0. then i fT I(M(v.Tt))f(u+t. Hence T TE f I( M(v. t) denote the density of F(•. T. Indeed. v < t < T} n M(0. where the last equality follows from ST < St on M(0.T) occurs. If ST < 0. T Theorem 2 . T)) dv = TEST = T fP(ST < x) dx T T NT 1 f P(ST < x) dx = 1 f P Ui T .2. It is then clear from the cyclical nature of the problem that this holds irrespective of whether M(0. ST > 0. T T o i =1 Let f (•. We claim that if M(0.T)f(I z /)(0. T)) dv. we can take v E (w E. It follows that if M(0 .. T]. For example. .2 10(u.xdx. we can write M(v. T) occurs. v).
u + dt] and there is no upcrossing of level u after time t.3 Define r_ (z) = inf It > 0 : St = z}. The proof is combined with the proof of Theorem 111.u+dt]).106 CHAPTER IV.v.T)+ J0 T (1V.p. ST_ _ z} . Hence P(ST<u) = 1 .(0.z.2. Let Z be a r. {S t > z. define St = ST .T) = C(z. Then P(T(0) E • I T(0) < oo) = P(T_ (Z) E •).2 .T) = {St < 0.t). Proof of Theorem 111. PROBABILITY OF RUIN IN FINITE TIME u Q II T Figure 2. ST_ _ z}.ST_ t_ and let A(z. Proposition 2. 0 < t <T.Tt))P(StE[u. u which is the same as the assertion of the theorem.T) = . C*(z. {St > . E [t. 0 < t < T. O(T . The following representation of T(0) will be used in the next section. z > 0. which occurs w. 0 < t <T . which is independent of St and has the stationary excess distribution B0.2. ST_ _ z}. t + dt] occurs if and only if St E [u. For a fixed T > 0.b(u.2 Here o. 2.
z + dz]. Fig. ST(o)_ E [z. u which is the assertion of Theorem 111.T))f3B(z) dz dT.1) yields P(ST(o)_ E [z.2. and since {St}o<t<T.T). 2. z + dz].T + dT]. Hence integrating (2. {St }o<t<T have the same distribution .T)) = P(Cx. T(0) < oo) B(y B(z) + z) f3B(z) dz = 3 f °^ B(y + z) dz = f3 + x v f B(z) dz. A(z.T(0)<oc) = f x F(U > y + z U > z) P(Sr(o)_ E [z. z + dz]. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Then 107 P(r(0) E [T.T))dT = Off(z) dz P(T_ (z ) < oo) = 3B(z) dz.3. Figure 2. r(0) < oo) = 3R(z) dz JP(C(z. Thus P(Sr(o)_>x. T + dT] I S7(o)_ E [z.3 But by sample path inspection (cf. It follows by division by P(ST(o)_ E [z. Proof of Proposition 2.T)). we therefore have P(A(z.1) that P(T(0) E [T.1) z T .2. T(0) < oo) = OR(z) dz in (2. (2. z + dz]) = P(A(z.ST(o) >y. 7( 0) < oo) = P (C(z)) dT. z + dz].T) = C*(z. .3).2.
1 and the present proof is in the spirit of Ballot theorems. cf. 2.5a). T(0) < oo) 0 = dT f 0 P(C(z))P(Z E [z.1) where a > r. who instead of the present direct proof gave two arguments. r(0) < oo.(3(B[r( a)] . Proof Optional stopping of the martingale I er (a) 9 t. a martingale proof is in Delbaen & Haezendonck [103]. because of77>0.3 was noted by Asmussen & Kl(ippelberg [36].c(r(a)) l = l er( a)se+at } u yields 1 = eyr(a)Eear(y). z + dz].5 and one upon excursion theory for Markov processes (see IX.1 Eear( y) = eyr(a).r(a).1) .2. Notes and references For Theorems 2. 3 Laplace transforms Throughout in this section. Tak'ecs [359].3. I L Let ga(x) be the density of the measure E[ear(°).108 Hence CHAPTER IV. one based upon a result of Asmussen & Schmidt [49] generalizing Theorem 11. In the setting of general Levy processes. Lemma 3 . ^(0) E dx] (recall that ^(0) = Sr(o)) and write ga[b] = f OD ebxga(x) dx. see in addition to Prabhu [293] also Seal [326].s. PROBABILITY OF RUIN IN FINITE TIME ]P(7(0) E [T. z + dz]. [329]. r(a) denotes the solution < 'Yo of the equation a = ic(r (a)) = . Theorem 2. T(0) < oo) = dTP(T_(Z) E [T. some relevant references are Shtatland [338] and Gusak & Korolyuk [181]. Lemma 3.2 ga(x) = Qexr(a) f "o eyr(a)B(dy) x . Note that T_ (y) < oo a. Proposition 2.T+dT]). Let T_ (y) be defined as Proposition 2.T + dT] T(0) < oo) dT f ' P(C(z))P(Sr( o)_ E [z.6. (3.(yo).1.
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)) . Further by Theorem 111.x)(a) B(dy)• Lemma 3 .T(0) < oo] = 20[b] = za[b] (9a[b] 9a[0])/b 1 . E[ear (o) I T(0) < oo . Hence eb"du E[eaT(").x)ga (x) dx where za(u) = f.2.3. b . the result follows after simple algebra. r(u) < oo). (Laplace transform) of the ruin Corollary 3. u . time T(u): u u Here is a classical result : the double m. rr(0) < oo) = 1_ r(a) Proof Let b = 0.r(a) = a [B[b] B[r(a)]] .g.f.ga [b] 0 TO Using Lemma 3. Z = y] = EeaT.ic(b)/b x(b) + a eb"E[eaT(" ).3 ga[b] = c(b) Proof + b + a .5 f 00 o a/r(a) .ST(o)_ just before ruin .ga [b] 1 . Corollary 3.3.1] evr(a)B(dy)[ b . T(u) < oo] du = Proof Define Za(u) = E [eaT(" ).r(a) The result now follows by inserting /3B[s] = ic( s) +/3+ s and ic(r(a)) =a. LAPLACE TRANSFORMS 109 Proof Let Z be the surplus .2 P(Z E [y. y + dy]. £(0) E dx) = /3B(x + dy) dx and hence ga(x) = f e r)/3B(x + dy) _ /3 f x e(v.(v) = ev''(a). Then by Proposition 2.4 E[eaT (o).3.°° ga(x)dx. (u .r(a) b . It is then easily seen that Za(u) is the solution of the renewal equation Za (u) = za (u) + fo Z.
Then given r(u) < 00. Proposition A1.e.2 Assume ri < 0.1)Er(u) . (u) t.r(u) = Er(u) • ES.s. By Proposition 111.mu D 2 4 N(0..1) i. T(u)/u mL as u + oo. t T(u) T(u) T(u) t m = lim = lim = lim Utioo u + Sr(u) u+oo S. for any m T(u) u . mu ) ( 0 m < ML '(u) 1 m > rL.2. = (p . Then as u * oo. u 1 ET(u) 1 p1 u where Pw2 = 311B)m3• 7(u) . For the second . cf. T(u) a. PROBABILITY OF RUIN IN FINITE TIME 4 When does ruin occur? For the general compound Poisson model. Theorem 4 . For the proof. the known results are even less explicit than for the exponential claims case. (4.1. where _ 1 _ 1 1 C ML w(ry) 6B'[7J 1 . and take basically the form of approximations and inequalities. The first main result of the present section is that the value umL. for any c > 0 P( Further.6. That is. i.1 Assume 77 > 0. This proves the first assertion of (4. St/t 1 1/m.(u ) = o(u) a. and hence a.h(u.s. we need the following auxiliary result: Proposition 4. uoo u using e. P = /3µB > 1. Later results then deal with more precise and refined versions of this statement.110 CHAPTER IV.mL > E T(u) < 00 ) 40.3).00 St = lim ..w ) v/. note that by Wald's identity u + EC(u) = ES.UProof The assumption 11 < 0 ensures that P(T(u) < oo) = 1 and r(u) a4' oo.3LELU 1 1p' is in some appropriate sense critical as the most 'likely' time of ruin (here C is the CramerLundberg constant).
5) St . Tu) T( u) .2 of [86]) and (4.mL U > E.1).4. For (4 .mu (2) '• m3/2 µB 7 .1).r(u)/m T(u) ti µB2) Z.7 6  11 Proof of Theorem 4.g. . proving (4. WHEN DOES RUIN OCCUR? and that Ee(u)/u a 0. 4a Segerdahl's normal approximation We shall now prove a classical result due to Segerdahl. T (u) < 00 J 0(u) e7'PL U \ I T u) . this can be rewritten as u + 1(u) .3. note first that ( Proposition 111.1). the result comes out not only by the present direct proof but also from any of the results in the following subsections. cf. T) for T which are close to the critical value umL). apB ) .6.1 (by considering 0(u. Notes and references Theorem 4.1 The l. the same conclusion holds with t replaced by r(u). 4). again Proposition A1. PL (•)+ 0. which may be viewed both as a refinement of Theorem 4. T(u) < oo f / 00) e7uE L [e_7 (t1).6µB2) Z v m (3µB2) Z. implying T(u) .h. According to Anscombe' s theorem (e.t/m D (2) 111 . If Z .mL >E By Proposition 4.2) follows immediately from u (4.1) is T (u)  U mL P( T (u) < I > E. though it is not easy to attribute priority to any particular author. and (4.3). 1'r(U) . and as a timedependent version of the CramerLundberg approximation. Thus.^ N (o.mu m .s.2. Theorem 7. of (4.1 is standard.1.N(0.
e(u') oo w .4).f ( (oo)) .v. we get E[ T (u) .L+YWLV'U) . Proof Define u' = u .(u. Hence Ef (Vu )) 9 (T(u. letting Z be a N(0. (oo.t.3). Using ( 4. Let h(u) = E f (^(u)).a C4'(y )• ( 4.mul h(oo)Eg(Z).6) whenever f.ul/4. PROBABILITY OF RUIN IN FINITE TIME Corollary 4.ST( u') = u1/4 .112 CHAPTER IV.w2) r. then e(u) and r(u) are asymptotically independent in the sense that. with w2 as in (4. S( u ) < ul/4] < ET(ul / 4) = O(ul/4). we need the following auxiliary result: Proposition 4. one has 9 (r(u)_rnu) Ef (^(u)) * E. and thus in (4. we can replace T(u) by r(u').r.T(u') given F.) is readily seen to be degenerate at zero if ST(u•) > u and otherwise that of T(v) with v = u . using that ul/4 . O . oo).4 (SIAM'S LEMMA) If 71 < 0. resp .5) For the proof.l:(oo) (recall that rt < 0).))I h(ul /4  ^(u)) I(6 (u') C ) f < ul /4 + f(e(u') . e'°'/b (u.um.T ( u')] = E[ T ( 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). g are continuous and bounded on [0. P because of ^(u') . Then h(u) 4 h(oo) = E f (6(oo)).^(T(u')). Then the distribution of T(u) . oo ). E9(Z) (4.u1/4)I(S(u') > u1 /4) h(oo) + 0. and similarly as above we get E[f(^(u)) I Fr(u.6).VU T.)mu \ h(oo)Eg (r(ul) . Then for any y.
3 in terms of Edgeworth expansions .(ay) = 17 7y = ay . For refinements of Corollary 4.7) whenever u is large and ly(T)l moderate or small (numerical evidence presented in [12 ] indicates . WHEN DOES RUIN OCCUR? Proof of Corollary 4.4) in the last. Cf. see also von Bahr [55 ] and Gut [182].1. Theorem 4. 3 is due to Segerdahl [333].5) and solve for y = y(T). Segerdahl 's result suggests the approximation b(u. CL Fig. 4b Gerber's time. that for the fit of (4.3 ery"z/i(u .7) To arrive at this . oo ) as u * oo. 10) '5(u) . u needs to be very large). however . y u) < . just substitute T = umL + ywL in (4. The present proof is basically that of Siegmund [342]. also Hoglund [204].z/)(u . see Asmussen [12] and Malinovskii [254]. in practice one would trust (4. PL(T(u ) < umL + ywL) 113 4 C4(y).umL wI V"U u (4.7) to be good.dependent version of Lundberg's inequality For y > 0. yy by 1 K. umL + ywL f) = e"P(T (u) < umL + ywL) = EL [e7V "). where we used Stain's lemma in the third step and (4. T(u) < umL + ywL f. y u) < e 7v" .9) ( 4 .oo. For practical purposes . e7v" y < ^'(7) (4 . 0. define ay.T) Ce7"4 (T .5 '(u . Thus .8) Note that ay > 7o and that 7y > •y (unless for the critical value y = 1/ML). Notes and references Corollary 4 .4. ELe7E (") . y > k'(7) .yK(ay)• (4. .7) to be valid is that T varies with u in such a way that y(T) has a limit in (. The precise condition for (4.
Then ic(ay) > 0 (see Fig .6 It may appear that the proof uses considerably less information on ay than is inherent in the definition (4.Y' (u. From the proof it is seen that this amounts to that a should maximize ayic(a). u Differentiating w. For a different proof. and hence t. f Some urther discussion is given in XI.b (u. and generalizations to more general models are given in Chapter VI. Hoglund [203] treats the renewal case.yu ) = eayuEav [e .r. who used a martingale argument. yy is sometimes called the timedependent Lundberg exponent. the point is that we want to select an a which produces the largest possible exponent in the inequalities.h(u. if y > 1/ic'(y). 0.t. yu 11 < T(u) < oo j < eayu +Y UK(ay) Remark 4. Numerical comparisons are in Grandell [172 ]. yu < T (u) < oo 1 l e ayuEav [eT ( u)K(ay). the bound a7y° turns out to be rather crude . dy) Notes and references Theorem 4 . In view of Theorem 4.6. yu) < C+(ay)e7a„ where l C+(ay) = sup f 00 eayR(xy)B( . which shows that the correct rate of decay of tp(u.1).9): Proposition 4. we arrive at the expression in (4.114 CHAPTER IV. yu ) = < eayuEay [eay^ ( u)+T(U)K ( ay). An easy combination with the proof of Theorem 111.8). see MartinLM [257] . a.8).7 i. T(u) < yu] < eayu + yUr(ay) Y < eayuEav [ eT(u)K(av )L T(u) < yu} Similarly. which may be understood from Theorem 4.5.8 below . However.2. . we have rc(ay) < 0 and get (u) .v"U.3 yields easily the following sharpening of (4. 5 is due to Gerber [156 ].ay4(u)+ T(u)K(ay ). yu) is e 'Yyu/ . PROBABILITY OF RUIN IN FINITE TIME Proof Consider first the case y < 1/K'(y).
c(&) = ic(ay) is < 0.13) .ay a.yyu y l ay I 21ry/3B" [ay] V fU_ u + 00. we have ryas = ay . i.'(y ). yu) = e. then ay > 0..ay and get Ea e ayf (00) y _ 'Ya( ayKal lay C 1 .ay y 'Yay  ay .6 with P replaced by Pay and FL by Pay.ayuEay f eay^ ( u)+T(u)K(ay). (0) r1 (a) ' I. it is instructive to reinspect the choice of the change of measure in the proof. This idea is precisely what characterizes the saddlepoint method.i(u.2 yields EaT(u) u u r.4. The traditional application of the saddlepoint method is to derive approximations. if we want EaT(u) . not inequalities.yu) c ay .e. T(u) < yu] . For any a > yo.8 If y < 1/ic'(ry). (4.: T.e.(u.ayC() . We thereby obtain that T is 'in the center' of the Padistribution of T(u). (4. then the solution &y < ay of . u 4 oo.^3 ]1/ Bay [lay . Proposition 4. yu ) eaauEaye . and b(u.. the formula 0(u. yu ) ayay e ryyu ayay 27ry/3B"[ay] u Proof In view of Stam 's lemma. the choice of ay.5. and in case of ruin probabilities the approach leads to the following result: Theorem 4 . and ii(u) . [eT(u )K( ay). WHEN DOES RUIN OCCUR? 115 4c Arfwedson's saddlepoint approximation Our next objective is to strengthen the timedependent Lundberg inequalities to approximations.z..12) < yu] Here the first expectation can be estimated similarly as in the proof of the CramerLundberg ' s approximation in Chapter III. Ea . Using Lemma 111. T(u) suggests heuristically that l t/.11) ' If y > 1/ r . As a motivation. (4. then the relevant choice is precisely a = ay where y = T/u.
1.a)2 .1)3 = y3/3B"[ay]. Then ic(a) = .a. .ay ) r. and the equation ic'(a) = 1/y is easily seen to have .13).1B[ay]1 ) y(ay .c(ay)ul/2W p 2ir = eyu(ay) dz 1 rc(ay ) 2.ay + ayl /BLay] . T(U) < yu] = eyuk (ay)E''ay (ek(ay )"1/2WV. The proof of (4.4).ay) ay +.I ay &y a ^c'(ay) a (1 +.12) is 0 entirely similar.1)3 = (jB"[ay]l (Pay . it seems tempting to apply the normal approxiyu + ul/2wV. The difficulties in making the proof precise is in part to show (4.11) follows. we get heuristically that Eay Ler (u)r(ay).(ay) _ y(ay . (ay) J0 1 K(ay )u 1 00 c2(x) dx /2 w 1 ezcp(z /( k(ay)u1 /2w)) dz /O° _ 1 1 J e Z . PROBABILITY OF RUIN IN FINITE TIME ry I i . 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). V < 01 Ir 00 er(ay)"1'2"'x eyur.7ruw2 Inserting these estimates in (4.3(5/(S . Example 4.(j (1 . where V is normal(0.l'B)y /(Pay . Writing r(u) and W2 = I3ay{.a) .1) .13) rigorously.13).B[ay] /ay &y y(ay . a nr=.1) under Pay mation (4. i B[7ay .9 Assume that B(x) = eay.ay)K(ay) ay ayI&YI For the second term in (4. (4.116 CHAPTER IV.c'(a) _ /3a/(8 .
then { __ . c a 00./4 ^y for 1/i (u.ay)3 0 3/2 and (4.1) .5.tcp) Lo {Wo ( t)}t>0 . The mathematical result behind is Donsker's theorem for a simple random walk {Sn}n=o. 2 = Var(Si ) the variance. yu) when y < 1/ic'('y) = p/1 . (5. .11) gives the expression '31/4 ( .i )( v s vc ('3 + s _2 / .because the c.3+52 1+/351/y' sy 7 B ii[ay] 25 _ 251/2(1 + y)3/2 (5 . 0 Notes and references Theorem 4. A related result appears in BarndorffNielsen & Schmidli [59].. in discrete time: if p = ES.p. 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 =.8 is from Arfwedson [9]. DIFFUSION APPROXIMATIONS solution ay=5 117 V 1 (the sign of the square root is . is undefined for a > 5).g. 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. and next to note that such an approximation in particular implies that the first passage probabilities are close.. is the drift and o.1. y) a''y" L '3 _ fl ) 51 /4(1 +1IY)3/4 \.= (s.f..
and this can be obtained under the assumption that the safety loading rt is small and positive.a = Snp) and the inequalities Sn )C .1.3) whenever c = cp f oo as p 1 p.. Indeed . Mathematically.. However.t} _ {W_1(t)} . a2 =/3µB2) Proof The first step is to note that { WC (St P) . Letting c = a2/pp. cf.1 below). Lemma 111. 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. where p is the critical premium rate APBTheorem 5 . for the purpose of approximating ruin probabilities the centering around the mean (the tcp term in (5.p.1) with S. this is an easy consequence of (5. n/c < t < (n + 1)/c.7c). (5. PROBABILITY OF RUIN IN FINITE TIME where {W( (t)} is Brownian motion with drift S and variance (diffusion constant) 1 (here 2 refers to weak convergence in D = D[0.1 As p J. St = EN` U= .p/c < St(p) < S((n+l)/ c + Pp/c.118 CHAPTER IV.z } {W_1(t )}t>o (5. such that the claim size distribution B and the Poisson rate a are the same for all p (i. It is fairly straightforward to translate Donsker's theorem into a parallel statement for continuous time random walks (Levy processes). We want an approximation of the claim surplus process itself.e.1)) is inconvenient. p. + {Wo(t ) .tp).2) t>o where p = pp = p .3.3) takes the form LI S(P) { a2 to2/µ2 + t LI S (P) { a2 ta2/µ2 {W0(t)}. of which a particular case is the claim surplus process (see the proof of Theorem 5. oo)). we have o {i!t s: .tcpp) y = { WC (Sct) pct) } {Wo( t)}t>o (5. we shall represent this assumption on 77 by a family {StP) L of claim surplus processes indexed by the premium rate p. 0 . and consider the limit p j p.
this implies P sup 0<t<T a 12 Stu2 /µ2 > u 4 P ( sup W_1( t) > u O<t<T But the l.1.h. (5. is IG(T. C. 196. (ua2 To2 op \ IPI > IG ( T . ^ p2 Proof Since f 4 SUP0<t<T f (t) is continuous on D a. since ti(u) has infinite horizon . the continuity argument above does not generalize immediately. Corollary 5. For practical purposes . 199.r. .. any probability measure concentrated on the continuous functions. has a continuous distribution. the continuous mapping theorem yields sup W Sz2 to lP 4 sup Wi(t)• O<t<T O<t<T a2 Since the r.h. Corollary 5 .. u).s.2 suggests the approximation u 0(u. [169] or [APQ] pp. u) of r( (u) (often referred to as the inverse Gaussian distribution) is given by IG(x.h.( ^ I + e2( \ I . and in fact some additional arguments are needed to justify (5.5). TS(u)=inf{t>0: WW(t)>u}. u) =PIT( (u) < x) = 1 . 1.Ta2 /p2). Because of the direct argument in Chapter III. w. (.2 As p j p. ulpI/a2).8 or [APQ] p.e.1 I 7= . (5.6) This is the same as the heavy traffic approximation derived in III.T) IG(Tp2/ a2). However.6) from Theorem 5. and the r. we obtain formally the approximation V.f I \\\ J \ (5.t.1.5. DIFFUSION APPROXIMATIONS Now let Tp(u) = inf{t>0: S?)>u}.7c. we omit the details .s.s. see Grandell [ 168].4) Note that IG(. u) is defective when < 0.1 .5) Note that letting T * oo in ( 5.(u) ti IG(oo. ulpl /a2) = e2"1µl / or2.. 263) that the distribution IG(•. 119 It is wellknown (Corollary XI. ('. is 1/ip (ua2 /IpI.u).
The picture which emerges is that the approximations are not terribly precise. See for example Billingsley [64].Po = 09µB6 .1 and Section VIII. that 00 4090. and two further standard references in the area are Grandell [168]. as an example of such a generalization we mention the paper [129] by Emanuel et al. PROBABILITY OF RUIN IN FINITE TIME Checks of the numerical fits of (5. Then as 0 _+ 90. on the premium rule involving interest. pt? 4 peo. e.t.Pe. Michna & Weron [152] suggested an approximation by a stable Levy process rather than a Brownian motion. 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. for more general models it may be easier to generalize the diffusion approximation than the CramerLundberg approximation.g.5) combined with the fact that finite horizon ruin probabilities are so hard to deal with even for the compound Poisson model makes this approximation more appealing. 0) { 2 StQ2 /µ2 D { W_ i(t)}t>o t>o D 2 where p = pe = pe . Furrer. [169]. in Asmussen [12]. However. The proof is a straightforward combination of the proof of Theorem 5.5) and (5. Further relevant references in this direction are Furrer [151] and Boxma & Cohen [75]. the claim size distribution B9 and the premium rate p9 depends on 0. in the next subsection we shall derive a refinement of (5.6) therefore does not appear to of much practical relevance for the compound Poisson model. the B9. and which is much more precise. In view of the excellent fit of the CramerLundberg approximation. We conclude this section by giving a more general triangular array version of Theorem 5. Theorem 5. as 0 * 00 and that the U2 are uniformly integrable w. . B0 * Boo.6 of [APQ]. the simplicity of (5.120 CHAPTER IV.3 Consider a family {Ste) } oc claim surplus processes indexed by a parameter 9. such that the Poisson rate Oe. In contrast.r. we have ^A. For claims with infinite variance. However. The first application in risk theory is Iglehart [207]. Assume further that 039µB6 < pe.00µB6 + 0.. pe . (5. in particular for large u.1. All material of this section can be found in these references.6) are presented.
this means the following: 1. claim size distribution B . Then EOU' = Boki[0] = Biki[eo]/E[9o] and "(s) = k(sBo)k(9o).1) .90) and the given risk process corresponds to Poo where 90 = 'yo. whereas there we let the given 3B.c(s) = . In terms of the given risk process with Poisson intensity . 3. let P9 refer to the risk process with parameters Q9 = QoB0[9] = QB[9 9o].Q (B[s] . Bo(dx) = B[eo]B(dx). . The objective of the corrected diffusion approximation is to take this and other deficits into consideration.ao (0) _ /c(s + 9 .6. Determine yo > 0 by r. Let PO refer to the risk process with parameters e9oz Qo = QB[90]. Since Brownian motion is skip free. 77 = 1/p .1 > 0. and we are studying b(u.4. and we want to consider the limit 77 10 corresponding to Oo f 0. 9o T 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. PB('r(u ) < oo) < 1 for 9 < 0.T) = Peo(r(u) < T) for 90 < 0. which we have seen to play an important role for example for the CramerLundberg approximation . Then r. P9(r (u) < oo) = 1 for 9 > 0. risk process with safety loading 77 > 0 correspond to 9 = 0 . B9(dx) =Bale] Bo(dx) e9z keo)z = B[9 . .90) ./c(9 . However . 0(0) = 0.'(yo) = 0 and let 90 = 'Yo. 77 is close to zero. In this setup. The setup is the exponential family of compound risk processes with parameters ( B9 constructed in III.90] B(dx).9(s) = Ico ( s + 9) . this idea ignores (among other things) the presence of the overshoot e(u). it is more convenient here to use some value 9o < 0 and let 9 = 0 correspond to n = 0 (zero drift).s and p = /3µB < 1. this is because in the regime of the diffusion approximation . 2. For each 9.6.
_ ^(u) = ST .u. means up to o(u1) terms): . u) = IG(x/u2.Varo S1 = f30Eo U2 = S1.1) IG(x.(y) = 0. . The first step in the derivation is to note that µ = k (0) = r0 (00) . u) = euh(a . Vargo S.3) this implies (take u = 1) Ego exp { .7(u)/u2} eh(A.. . (. (01. for brevity.T) 1+u2 (6. the solution of r. i.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. and Si = QoEoU2 = Q B"'['Yo Eo U3 ]. S2 = 3E0U2 Bier [Yo] 3B"[Yo] Write the initial reserve u for the given risk process as u = C/Oo ( note that C < 0) and. (. (6.2) . The corrected diffusion approximation to be derived is (u.1) . 0o to. (U. C) = 2A + (2 .. C .. IGu+u2. C. () where h (A. bl IG(t81. PROBABILITY OF RUIN IN FINITE TIME Recall that IG(x.C.. 1) • Since L eatIG (dt. Theorem 5. tu2 ) i IG (t. 9otc0" (0) = 0061 = ul.122 CHAPTER IV.() The idea of the proof is to improve upon this by an O (u1) term (in the following. One has (6.e. 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. write r = T(u).2' where as ususal ry > 0 is the adjustment coefficient for the given risk process.S.
the r.s.s. 1.1 As u + oo. distributed as Z . 6 . of (6. we have p =. 1% in (2) and (4). In ( 1) and (2). in (3) and (4).5) Once this is established .f.h.z . . however.4.v. we get by formal Laplace transform inversion that C 2 u.v. . of a (defective) r. (6. Note. . the formal Laplace transform inversion is heuristic: an additional argument would be required to infer that the remainder term in (6.3).1 below is exact. CORRECTED DIFFUSION APPROXIMATIONS 123 Proposition 6. But the Laplace transform of such a r.5) according to (6.d. is the c.2). it holds for any fixed A > 0 that Ego exp { Ab1rr(u)/u2} . u is Eeazead2/++ Eeaz[1 + ab2/u] where the last expression coincides with the r. A numerical illustration is given in Fig.3. 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. which is based upon exponential claims with mean µB = 1.'yu /2)(1 + b2/u)} + Aug 1I J .h. bl I IG I t +2 . 9o T 0 in such as way that C = Sou is fixed. To arrive at (6. calculated using numerical integration and Proposition 1.2 ).52/u where Z has distribution IG (•. just replace t by Tb1/u2. .2) is indeed o(u1). p = 0. The initial reserve u has been selected such that the infinite horizon ruin probability b(u) is 10% in (1) and (3).3 = 0.1 + 629. that the saddlepoint approximation of BarndorffNielsen & Schmidli [59] is a serious competitor and is in fact preferable if 77 is large] . that whereas the proof of Proposition 6. and the dotted line the corrected diffusion approximation (6.7. The solid line represents the exact value .exp { h(A.1 + u2 I Indeed.ry2 . however .6.
.08 0.4 may not be outstanding but nevertheless. the fit at p = 0.T) 111 0. and all of the numerical studies the author knows of indicate that its fit at p = 0.. see Asmussen [12].111 W(U.02 I 90 120 160 2W A0 Z WT 40 80 120 160 100 240 280 T 111 WI.00 0.OOIi O.114 0.(061 0.7.1 proceeds in several steps.011 L1 60 T IM 11.T1 00.TI CHAPTER IV.05{ 0.07 0.199 0. (Inc 0s 0. A51 7(SAT 3 3 h(X.T) 0.124 0.2 e.W21 0.01 0.0 0. For further numerical illustrations.1 W IU.08 a. Note that the ordinary diffusion approximation requires p to be close to 1 and '0 (u) to be not too small. The proof of Proposition 6.() Lemma 6.1 It is seen that the numerical fit is extraordinary for p = 0. Similarly. OM 0.7 or at values of Vi(u) like 1% is unsatisfying. PROBABILITY OF RUIN IN FINITE TIME 0..EB 0 p ex p ( 7 S h ^)u . BarndorffNielsen & Schmidli [59] and Asmussen & Hojgaard [34].^) .19)2 11 20 20 i0 T 1n0 Figure 6. it gives the right order of magnitude and the ordinary diffusion approximation hopelessly fails for this value of p.aa1 .u2 2u3 (e .
T (co (8/u) . (6.2u (B3 .4 Ea. () 62 Eeo exp u u2 J .(3)Eea LauT exp i 3J . in Lemma 6. 1 = PB(T < oo) = Eo0 exp 125 {(B .. 3 lim Eof (u) = EoC(oo) = a2 Ep = 3EoU2 uroo Proof By partial integration .7) 2 2 ..61a2T (B3 ..4) that the r.co (e) ..2 behaves like C l Eeo eXp r _ ^81T 1 Sl u2 1 u 2u3 [1+h(AC) S . exp ue } al 1J 3 exP I [2). CORRECTED DIFFUSION APPROXIMATIONS Proof For a>0.1) h(A.co ((/u)) } Let 8 = (2a + (2)1/2 = h(). the result follows.+ h (A. + a1b2 + . 1 / Po(C(0) > y) dy EoC(0) x k EDUk + 1 k Eo[(0)k+1 EoC(0) _ (k + 1)EoU' EoC(^) _ (k + 1) Eo£(0) Lemma 6 .r0 (00)) } Replacing B by 8/u and Bo by C/u yields e(B() = E eo exp { (e .h.00)(u +C)  'r (.C2 = 2).s.6.C)C/u . the formulas Po(C(0) > x) Po(C(co) > x) imply 1 °° Po(ST(o) > x) = EIU fIP0 (U>y)dy .3 EoU2 + 103OoEoU3 + " 2 6 Using d2 . C) 1 1 + u2/ 111 + 2u CZ Z  (2A + ()1/2 J 1 Proof It follows by a suitable variant of Stam's lemma (Proposition 4. (6. () + C and note that 2 KO (0) = 102.(3) J t _ aa1T l + eh(A.6) u U3 Lemma 6 .
l Lemma 6 .\+ (2 (3 e 2u [ (2. 2 + 00 = .2.e h(aS)h (^^ 262 exp {_h(. [2+ (2 .126 CHAPTER IV. There are two reasons for this : in this way.1 (y/2 + Oo)u . yu/2) h(A.2) for O(u) (indeed . we get the correct asymptotic exponential decay parameter ^/ in the approximation ( 6.h. yu/2) 11+ 62 I} S 1 \\\ u/11 l 62 (3 2u 2A Proof Use first (6.S) d e 62 .. and the correction terms which need to be added cancels conveniently with some of the more complicated expressions in Lemma 6. Thus a2 y = 290 + O (u2). and inserting this and 9o 2 = S/u on the r. 0 The last step is to replace h(A.\ + () 1 2 / .6  d h(A. () . C) ( 1+ u2 The result follows by combining Lemma 6 .6) and 7co (Oo) = ico('y + Bo) to get 0 = 21 (^/2 + 2y90) + 1112 (_Y3 + 3_Y200 + 3y9o) + O(u4). 2 and (6.(2A + ()1/21 exp S h(A.() I 1 + u2 ) y . 5 exp { _h(A) (1 + / y u J)) exp 1.2u [2A+ (2 3 . letting formally T * oo yields 7/)(u) C'e7u where C' = e7a2). yields +90 62 0 + O(u 3) 2u2 +O(u 3).4. () by h(\. Thus by Taylor expansion around ( = 90u. we get h(A. PROBABILITY OF RUIN IN FINITE TIME The last term is approximately (e 3 (3) 27.s.2 (^/2 + 3y9o + 390) + O(u3).() . yu/2).x.7) and using eh(a.h (A.
HOW DOES RUIN OCCUR? exp { h (x. 'yu/2) 127 ( i+ M pz^ exP { h (A.T) has not been carried out and seems nontrivial. and to the Markovmodulated model of Chapter VI in Asmussen [16]. Hogan [200] considered a variant of the corrected diffusion approximation which does not require exponential moments. 0 1 Proof of Proposition 6.e. The adaptation to risk theory has not been carried out. the approach to the finite horizon case is in part different and uses local central limit theorems. that is.1 (y/2 + Oo)u )} 1 (i + U ) [2+ C2 2u 62 S Pt^ exP { J 62(2 exp { h(A.5 in Lemma 6. The answer is similar: the process behaved as if it changed its whole distribution to FL. this case is in part simpler than the general random walk case because the ladder height distribution G+ can be found explicitly (as pBo) which avoids the numerical integration involving characteristic functions which was used in [345] to determine the constants. the same as for the unconditional Lundberg process. We shall now generalize this question by asking what a sample path of the risk process looks like given it leads to ruin. the analogous analysis of finite horizon ruin probabilities O(u.1: Just insert Lemma 6. . () (i+a ) 2A + (2 .7. In Siegmund's book [346]. ()} 3 h (A. ()} . 7 How does ruin occur? We saw in Section 4 that given that ruin occurs. i.(i+ 62 exP{ h(A. the 'typical' value (say in sense of the conditional mean) was umL. () I 1 + u 2 ) } S 1 . with the translation to risk processes being carried out by the author [12]. Fuh [148] considers the closely related case of discrete time Markov additive processes. The corrected diffusion approximation was extended to the renewal model in Asmussen & Hojgaard [34]. u Notes and references Corrected diffusion approximations were introduced by Siegmund [345] in a discrete random walk setting. () I 1 + u2 )I 2u L 2A+C2_(2 exp { _h. His ideas were adapted by Asmussen & Binswanger [27] to derive approximations for the infinite horizon ruin probability 'i(u) when claims are heavytailed.4.
3L and the claim size distribution from B to BL. P(u) and rate = aL w. ^(u) is exponential with rate 8 w.3 to . . the Poisson rate changes from . F(u)c] ti e' ru]PL (F(u)`) > 0. (u) and satifying PL(F(u)) * 1. Then also P(u)(F(u)) + 1.(u)_ and similarly the denominator is exactly equal to Ce7u.vi(u) Ce'Yu Corollary 7. F(u)c] P(r(u) < oo) ?P(U) < EL[e7u.T. PROBABILITY OF RUIN IN FINITE TIME changed its arrival rate from 0 to /3L and its claim size distribution from B to BL.128 CHAPTER IV.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. Proof P(u) (F(u)c) = F(flu)c. In the exponential case.(u)_ and ^(u) are independent . Recall that .t. FL As example.1. Note that basically the difference between FT(u) and .r. stating roughly that under F(u).TT(u) _measurable.FT(u) = o' (T(u ). so in the in the proof.F.ST(u)}t> o is just an independent copy of {St}t>o).. {ST(u)+t . r(u) < oo) . we give a typical application of Theorem 7. Theorem 7 .1 Let {F(u)}u>0 be any family of events with F(u) E F.r. 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.2 If B is exponential.. {St}0< t<T(u)) Proof Write e'rsr(u ) = e'rue'r£(u).t. In fact. . then P(u) and FL coincide on . + TMOO ).EL[e7S. .(u) is not . the numerator becomes e'ruELe7^ (u)PL(F( u)t) = e7uCFL (F(u)°) when F(u) E .(u)_ is that i. u * oo. Recall that 13L = (3B[ry] and BL(dx) = e'rxB(dx)/B[7].F.(u). and let M(u) be the index of the claim leading to ruin (thus T(u) = Ti + T2 + .
however. Proof For the first assertion. From a mathematical point of view.(1 . see further XI. This is currently a very active area of research. the queueing results are of a somewhat different type because of the presence of reflection at 0. A somewhat similar study was carried out in the queueing setting by Anantharam [6]. .eaLx. the subject treated in this section leads into the area of large deviations theory.3 M(u) pcu) 1 . Notes and references The results of the present section are part of a more general study carried out by the author [11]. who also treated the heavytailed case.eALx) M(u) k=1 u The proof of the second is similar.7. HOW DOES RUIN OCCUR? Corollary 7. 129 M(u) >2 I(Tk < x) M(tu) p(u) M(u) >2 I(Uk < x) BL(x). take I(Tk < x) .3.
This page is intentionally left blank .
with common distribution B. .(1. In the socalled zerodelayed case. We use much of the same notation as in Chapter I. the distribution Al of T1 is A as well. .1 (T1 = a1). Thus the premium rate is 1. . the claim sizes U1. U2... and the one corresponding to T1 = s by 1/i8 (u). {St} is the claim surplus process given by I. are i. Proposition 1.Then no matter the distribution Al of T1i B. Then the arrival process is stationary which could be a reasonable assumption in many cases (for these and further basic facts from renewal theory. Var(St) = 11Ba2A + I4AaB lim t goo t PA 131 . and M is the maximum of {St}.1).Chapter V Renewal arrivals 1 Introduction The basic assumption of this chapter states that the arrival epochs O'1. T3. r(u) the time to ruin.i. see A.. the one corresponding to the stationary case by 00)(u).d.1 Define p = !µ.. The ruin probability corresponding to the zerodelayed case is denoted by 1/'(u).. the Tn are independent.Q.7).. with the same distribution A (say) for T2. . D'2. AA t*oo lim St = lim ESt t tioo t = p . with Nt = # {n: Un <t} the number of arrivals before t. A different important possibility is Al to be the stationary delay distribution A° with density A(x)/µA. of the risk process form a renewal process: letting Tn = Qn ..1.
The renewal model is often referedd to as the Sparre Andersen process. RENEWAL ARRIVALS lim E [St+a . 2).Nt] * a/PA.St] = a(p . However . Nt ESt = E E UI Nt t = ENt•pB .132 Furthermore for any a > 0. This has a direct physical interpretation (a large portfolio with claims arising with small rates and independently). one could imagine that the claims are recorded only at discrete epochs (say each week or month) and thus each U.t. Here are two special cases of the renewal model with a similar direct interpretation: Example 1.1) ENt/t + 1/µA. t 4oo Proof Obviously. say at a. the definition 77 = 1/p . Example 1 . Nt + EVar U. but the arrival rate in the ON state is . such that no arrivals occur in the off state.2 (DETERMINISTIC ARRIVALS) If A is degenerate.0 > 0. Thus. From this ( 1. Proposition 1. we get similarly by using known facts about ENt and Var Nt that Nt Var(St) = Var E Nt U. stating that E[Nt+a . the . and (1 .1 of the safety loading appears reasonable here as well.1) follows .1). 3) follows similarly by Blackwell 's renewal theorem.a is really the accumulated claims over a period u of length a. Sparre Andersen whose 1959 paper [7] was the first to treat renewal assumptions in risk theory in more depth. s + t µA PA 0 Of course. after E. Nt = Var(PBNt) + E(4Nt) Q2 2 0`2 A tpB B + o(t). If the environment is Markovian with transition rate A from on to off and u from OFF to ON. OFF. CHAPTER V. A.3 (SWITCHED POISSON ARRIVALS) Assume that the process has a random environment with two states ON. For (1 . by the elementary renewal theorem (cf.1 gives the desired interpretation of the constant p as the expected claims per unit time. The simplest case is of course the Poisson case where A and Al are both exponential with rate 0.
4) w.T between a claim U and an interarrival time T. INTRODUCTION 133 interarrival times become i. .2.. Therefore. if for nothing else then for the mathematical elegance of the subject. 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 .r. u For later use. (an arrival occurs necessarily in the ON state. as follows easily by noting that the evolution of the risk process after time s is that of a renewal risk model with initial reserve U1 . (1. the relevance of the model has been questioned repeatedly.4) fo Indeed.4) with phase space {oN. arrival times. we note that the ruin probabilities for the delayed case T1 = s can be expressed as in terms of the ones for the zerodelayed case as u+8 z/i8(u) = B(u + s) + '( u + s . initial vector (1 0) and phase generator 11 However.. and the present author agrees to a large extent to this criticism. The following representation of the ruin probability will be a basic vehicle for studying the ruin probabilities: Proposition 1.oFF}. U1 .1.s.. the fundamental connections to the theory of queues and random walks. integrate (1. in general the mechanism generating a renewal arrival process appears much harder to understand... we have From this the result immediately follows..d. For the stationary case. and then the whole process repeats itself).i.} with {S(d)} a discrete time random walk with increments distributed as the independent difference U . we feel it reasonable to present at least some basic features of the model.4 The ruin probabilities for the zerodelayed case can be represented as 0(u) = P(M(d) > u) where M(d) = Max {Snd) : n = 0.s > u) of ruin at the time s of the first claim whereas the second is P(r(u) < oo.1. Ao.s < u).t. However. A is phasetype (Example 1. and for historical reasons. S o<t<oo n=0.y)B(dy). More precisely.1. Proof The essence of the argument is that ruin can only occur at claim times.
If . the claim surplus process are given by Nt Nt Rt = u+^U.a*PB• > 1.Ut. Using Lundberg conjugation .134 CHAPTER V. with common distribution B* (say) concentrated on (0. resp . the remaining part of the prepayment (if any ) is made available to the company. the claims and the premium rate are negative so that the risk reserve process . The initial reserve is obtained by prepayments from the policy holders.1 r* (u) One situation where this model is argued to be relevant is life annuities.3* pB. St = t . The compound Poisson model with negative claims We first consider a variant of the compound Poisson model obtained essentially by signreversion . U Figure 2. At the time of death . b=1 !=1 where {Nt } is a Poisson process with rate .0* (u) = P (rr* (u) < oo) where rr* (u) = inf It > 0: Rt < 0} ) . we shall be able to compute the ruin probabilities i(i* (u) for this model very quickly (.3* (say ) and the U. < 1. A typical sample path of {Rt } is illustrated in Fig.1) .1 If.d. i. 00).1) +ry. Theorem 2 . are independent of {Nt} and i. RENEWAL ARRIVALS 2 Exponential claims. (2.1. each of which receive a payment at constant rate during the lifetime . then 0 * (u) = 1 for all u > 0. A simple sample path comparison will then provide us with the ruin probabilities for the renewal model with exponential claim size distribution. That is . t. 2. then 0*(u) = e 'r" where ry > 0 is the unique solution of 0 = k*(ry) = *(B*[ry] .
Hence y exists and is unique. (a) is*(a) (b) . Fig.(a) 7 Figure 2. then by Proposition 111. 2. If I3*pB* < 1.1.UB..2 sup St = inf St = 00 t>o t>o and hence 0* (u) = 1 follows.2(b). cf. 2.3*.2.2 Assume now .2(a). Since ic'(0) < 0.s. 0 Now return to the renewal model. Define T_ (u) = inf It > 0 : St = u} . and thus 1 = P(T. > 1 . 0) and has typically the shape on Fig.(u ) < oo) = E {e7sr_ (u). B*. Proof Define 135 St =u .Rt. Then { St } is the claim surplus process of a standard compound Poisson risk process with parameters 0 *. of {St} is c(a) = is*(a7). Hence T_(u) < oo a. Then the function k* is defined on the whole of (oo. EXPONENTIAL OR NEGATIVE CLAIMS [Note that r. and the Lundberg conjugate of {St} is { St } and vice versa. the safety loading of { St} is > 0. St=Rtu=St.* (a) = log Ee'st I.f. T_ (u) = inf { t > 0 : St = u 'r* (u). B* [7] and let {St} be a compound Poisson risk process with parameters .g. Then the c. . Let B(dx) = ^e7x B*(dx).0. B. T_ ( u) < 001 e7"P(T_ (u) < oo) = e"V)* (u).
Now the value of {St*} just before the nth claim is To +T1* +. 2.1....Y a I.. To + M(d) in the notation of Proposition 1.a) = 1 .2)....4 goes as follows: define 7r+ = P(M(d) > 0) and consider {St*} only when the process is at a maximum value. u Hence P(M(d) > u) _ 1r+e'r".136 CHAPTER V..Tr+.e. Taking m.Tn} n=0.1. A variant of the last part of the proof.. According to Theorem 2..1) + ry = 0 which is easily seen to be the same as (2. To + max {Ul+•••+UnTI. RENEWAL ARRIVALS Theorem 2 . and from Fig . the failure rate of this process is y.1 means that M* is exponentially distributed with rate ry.Un } = max St = t>0 n=0.Ui . the distribution of M(d) is a mixture of an atom at zero and an exponential distribution with rate parameter ry with weights 1 ..2) 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 .f. T2 = U2.• • • .)(u) _ 1r+e7" where ry > 0 is the unique solution of 1 = Ee'Y(uT ) = S 8 A[. we get Ee'M(d) = Ee°M* _ Y/(. Then B* = A.'s and noting that V)*(u) = P(M* > u) so that Theorem 2. and 5PA > 1.Ti = U1. with the probability that a particular jump time is not followed by any later maximum values being 1 . and (2 . 1) means that 8(A[ry] . respectively..4.2 If B is exponential. alternatively termination occurs at a jump time (having rate 8). then . 3* = 6.1 it is seen that ruin is equivalent to one of these values being > u. with rate S (say).7r+ 7r EeTo b/(Sa) + +.•.u+ and lr+.+Tn U1 Un. Hence M* max {To + Ti + • • • + Tn ..g. which has the advantage of avoiding transforms and leading up to the basic ideas of the study of the phasetype case in VIII.. However. and hence the failure rate .1.Y] (2.
. Hence the failure rate of M(') is 6(1 ..7r+) and hence r+ = 1. we have ] A[a )3] E«d'efl' = Bad> [a] A ad> [Q] = B[a +. hence exponential with rate b. 111.6. : S(d) > u} . the imbedded discrete time random walk and Markov additive processes.2.7r+). 0 3 Change of measure via exponential families We shall discuss two points of view. This follow since. Thus a ladder step terminates at rate b and is followed by one more with probability 7r+. Putting this equal to y. a ladder step is the overshoot of a claim size.B(dx). 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^. However.y/b.5.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] . B^d) where Aad> (dt) = ^[ a] A(dt).3 A[a] B[a] F( d) [a +)3] F(d) [a] = Fad) [^] Letting M(u) = inf in = 1. resp.7r+) = ry and hence P(M(d) > u) = P(M(d) > 0)e7u = 7r+e'r". consider instead the failure rate of M(d) and decompose M(d) into ladder steps as in II.2.. the relevant exponential change of measure corresponds to changing the distribution F(d) of Y = U . CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 137 is b(1.3.. 3a The imbedded random walk The key steps have already been carried out in Corollary 11.4. The probability that the first ladder step is finite is 7r+. Furthermore. Bads (dx) = . we see that ry = 6(1. which states that for a given a. letting P(d) refer to the renewal risk model with these changed parameters .
1). 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. 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: Proposition 3. and claim (a) follows immediately from this and e (u) > 0.138 CHAPTER V.r..2 In the zerodelayed case.(u) ..3 For the delayed case Tl = s.T is nonlattice. In fact. Proof Proposition 3. 7µA . provided the distribution F of U . let 7 > 0 be the solution of r.p)/($B'[7] . ik. We have the following versions of Lundberg' s inequality and the CramerLundberg approximation: Theorem 3 . (d) (7) _ 0.4. Corollary 3. For the stationary case. i . in the easiest nonexponential case where B is phasetype. (a) '(u) < eryu. O(u) = eauE (d)ea{ (u)+M(u)K (d)(a) . to converge in distribution since p(yd) (r(0) < oo) = 1 because of r (d)' (y) > 0.C8e7u where Cs = Ce78B[7]. E(d)e 1' (u). For claim (b).t. VIII. RENEWAL ARRIVALS be the number of claims leading to ruin and ^(u) u = SM(u) .u the overshoot .. (b) V)(u) .e.Ce"u where C = limu.C(°)eryu where C(O) = C0[7] . 00)(u) . just note that F7(d) is nonlattice when F is so . This is known to be sufficient for ^(O) ]p (d) ([APQ] Proposition 3. 187) and thereby for ^(u) to be nonlattice w.2 p. Consider now the Lundberg case.1) is explicit given 7. cf.
(u + s . For s > 0. To determine boundary 0. 0) = tc(a)h(s). Sdt) = h(s . we get r u +8 e"8(u) 139 e7uB(u + s) + 4 0 + L 00 J e7(v8)e7(u+8v). 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] . The underlying Markov process {Jt} for the Markov additive process {Xt} = {(Jt. we invoke the behavior at the 1 = h«(0. 3b Markov additive processes We take the Markov additive point of view of II. where G is the infinitesimal generator of {Xt} = {(Jt. (s./c.1) (normalizing by h(0) = 1). Equating this to tch(s) and dividing by h(s) yields h'(s)/h(s) _ .dt ) eadt = h ( s) . Here K. y) = e°yh(s).3. h(s) = e(a +x( a))8 (3. B(x) = o(e7x) and dominated convergence.1) = C(O). 0 0 ..y) B(dy) 0 For the stationary case. we look for a function h(s) and a k (both depending on a) such that Gh.Sdt] = Ee'uh(T) means 1 = f ' e°^B(dy) f ' h( s)A(ds). IPA 0 Of course. E8h0 (Jdt.5.(s.a .h'(s).0 ) = Eo[ha ( Jdt.(°) ( Ce8B[7] Ao(ds) similar manner.St)} can be defined by taking Jt as the residual time until the next arrival. delayed version of Lundberg's inequality can be obtained in a e7u. The expressions are slightly more complicated and we omit the details. Let P8f E8 refer to the case Jo = s. 0) = ah (s) .. St)} and h. According to Remark 11.5.dt(ah ( s) + h'(s)) so that Gha ( s. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES Proof Using (1.9.4).
.rc(a)] B[a] A[a . Further. Ba where Aa (dt) .c(a)] which shows that U1. 5 For the compound Poisson case where A is exponential with rate .tK( a)e. An important exception is. (3.S„):0<v< )be Lt = eaSt tK(a) h(Jt) = east ..2) means 1 = B[a]/3/(/3+a+rc (a)). Remark 3 .e. .rc(a)] = B = Ba[13]Aa[5]. . An easy extension of the argument shows that U1.a . . .8. however. St)}too by letting the likelihood ratio Lt restricted to Yt = a((J. Ba(dx) = B(dx).s governing {(Jt. we can now for each a define a new probability measure Pa.( a+r' (a))(Jt s) h(s) where c(a) is the solution of (3. (3. [a + /3] A[b . J n+1 u are independent with distribution Aa for the Tk and Ba for the Uk.a . rc(a) = 0 (B[a] . resp. A[a .rc(a)] = 1. Ease AU1+6T2 [ AU1+6T2 = Ea a LT.c(a)] B[a] Proof Pa.s and P(d). T2.140 CHAPTER V.. RENEWAL ARRIVALS B[a]A[a . [e1U1 + 6T2ea ( U1s)stc ( a)e(a+K(a ))( T2s)I B[a +. .4 The probability measure Pa. ] = E...5.2) As in 11.13]A[b .s is the probability measure governing a renewal risk process with Jo = s and the interarrival distribution A and the service time distribution B changed to Aa.s(Jo = s) = 1 follows trivially from Lo = 1. U. Note that the changed distributions of A and B are in general not the same for Pa. = J8 = T2..2). since JT. resp.e(«+k(a))t esy A(dt). Aa as asserted . T2 are independent with distributions Ba.1)a in agreement u with Chapter III. i. the determination of the adjustment coefficient ry where the defining equations rc(d) (ry) = 0 and rc(ry) = 0 are the same. . Proposition 3.
T(u) < yu h(JT(u)) < eayu+yuk(ay ) ( Eia y Le(a(+k(ay))s v.5.ay+ray))TM(. yu) = F'ay. . defined as the single server queue with first in first out (FIFO.. and U„ the service time of customer n. The actual waiting time Wn of customer n is defined as his time spent in queue (excluding the service time). and define yy = ay . A(ds). say finite horizon ruin probabilities where the approach via the imbedded random walk yields results on the probability of ruin after N claims. 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.g. see in particular Dassios & Embrechts [98] and Asmussen & Rubinstein [45]. the amount of . Label the customers 1. see e.(u. The virtual waiting time Vt at time t is the residual amount of work at time t. yu). 2. Notes and references 4 The duality with queueing theory We first review some basic facts about the GI/G/1 queue.r. not after time T.)+1 e J j e(ay+w(ay))8 e .rc( ay)] = e(aa+(r ))sb[a ]e7yu L y1 In particular. or FCFS = first come first served) queueing discipline and renewal interarrival times.4. Then J(rr(u)) = TM(u)+1 and hence Ws(u. XII.1 and n. u The approach via the embedded random walk is standard.s eaysr(")+r(u ) K(ay) h (s) .t. The claim for the zerodelayed case follows by integration w.4. which is the same as the asserted inequality for 0.5 Proposition 3. for the zerodelayed case zp8(u. . yu ) < e7yu.yu ) e7vu A[ay .. Then "^ e(ay+w(aY))8 Ys(u. Let M(u) be the number of claims leading to ruin . For the approach via Markov additive processes. that is. the time from he arrives to the queue till he starts service. Using the Markov additive approach yields for example the following analogue of Theorem IV. [APQ] Ch.y yuAa y [ay + K(ay) .yx(ay). that is. and assume that T„ is the time between the arrivals of customers n . . it is easily seen that ic(ay ) > 0.4.6 Let y < let ay > 0 be the solution of ic'(ay) = 1/y. Proof As in the proof of Theorem IV.
1). Then P(r(u) < T) is the probability z/iiNi (u) of ruin after at most N claims. 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. Wn converges i n distribution to a random variable W. since customer n arrives at time on.2 Let Mnd) = maxk=o. and combining with (4."^ Vi(N) (u). we have Wn = Van(left limit). 0 Applying Theorem 11. but we shall present a slightly different proof via the duality result given in Theorem II. the proposition follows. . and we have P(W > u) = V.1 and Corollary 11.. (u). Thus.1. It then jumps to VQ„ .• • • Tk ). on + Tn) the residual work decreases linearly until possibly zero is hit. 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. Also {Zt}o<t<T evolves like the leftcontinuous version of the virtual waiting time process up to just before the Nth arrival.142 CHAPTER V.4. and obviously z/'(u) = limN. the waiting time a customer would have if he arrived at time t. an+1) = [on. (4.3.4): Proposition 4. in which case {V} remains at zero until time on+1. but interchanging the set .2) (b) as t * oo.+ Un. If W1 = 0. Vt converges in distribution to a random variable V. whereas in [On . Then: (a) as n + oo. Thus Vos}1 _ = (Wn + Un .3 Assume rl > 0 or. Let the T there be the random time UN.Tn)+.1) The following result shows that {Wn} is a Lindley process in the sense of II.4.. The traffic intensity of the queue is p = EU/ET. equivalently. (4.4: Proposition 4. p < 1.. we get: Corollary 4. and we have P(V > u) = ?/iiol(u).3) Proof Part (a) is contained in Theorem 11. (4. then Wn v M. equivalently.Tn)+ Proof The amount of residual work just before customer n arrives is VQ„ .2.1.n1 (U1 +• • •+Uk Tl ..4.1 Wn+1 = (Wn + U.
2).T*)+. Letting n oo in Corollary 4. namely W1 = 0 in (a) and Vo = 0. u Now return to the Poisson case .4.y)F(dy). which implies the convergence in distribution and (4. T] form a stationary renewal process with interarrival distribution A. (4. T1 . convergence in distribution hold for arbitrary initial conditions . T1.. where U*.1. by an obvious reversibility argument this does not affect the distribution . It follows that P(WN > u) =.T* = y yields K(x) = P ((W + U* . hence (since the residual lifetime at 0 and the age at T have the same distribution .T*)+ < x) = P(W + U* .T* are independent and distributed as U1. Then K(x) = J x00K(x . but this requires some additional arguments (involving regeneration at 0 but not difficult) that we omit. Corollary 4..le) the same is true for the timereversed point process which is the interarrival process for { Zt}o<t < T• Thus as before . Then the arrivals of {Rt} in [0.5) Proof Letting n . resp . Ti) and similarly for the U.. Then the corresponding queue is M/G/1. conditioning upon U* . we get W = (W + U* . For part (b). In fact . However. { 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). and we get: . i...2.T* < x) fK(x_y)F(dy) (x > 0 is crucial for the second equality!). TN) with (TN..4 The steadystate actual waiting time W has the same distribution as M(d).. x > 0.(N)(u) has the limit tp(u) for all u.. THE DUALITY WITH QUEUEING THEORY 143 (T1.Ao in (b). (4.5 (LINDLEY'S INTEGRAL EQUATION) Let F(x) = P(U1T1 < x).. K(x) = P(W < x). A. cf. Hence for x > 0. we let T be deterministic . as WN.e..4) Tlim F(s) (VT > u) = limo P(s) (r(u) < T) = '+^io) (u)• 0 It should be noted that this argument only establishes the convergence in distribution subject to certain initial conditions. we obtain: Corollary 4.oo in Proposition 4. and hence in particular ZT is distributed as the virtual waiting time just before the Nth arrival.
5) is in fact a homogeneous WienerHopf equation. Proof For the Poisson case. . see e.g. 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. VIII).6 For the M/G/1 queue with p < 1.5) to hold for all x E R and not just x > 0). implying P(W > u) = P(V > u) for all u. 0 Notes and references The GI/G/1 queue is a favourite of almost any queueing book (see e . RENEWAL ARRIVALS Corollary 4. the zerodelayed and the stationary renewal processes are identical. That is. Note that (4. The equation (4. Hence '(u) = Ali(°)(u). W v V. Cohen [88] or [APQ] Ch.144 CHAPTER V. the actual and the virtual waiting time have the same distribution in the steady state.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]. Asmussen [24] and references there.g.
M = supt>o St. 145 Oj( u. The ruin probabilities with initial environment i are '+ki(u) = pi(T(u ) < oo) = Pi (M > u). dv. here it exists whenever A is irreducible which is assumed throughout.)iJEE and its stationary limiting distribution by lr. .f pi. As in Chapter I. • The premium rate when Jt = i is pi. • Claims arriving when Jt = i have distribution Bi.T) = Pi (T(u) < T).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 .. and can be computed as the positive solution of WA = 0. Thus. Ire = 1. {St} denotes the claim surplus process. N. i=1 0 and r(u) = inf It > 0: St > u}.(3i when Jt = i. {Jt} describes the environmental conditions for the risk process. t St = E Ui . The intensity matrix governing {Jt} is denoted by A = (A.
1) iEE Then pi is the average amount of claims received per unit time when the environment is in state i. Proposition 1. Example 1. i and corresponding arrival intensities Qn. T(=)). P = E 7riPi. we could distinguish between normal and icy road conditions.146 CHAPTER VI. Assume similarly that the sojourn time in the normal state has distribution A(n) which we approximate with a phasetype distribution with representation (E('). cf. t(i) = T(')e are the exit rates. = iii when j E E(i). with rates Ani and Ain. the operational time argument given in Example 1. /3 = Nn when j E E(n).5 below. and assume that weather conditions play a major role for the occurence of accidents. For example. this is no restriction when studying infinite horizon ruin probabilities. the intensity matrix is A OW) T(i) T(n) t(n)a(i) where t(n) = T(n)e. Thus. a(i). f3i and claim size distributions Bn.2. meaning that accidents occuring during icy road conditions lead to claim amounts which are different from the normal ones. assume that the sojourn time in the icy state has a more general distribution A(i). Bi. We let p Pi = /ji/AB.1 Consider car insurance. say.. say. in blockpartitioned form.2 (ALTERNATING RENEWAL ENVIRONMENT) The model of Example 1. r^ = P (1. and we have f3. we can approximate A(i) with a phasetype distribution (cf. and p is the overall average amount of claims per unit time. Unless otherwise stated.a(').14. we shall assume that pi = 1.11 below.4) with representation (E(i). An example of how such a mechanism could be relevant in risk theory follows. one expects that 3i > on and presumably also that Bn # Bi. According to Theorem A5.1 implicitly assumes that the sojourn times of the environment in the normal and the icy states are exponential. which is clearly unrealistic.T(n)). Cl The versatility of the model in terms of incorporating (or at least approximating) many phenomena which look very different or more complicated at a first sight goes in fact much further: Example 1. u . cf. MARKOVIAN ENVIRONMENT where as usual Pi refers to the case Jo = i. respectively. Then the state space for the environment is the disjoint union of E(n) and E(i). leading to E having two states n. Example 1 .
This amounts to a family (A(")) ?CH Of sojourn time distributions... n8}. Qi = . and 1/ii(u) = t/ii(u).Q.3i. say. (9) where q = CHI.3. Example VIII.. such that the icy period is of two types (long and short) each with their sojourn time distribution A('L). depending only on 77. One way to model this would be to take A(') to be Coxian (cf. let T 9(T) = f pi. but assume now that the arrival intensity changes during the icy period. say it is larger initially. the parameters are ^ij = aid/pi. 1 . St = SB=(t).a(n). The simplest model for the arrival intensity amounts to . Approximating each A('?) by a phasetype distribution with representation (E('l). i ) : n E H. and similarly for the normal period.3i/pi. i8f n1..3 Consider again the alternating renewal model for car insurance in Example 1.1. In the car insurance example. u Example 1. w... 0 Then (by standard operational time arguments) {St } is a risk process in a Markovian environment with unit premium rate. iq (visited in that order) and letfOil >. is the probability that a long icy period is followed by a short normal one. t(n) = T("i)e. A('^). T(1) +w11t(1)a(1) w12t (1)a(2) w21t(2)a(1) w1gt(1)a(9) w2gt ( 2)a(q) T(2) +w22t( 2)a(2) A = wg1t(9)a(1) wg2t(9)a(2) .>. MODEL AND EXAMPLES 147 Example 1 .tEH is a transition matrix.5 (MARKOVMODULATED PREMIUMS) Returning for a short while to the case of general premium rates pi depending on the environment i..1. resp.2. the state space E for the environment is { ('q. it = Jel(t). and ...p.T(n)). we assume again pi = 1 so that the claim surplus is Nt St = ?Ui_t. Indeed. . such that a sojourn time of type rt is followed by one of type c w. where W = (w. T(9) +wggt(9)0. 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.. Then for example wi. dt. i E E(n) }. u From now on.4) with states i1.n.J017.j = .. u Example 1 . one could for example have H = {i1.. .
e(A(Oi)d'sg)xe. More precisely. JT1 = j) = Qj • e. .5. )3*. In particular.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.148 CHAPTER VI.(3iBi(dx). vi(dx) = .(Qi)diag)• More precisely. Nt Nt a . o = 0.A . Pi (Ti E dx. B* = 1 /^* Bi. the empirical distribution of the claims is B*. t l=1 Note that the last statement of the proposition just means that in the limit. dx. qij = 0 in the notation of Chapter 11. Next we note a semiMarkov structure of the arrival process: Proposition 1. the Markov additive structure will be used for exponential change of measure and thereby versions of Lundberg's inequality and the CramerLundberg approximation. .7 The Pidistribution of T1 is phasetype with representation (ei. MARKOVIAN ENVIRONMENT We now turn to some more mathematically oriented basic discussion. 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. one can associate in a natural way a standard Poisson one by averaging over the environment. The key property for much of the analysis presented below is the following immediate observation: Proposition 1.8 As t oo. we put )3* = E 7fi/3i. A remark which is fundamental for much of the intuition on the model consists in noting that to each risk process in a Markovian environment. iEE iEE )3 These parameters are the ones which the statistician would estimate if he ignored the presence of Markovmodulation: Proposition 1. N > 1(Ul < x) a4 B*(x). 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).
.1. 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$. y Ni) i Nti) t a.. Bi. 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. has distribution (7ri)3i //3*)iEE and is independent of Ti. Conditioning . Bi. Then it is standard that ti lt '4' iri as t > oo. zli( (u) .(/3i)aiag). aA... e. this converges to the exponential distribution with rate 0* as a * oo. cf. A. iEE 13 A different interpretation of B* is as the Palm distribution of the claim size. Example 11.7. Bi(x).9 Consider a Markovmodulated risk process {St} with param eters Ni. Proposition 1. and furthermore in the limit JT.* (u) for all u. i. Nt a' t t iEE Also.6. oo) as a 4 oo. Hence Nt'> a .aA . By Proposition A5. and let {St °i} refer to the one with parameters Pi. 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 . N + oo Hence 1 Nt 1 N`+) Nits Nt E I ( Ut <.. Proof According to Proposition 1. In particular. B*. However . the Fidistribution of T1 in {St(a ) } is phase type with representation (E. Then {St)} + {St*} in D[0. we may view Nt`i as the number of events in a Poisson process where the accumulated intensity at time t is Niti. In particular. the limiting distribution of the first claim size U1 is B*.2.x) = Nt E > I (Uk) X) Nt Bi(x) 1=1 iEE k=1 iEE 1: t5 Bi( x) = B*(x). {St} to the compound Poisson model with parameters 0 *. ^j 7riNi.4. given {Jt}0<t<0.
there are p = 2 background states of {Jt}.. s 5 in state 2.10 Let E_{1.. 132=2. 9 . T) for all u and T. thick.2 +2 2 = 3. Continuing in this manner shows that the limiting distribution of (T. and (at least when a is small such that state changes of the environment are infrequent). is as in {St }. which also yield O(a) (u. marked by thin. the company even suffers an average loss. lines in the path of {St}.1.).s = o in state 1 and with intensity 1 .FT.g. we first get that 3 (3* = 2. > 1. those of type E7 with intensity z s = 5 in state 1 and with intensity z . state 1 appears as more dangerous than state 2. A= (  a a ) \ a a 5 5 J 9 3 2 a1=2. From this the convergence in distribution follows by general facts on weak convergence in D[0.1 of [145]. B1=3E3+2E7. U2) are independent of . U. oo)..150 CHAPTER VI. T) + ?P* (u. On Fig.2}. resp. 1. from Theorem 3. the overall drift is negative since it = (2 2) so that p = 71P1 + 112P2 = 7.. Computing the parameters of the averaged compound Poisson model. Claims of type E3 arrive with intensity 2 . Thus.2.=1.31µB 2 = 2 5 3 7 70 Thus in state 1 where p. B2=1E3+4E7. since E3 is a more dangerous claim size distribution than E7 (the mean is larger and the tail is heavier). That is. we may imagine that we have two types of claims such that the claim size distributions are E3 and E7. 1. 0 Example 1. The fact that indeed 0(a) (U) 3 0* (u) follows. shows similarly that in the limit (T2.....l3* and U2 having distribution B*. 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 = . with T2 being exponential with rate . the paths of the surplus process will exhibit the type of behaviour in Fig. e. MARKOVIAN ENVIRONMENT upon FT.1 with periods with positive drift alternating with periods with negative drift.. 5 5 where E5 denotes the exponential distribution with intensity parameter 5 and a > 0 is arbitrary.s = 1o in state 2.2.
01 /.11 (a ) ESt/t * p .1 a. That is.1). MODEL AND EXAMPLES 151 Figure 1.1 Thus. (b) St/t * p . Proof In the notation of Proposition 1..1) of the safety loading is (as for the renewal model in Chapter V) based upon an asymptotic consideration given by the following result: Proposition 1. note first that EN Uk')/N a4' µgi. a fraction r. = P.(3.1. iEE .s.1. we have E[St + t I (t(i))iE EI = E t(i)OW = iEE t(i)Pi• iEE Taking expectations and using the well known fact Et(i)/t * 7ri yields (a). For (b).3* = 3/4 of the claims occur in state 1 and the remaining fraction 1/4 in state 2. 0 The definition (1. t * oo. the averaged compound Poisson model is the same as in III. Hence (i) Nti) 1 U(i) k' N(i)k=1 E t 4 St + t = iEE Nt t 1: 7ri Qi µs. t + oo. Hence B* = 415E3+5E7/ +4 ( 51E3 +5 E7) = 1E 3 +2E7.8.
and . Proof The case 77 < 0 is trivial since then the a. See Meier [258] and Ryden [314]. with X2. PB. and hence M = 00.g. and hence 1/ii(u ) = 1 for all i and u. The case 77 > 0 is similarly easy.Jt=i}.1 and the Corollary are standard. Proposition 1. X3. [APQ]. also + ..a form a renewal process . . having the Pidistribution of X. and hence oscillates between 0o and oo so that also here M = oo. .. and hence wn /n a4. and so on.. w2=inf {t>w1:Jt_#i.s.2(a) p. [212]. Since the X„ are independent .\ i and EiX1 Ei f 13 J. 0 Notes and references The Markovmodulated Poisson process has become very popular in queueing theory during the last decade.0i(u) < 1 for all i and u. s. Eiw. limit p . see the Notes to Section 7. 38) Eiw1 = 1/ir.1 jEE = (p . The mainstream of the present chapter follows [16].. with some important improvements being obtained in Asmussen [17] in the queueing setting and being implemented numerically in Asmussen & Rolski [43]. and involves a version of the .. [302]. In risk theory.1)Eiw = 0. X2 =SW2 So. then M = 00 a. Now obviously the w. 136 or A.152 CHAPTER VI. There seems still to be more to be done in this area. X 1 =Sty. then M < oo a.1 of St / t is > 0. MARKOVIAN ENVIRONMENT Corollary 1.. Now let r) = 0.12 If 77 < 0. If 77 > 0. 2 The ladder height distribution Our mathematical treatment of the ruin problem follows the model of Chapter III for the simple compound Poisson model. let some state i be fixed and define w=wl=inf{t >0:Jt_#i. Theorem II.Jt=i}.4..Eiw o'o Eiw • E ^ifjµs. n n Thus {SWn l is a discrete time random walk with mean zero. and a more comprehensive treatment in Asmussen [16].ld. The proof of Proposition 1. [315]... EiX = 0. Statistical aspects are not treated here. + Xn SWn ](1 a .s. see [APQ] p. dt .1(b) is essentially the same as the proof of the strong law of large numbers for cumulative processes. Then by standard Markov process formulas (e.. some early studies are in Janssen & Reinhard [211].
Thus.2) R(dx)S((y .2. j. B* in Section 1. 00 (2. Define the ladder epoch T+ by T+ = inf It : St > 0} = r(0).1 irG+(dy)e =.(u) = Pi(M < u) = e' E G+ (u)(I . by specializing results for general stationary risk processes (Theorem II .j. e.x. j.1) 0 (b) G+ (y.6.2(a) below ) where the ladder height distribution is evaluated by a time reversion argument.6*. •) II = JG+(i. see also Example II. which represents a nice simplified form of the ladder height distribution G+ when taking certain averages : starting {Jt} stationary. let G+(i. •)• kEE Also.4) we obtain the following result .Jr+ =j.g. j. (y.2 (a) The distribution of M is given by 00 1 .x). The form of G+ turns out to be explicit (or at least computable).a/i. k. •). oo)).dx). Proposition 2.Jt=j)dt. j.5. T R(i.j. For measurevalued matrices. . the definition of .A) = Pt(ST+ E A. j. oo) = J ao 0 G+(i. IIG+ II denotes the matrix with ijth element IIG+(i. T+ < oo) and let G+ be the measurevalued matrix with ijth element G+(i. oo)) = f R(i. cf. and S (dx) the measure valued diagonal matrix with /3 Bj(dx) as ith diagonal element. G+ is the matrix whose ijth element is E G +(i.. we get the same ladder height distribution as for the averaged compound Poisson model. THE LADDER HEIGHT DISTRIBUTION 153 PollaczeckKhinchine formula (see Proposition 2.3*B *(y)dy. but is substantially more involved than for the compound Poisson case . for i.IIG +II)e. Proposition 2. we define the convolution operation by the same rule as for multiplication of realvalued matrices. •) * G +(k.i.EA. 6. only with the product of real numbers replaced by convolution of measures. dx)/jBj(y .A) =ZI(St E.j E E. However . n=0 (2. Let further R denote the preT+ occupation kernel. That is.
3) We let {St*} be defined as {St}. resp. only with {Jt} replaced by {Jt } (the /3i and Bi are the same ).2 useful . St < S* for u < t.6.2) is just the same as the proof of Lemma 11. That is. {mx} is a non terminating Markov process on E. and that the environment is j at the nth when we start from i is e . and let further {my} be the Evalued process obtained by observing {Jt } only when {St*} is at a minimum value. we need as in Chapters II.3. the intensity matrix A* has ijth element * 7r ^i3 7ri and we have Pi(JT = j) = 7rj P2(JT = i)7ri (2. . mx = j when for some (necessarily unique) t we have St = x.1 The following observation is immediate: Proposition 2. From this (2. The u proof of (2. G+ the probability that there are no further ladder steps starting from environment j is e^ ( I . see Figure 2. III to bring R and G+ on a more explicit form .1 for an illustration in the case of p = 2 environmental states of {Jt}. hence uniquely specified by its intensity matrix Q (say). lines in the path of {St}. JJ = j. MARKOVIAN ENVIRONMENT Proof The probability that there are n proper ladder steps not exceeding x and (x)ej.IIG+II)e.154 CHAPTER VI. To this end . we need to invoke the timereversed version {Jt } of {Jt} .3 When q > 0. thick. To make Proposition 2. 0  x Figure 2.1) follows by summing over n and j. marked by thin.
0 mms1   ^O \ T. and S(dx) is the diagonal matrix with the f3iBi(dx) on the diagonal. For example the excursion of depth 2 has one subexcursion which is of depth 1.and a jump (claim arrival) occurs at time t.2 .. and the excursion ends at time s = inf {v > t : S. we recursively define the depth of an excursion as 1 plus the maximal depth of a subexcursion. The definitions are illustrated on Fig.(/3i)diag + T S(dx) eQx.*. An excursion of {St*} above level x starts at time t if St = x. and the excursion is said to have depth 1 if each of these subexcursions have depth 0. Q( n+l) _ ^. In general. 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). } is a minimum value at v = t. s]. If there are no jumps in (t.2.4 Q satisfies the nonlinear matrix equation Q = W(Q) where 0 co(Q) = n* .2. Otherwise each jump at a minimum level during the excursion starts a subexcursion. we say that the excursion has depth 0. Proof The argument relies on an interpretation in terms of excursions. 2. = x}. THE LADDER HEIGHT DISTRIBUTION 155 Proposition 2. Figure 2.0. {S.(/3i) diag. the sequence {Q(n)} A* defined by Q(O) = . corresponding to two subexcursions of depth 0. Furthermore. ( Q( n)) converges monotonically to Q.2 where there are three excursions of depth 1.
T+>t) _ ^iF 7ri (JJ =i. Then a jump to j (i.4). A) = L' U(j. Suppose mx = i. we first compute qij for i $ j. 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 .Qi + )%pij) Now just note that t pij and insert (2. StEA .156 CHAPTER VI. Fi(mh =i ) = 1 + =hflh+Qihpii+o(h) implies qii = 'iii /i +)3ipii. = j. 0)). By considering minimum values within the excursion. Similarly. Q = W(Q) follows. (2...s. p1^) Define a further kernel U by f U(i. the subintensity matrix of {min+i ) } is cp (Q(n)) = Q(n +l) which implies that qgj +1) = \!. it becomes clear that pij = r [eQh] 0 ij Bi (dy) • (2. Theorem 2 . i.5) A (note that we use A = {x : x E Al on the r. It follows that qij = A.St <S*.St EA. A). or through an arrival starting an excursion terminating with J.j +/3ipij.5 R(i. h.(01)diag = Q. 7rE Proof We shall show that Fi(Jt=j. Writing out in matrix notation . It is clear that { mini } is a terminating Markov process and that { mio) } has subintensity matrix A* .u< t). The proof of Q = W(Q) then immediately carries over to show that the subintensity matrix of {mil) } is cp (Q(o)) = Q(l).j. either due to a jump of {Jt } which occurs with intensity A= j. j.4) To show Q = cp(Q). Similarly by induction .6) . of the definition to make U be concentrated on (co. e. 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. A) = f Pi(mx = j) dx eie4xej dx A u (2.
Jo=i. e. {Jt }.. `` {K(n)} [the W(•) here is of course not the same as in Proposition 2. dt. To this end. S. it is readily checked that 7r is a left eigenvector of K corresponding to the eigenvalue 0 (when p < 1). 0 < u < t) = 7rjPj(Jt =i. consider stationary versions of {Jt}.7 It is instructive to see how Proposition 2.6 (a) R(dx) = eKxdx. Jt = i. and to obtain a simple solution in the .4]. x < 0.St <Su. (Jo = j. (b) for z > 0. where A is the diagonal matrix with 7r on the diagonal: Corollary 2. we shall see that nevertheless we have enough information to derive. We may then assume Ju=Jtu. Remark 2.1 can be rederived using the more detailed form of G+ in Corollary 2. and we let k be the corresponding right eigenvector normalized by Irk = 1.Qi)diag. St < St U.St EA.z+>t) = P.0<u<t) = P.. St E A..=StSt. u It is convenient at this stage to rewrite the above results in terms of the matrix K = 0'Q'A. the CramerLundberg approximation (Section 3).r. and get irPi(Jt =j. (c) the matrix K satisfies the nonlinear matrix equation K = W(K) where W( K) = A ( i) diag + fi J "O eKx S(dx). K( n (d) the sequence converges monotonically to K. oo)) = f o' eIXS((x + z.(.StEA. G+((z.t. 0 +1) = cp (K( n)) defined by K(o) = A .(Jt=j.. From Qe = 0.2. 0<u<t). THE LADDER HEIGHT DISTRIBUTION 157 from which the result immediately follows by integrating from 0 to oo w.S„<0.6 is hardly all that explicit in general.0<u<t. and this immediately yields (2.. St EA. oo))dx.g.6).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.
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
Chie ryu < Vi(u ) < C+hie 7u.8 ) Then for all i E E and all u > 0. yu < r(u) < 00] < hiav)C+o)( y)eavu+yuw(av) 0 Note that the proof appears to use less information than is inherent in the definition (3.j. as in the classical case (3. av 'i [h.. we have ic(ay) < 0 and get 'i(u) . (3.7.y)G+(z. yu) f h(av)e v avuE«v. dy)• o iEE jEE . for a vector <p(u) = (cpi (u))iEE of functions . r(u) < yu] hiay)C+ h=av)C+ o) (y)eayu+yuw(av). we let G+ * W(u) be the vector with ith component E(G+(i. if y > 1lk'(ry).(ay)}.5). However.3. r(u) yu o)(y)eavuEav. We further write G(u) for the vector with ith component Gi(u) = EiEE G+(i. oo)) and. 1 Similarly.V)i(u.00 su e7( ( 3.5) will produce the maximal ryy for which the argument works.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).j) * coj)(u) _ f u ^Pj(u .i I (a) exp {aye(u) + r(u)r. exp {e() + r(u))} .9) For the proof. (u. we shall need the matrices G+ and R of Section 2. 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. yu < r(u) < 00 h 4(u) < h(av)C+o)(y)eavuEav .i [e*(u)K(av). CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 165 hiav)e _avuE. hj P .i [eT(u)K(av ).
166 CHAPTER VI. = Eo G+ G. jEE u 0 j.7.x ) = Gi(u). Lemma 3 .j.ery(&u+x)Bj (dy) Bj(u Bj (u . 00 Thus C+ > hj f"o e7(Yu)G +(i. 00 f C_ hj f e(Y)G+(i. 0 G+(i. MARKOVIAN ENVIRONMENT Lemma 3 . dy) = aj f Bj(dy . dx ) Bj (u . Proof Write UN = EN G+ .x ) R(i.3jhj // f 00 R(i. Hence lim cp(n) exists and equals U * G.u Iv 2°)(u)I Pi(rr+(N + 1) < oo) + 0. if r+ (n) is the nth ladder epoch. Then iterating the defining equation ip(n+1) = G + G+ * V(n) we get W(N+1) = UN * G + G+N+1) * ^(o) However.u IMP:°) (u) I < oo. dy) 00 C+ ijhj f R(i.(0) ] (u) < sup Jt t. Then cpin)(u) sit (u) as n + oo.x) x) jEE 0 E Qj f jEE R(i.dy).x) jEE 00 u 0 //^^ C+E. j.& (u). n > oo. dx) 100 C .12 Assume sup1. j. dx). we have G *(N +1) * ^. j.13 For all i and u. j. just note that the recursion <p(n+1) = G + G+ * (p(n) holds for the particular case where cpin)(u) is the probability of ruin after at most n ladder steps and that then obviously u cp2n) (u) + t. and define W(n+1) (u) = G(u) + (G+ * tp(n))(u). dy) : 1(u) < C+ > hj u e(1tL)G+(i. U = U". dx) f e7( vu)Bj (dy .j. °O . _ To see that the ith component of U * G(u) equals ?Pi (u).
and assuming it shown for n. jEE estimating the first term in (3. and the proof of the lower one is similar.n) ( u . MT < u.13 Let first cp=°)(u) = C_ hie"u in Lemma 3. T) < C+(')' o)hi7u)e7ou8T . we have Vii (u) .3. taking cps°) (u) = 0. 14 Let yo > 0 be the solution of 'c'(yo ) = 0. we get Wo n +1) (u) = ? 7 i ( U ) + E J u gyp.MT<u.11) C_e7u 57 O+[i.M > u) = Pi(ST<u. dy) jEE u U +C_ hje7( yu)G jEE"" +(i. T) = Pi (7. j. 13 and the second by the induction hypothesis .10) C_ 1 f hje7(y.T) = Pi(M > u) . (3. (3. and let 8 = e'(70).8) with y replaced by yo and hi by h=7o ).10: Theorem 3 . Then 0< Vi (u )  0i(u. +i .u)G+(i. Indeed. j.(u) < T ) to 0i (u) which is different from Theorem 3. dy) jEE o (3.11). y]hj = C_ e7uhi. u The proof of the upper inequality is similar . 167 u Proof of Theorem 3.13.13) Hence.10 ) by Lemma 3 . from which the lower inequality follows by letting n * oo. dy) (3.12) Proof We first note that just as in the proof of Theorem 3. 9 for the last equality in (3. let C+(yo) be as in (3.y)G+(i. ST < u] < C+(yo)e7ouEi [h^7o)e70ST1 l T J = C h(7o)e7ou8T . and using Lemma 3 . j.ST).tpi(u. Here is an estimate of the rate of convergence of the finite horizon ruin probabilities 'i (u.Pi(MT > u) = Pi(MT < u.M>u) = Ei [VGJT (u . it follows that Vi(u) < C_(yo) h=70)e7ou. We claim by induction that then cpin) (u) > C_ hie7u for all n. this is obvious if n = 0. letting MT = maxo<t<T St.11. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES proving the upper inequality. j.
(4.5) Note that whereas (4. in part from the folklore principle that any added stochastic variation increases the risk. 0"(u) = P(M" > u)) Now consider the risk process in a Markovian environment and define i' (u) _ >iEE irioi(u). The Markov process {Jt} is stochastically monotone (4. Occasionally we strengthen (4.4) To avoid trivialities. It was long conjectured that 0* Vi..3p.168 CHAPTER VI..o. but that in general the picture is more diverse.. ". u > 0. The conditions which play a role in the following are: . we define the stochastic ordering by 0' < s. this correponds to the usual stochastic ordering of the maxima M'.3). and finally in part from queueing theory. The motivation that such a result should be true came in part from numerical studies..o.2) (4. where it has been observed repeatedly that Markovmodulation increases waiting times and in fact some partial results had been obtained.2) alone just amounts to an ordering of the states. [177]. we also assume that there exist i # j such that either /3i <.. we refer to .3) to B = Bi does not depend on i.. Further related discussion is given in Grigelionis [176]. Bp. (4.0.. B2 <_s. is the one for the Markovmodulated one in the stationary case (the distribution of J0 is 7r).33 or Bi 0 Bj.1) Obviously.3) Bl <_s. The results to be presented show that quite often this is so. <s. M" of the corresponding two claim surplus proceses (note that 0'(u) _ P(M' > u).o. (4. V)" if z/i'(u) <'c"" (u). where o*(u) is the ruin probability for the averaged compound Poisson model defined in Section 1 and .. < . 4 Comparisons with the compound Poisson model 4a Ordering of the ruin functions For two risk functions 0'. For the notion of monotone Markov processes. MARKOVIAN ENVIRONMENT Notes and references The results and proofs are from Asmussen and Rolski [44]. this is not the case for (4.31:5)32 .o.
x) of i and using Lemma 4.2. then j is the more risky state . also Proposition 2.6).4.6.x)dx _ /3*B*(u) + f u / ^ t=1 > 3 * B* ( ) + f (4.1. . p).10) and (4. Proschan & Walkup [140]. = b. we obtain (cf.3* f uB(x) z/^. it follows by a standard . 1:7riaibi > E 7riai i=1 i=1 j=1 The equality holds if and only if a1 = ... we need two lemmas.. The first is a standard result going back to Chebycheff and appearing in a more general form in Esary. the second follows from an extension of Theorem I1.r (JT(o) = i.. dx (4. then P P P 7rjbj. 7(0) < oo) = pirf+). Lemma 4 . 2 If al < .9 ) below).1 for the first term in (4.7) 7ri. note that (4. < bp and 7ri > 0 (i = 1.. < a.2. where 7r2+) = QiµBilri/p.r(u x)dx. and it is in fact easy to show that Vii(u ) < t/j(u) (this is used in the derivation of (4.9) (4. Comparing (4.9) follows by considering the increasing functions 3iBi (x) and Oi (u . Conditions (4. b1 < . 0 Here (4.. = aP or b1 = .. Theorem 4 .1) which with basically the same proof can be found in Asmussen & Schmidt [49]. Proposition 2.7) and Lemma 4. E 7r i Wi(u .5 (cf.1 Assume that conditions (4.x) dx u o i =1 i=1 (4.2)(4.x)B*(x) dx.r (Sr(o) E dx Jr(o) = i.3 for the second) *(u) _ /3 *B* (u) +.8) ^j Tri/iBd(x) . Conditioning upon the first ladder epoch. Proof of Theorem 4.4) is automatic in some simple examples like birthdeath processes or p = 2 .4) say basically that if i < j .2)(4.13* J0 u 0*(u ..* For the proof. Then V..6) 7r= fl*B*(u) + p> s=1 +) fu 0 b (u  x)Bt (x) /pB. Section 4. T(0) < oo) = Bi(x) dx/tcai . (b) P..... (4. ^i 7ri = 1. .10) Q*B*(u)+. COMPARISONS WITH THE COMPOUND POISSON MODEL 169 Stoyan [352]. 3 (a) P. Lemma 4 .4) hold..3iBi(x)YPi(u .
4) is essential (the present author conjectures it is).s. it will hold for all sufficiently large u. it is sufficient to show that 0'. Frey. Recall that .0*• i=1 But it is intuitively clear (see Theorem 3.(0) = V. (4.4 Assume that . let = ( 1/2 1/2 ) . (u) is not in general true: Proposition 4.. (u) may fail for some u. 01 = 103.h..8) we get P P '*' (0) = 3* + /3*1* (0) _ > lri'3qqi • E 7i/ipBi .3i. i=1 i=1 7'r(0) _ EFioiwi(0) .2.s. Proof Since 0. this ruin probability is /3iPBi. As is seen.6)..0.3µi < 1 for all i.(0) < b *'(0) for e small enough.6). Using (4.1 of [145] for a formal proof) that z/ii(u) converges to the ruin probability for the compound Poisson model with parameters .4 is not vacuous..* (0). u Here is a counterexample showing that the inequality tp* (u) < V). = 102.11) i=1` and that A has the form eAo for some fixed intensity matrix A0. that P P /^ 1r1NiµBi /^2 /^ ^i/ji pBi < 1il3i i=1 i=1 (4. µB. Then the l. Q2 = 1.170 CHAPTER VI.4 is the understanding of whether the stochastic monotonicity condition (4. Notes and references The results are from Asmussen. they are at present not quite complete. 4b Ordering of adjustment coefficients Despite the fact that V)* (u) < *. of order 101. For u = 0. and from this the claim follows.1 and Proposition 4.r (u ) fails for all sufficiently small e > 0. of (4. u To see that Proposition 4. What is missing in relation to Theorem 4. 0. Bi as e J. dominates the solution 0* to the renewal equation (4./3*.. except possibly for a very special situation .11) is of order 104 and the r. Rolski & Schmidt [32]. µB2 = 104. Then i/i*(u) < .h. MARKOVIAN ENVIRONMENT argument from renewal theory that tk.
This implies that A is strictly convex.)a. e.5.14) A„(O) iioo varXt t t By convexity.g..13) (4. in particular . The adjustment coefficient y* for the averaged compound Poisson model is the solution > 0 of rc*(ry*) = 0 where rc*(a) _ 13*(B*[a] .a. (4. it follows by Proposition A1.7) )i is convex with A'(0) = lim EXt tioo t = iEE 70i = 0. Lemma 4.2 we have (Ei[e"X'. cf.. (4.12) iEE Theorem 4. with strict inequality unless a = 0 or bi = 0 for all i E E. is nondegenerate unless bi does not depend on i E E. which in view of EiEE 1ibi = 0 is only possible if Si = 0 for all i E E. Hence if 5i 54 0 for some i E E.(a) > 0 for all a 0 0.Jt=kI A (the return time of k) where k E E is some arbitrary but fixed state. It is clear that the distribution of X.4(b) that the limit in (4. COMPARISONS WITH THE COMPOUND POISSON MODEL 171 the adjustment coefficient for the Markovmodulated model is defined as the solution y > 0 of ic(y) = 0 where c(a) is the eigenvalue with maximal real part of the matrix A + (rci(a))diag where rci(a) = ai(Bi[a] .5.5. Further (see Corollary 11. Then {(Jt.6 Let (di)iEE be a given set of constants satisfying EiEE iribi = 0 and define A(a) as the eigenvalue with maximal real part of the matrix A + a(bi)diag• Then )t(a) > 0. 0 .g.ld) with generic cycle w = inf{t>0: Jt_54 k.a = E irirci(a).1) .5 y < ry*. Jt = i])' EE = vA+n(6.13) implies A(a) > 0 for all a. Xt)} is a Markov additive process (a socalled Markovian fluid model.14) is nonzero so that A"(0) > 0. Now we can view {Xt} as a cumulative process (see A.1) . with strict inequality unless rci (y*) does not depend on iEE. Proof Define X= f &ids. Asmussen [20]) as discussed in 11. (4. and by Proposition II.4.
6. Hence rc (y*) > 0. Here we put a = 1/e. a = 1 in Lemma 4. note that y(a) + mins=1.5 is from Asmussen & O'Cinneide [40].. whereas the .Qi and Bi are fixed . Frey.16) Differentiating (4. (4. h depend on the parameter (e or a). 0 = ((ri(Y))diag + ery (4{('Y))diag)h + (A0 + e(?i'Y))diag)h'. MARKOVIAN ENVIRONMENT Proof of Theorem 4. Hence letting e = 0 in (4. h(0) = e. (4. 4c Sensitivity estimates for the adjustment coefficient Now assume that the intensity matrix for the environment is Ae = Ao/ e. Let bi = rci(y*). improving upon more incomplete results from Asmussen. Further a(1) = rc(y*) by definition of A(. the basic equation is (A + (rci(y))diag)h = 0.p yi and compute 8y 8a a=0 In both cases.. we have 7rh' = 0.e7r)1 (Ici(Y*))diage. Rolski & Schmidt [32]. Since ic is convex with rc'(0) < 0 .15) Normalizing h by 7rh = 0. Then > risi = 0 because of (4. multiply the basic equation by a to obtain 0 = (A0 + e(r£i(y))diag)h. h'(0) = (Ao ..) and rc (•).. and our aim is to compute the sensitivity ay e a E=O A dual result deals with the limit a 4 oo.172 CHAPTER VI.15) once more and letting e = 0 we get . Notes and references Theorem 4. this implies that the solution y > 0 of K(y) = 0 must satisfy y < y*. In the case of e.eir)h'(0). y. If rci(y* ) is not a constant function of i E E. where A. we get rc (y*) > 0 which in a similar manner implies that u y < y*.5. Thus y(e) * y* as e 10. The corresponding adjustment coefficient is denoted by ry(e).12) and rc*(y*) = 0.15) yields 0 = (Ii(y*)) diage + Aoh'(0) = (rci('Y*)) diage + (Ao .
We get 0 = (aAo + ( lc&Y))diag)h. (4. multiplying (4. Inserting (4.5.19) holds. 5 The Markovian arrival process We shall here briefly survey an extension of the model.17) by 7r to the left to get (4.19) Then 'y ^ ryl as a ^ 0 and we may take h(0) = el (the first unit vector). (4.8 If (4. . 0 = (Ao + ry'(ii(Y)) diag )h + (aAo + (Ki(7'))diag)h'. and may have some relevance in risk theory as well (though this still remains to be implemented)..17) (4.8 when ryi < 0 for some i is open.16) yields Proposition 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. Rolski & Schmidt [32]. Frey.i(7' *))diagh'(0). the intensity for such a transition (referred to as marked in the following) is denoted by Aii l and the remaining intensity .18).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).18) 0 = 27'(0)p+27r(rs. p. The analogue of Proposition 4. (4. THE MARKOVIAN ARRIVAL PROCESS 173 0 = 27'(0)(ri(`Y *)) diage + 2(ci('Y* )) diag h' (0) + Aoh" (0) .20) Letting a = 0 in (4. then 8a a=o All rci (0) Notes and references The results are from Asmussen. which has recently received much attention in the queueing literature.. i = 2. We assume that 0 < y < 7i. 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.
2 (SUPERPOSITIONS) A nice feature of the setup is that it is closed under superposition of independent arrival streams . This is the only way in which arrivals can occur. Thus . j(2) } be two independent environmental processes and let E(k). where qij is the probability that a transition i * j is accompanied by a claim with distribution. T). with common distribution B. and that are determined by A = A(l ) +A(2) where A is the intensity matrix the governing {Jt}.^) etc. but the point process of arrivals is not Poisson but renewal with interclaim times having common distribution A of phasetype with representation (v. A(1'k) A(2 k1). Jt2)) (2.2 for details). we use the convention that a1i = f3i where 3i is the Poisson rate in state i. and thus 1i = 0. Note that the case that 0 < qij < 1. Here are some main examples: Example 5 . We then let (see the Appendix for the Kronecker E = E(1) x E(2). we may let {Jt} represent the phase processes of the individual interarrival times glued together (see further VIII.174 CHAPTER VI. A(l) = tv.1 (PHASETYPE RENEWAL ARRIVALS) Consider a risk process where the claim sizes are i. II. A(l) = T.(13i )diag. MARKOVIAN ENVIRONMENT f o r a transition i + j by A . refer to notation) { Jt k) }. Jt = (Jtl). For i = j. B. and the marked transitions are then the ones corresponding to arrivals. A(1) = A . let { Jt 1) }.i. the claim surplus is a Markov additive process (cf. .6i ) diag.2).d. then {Nt} is a Markov additive process if and only if it corresponds to an arrival mechanism of the type just considered. The extension of the model can also be motivated via Markov additive processes: if {Nt} is the counting process of a point process. Bij = B. Indeed. u Example 5 . In the above setting. the Markovmodulated compound Poisson model considered sofar corresponds to A(l) = (. the definition of Bi is redundant because of f3i = 0. Bii = Bi . the definition of Bij is redundant for i i4 j. Again .2) A(1) = A(' 1) ® A(1. is neither 0 or 1 is covered by letting Bij have an atom of size qij at 0. that Bii = Bi . A ( 2) = A (2`1 ) ® A.4).
.. more recently. However .iN C17 AilO.1i2.iN = C27 All other offdiagonal elements of A are zero so that all other Bii are redundant.iN.. E = { WORKING. THE MARKOVIAN ARRIVAL PROCESS Bij. iN. say. after which it starts afresh. INVALIDIZED. Thus. possibly having a general phasetype sojourn time.5.. Example 5 ... • upon a claim. The versatility of the setup is even greater than for the Markovmodulated model.. RETIRED.}.kl is redundant). or. This means that the environmental states are of the form i1i2 • • • iN with il.. 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..iil. Similarly.. Hermann [193 ] and Asmussen & Koole [37] showed that in some appropriate . MARRIED.. e.. superpositions of renewal processes. Assume further that the ith policy leads to a claim having distribution Ci after a time which is exponential. In fact . Easy modifications apply to allow for • the time until expiration of the kth policy is general phasetype rather than exponential. i2i . where ik = 0 means that the kth policy has not yet expired and ik = 1 that it has expired. and that the policy then expires. assume that there is a finite number N of policies. the idea of arrivals at transition epochs can be found in Hermann [193] and Rudemo [313]. Bilo.iN = a2. WIDOWED.. the kth policy enters a recovering state..iN.. In this way we can model. DEAD etc. u Example 5 .iN are zero and all Bi are redundant. iN. claims occur only at state transitions for the environment so that AN2. E 10. iN = all BOi2.g.. 11.iil... 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 . as the Markovian arrival process ( MAP).. all Al i2.3 (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).1i2 . with rate ai.kj = Bik) B13 4k = Bak) 175  (the definition of the remaining Bij... DIVORCED.4 (A SINGLE LIFE INSURANCE POLICY ) Consider the life insurance of a single policy holder which can be in one of several states.
let the period be 1. Thus at time t the premium rate is p(s + t). 1). . one limitation for approximation purposes is the inequality Var Nt > ENt which needs not hold for all arrival streams. By periodic extension. 0 < t < 1. Some main queueing references using the MAP are Ramaswami [298]. 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 . we may assume that the functions /3(t). Without loss of generality. a claim arrives with rate /3(s + t) and is distributed according to B(8+0 .p)/p. We denote throughout the initial season by s and by P(8) the corresponding governing probability measure for the risk process. Lucantoni [248]. we talk of s as the 'time of the year'. )3 t 1 J (6. For the Markovmodulated model.176 CHAPTER VI. [248].1) Then the average arrival rate is /3* and the safety loading rt is 77 = (p* . for s E E = [0. Lucantoni et at. 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. where i f00 xB(°) (dx) _ . Neuts [271] and Asmussen & Perry [42]. from an application point of view. but now exhibiting (deterministic) periodic fluctuations rather than (random ) Markovian ones. Sengupta [336]. • Claims arriving at time t of the year have distribution B(t). Let 1 1 /3* _ f /3(t) dt. p * = 0 p(t) dt.2) Note that p is the average net claim amount per unit time and µ* = p//3* the average mean claim size. 1). • The premium rate at time t of the year is p(t). one needs to assume also (as a minimum) that they are measurable in t. continuity would hold in presumably all reasonable examples. B* = J f B(t) ((*) dt. The basic assumptions are as follows: • The arrival intensity at time t of the year is 3(t) for a certain function /3(t).3*µs • p = f /3(v) dv 0 0 (6. p(t) and B(t) are defined also for t t [0. Obviously.
1 As an example to be used for numerical illustration throughout this section. the average compound Poisson model is the same as in III. Thus . p* as an averaged version of the periodic model.3) Note that A enters just as a scaling factor of the time axis. the discussion in 111. The claim surplus process {St } two is defined in the obvious way as St = ^N° Ui . It is easily seen that . and thus the averaged standard compound Poisson models have the same risk for all A.3(t) = 3A(1 + sin 27rt). let . The arrival process {Nt}t>0 is a timeinhomogeneous Poisson process with intensity function {/3(s + t)}t>0 . and we recall from there that the ruin probability is 24 1 *(u) _ 3 5eu + 35e6u. In contrast.w(t).(3. We u assume in the rest of this section that p(t) . of the periodic model as arising from the compound Poisson model by adding some extra variability. u Remark 6 .1. for Markovmodulated model typically the adjustment coefficient is larger than for the averaged model (cf. In particular. (6. or. one may think of the standard compound Poisson model with parameters 3*. Many of the results given below indicate that the averaged and the periodic model share a number of main features. it turns out that they have the same adjustment coefficient. Thus.2 Define T 6(T) = p(t ) dt. The behaviour of the periodic model needs not to be seen as a violation of this principle. respectively.8. not random. Example 6 . St = SeI(t).10. we shall see that for the periodic model increasing A increases the effect of the periodic fluctuations.t.w(t)) dt). since the added variation is deterministic. B*.6. RISK THEORY IN A PERIODIC ENVIRONMENT 177 In a similar manner as in Proposition 1.1) and Example 1. 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 . equivalently. 0 Then (by standard operational time arguments ) {St} is a periodic risk process with unit premium rate and the same infinite horizon ruin probabilities. Section 4b).3* = 3A.9). in agreement with the general principle of added variation increasing the risk (cf. p* = A whereas B* is a mixture of exponential distributions with intensities 3 and 7 and weights 1/2 for each (1/2 = ff w(t)dt = f o (1. the conditional distribution . In contrast.
a . i.178 CHAPTER VL MARKOVIAN ENVIRONMENT of U. but it turns out to have obvious benefits in terms of guidelining the analysis of the model as a parallel of the analysis for the Markovian environment risk process.a) = exp { . of the averaged compound Poisson model (the last expression is independent of s by periodicity). Notes and references The model has been studied in risk theory by.s . [101] . 6b Lundberg conjugation Motivated by the discussion in Chapter II. As usual. and define h(s.1]) . J Theorem 6 . see the Notes to Section 7). t + dt] or not. we start by deriving formulas giving the m. To this end.tc* (a)] dv then h (.f.f. [44] (the literature in the mathematical equivalent setting of queueing theory is somewhat more extensive. 1).5. The claim surplus process {St} may be seen as a Markov additive process. let f 8+1 tc *(a) _ (B* [a] . . (6. a) is periodic on R.g.. a) etw*(a) h(s+t.8).adt +. and the ruin probabilities are 0(8) (U) = P(s )(r(u) < 00).a) Proof Conditioning upon whether a claim occurs in [t.east B(8+t) [a] east . of the claim surplus process..(3(s + t)dt)e«St adt + /3(s + t)dt .4) At a first sight this point of view may appear quite artificial. e.T) = P(8)(r(u) <T). 3 E(8)eaSt = h(s.^8 [.(1 . we obtain E.5 (see in particular Remark 11.3(v)(B(vl [a] .1) dv . with the underlying Markov process {Jt} being deterministic period motion on E = [0. 0 (5)(u.a.. given that the ith claim occurs at time t is B(8+t).1) . with some variants in the proofs.3(s + t)dt[B(8+t)[a] .1) a = J8 .a be the c. r(u) _ inf It > 0 : St > u} is the time to ruin .(8) [eaSt+dt I7t] = = (1 .g.al. Jt = (s + t) mod 1 P(8) . Dassios & Embrechts [98] and Asmussen & Rolski [43]. Daykin et.Q(v) (B(„) [a] .g. The exposition of the present chapter is basically an extract from [44].e.
1]) . St)} and .6.5. a) = h(s. at + f log h(s + t.1)dv l og E(8) et where atetk•(a) h(t. we can write Lo Jt. 0) exist and are finite.(e) Let = h( h(Jo. + v)(B([a] .5 The formula for h(s) = h(s. According to Remark 11. RISK THEORY IN A PERIODIC ENVIRONMENT E(8)east+ dt d Et.t} is a multiplicative functional for the Markov process { (Jt.t}t>o = h(s. a) et.s. a) = exp I f t3(v)(kv)[a) .t. h(s + t. a).1]) .3.3(s + t)[D(8 +t)[a] .4 For each 0 such that the integrals in the definition of h(t .* (a) h(s.2.0(s + t)dt[B(8+t)[a] .(8)east 179 = = = = = E(8)east (1 .log h(s.. u Remark 6.1]. E (8)east (a +. it then suffices to note that E(8)Le.6 . dt log E(8)east a + f3(s + t) [B(8+t) [a] .1)dv  o h(t. St)} . a) h(s + t.c* (e) {Le.adt +.9 as follows.9) eastt. a) . With g the infinitesimal generator of {Xt} = {(Jt. a) Thus E(8)east = h(s + t.t = 1 by Theorem 6. 0) P(8)a. Proof In the Markov additive sense of (6.4). B) eoSt t. a) as well as the fact that rc = k` (a) is the correct exponential growth rate of Eeast can be derived via Remark 11. a) Corollary 6. so that obviously {Lo. 9) is a P ( 8)martingale with mean one.
g. That is. ( iii) use approximations with piecewiese constant /3(s).7 When a > yo. St)} with governing probability measures Fes).'y).3(v)( Bi"i [a] . y solves n* (y) = 0. Now define 'y as the positive solution of the Lundberg equation for the averaged model. (ii) use Markovmodulated approximations (Section 6c).3(s)h(s) + h'(s) +. That is.60(t) = a(t)B(t)[0]. it follows by Theorem II. the restrictions of Plsi and Pest to Ft are equivalent with likelihood ratio Le. [70] . we put for short h(s) = h(s.3(s)ks)[a]h(s)} ah(s) 13(s)h(s) + h'(s) +. see [44] for 11 a formal proof.6 The P(s).5 that we can define a new Markov process {(Jt.a . When a = y. ry)) at which n* (a) attains its minimum. Sdt) = h(s + dt) eadt (1 . Proof (i) Check that m. B(s). Proposition 6. 0) = h(s) + dt {ah(s) . A further important constant is the value yo (located in (0.3(s)dt • B(s)[a]h(s) = gha(s.T.3. yo is determined by 0 = k* (70) = QB*.f. P(s) (T(u) < oo) = 1 for all u > 0.2.6 ( s ) exp { 0( s )&s) [a] + tc . (iv) finally. Equating this to rch (s) and dividing by h(s) yields h(s ) = h(s) = a + .y) = eayh(s).0) = Kh(s). .1.1) . Lemma 6 . That rc = is*(a) then follows by noting that h(1) _ u h(0) by periodicity.tc] dv} (normalizing by h(0) = 1). correspond to a new periodic risk model with parameters ex . For each 0 satisfying the conditions of Corollary 6. say. However. as above E (s) ha(Jdt. of St is as for the asserted periodic risk model. J s [. 0 < s < 1.3(s)B(s) [a]h(s).(3(s)dt) +. the requirement is cha(i.4. Proposition 6.180 CHAPTER VI. such that for any s and T < oo. MARKOVIAN ENVIRONMENT ha(s. Bet)(dx) = ^ B(t ) (dx). cf.
^(u) = ST(u) . T(u) < (6.2. (6. The relevant likelihood ratio representation of the ruin probabilities now follows immediately from Corollary 11. xEJ 0 (s)b(8)(x) > 0. have components with densities b(8)(x) satisfying inf sEI. which is not used elsewhere in the book. 1).7) h(B(u). we get: . Lemma 6 .1. Corollary 6.9 Assume that there exist open intervals I C [0. 9(u)) for any bounded continuous function (e. RISK THEORY IN A PERIODIC ENVIRONMENT Proof According to (6. the Markov process {(^(u).9) and noting that weak convergence entails convergence of E f (^(u). T) = h(s. and we refer to [44]. u which is > 1 by convexity. and no matter what is the initial season s. a)e«uE (a iP(s) (u) = h( s)e7uE(` ) h(O(u)) To obtain the CramerLundberg approximation from Corollary 3.4. Wu).g. The proof involves machinery from the ergodic theory of Markov chains on a general state space.8 The ruin probabilities can be computed as (u)+T(u)k'(a) ^/i(8) (u. 0(u)) * (b(oo). a) e«uE(8 ) e «^ .2). a) TI h(9(u). we need the following auxiliary result . e(cc)) Letting u > oo in (6.1) the distribution of (l: (oo).10) Then for each a. a) a > ry0 (6.9(u))} u>0. q) = eryx/h(q)).9) 0(')(u) = h(s. the mean number of claims per unit time is p« 181 = Jo 1. f (x. J C R+ such that the B(8). s E I.8) (6. considered with governing probability measures { E(8) }E[ .6. say s0.6(v) dv Jo ' xe«xB (°) (dx) r^ xe«xB'(dx) = Q'B' [ a] = ^' J 0 = ^c"'(a) + 1. has a unique stationary distribution. B(oo)). Here and in the following.u is the overshoot and 9(u) = (T(u) + s) mod 1 the season at the time of ruin.
11 7/'O (u) < C+°)h(s) ery". it does not seem within the range of our methods to compute C explicitly. where C(o) = 1 + info < t<i h(t) . MARKOVIAN ENVIRONMENT Theorem 6.11) gives an interpretation of h(s ) as a measure of how the risks of different initial seasons s vary.1. 10 shows that certainly ry is the correct Lundberg exponent. A=1/4 A=1 A=4 0 Figure 6. elementary calculus yields h(s) = exp { A C 2^ cos 2irs  4^ sin 21rs + 11 cos 41rs .1 In contrast to h.11) Note that ( 6. we obtain immediately the following version of Lundberg ' s inequality which is a direct parallel of the result given in Corollary 3. At this stage .10) of Lemma 3. where e. Among other things. (6.9). Theorem 6 .16.W. Noting that ^(u) > 0 in ( 6. illustrating that the effect of seasonality increases with A.ir) } Plots of h for different values of A are given in Fig.Ch(s)ery".10 Under the condition (6. 1. this provides an algorithm for computing C as a limit. Vi(8) (u) .1. For our basic Example 6 .6 for the Markovmodulated model: Theorem 6 . 6.) C = E1 h(B(oo)) u + oo. which may provide one among many motivations for the Markovmodulated approximation procedure to be considered in Section 6c.182 CHAPTER VI.
4.(ay) > 0 when y < 1/ic' (7). . in our basic example with A = 1.7e . 1 (6.167r I Cu. where ay is the unique solution of W(ay) =y• (6.g.47r sin 27rs + 167r cos 47rs .w) .13 Let = 1 B(t) C o<tf i h(t) 2no f °O e'r(Yx)B( t) (dy)' (x) x 1 B(t) (x) C+ = sup sup o<t<i h ( t) xo J. e7 ( yx)B(t)(dy) > Then for all $ E [0.42 so that 183 tp(8) (u) < 1. We state the results below.12) As for the Markovian environment model.13) Elementary convexity arguments show that we always have ryy > Y and ay > ry.0(8) (u+ yu) (6.(8) (u. whereas ay < y. Theorem 6 . e.w)e4u dx 9w + 7(1 . Just as in IV.15) The next result improves upon the constant C+) in front of eryu in Theorem 6. we substitute T = yu in 0(u.14) < C+)(y)h(s) e7yu. Theorem 6. ay) • (6. RISK THEORY IN A PERIODIC ENVIRONMENT Thus. 1 ) and all u > 0.7x j dx _7x } _ 6w + 6(1 .. T) and replace the Lundberg exponent ry by ryy = ay . we first note that the function fu° exu {w • 3e .17) (6. (ay).42 • exp {J_ cos 27rs . (6. we obtain Co) = 1. Consider first the timedependent version of Lundberg's inequality.12 Let 00)(y) 1 Then info < t<i h(t.(s)(u) < C+h(s)e7".13 to our basic example. the proofs are basically the same as in Section 3 and we refer to [44] for details.yr.w)e4u .6. #c( ay) < 0 when y > 1/tc'('y). yu) 000 (u) .w ) • 7e u{w • 3e3x + ( 1 .11 as well as it supplements with a lower bound. r.16 In order to apply Theorem 6. C_h(s)e7u < V.3x + (1 . Lundberg's inequality can be con siderably sharpened and extended.
1 sin 27rs + 1 cos 47rs . with s the initial season.16) with 'y replaced by yo and h(t) by h(t. .19 I eu. . 14 Let C+('yo) be as in (6. such a deterministic periodic environment may be seen as a special case of a Markovian one (allowing a continuous state space E = [0. Of course. . and in fact. MARKOVIAN ENVIRONMENT attains its minimum 2 /3 for u = oo and its maximum 6 /(7 + 2w) for u = 0. n}. and let 8 = er' (Y0).(8)(u. 1/i18 1 s (u) > 0. The idea is basically to approximate the (deterministic) continuous clock by a discrete (random) Markovian one with n 'months'.9 3 0<8<1 p 27r 47r 167r 161r 2 _ _e.66.20).T) < C+('Yo)h( s.L sin 27rs + 1 I cos 47rs . 1) for the environment). exp 2^ cos 21rs . Some of the present proofs are more elementary by avoiding the general point process machinery of [44]. much of the analysis of the preceding section is modelled after the techniques developed in the preceding sections for the case of a finite E. Finally.4^ sin 2irs + 16^ cos 41rs . but thereby also slightly longer. Then 7oudT .'Yo)e (6..18) Notes and references The material is from Asmussen & Rolski [44].013.184 CHAPTER VI.I eu. completing a cycle . we have the following result: Theorem 6 . the nth Markovian environmental process {Jt} moves cyclically on {1.181 s(u) < 1. Thus.\ = 0 ..013.0. 1). 0 <'p(8)(u ) .16. 6c Markovmodulated approximations A periodic risk model may be seen as a varying environment model. Thus C_ = 2 inf ex cos 2irs .1 sin 2irs + 16_ cos 47rs .cos 27rs . C+ = 1.g.66. yo).19 } 0 <8<1 8 + cos 21rs Thus e. for A = 1 (where 3 e0.\ 3 C+ = sup 6 exp { A (. where the environment at time t is (s + t) mod 1 E [0. This observation motivates to look for a more formal connection between the periodic model and the one evolving in a finite Markovian environment.20 •exp { 2n cos 27rs .
. To this end. it is desirable to have formulas permitting freely to translate from one setting into the other. so that the intensity matrix is A(n) given by n n 0 ••• 0 0 n n ••• 0 A(n) _ (6. but others are also possible. since the settings are equivalent from a mathematical point of view. AE= Aii'r?/7ri• The arrival intensity is /3i when Jt = i.7.21) which serves as an approximation to 0(1)(u) whenever n is large and i/n s. (6. A be the parameters defining the risk process in a random environment and consider a queueing system governed by a Markov process {Jt } ('Markovmodulated') as follows: • The intensity matrix for {Jt } is the timereversed intensity matrix At _ A ())i. M(n) = Supt>o Stn). Notes and references See Rolski [306]. Thus. Bi. and the ruin probability corresponding to the initial state i of the environment is then Y'yn)(t) = F (M(n) > t). DUAL QUEUEING MODELS 185 within one unit of time on the average .jEE of the risk process.1 ((i 1)/n) ) and Bni = B . This queue is commonly denoted as the Markovmodulated M/G/1 queue and has received considerable attention in the last decade. 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. z/'i (u.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. Let 0j. We want to choose the /3ni and Bni in order to achieve good convergence to the periodic model. one simple choice is Oni = 0( i . We let {Stn)} (6. 7 Dual queueing models The essence of the results of the present section is that the ruin probabilities i/ (u).20) be the claim surplus process of t>o the nth approximating Markovmodulated model.
(VT > u I JT = 2). (7. and for (7.3) 7ri where (V. Taking probabilities and using the stationarity yields 7riPi(T(u) < T..P(V > u.2) and use that limF (VT > u. let T .1). For (7. ii (u) = it /3 P(W > u. • The queueing discipline is FIFO.3.2 The relation between the steadystate distributions of the actual and the virtual waiting time distribution is given by F(W > u. (7. The actual waiting time process 1W1.. J* = i) = P.oo in u (7. JT = j) = LjPj (VT > u.T(V > u I J* = i). {Jt }o<t<T• Then we may assume that Jt = JTt.2). JT = j} and {VT > u. J*) is the steadystate limit of (Vt.1) over j.. In particular. . JT = j) = 7rjPj(VT > u. and the virtual waiting time (workload) process {Vt}too are defined exactly as for the renewal model in Chapter V. Proof Consider stationary versions of {Jt}o<t<T. Jo = j. I* )3i P(V > u. .4) where 0* = >jEE 7rj/3j. Proposition 7. JJ = i).3). 2 . The first conclusion of that result then states that the events {T(u) < T. just sum (7. JT = i) = P(V > u. Jo = i.0i (u . Then Pi(T(u) < T. JT = i} coincide. J* = i) for all j. T) = 7ri 1 P.186 CHAPTER VI.n(VT > u. JT = i) = 'P. (7. JT = Z).1 Assume V0 = 0. I* = i). MARKOVIAN ENVIRONMENT • Customers arriving when Jt = i have service time distribution Bi. Proposition 7. 0 < t < T and that the risk process {Rt}o<t<T is coupled to the virtual waiting process {Vt}o<t<T as in the basic dualitylemma (Theorem 11. Jt ).1) 7ri In particular. J* = i).=1 .1) follows. (7. Now let In denote the environment when customer n arrives and I* the steadystate limit. and (7.2) Oi(u) = 1.
we have 1: I(W. if T is large. . J* = i) see W > u. P(W >u.3). In the setting of the periodic model of Section 6. a paper relying heavily on classical complex plane methods. Taking the ratio yields (7. see Regterschot & van Doorn [123]. u Notes and references One of the earliest papers drawing attention to the Markovmodulated M/G/1 queue is Burman & Smith [84]. I*) with the timeaverage . I* = i. [243].I *=i). With {Vt} denoting the workload process of the periodic queue.7) of that paper. and Rolski [306]. and of these.o. p < 1 then ensures that V(*) = limNloo VN+9 exists in distribution. P(1')(r(u) < oo) = P(')(00) > u). DUAL QUEUEING MODELS 187 Proof Identifying the distribution of (W. n=1 N However.I. 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.7) (7. N * oo. and further references (to which we add Prabhu & Zhu [296]) can be found there. that /3(t). (7. >u. Lemoine [242].T)(T(u) <T) = P(8)(VT > u). P(..4) and (7.1 is from Asmussen [16]. a general formalism allowing this type of conclusion is 'conditional PASTA'. with (7. Proposition 7.g.4).5) follows from (7.l. B(t) have been periodically extended to negative t). The first comprehensive solution of the waiting time problem is Regterschot & de Smit [301]. T].6) (7.7. on average /32TP(V > u. see in particular Harrison & Lemoine [186].3) improving somewhat upon (2. and one has PI'>(rr(u) < T) = P('_T)(VT > u). on average 0*T customers arrive in [0. and (7. The relation (7.8) For treatments of periodic M/G/1 queue. A more probabilistic treatment was given by Asmussen [17].4) can be found in Regterschot & de Smit [301].=i) a4.
This page is intentionally left blank .
finite horizon. t] are Nt At = Ui (1.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 . and the evolution of the reserve may be described by the equation Rt = u .T) = F(T(u) < T). the aggregate claims in [0. resp . i&(u. 189 .6. with common distribution B and independent of {Nt}. . {Rt} moves according to the differential equation R = p(R).At + p(R8) ds...i. Thus. are i. However .T) = FloinfTRt< OIRo=u1 denote the ruin probabilities with/initial reserve u and infinite.2) tk(u.1) (other terms are accumulated claims or total claims). the premium charged is assumed to depend upon the current reserve Rt so that the premium rate is p(r) when Rt = r. Zt As earlier. z/i(u) = F IinffRt< 0IRo=u 1 (1. U2. Thus in between jumps. and that the claim sizes U1.d. and T(u) = inf {t > 0 : Rt < u} is the time to ruin starting from Ro = u so that '(u) = F(T(u) < oo).
Hence in terms of survival probabilities. rather than when the reserve itself becomes negative. or o(u) < 1 for all u.p2. say at interest rate b.3 (ABSOLUTE RUIN) Consider the same situation as in Example 1. Example 1. Another could be the payout of dividends: here the premium paid by the policy holders is the same for all r.p/S) r > p/S p5(p/5r) 0<r<p/5 Then the ruin problem for {Rt } is of the type defined above. Assume 0(u) < 1 for some u. P(r) _ p + e(r . and the probability of absolute ruin with initial reserve u E [p/S.e. oo) is given by i (u + p/S). when x > p/S. However. pi > p2 and p(r) = One reason could be competition. No tractable necessary and sufficient condition is known in complete generality of the model.i(u) = 1 for all u. In this situation. we get p(r) = p + er. That is.4 Either i.1 Assume that the company reduces the premium rate from pi to p2 when the reserve comes above some critical value v. If Ro = v < u. i. A basic question is thus which premium rules p(r) ensure that 'O(u) < 1. Thus at deficit x > 0 (meaning Rt = x).'(u)) > 0 so that V'(v) < 1. we can put Rt = Rt + p/S. 1 . but when the reserve comes above v. that {Rt} will reach level u before the first claim arrives.190 CHAPTER VII. it seems reasonable to assume monotonicity (p(r) is u . the payout rate of interest is Sx and absolute ruin occurs when this exceeds the premium inflow p. Example 1. but assume now that the company borrows the deficit in the bank when the reserve goes negative. Proof Obviously '(u) < ilb(v) when u > v. dividends are paid out at rate pi . RESERVEDEPENDENT PREMIUMS The following examples provide some main motivation for studying the model: Example 1 .2 (INTEREST) If the company charges a constant premium rate p u but invests its money at interest rate e.Vi(v) u > e(1 . Proposition 1. where one would try to attract new customers as soon as the business has become reasonably safe. there is positive probability. say e.2. Now return to the general model.
{ Vt} remains at 0 until the next arrival).1. (1. instead of (1. We next recall the following results. (b) If p(r) > /3µB + e for all sufficiently large r and some e > 0. let uo be chosen such that p(r) < p = /3µB for r > uo. Starting from Ro = uo. 296297): Theorem 1. This is basically covered by the following result (but note that the case p(r) .I3IB requires a more detailed analysis and that µB < oo is not always necessary for O(u) < 1 when p(r) 4 oo.5 (a) If p(r) < /. Here {Vt}two is a storage process which has reflection at zero and initial condition Vo = 0.3.. In case (a). appealing to Proposition 111. if and only if V)(u) < 1 for all u.3µB for all sufficiently large r. Then 0(u) = P(V > u).1.2) for r sufficiently large so that p(oo) = limr. Proposition I1I. that u zPp(u . and hence by a geometric trials argument.1. one can couple the risk process and the storage process on [0. Theorem 1.b(u. B and (constant) premium rate p. Let Op(u) refer to the compound Poisson model with the same 0.2) we have t Vt = At .6) . the probability that Rt < uo for some t is at least tp(0) = 1 (cf. Hence ik(u) < 1 for all u by Proposition III.2(d). cf.4.uo) < 1. That is.T) = P(VT > u).6 For any T < oo.2 once more. . INTRODUCTION 191 decreasing in Example 1.4) 0 and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0. hence Rt < uo also for a whole sequence of is converging to oo.f p(Vs) ds. which was proved in 11.1. obviously infu<uo z/'(u) > 0.+ p(r) exists. then ?(u) = 1 for all u. (1. we have z/i(u) <p(u . {Vt} decreases at rate p(v) when Vt = v (i.e.5) and the process {Vt} has a proper limit in distribution . say V.2(d)). In case (b). Proof This follows by a simple comparison with the compound Poisson model. In between jumps. (1. V = p(V)). T] i n such a way that the events {r(u) <T} and {VT > u} coincide. and P(Rt + oo) > 0. let uo be chosen such that p(r) > p = 0ILB + e for r > uo. However. [APQ] pp. In particular. then l/i(u) < 1 for all u. Then if u > no.uo) and.1 and increasing in Example 1.o(uo) = 1 so that t/'(u) = 1 for all u by Proposition 1.
say.8) Proof In stationarity.y)g(y) dy. In view of the path structure of {V t }. one having an atom at 0 of size 'yo. (1. t + dt] if and only if Vt E [x. the flow of mass from [0. (1. It follows in particular that 0(u) = fg(Y)dy.Sx}. say. t + dt] can be neglected so that a path of {Vt} corresponds to a downcrossing in [t. u Define ^x 1 w(x) Jo p(t) dt. say when {Vt} is in state y.6 applicable. we arrive at the desired interpretation of the r. yo ^ 1 + oo Q exp {.8) is the rate of downcrossings (the event of an arrival in [t. Jo AX) (1. and is succesful if the jump size is larger than x . and the other being given by a density g(x) on (0.6x and that w(x) < oo for all x > 0.8 Assume that B is exponential with rate b. the l.7 p(x)g(x) = tofB (x) + a f (x . B(x) = e.h.y. of (1. Now obviously. oo).8) as the rate of upcrossings. this means that the rate of upcrossings of level x must be the same as the rate of downcrossings. x] to (x.Qw(x) . RESERVEDEPENDENT PREMIUMS In order to make Theorem 1. 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. where g(x) = p( ^ exp {.7) Proposition 1.Sx} dx. say if p(r) goes to 0 at rate 1 /r or faster as r j 0. x + p(x)dt]).s. for the storage process {Vt}.8) as g(x) = p 1 {yo13e_6x +. oo) must be the same as the flow the other way.192 CHAPTER VII. Corollary 1. It is intuitively obvious and not too hard to prove that G is a mixture of two components. Oeax f x e'Yg (y) dy } = p) eaxa(x) . Note that it may happen that w (x) = oo for all x > 0.9) Proof We may rewrite (1.s.h.6w(x) . of (1. An attempt of an upcrossing occurs as result of an arrival. we thus need to look more into the stationary distribution G. Considering the cases y = 0 and 0 < y < x separately. Then the ruin probability is tp (u) = f' g(y)dy.
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
13/E).3a/ (5 .f. As an example.e. .3 is also valid if {Rt} is obtained by adding interest to a more general process {Wt} with stationary independent increments.x) dx e.pa.5 The analysis leading to Theorem 2. i. assume that {Wt} is Brownian motion with drift µ and variance v2.V < x)]0 + f P(V > p/E ) + eby fv (p/E . The process {St} corresponds to {Wt} so that c(a) or2a2/2 . /^ u Example 2 . From ic(a) = .b P/E dx /' P/ ' (p/  x)p/e 150/f I' (/3/E) (6P1'E.01'E) + (p/E)al aO l febP/E } IF (0 /0 jF From this (2.3/E) By the memoryless property of the exponential distribution.2y +µ ) dy . of Z is IogH[a] = f ytc(y)dy = e fa (0.a) .200 CHAPTER VII.V. RESERVEDEPENDENT PREMIUMS u from which (2. it follows that logH[a] = f 1 c(y)dy = 1 f '(pa/(a +y))dy f 0 0 Ey R/E 1 [pa + )3log 8 ./3 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.2) follows by elementary algebra.g.2) follows by elementary algebra. RT(u) has an exponential distribution with rate (S) and hence E [H(RT(u))I r(u) < oo] L Pe6'r (P/C .pa.13 /E) r (. r (j3/E) In particular. Proof 2 We use Theorem 2. with density x(3/e1aQ/e fV (x) _ e 6X ' x > 0.3. and the c. where V is Gamma(b. then {Rt} is the diffusion with drift function p+Ex and constant variance a2.
3. 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).. Delbaen & Haezendonck [104]. Paulsen & Gjessing [286] and Sundt & Teugels [356]. 3 The local adjustment coefficient.e. THE LOCAL ADJUSTMENT COEFFICIENT _ Q2a2 pa 4e E 201 I. [129] and Harrison [185]. A r. write Vi* (u) for the ruin probability etc. or to nonlinear premium rules p(•). [129].2 is a special case of a perpetuity. Further studies of the model with interest can be found in Boogaert & Crijns [71]. Corollary 2. Z is normal (p/E. It must be noted. The formula (2.p*.. se e. 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. Paulsen [281].g. not even Erlang(3) or H3. that the analysis does not seem to carry over to general phasetype distributions.8.Y*p* W*(u) < ery*u = 0. [283]. it follows that the ruin probability is Cu) H(u) H(0) 11 Notes and references Theorem 2. write y* for the solution of the Lundberg equation f3(B[ry *] . Goldie & Griibel [167].v. as in the proof of Proposition 2. Some of these references also go into a stochastic interest rate. Paulsen & Gjessing [286] found some remarkable explicit formulas for 0(u) beyond the exponential case in Corollary 1.g. Gerber [157] p.3 is from Harrison [185].4 is classical. and since RT = 0 by the continuity of Brownian motion. see e. Emanuel et at. for a martingale proof.d. [282].3) was derived by Emanuel et at. 134 (the time scale there is discrete but the argument is easily adapted to the continuous case). and recall Lundberg 's inequality . Q2/2E).1) .i. of the form Ei° p"X" with the X„ i. [357]. Gerber [155]. however.
we will use the local adjustment coefficient 'y(x). When u > uo. (3. a first step is the following: Theorem 3 . Then lim sup u>oo u and e E''p(r) + 0. 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)..e.1) . Let y* < So.E). Proof of Theorem 3. which in turn by Lundberg's inequality can be bounded by ery*(1E)" Hence limsup„.e.ap(x).>o 7(x) > 0. x * oo. as solution of the equation n(x. c(. and that p(x) * oo. it holds that f3[s] T oo.w (u) J dt > c(3)eeu v 1 p(u+ t) . 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. i. x>0 (3. 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 . choose c(. and (for simplicity) that inf p(x) > (3µs . a) = f3(B[a] .C*ef*". For the last asssertion .*(u) . The steepness assumption and p(x) + oo ensure 'y(x) * So. Letting first E * 0 and next ry * T 5o yields the first statement of the theorem.1. let p* be a in (3.'y ( x)) = 0 where r.1 Assume that for some 0 < 5o < oo. e(1o+e)2 (x ) u > 00. choose uo such that p( x) > p* when x > u0E.2) such that p(x) < c(. 1) and for a given E > 0. 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) . (x.1. (3.log '(u)/u < ry*(1 .4) we assume existence of y(x) for all x.i)eex. log ?i(u) < < 00 JO .3) When trying to extend these results to the model of this chapter where p(x) depends on x. then log u (u) In the proof as well as in the remaining part of the section .1 ). B(x) > C(2)e(ao+f)x for all x.202 CHAPTER VII. as will hold under the steepness assumption of Theorem 3.. i. oo for all E > 0. If 60 s f 6o.5) which implies inf. RESERVEDEPENDENT PREMIUMS and the CramerLundberg approximation V.
the limit is not u + oo but the slow Markov walk limit in large deviations theory (see e.6) The second main result to be derived states that the bound in Theorem 3. Then limelog l/ie (u) = I(u). {Rte)} defined as in (1.v. let 0e (u) be evaluated for the process only with 3 replaced by /0/e and U.1 only presents a first step. ruin will occur if the claim is at least u + v. the asymptotics u * oo and c . (u) = O(u/e). . then Rte) = CRtie for all t so that V).13 below holds. (3.3 Assume that either (a) p(r) is a non decreasing function of r.7) CIO Remarks: 1. 2 Assume that p(x) is a nondecreasing function of x and let I(u) = fo ry(x)dx. The rest of this section deals with tail estimates involving the local adjustment coefficient.' (u) < eI("). 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. Condition 3. Theorem 3 . The first main result in this direction is the following version of Lundberg's inequality: Theorem 3 . Bucklew [81]).ea°/(ecf1)). The form of the result is superficially similar to the CramerLundberg approximation. (3. one can then assume that e = 1 is small enough for Theorem 3.e. However. the result is not very informative if bo = oo. noting that in many cases the constant C is close to 1.. or (b) Condition 3.2). UJU > x cannot have a much heavier tail than the claim U itself.3. 3. 2.4)eE" Given such an arrival.(u) > so.g. If p(x) = pis constant . and in particular. I. Then . Therefore the probability that a claim arrives ( before the reserve has reached level u + v is at least c(.0 are the same.3 to be reasonably precise and use e` (u) as approximation to 0 (u). u Obviously. which essentially says that an overshoot r. 3) = (1 .2 is also an approximation under appropriate conditions. Theorem 3. by cU2. THE LOCAL ADJUSTMENT COEFFICIENT 203 where c. For e > 0.13 is a technical condition on the claim size distribution B. and hence '(u) > c(4)eeuc( 2)e(do+e)u The truth of this for all e > 0 implies lim inf log V.
8 in terms of I(u) when the claims are exponential: Example 3 . Then y(x) = b . However.bx} dx fo 00 1 + [exp {/(3w(x) . 3.7) is only captures 'the main term in the exponent' but is not precise to describe the asymptotic form of O(u) in terms of ratio limit theorems (the precise asymptotics could be logI(u)e1(U) or I(u)"e_I(u). As typical in large deviations theory.bx}]o + b /' oo exp low (x) . we get = 1+ J" AX) exp {(3w (x) .bx} dx 1+0.3.exp {/33w(u) . 3a Examples Before giving the proofs of Theorems 3. and r j 1 'Yo v(x)dx = bu  a J0 p(x)ldx = Integrating by parts. say. First. rather than eI(u)). we consider some simple examples.204 CHAPTER VII. One would expect the behaviour in 2) to be important for the quantitative performance of the Lundberg inequality (3.bx} dx oo exp low(x) bx dx 70 Ju r oo = b J exp low (x) .bu}. we show how to rewrite the explicit solution for ti(u) in Corollary 1.6).bx} dx = 1 + J0 dodx(x) exp {.3.(x) dx.(iw(x) .2. u .1 + b f e. RESERVEDEPENDENT PREMIUMS 4.4 Consider again the exponential case B(x) = eax as in Corollary 1. the logaritmic form of (3.bx} dx . 5. J0 ^oo g(x ) dx f AX) lexp IOW (X ) bx + b u 1 exp low(x) .8. it is formally needed only for Theorem 3.(3/p(x).
Similarly. BE * 0. and (3.BE. In particular. It is well known that (see Theorem XI.10) where AE = e log 000 e. Be = e log U000 e. ry(x /b I u o e f0 °° e  e.f y(x)dxd y If 7(x) is increasing . and (3. applying the inequality 7(x + u) > 7(x) yields immediately the conclusion of Theorem 3. 0.I ( u) fool.3. (3. For Theorem 3.9 ) 11000 eI(v)dy f000 e. (3. we get r 00 e.I ( v )dy fo +u) dxdy .9) yields e log .1.fo 7(x) dx /E dy > av 'yo /Edy = E (1 .2(X) = ev2(x) so that 7e(x) = 7(x)/e.fa 7(x+u)dx/Edy o The analogue of (3. 191195) that 1P (U) = fu0 eI(v)dy = eI(u) follo e. > lime log e = 0 and AE * 0. THE LOCAL ADJUSTMENT COEFFICIENT and hence 205 f°° eI(v )dy . 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. u .3.2.10 or Karlin & Taylor [222] pp. oo) with drift µ(x) and variance a2 (x) > 0 at x. (X) = µ(x). in the definition of AE converges to 0.fory(x+u)dxdy ( 3.8) 7(x)dxdy 11 We next give a direct derivations of Theorems 3. (u) = I(u) + AE .e.5 Assume that {Rt} is a diffusion on [0...ev 0 O /E) J0 70 70 Yo This implies lim inf A.5) is infx>o 7(x) > 0 which implies that f °O .fo 7(x)dx/Edy f . IE(u) = I(u)/e. note first that the appropriate slow Markov walk assumption amounts to u. the integral is bounded by 1 eventually and hence lim sup AE < lim sup a log 1 = 0.3 in the particularly simple case of diffusions: Example 3. 70 > 0 such that 7(x) < 7o for y < yo. Choosing yo.2. 3.0.7) follows.1/8 .
10) holds if we redefine AE as AE = flog (j °° efo 7(x)dx/edy _ E/5 I and similarly for B. that the interchange of the slow Markov walk oo is not justified and in fact. _ .5.7o C 15 I I.e. however .4. RESERVEDEPENDENT PREMIUMS The analogue of Example 3. Of course.1 3./3 1 AX dx.. so our approach is to determine standard functions Gl (u). E+o e*O By (3.0) = 0. .Q/p*. . Further. we have 5 > 7o and get lim inf AE > lime log e . .6/p* so that u 1 I (U) = bu . lim sup Af < lim sup c log(1 ..(u) oo.+1 (u) = o( 1). As in Example 3. G. 0. the slow Markov limit a * 0 and the limit u walk approximation deteriorates as x becomes large. .. 0 Now (3. G. I(u ) = G1(u) + . Ignoring 1/5 in the formula there. Then the solution of the Lundberg equation is y* = b . G.6) exactly as in Example 3.5 for risk processes with exponential claims is as follows: Example 3 . this leads to (3.5.0. (u) representing the first few terms in the asymptotic expansion of I(u) as u + oo. I. ) Note that this expression shows up also in the explicit formula for lk(u) in the form given in Example 3.7) follows just as in Example We next investigate what the upper bound / approximation aI (°) looks like in the case p(x) = a + bx (interest) subject to various forms of the tail B(x) of B. . the results are suggestive in their form and much more explicit than anything else in the literature. Thus 7e(x) _7(x)/e and (3.206 CHAPTER VII. the slow Markov walk assumption means 5E = b/c. 7(x) is typically not explicit.5.. > .0/e. Nevertheless .6 Assume that B is exponential with rate S.5) and 7* = 5 . + Gq(u) + o(G9(u))• Gi (u) It should be noted .
y = 2 if B is uniform on (0. c4 = c2b 1/'/(1 .11) with a > 0. THE LOCAL ADJUSTMENT COEFFICIENT Example 3 . and hence (3. B[s] = 1 + s exB(x)dx = 1 +c1SF(a) ('+o(')) (S .1/a).s)C' f "o o as s T S.7 Assume that B(x) . More precisely. e. 1. the typical case is a = 1 which holds . For example.1) leads to (S7T N Ocp a. This covers mixtures or convolutions of exponentials or.. 1) and 17 = k + 1 if B is the convolution of k uniforms on (0.x)n1. Hence (3.C2p* C2 = (3clr( a))11'.3.1/k). 2.1 =$ f cse8 Sn f e"B(x)dx = e8 Jo s eIB ( 1 . say 1 is the upper limit and B(x) . u Example 3 . .1) leads to . x T 1. ry* loge*+ g7loglogp*. 77 = 1 if B is degenerate at 1.clxale5x 207 (3. in the phasetype case .3cse7*I7(77) .Y . .4) or gamma distributions.ry*°p*. It follows from (3.11) that b[s] * co as s f S and hence 7* T S as p* + oo. phasetype distributions (Example 1.cs(1 . Here B[s] is defined for all s and B[s] . I(u) Pt.8). fu I(u) Su .:. if the phase generator is irreducible ( Proposition VIII.g.c2 Su a dx ) Su a<1 Su .y/s)dy sn 1 1 f ' evy'7ldy = cse8r(T7) as s T oc. u(logu + r7loglogu).8 Assume next that B has bounded support. more generally.12) with y > 1. (3.c3 logu a= 1 J 0 a + bx 1/ ( c4ul 1/° a > 1 where c3 = c2 /b.
e7o ( u)(ul+u r. assume that B(x) CO x2/2c7. RESERVEDEPENDENT PREMIUMS Example 3 .r„(Ti).11) and (3. 1 = E. 3b Proof of Theorem 3. .Ote7o( u)(u.r^. of U1 + v .14) for the m .4) is the formula h logEues ( Rhu) .log p*.4) of the local adjustment coefficient is not the only possible one: whereas the motivation for (3. 7 * .f.u . x f oo .3 (B[s] .c8 log .1 Cgs o"O 0 esxex2/2c7 dx = cgsec782/2 f . ec78)2/2c7 dx C7 .ru(TI)) . The assumption implies that ru(t) .(T1)) > Ee7o(u)(ul+vr»(Ti)). (3. h 10. This leads to an alternative local adjustment coefficient 7o(u) defined as solution of 1 = Ee''o(u)(vi+u .g. (3.10 Assume that p(x) is a nondecreasing function of x.12).sp(u).(t))dt.208 CHAPTER VII.9 As a case intermediate between (3. g.B[7o (u)] . (b) 'y(x) <'Yo(x)• Proof That 7(x) is nondecreasing follows easily by inspection of (3. h].u is a nondecreasing function of u.Ul up to the first claim (here ru (•) denotes the solution of i = p (r) starting from ru(0) = u). of the increment in a small time interval [0.f. 1 0 3e. I (u) c8u log u 0 where c8 = 2/c7. Hence for u<V.4).css 2%rc7eC782/2.2 We first remark that the definition (3. one could also have considered the increment ru (T1) .13) We get b[s] .15) Proposition 3.1) .•. (3. By convexity of the m . Then: (a) y(x) and 7o(x) are also nondecreasing functions of x. this is only possible if 7o(v) 2 7o(u)• ..
(n+l) (u) 1 . note that the assumption implies that ru(t) .4) considered as function of 7 is convex and 0 for y = 0. Separating after whether ruin occurs at the first claim or not.1) . Hence „/. this is only possible if yo(u) > 7(u). Then (u) < efo Yo(x)dx.17) from which the theorem follows by letting n + oo. Hence 1 = EeYo(u)(U1+uru(T1)) < E. fa 7o(y)dy < u7o(u) < xyo (u) for x > u.10(b): Theorem 3.11 Assume that p(x) is a nondecreasing function of x. The case n = 0 is clear since here To = 0 so that ik(°)(u) = 0.7o (u)p(u)• Since (3.16) Proof Define 411(n)(u) = P('r(u) < on) as the ruin probability after at most n claims (on = TI + • • • + Tn).(n+1) (u) efo Yo(x)dxI^"Q exyo( I u u)Fu(dx )+ J . the case of 7 then follows immediately by Proposition 3.Fu(u ) + J ^(n)(u .u > tp(u).es'Yo(u)Fu(dx)} o0 e fo yo( x)dx j.2 in terms of 7o.u[70(u)] fo eyo(x)dx .e70(u)(U1P(u)T1) 209 0 + 7o(u)p(u)' 0 <_ 00['Yo( u)] .ru(T1) < x). (3. we obtain „I. it is easily seen that fu x7o(y)dy < xyo (u).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. Also. u We prove Theorem 3. We shall show by induction that (' Y'(n) (u) < e fo 'Yo(x)dx (3.17) shown for n and let Fu(x) = P(U1 + u .3. THE LOCAL ADJUSTMENT COEFFICIENT For (b). Assume (3.
10(a) for some of the inequalities. x + x/n] by two classical risk processes with a constant p and appeal to the classical results (3.n) > k =1 II v ^k n. (u) < I(u). 3.3/e and U.3). 0. in accordance with the notation i/iE (u).210 CHAPTER VII. (3. W O . Lemma 3.n <Z auk}l.2).n = ku. Also.n.x/n.E ( u/n) ^•e.n) pn niE (u /n) n n_1 n.n u k}1.n AX).3 The idea of the proof is to bound { R( f) } above and below in a small interval [x .3). RESERVEDEPENDENT PREMIUMS where the last identity immediately follows from (3. For these reasons. P k. {RtE)} (starting from u = un. (u). Further.12 lim sup4^o f log O. C*e where the first equality follows by an easy scaling argument and the approximation by (3..2..n so that n. for either of Theorems 3. (un2. we have chosen to work with y(u) as the fundamental local adjustment coefficient.E (u/ n) Y'E (un .10(b ) that the bound provided by Theorem 3.15). Proof For ruin to occur.. the value of {R(E)} at the time of downcrossing is < unl.E (u) denote the ruin probability for the classical model with 0 replaced by . yo(u) appears more difficult to evaluate than y(u).11 is sharper than the one given by Theorem 3. and here it is easily seen that yo(u) ..n.nbe C*.n. by €U=. However.n = sup p(x).I. the probability that ruin occurs in the CramerLundberg model with p* = pn. we used also Proposition 3.: y(u). pk n = uk_l. Let Ck.n. 3c Proof of Theorem 3. The probability of this is at least n n. y* evaluated for p* = Pk.. in . given downcrossing occurs. resp. Y*u /E. To this end. 0 It follows from Proposition 3.n inf n uk1.E (u/n).11 be reasonably tight something like the slow Markov walk conditions in Theorem 3.. 0.3 is required.n) must first downcross unl. ryk. op*.n (starting from u/n) without that 2u/n is upcrossed before ruin. and. let Op*.2. define uk.e (u) = v'.E (u/n) Now as e .
in obvious notation one has tC (x) = y(x)/e.F (2u/n).18) (ii) the family of claim overshoot distributions is stochastically dominated by V. It follows that n log V'C (u) k =1 log Ypk. y > 0 it holds that F(U>x +yIU>x) B(x + y) < F (V > y). we need the following condition: Condition 3. (u) CIO < Letting n 4 oo and using a Riemann sum approximation completes the proof. Indeed .E (u/n) OP +^p•. since ry' is an increasing function of p'. 40 Combining with the upper bound of Lemma 3.n <X<Uk.n + 0(1). so that Theorem 3.nu /fn( 1 Ck  e. 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. THE LOCAL ADJUSTMENT COEFFICIENT particular.7k..nk=1 limsupelogv).12 completes the proof. also ryk. uk_1.13 There exists a r.a( u)z. /' (u/n) 'T nk. *p•.2 gives 7PE (u) < eIi"i/f = lim inf Clog 0E (u) > I (u).... (3. v. 211 Clearly. for all x .n = sup ?'(x). B(x) (3.n. i. k=1 k=1 n u _ nE7 k. 3 now follows easily in case (a). 0 with n and u fixed.n cE (2u/n) Ck ne7k.E (u/n) Op•. .e.3. ne7k. 11 Theorem 3.E (urn) < \ *I. In case (b).i.19) .n .! (u/n) n n m 7k.nu /En) o(1)).nu/en(1 + where o(1) refers to the limit e .log Ck.
u /en 0(i) _n so that E2 < e2ryl nu/En0(1). (u/n ..E (0) cf. v ) = v .QEU 1 .EV) = El + E2. P (T(E) (u.2y 1 ' . T() (u. infx>2u /n P(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.V) = e71 nu/Eno(l) (using (3.n V.^(E) (u. (u/n .1 n.R<) (u v).eV) • P (T(E) (u.of:>2 in n(x). (3.EV) = e. V < u/En] + P(V > u/En) (u/En .5) and the standard formula for b(0). v ) = inf { t > 0 : R(c ) < v R) = u } . u/n) < oo] E [OE (u/n . (R. u/n) < oo) . For E2. N with EN < 1 = infx>2u/nA(x) = 0(1).18) for the last equality).. u/n) < oo) EV).E(E) (u. we first note that the number of downcrossings of 2u/n starting from RoE) = 2u/n is bounded by a geometric r. ) (u u /n)) .v.E (u/n .212 CHAPTER VII. Write EO.( . . T(E) (u. u/n)) I T(E) (u.EV) = EiI 1 . u/n) < oo] .nu /EnE [e71. The probability of ruin in between two downcrossings is bounded by Epp . RESERVEDEPENDENT PREMIUMS To complete the proof. Then Y'E (u) ^(E) (u. Ei + E2 < e71.n < ery1.nu/En0(1) . u/n)) . let v < u and define T(E) (u.^'' = E [ .. u /n) < oo] l = = < E [OE (u/n . Then the standard Lundberg inequality yields El < E?..E (2u/n .
Comparing these references with the present work shows that in the slow Markov walk setup.13.7(x)) (3.20) (with ic(x.r.J r(Rs)p(R. Djehiche [122] gives an approximation for tp(u. one can in fact arrive at the optimal path by showing that the approximation for 0(u. the approximation (3. . Whereas the result of [122] is given in terms of an action integral which does not look very explicit. 0 ) (= p(x) .)ds + Y(R2.7) then comes out (at least heuristically) by analytical manipulations with the action integral.7) for ruin probabilities in the presence of an upper barrier b appears in Cottrell et al. the results are from Asmussen & Nielsen [39]. the risk process itself is close to the solution of the differential equation r(x) _ r (x.T) = P „(info<t <T Rt < 0) via related large deviations techniques.3EU) (3.4) and the prime meaning differentiation w.21) to pass from u to 0. where the key mathematical tool is the deep WentzellFreidlin theory of slow Markov walks (see e .u/n) < oo) CI  > u n n ryi n' i=1 Another Riemann sum approximation completes the proof. 0 and b T 00 are interchangeable in the setting of [89].) = exp . whereas the most probable path leading to ruin is the solution of r(x) _ k (x.)Ui } . they also discuss simulation based upon 'local exponential change of measure' for which the likelihood ratio is ( /'t /'t Ns Lt = exp S .3. u/n) < oo { 40 )I U nryl n+liminfelogP (T(')(u.1.=1 J An approximation similar to (3. it might be possible to show that the limits e . (u) 40 213 lim inf e log(Ei +E2) + logP (r(`) (u.J y(Rs)dR. 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. [89]. Typically.g. T) is maximized over T by taking T as the time for (3.t. s). THE LOCAL ADJUSTMENT COEFFICIENT Hence lim inf e log Ali.21) (the initial condition is r(0) = u in both cases). l o JJJ o . Bucklew [81]). Similarly. s) as in (3. u Notes and references With the exception of Theorem 3.
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.g. however.. e. .214 CHAPTER VII. RESERVEDEPENDENT PREMIUMS the simplest being to require b[s] to be defined for all s > 0 (thus excluding . For different types of applications of large deviations to ruin probabilities . see XI. the exponential distribution ). We should like.3.
we write Pv for the case where Jo has distribution v so that Pv = KER viPi• 215 . P. then the problem may admit an algorithmic solution involving a reasonable degree of computational effort if one allows for the more general assumption of phasetype structure. F (Jt = A eventually) = 1 for all i E E 1 and where all states i E E are transient. This implies in particular that the intensity matrix for { it } can be written in blockpartitioned form as T 0 0 . and not in other cases. A proper knowledge of phasetype distributions seems therefore a must for anyone working in an applied probability area like risk theory. 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. Note that since (1. if a problem can be solved explicitly when the relevant distributions are exponentials. 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 often write p for the number of elements of E.1) is 'Here as usual . refers to the case Jo = i. if v = (vi)iEE is a probability distribution.Chapter VIII Matrixanalytic methods 1 Definition and basic properties of phasetype distributions Phasetype distributions are the computational vehicle of much of modern applied probability. A distribution B on (0. on Eo = E U {A} where A is some extra state which is absorbing. Typically. that is.
MATRIXANALYTIC METHODS the intensity matrix of a nonterminating Markov process. an exponential distribution with rate parameter . i. and the phasetype distribution is the lifetime of a particle with constant failure rate /3. 0 2this means that tii < 0. and we have t = Te. Thus the phasetype distributions with p = 1 is exactly the class of exponential distributions. the exit rates ti and the transition rates (intensities) tij: tj 3 aj ai i ti tk tjk FkJ ak Figure 1.3. t1 = /3. a.0 = t11. The initial vector a is written as a row vector. 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.216 CHAPTER VIII. i. j. B(t) = Fa(^ < t ). C is the lifetime sup It > 0 : Jt E E} of {Jt}. E = {i. Then a = a1 = 1. k}. (1. that is. A convenient graphical representation is the phase diagram in terms of the entrance probabilities ai. Here are some important special cases: Example 1 . We now say that B is of phasetype with representation (E. T is a subintensity matrix2.1 Suppose that p = 1 and write .e.1 The phase diagram of a phasetype distribution with 3 phases. In particular. Equivalently. tij > 0 for i 54 j and EjEE tij < 0 .e.2) The interpretation of the column vector t is as the exit rate vector. the ith component ti gives the intensity in state i for leaving E and going to the absorbing state A. T) (or sometimes just (a.T)) if B is the Padistribution of the absorption time C = inf{t > 0 : it = A}.
p}. 0 0 0 T= t= 0 ••• S S 0 0 0 0 0 0 .. the EP distribution may be represented by the phase diagram (p = 3) Figure 1.2 The Erlang distribution EP with p phases is defined Gamma distribution with integer parameter p and density bp XP1 6x (p.. .. .. so that the density is P E ai6ie6.2 corresponding to E = {1.. 0 •. 0 S 6 Example 1. 0 ••• 0 0 Sp1 0 0 t= 0 0 00 •.x i=1 Thus E _ Si 0 T 0 S2 0 0 .. 00)) S s o . 0 SP 0 and the phase diagram is (p = 2) ... 0 0 0 0 S 6 .3 The hyperexponential distribution HP with p parallel channels is defined as a mixture of p exponential distributions with rates 51. .. PHASETYPE DISTRIBUTIONS 217 Example 1... 6.. a = (1 0 0 .1)!e Since this corresponds to a convolution of p exponential densities with the same rate S. ... . .1.
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 0 Example 1 .1 tP1 1 Figure 1. the Erlang distribution is a special case of a Coxian distribution. The basic analytical properties of phase type distributions are given by the following result .e. j E E. (c) the m. the backwards equation for {Jt} (e. E t ikp kj = kEE kEE 3For a number of additional important properties of matrixexponentials and discussion of computational aspects . and is defined as the class of phasetype distributions with a phase diagram of the following form: 1 617 ti t2 2 b2. T). . 5 Let B be phasetype with representation (E.t2 yt bP.218 CHAPTER VIII.f is B (x) = 1 . [APQ ] p. i. Recall that the matrixexponential eK is defined by the standard series expansion Eo K"/n! 3.g. (b) the density is b(x ) = B'(x) = aeTxt.4 For example. dp. 36) yields s d.g.d. a.1)"n! aT"e.aeTxe. MATRIXANALYTIC METHODS Figure 1. Then for i . Theorem 1 . Then: (a) the c.4 (COXIAN DISTRIBUTIONS) This class of distributions is popular in much of the applied literature.f.3 . p:. ds^ = ds' = ttlaj + tikpkj. B[s] = f0°O esxB (dx) is a(sI T)lt (d) the nth moment f0°O xnB(dx) is (. the restriction of P8 to E. see A.
B(x) = 1'a (( > x) = P.5) as hi(tii + s) = ti  t ij hj.f.g.g. this means in vector notation that (T + sI)h = t.tii tii . For (c).p.1. and since b[s] = ah. d" dsn a (. the rule (A.n lt . 1. for n = 1 we may put ki = Ei( and get as in (1. h = (T + sI)1t.s) is the m . Alternatively.jEE B'(x) _ cx Pxe = aeTxTe = aeTxt (since T and eTx commute). PHASETYPE DISTRIBUTIONS 219 That is. After that. and since obviously P° = I.T) 't = (.6) . the solution is P8 = eT8.5) Indeed .tii is the rate of the exponential holding time of state i and hence (tii)/(tii . = aPxe. j#i jEE tijhj + his = ti.T) 1t. Since 1 .f. d8 P8 = TP8. i.g.tii we go to A. ti/ .1 ) n +l n ! a (s I + T ) .n1t = (1)nn!aTn1Te (1)nn! aTne.. in which case the time to absorption is 0 with m .f. Then h tit ti + ti3 h j . define hi = Eie8S. i. hj . and (b) then follows from 1: aipF.g.5) ki = 1 + tii L jj:Ai tii (1. we i w. . Rewriting ( 1. tij / .e.p. (1) n+1n!aT . or w.12) for integrating matrixexponentials yields B[s] = J esxaeTxt dx = a ( f°°e(81+T)dx ) t a(sI ..s j# tii i (1. we arrive once more at the stated expression for B[s]. Part (d) follows by differentiating the m. of the initial sojourn in state i.s I . (Jx E E) = this proves (a).tii and have an additional time to absorption either go to state j which has m .f. B(n)[0] = _ Alternatively.
are idempotent. we get the density as 9 9 6 (1 1) 10 7 1 0 10 2 aeTyt = e x . 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. Example A3. there are some examples where it is appealing to write T on diagonal form. making the problem trivial. 0 Example 1.h.7) the diagonal form of T is 9 9 1 9 T 10 7 10 70 1 10 6 10 7 0 70 9 1 10 where the two matrices on the r. another the case p = 2 where explicit diagonalization formulas are always available. Consider for example 3 9 a= (2 2).6 Though typically the evaluation of matrixexponentials is most conveniently carried out on a computer. MATRIXANALYTIC METHODS which is solved as above to get k = aTle."n! ( ( l 2 2 ) 17 9 0 \ 1 / 10 10 32 n! 35 6" +n!353 Similarly.220 CHAPTER VIII. see the Appendix.s. T= 2 111 so that 2 2 Then (cf. One obvious instance is the hyperexponential distribution.
then the matrix m. 0 Sometimes it is relevant also to consider phasetype distributions. . 5 and serves at this stage to introduce Kronecker notation and calculus (see A. There are two ways to interpret this: • The phasetype distribution B is defective.T). or one just lets U be undefined on this additional set. • The phasetype distribution B is zeromodified.f.4.hall. i.4b for definitions and basic rules): Proposition 1. where the initial vector a is substochastic.e a mixture of a phasetype distribution with representation (a/llall. T) is then defined to be oo on a set of probability 1.e 11BIJ = 1laDD < 1. and in fact one also most often there allows a to have a component ao at A. < 1.T) with weight hall and an atom at zero with weight 1 .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). i.29) and Proposition A4. This is the traditional choice in the literature. PHASETYPE DISTRIBUTIONS 1 10 7 10 221 9 6 70 7 9 10 2 +e 6x (1 11 2 2 35ex + 18e6x 35 The following result becomes basic in Sections 4. B[Q] of B is f3[Q] = J e'1zB(dx) _ (v (9 I)(T ® Q)1(t ® I).7) Proof According to (A.g. (1.7 If B is phasetype with representation (v. hail = E=EE a.11aDD. a random variable U having a defective phasetype distribution with representation (a.
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. O'Cinneide [276] gave a necessary and sufficient for a distribution B with a rational m. Using B(x) = aeTxe . No satisfying . 2.f.. the Erlang case). Here is a sufficient condition: Proposition 1. v. the result follows (with C = (ah)(ve)). See in particular the notes to Section 6. The Erlang distribution gives an example where k > 0 (in fact.8) Proof By PerronFrobenius theory (A..222 CHAPTER VIII. Other expositions of the basic theory of phasetype distributions can be found in [APQ]. x * oo.1 of the Appendix.f. Schmidli. the text is essentially identical to Section 2 of Asmussen [26]. cf. but todays interest in the topic was largely initiated by M.hve7x.4c). MATRIXANALYTIC METHODS la Asymptotic exponentiality Writing T on the Jordan canonical form.q be the eigenvalue of largest real part of T. and we have eTx . Neuts. but the relevant T is not irreducible. Rolski. cf. it is easily seen that the asymptotic form of the tail of a general phasetype distribution has the form B(x) _ Cxkenx.g. (1. 1. 0 Of course. assume that T is irreducible . B(x) . one has k = 0. (or Laplace transform) are often used where one would now work instead with phasetype distributions.Ce7'.8 are far from necessary ( a mixture of phasetype distributions with the respective T(') irreducible has obviously an asymptotically exponential tail. In older literature. let .g. Schmidt & Teugels [307] and Wolff [384]. Example A5. here k = p1). see his book [269] (a historical important intermediate step is Jensen [214]). h can be chosen with strictly positive component.8). we give a criterion for asymptotical exponentiality of a phasetype distribution B. In Proposition A5. h be the corresponding left and right eigenvectors normalized by vh = 1 and define C = ah • ve . the conditions of Proposition 1. i is real and positive. Notes and references The idea behind using phasetype distributions goes back to Erlang. where C. Then the tail B(x) is asymptotically exponential. distributions with a rational m.8 Let B be phasetype with representation (a. .F. Lipsky [247]. T). All material of the present section is standard. let v. 77 > 0 and k = 0. but in many practical cases. not only in the tail but in the whole distribution.
1) Proof Let {Jtk)} be the governing phase process for Uk and define {Jt} by piecing the { J(k) } together. n=O We may think of the U. +UnEA} 00 = EEI(U1 +... is Markov and has two types of jumps ..d. U1<t < U1+U2. as the lifetimes of items (say electrical bulbs) which are replaced upon failure.: U1 + . U2.+UnEA). or the density is available ) is.t. JtJt1) Then { 0<t<U1 . what is the smallest possible dimension of the phase space E? 2 Renewal theory A summary of the renewal theory in general is given in A. and U(A) is then the expected number of replacements (renewals) in A.r.T). Lebesgue measure. we refer to U as the renewal measure.1 Consider a renewal process with interarrivals which are phasetype with representation (cr. Jt={Jt?ul}. . we denote the density by u(x) and refer to u as the renewal density. but is in part repeated below.1. Then the renewal density exists and is given by u(x) = ae(T+ta)xt. The explicit calculation of the renewal density (or the renewal measure) is often thought of as infeasible for other distributions. if U is absolutely continuous on (0... 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 +.. (2. Let U1.. oo) w.f. be i. For this reason..g.i.. the problem has an algorithmically tractable solution if B is phasetype: Theorem 2. RENEWAL THEORY 223 algorithm for finding a phase representation of a distribution B (which is known to be phasetype and for which the m. but nevertheless. A related important unsolved problem deals with minimal representations: given a phasetype distribution .. with common distribution B and define4 U(A) = E# {n = 0..1 of the Appendix... If B is exponential with rate 0. known. the renewals form a Poisson process and we have u(x) = 0.2. however. . + U0 is 0 . .
and hence ( 2. the phasetype assumptions also yield the distribution of a further quantity of fundamental importance in later parts of this chapter . The renewal density at x is now just the rate of jumps of the second type.T) where v = aT1 /µB. Then: (a) the excess life t(t) at time t is phasetype with representation ( vt. IIafl < 1. However. the lifetime of the renewal process.1 Corollary 2. that is. + Uit_1 where s.IIBII which is > 0 in the defective case. This is defined as U1 + . define the excess life e(t) at time t as the time until the next renewal following t. since Uk = oo with probability 1 . and the distribution of Jx is ae ( T+t«)x. .2 Consider a terminating renewal process with interarrivals which are defective phasetype with representation (a. Corollary 2. i.1. u The argument goes through without change if the renewal process is terminating. Hence ( 2. Then the lifetime is zeromodified phase type with representation (a. this is welldefined..1) follows by the law of total probability. is the first k with Uk = 00. 2. MATRIXANALYTIC METHODS of the last type from i to j occurs at rate tiaj .T). and the jumps of the first type are governed by T. as the time of the last renewal. Proof Just note that { it } is a governing phase process for the lifetime. see Fig.3 Consider a renewal process with interarrivals which are phasetype with representation (a. i. u Returning to nonterminating renewal processes . Hence the intensity matrix is T + ta. B is defective . which is phase type with representation (v.224 CHAPTER VIII. Equivalently. T).. and let µB = aTle be the mean of B. which is ti in state i.T) where vt = ae (T+ta)t .1) remains valid for that case. the density is veTxt = B(x)/µB. . fi(t) U2 U1 . (b) £(t) has a limiting distribution as t * oo.T + ta).e.e.U1 U3 U2 U3 U4 Figure 2.
RENEWAL THEORY 225 Proof Consider again the process { Jt } in the proof of Theorem 2. Al. Hence in (b) it is immediate that v exists and is the stationary limiting distribution of it. T1 and eTx commute.q2. 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). we get B(x) aeTxe aT1eTxTe µB µB PB = veTxt. The time of the next renewal after t is the time of the next jump of the second type. Here are two different arguments that this yields the asserted expression: (i) Just check that aT1/µB satisfies (2.6. we first compute the stationary distribution of Q.e. The renewal density is then aeQtt = (al a2) ( 7i 7"2. (ii) First check the asserted identity for the density: since T. the unique positive solution of ve = 1.2.2) v(T + ta) = 0. hence e(t) is phasetype with representation (vt.) ( t2 ) .1. u Example 2 .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. According to Example A3.4 Consider a nonterminating renewal process with two phases. Next appeal to the standard fact from renewal theory that the limiting distribution of e(x) has density B(x)/µB. cf.e. i. = qz ql (x1 xz) = ql + qz ql + q ' and the nonzero eigenvalue A = ql . (2.
226 CHAPTER VIII. )t (51 . MATRIXANALYTIC METHODS e. Hence 7r = (1/2 1/2).e2bt) 13 Example 2 .t2) .(biaz + aza.4 yields the renewal density as u(t) = 5152 e.6 Let B be hyperexponential. The present treatment is somewhat more probabilistic. A = 25. and Example 2.a27rl) (tl .5 Let B be Erlang(2).4 yields the renewal density as u(t) = 2 (1 .tl) 7r2t2 + eat (a17r2 .`t (al a2) + C 11 172 ir12 / \ t 2 ) r1 (7r1 7r2) ( t2 7rltl + J + eAt (al a2) ( 71(t2 . and Example 2. Then Q= 0 55 )+(1o)=( j ad ). Then _ Q Hence 51 0 0 52 + 51 52 _ 5152 51a2 ) (al a2) 52a1 62a1 Slat + 52a1 51a2 51a2+52a1 A = 51a2 .t2) 1 + eat (a17r2 .a27r1) (t1 . t1B 0 Example 2 .52) 25152 51x2+5251 51a2+5251 Notes and references Renewal theory for phasetype distributions is treated in Neuts [268] and Kao [221].52a1. .
use the phasetype representation of Bo. Corollary 2. G+(. T). and if there is a subsequent ladder step starting in j whic occurs w. 3. The essence is contained in Fig. we see that the ladder height Sr+ is just the residual lifetime of the Markov process corresponding to the claim causing upcrossing of level 0.2. i. T). We asssume that B is phasetype with representation (a. a+j. marked by thin and thick lines on the figure.3. For (b). (b) V.(u) = a+e(T+tQ+)u Note in particular that p = IIG+II = a+e.i. itself phasetype with the same phase generator T and the initial vector a+ being the distribution of the upcrossing Markov process at time ST+_. which occurs at rate ti. Corollary 3. however. the Markov processes representing ladder steps can be pieced together to one {my}.1 Assume that the claim size distribution B is phasetype with representation (a. {St} the claim surplus process. Thus the total rate is tip + tia+. and M is zeromodified phasetype with representation (a+. Then each claim (jump) corresponds to one (finite) sample path of the Markov process. add a more selfcontained explanation of why of the phasetype structure is preserved.f3aT1. r(u) the time of ruin with initial reserve u. . The stars represent the ladder points ST+(k). Now just observe that the initial vector of {mx} is a+ and that the lifelength is M. Within ladder steps. Then: (a) G+ is defective phasetype with representation (a+. 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. Next.p. T) where a+ is given by a+ = . B the claim size distribution.1 on the next page. 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. we shall. cf.) = F(ST(o) E •. T(0) < oo) the ladder height distribution and M = supt>o St. Considering the first.e. represent the maximum M as the lifetime of a terminating renewal process and use Corollary 2. Since the results is so basic. and rewriting in matrix form yields the phase generator of {my} as T + ta+. Here we have taken the terminating Markov process underlying B with two states. T + to+). Proof The result follows immediately by combining the PollaczeckKhinchine formula by general results on phasetype distributions: for (a). with 0 denoting the Poisson intensity.3.
3. Figure 3. 0 Example 3.M {mx} ST+(2)  S ..1 .. MATRIXANALYTIC METHODS t .228 CHAPTER VIII..7)diag so that a+ = QaT 1 = 3 ( 3 2 2) 0 3 9 2 14 7 2 11 2 T+ta+ = 3 0 07/+( 7I \ 2 14 .2 Assume that .1 This derivation is a complete proof except for the identification of a+ with .Q = 3 and b(x) = . This is in fact a simple consequence of the form of the excess distribution B0.QaT1. T = (3 .t t d kkt S. 3e3x + . 7e7x 2 2 Thus b is hyperexponential (a mixture of exponential distributions) with a (2 2 ). see Corollary 2.
the discussion around Fig. 0(8) (u) (recall that z/i(u) refers to the zerodelayed case and iY(8) (u) to the stationary case). It is notable that the phasetype assumption does not seem to simplify the computation of finite horizon ruin probabilities substantially. For further more or less explicit computations of ruin probabilities. see Shin [340]. 3. We assume p = PB/µA < 1 and that B is phasetype with representation (a.j). We shall derive phasetype representations of the ruin probabilities V) (u). The parameters of Example 3. For the compound Poisson model. Fig. 3. (a) G+ is of phasetype with representation (a+. 4 The renewal model We consider the renewal model in the notation of Chapter V. T) for some vector a+ = (a+. we encounter similar expressions for the ruin probabilities in the renewal.6). The result carries over to B being matrixexponential.1 which does not use that A is exponential) by noting that the distribution G+ of the ascending ladder height ST+ is necessarily (defective) phasetype with representation (a+.1 can be found in Neuts [269] (in the setting of M/G/1 queues. cf.T). For an attempt.4.1): Proposition 4.1 In the zerodelayed case. the duality result given in Corollary 11. THE RENEWAL MODEL This is the same matrix as is Example 1. see Section 6. but that such a simple and general solution exists does not appear to have been well known to the risk theoretic community. if we define {mz} just as for the Poisson case (cf. T). his derivation of +'(u) is different.2 are taken from Gerber [157]. That is. this was obtained in Section 3. but there the vector a+ is not explicit but needs to be calculated (typically by an iteration).and Markovmodulated models. and the argument for the renewal case starts in just the same way (cf.6. see Stanford & Stroinski [351] . In the next sections.^(u) = a+e( T+ta+)ue = 24eu + 1 e6u 35 35 0 Notes and references Corollary 3. where a+ is the (defective) . with A denoting the interarrival distribution and B the service time distribution. 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 .4.
230 distribution of mo. where B0 is the stationary excess life distribution corresponding to B.*'} is Markov with the same transition intensities as {mx}.3 a+ satisfies a+ = V(a+).3. We have now almost collected all pieces of the main result of this section: Theorem 4 . but with initial distribution a rather than a+. B0 is phasetype with representation (aT1/µa. where u w(a +) = aA[T + to+) = a J0 e(T+t+)1A(dy).1). 4. (4. In fact. it follows by integrating y out that the distribution a+ u of mo is given by the final expression in (4. Hence by Theorem 11. MATRIXANALYTIC METHODS (b) The maximum claim surplus M is the lifetime of {mx}.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). Since the conditional distribution of my given T1 = y is ae4y. CHAPTER VIII. (c) {mx } is a (terminating) Markov process on E. Proof Obviously.*} from {St+y .T). Fig.1) Proof We condition upon T1 = y and define {m. the form in which we derive a+ for the renewal model is as the unique solution of a fixpoint problem a+ = cp(a+). obviously mo = m.4 Consider the renewal model with interarrival distribution A and the claim size distribution B being of phasetype with representation (a. Also.T)• Proposition 4. Then .5.1.T).Sy} in the same way as {mx} is defined from {St}. Then {m. which for numerical purposes can be solved by iteration. where a(8) = aT1/PA.6. with intensity matrix Q given by Q = T + to+. cf. the calculation of the first ladder height is simple in the stationary case: Proposition 4. the Palm distribution of the claim size is just B. Nevertheless. The key difference from the Poisson case is that it is more difficult to evaluate a+. G(') = pBo. But by Corollary 2.
The second follows in a similar way by noting that only the first ladder step has a different distribution in the stationary case..2) where a+ satisfies (4.1 .0.2.. . THE RENEWAL MODEL 231 . by a+ = lim a +n) where a+°) .1(b)..1).1) and a(8) _ aT. In particular .0) is an increasing function of /3. It remains to prove convergence of the iteration scheme (4. a+ can be computed by iteration of (4. . the maximum claim surplus for the stationary case has a similar representation as in Proposition 4.. only with initial distribution a(*) for mo. a+l ) = cp (a+°)) .3) (defined on the domain of subprobability vectors .^(u) = a+e ( T+ta+)xe.4. I {mx} . (4. Hence ^p(. The term tf3 in cp(i3) represents feedback with rate vector t and feedback probability vector (3.1/pA.•.e.M. thus . (4. Furthermore .3).1 by noting that the distribution of mo is a+.2 ) follows from Proposition 4.3) Proof The first expression in (4. i y ^ T1= y `•r Figure 4. a+) > 0 = a+o) implies a+) _ (a+) > W (a+)) = a+) . a+2) = ^p (a+l)) . and that this is given by Proposition 4. i.^(8)(u) = a ( 8)e(T+ta +) xe.
7+ ]. MATRIXANALYTIC METHODS and (by induction ) that { a+ n) } is an increasing sequence such that limn.4) makes sense and provides an analytic continuation of F[•] as long as s ¢ sp(T). 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). (4. let F be the distribution of U1 . Then each subexcursion of {St+Tl .f. Similarly. this implies that ahA[s] # 0.T)'t • A[s] (4. and let &+". and hence we may assume that h has been normalized such that ahA[s] = 1.. Then (4. For n = 0.5) Since s $ sp(T).T)1t. Then s is an eigenvalue of Q = T + ta+ if and only if 1 =. . Assume the assertion shown for n . Thus . we use an argument similar to the proof of Proposition VI. (4.) = P(mTl = i. Theorem 4.232 CHAPTER VIII.5 Let s be some complex number with k(s) > 0.2.4. F[s] being interpreted in the sense of the analytical continuation of the m. a+ ) exists . Thus by (4. the corresponding right eigenvector may be taken as (sI . In that case.T1. To prove the converse inequality.ST. Then e4'h = e82h and hence sh = Qh = (T + taA[Q])h = Th + A[s]tah. n) &+n) T a+. both quantities are just 0 .1 arrivals (n arrivals are excluded because of the initial arrival at time T1 ).1.g. which links together the phasetype setting and the classical complex plane approach to the renewal model (see further the notes).4) whenever EeR(S)U < oo.5) yields h = (sI .4). the normalization is equivalent to F(s) = 1. 0 = a+) < a+ yields a+) _ (a+0)) (a+) = a+ (n and by induction that a(n) < a+ for all n .} can contain at most n . Then F[s] = a(sI .T)It. with B[s]. 0 0 We next give an alternative algorithm. However. s ¢ sp(T). Let Fn = {T1 + • • • + Tn+1 > r+}be the event that {my} has at most n arrivals in [T1. To this end. limn4oo a ) < a+. so to complete the proof it suffices to show that &+ < a+) for all n. Obviously.P[s] = A[s]B[s]. Proof Suppose first Qh = sh. Fn ).
' that the equation F(s) = 1 has d distinct roots p1. hd. Further. This gives d roots 'y. letting vi be the left eigenvector of Q corresponding to pi and normalised by vihi = 1 . (4. Let d denote the number of phases.T)lt + t = s(sI .. Q has diagonal form d d Q = dpivi®hi = dpihivi. the classical algorithm starts by looking for roots in the complex plane of the equation f3[y]A[ry] = 1.. . ..type with representation (a+. 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 6.5(c) means that a+(sI T)1t = 1. Then G+ is phase.. Corollary 4. 0). in turn.. and hence by the WienerHopf factorization identity (A. yd satisfying R(ryi) > 0.9) we have G+[s] = 1 which according to Theorem 1. t(ry) > 0. W v M(d) in the notation of Chapter V). hd. 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 ). 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. Pd with corresponding eigenvectors hl. Pd in the domain ER(s) > 0 . we have IG_ [s] I < 1 . pdhd.lt we get Qh = (T + to+)h = T(sI .. T) with a+ = a(QT)/at. ..6.6 Suppose u < 0... This immediately implies that Q has the form CD1 and the last assertion on the diagonal form .... As in Corollary 4. Q = CD1 where C is the matrix with columns hl. we get at a(Q . . Notes and references Results like those of the present section have a long history.5. and define hi = (piI . . In older literature .. Hence with h = (sI T).6) i=1 i=1 Proof Appealing to Theorem 4.. D that with columns p1 hl.T)lt = sh..T)It. THE RENEWAL MODEL 233 Suppose next F(s) = 1. Since R(s) > 0 and G _ is concentrated on (oo. .T) = 1 ata+ = a+. the matrix Q in Theorem 2.1 has the d distinct eigenvalues . .p1i .. and the solution is .4. Given T has been computed..
It turns out that subject to the phase. see Neuts [269].. Here phase. the ruin probability can be found in matrixexponential form just as for the renewal model. is phasetype. In queueing theory. The exposition here is based upon [18].g. [119].contained derivation).) d (see. [270] and Latouche & Ramaswami [241]. involving .F. starting around in 1975. The distribution of W comes out from the approach but in a rather complicated form . which contains somewhat stronger results concerning the fixpoint problem and the iteration scheme. For further explicit computations of ruin probabilities in the phasetype renewal case . the background Markov process with p states is {Jt}. The number of elements of El=> is denoted by q. and the distribution of an arrival claim is B. 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 arrival rate in background state i is a. That is . In risk theory. This complex plane approach has been met with substantial criticism for a number of reasons like being lacking probabilistic interpretation and not giving the waiting time distribution / ruin probability itself but only the transform. Neuts and his students. and appears already in some early work by Wallace [377].type assumptions are basic. MATRIXANALYTIC METHODS d F 1 + a J e°" ip(u) du = Ee°w = 11(t. with representation say (a(' ).. E(t)). Asmussen & O'Cinneide [ 41] for a short self. e. The matrix.. the fixpoint problems look like R=Ao+RAI+R2A2+ . Numerical examples appear in Asmussen & Rolski [43].type assumption .4.234 then in transform terms CHAPTER VIII. For surveys . an alternative approach (the matrixgeometric method ) has been developed largely by M. see Dickson & Hipp [118]. the intensity matrix is A and the stationary row vector is ir . a pioneering paper in this direction is Tacklind [373]. We assume that each B. where R is an unknown matrix. whereas the approach was introduced in queueing theory by Smith [350]. The solutions are based upon iterations schemes like in Theorem 4. 5 Markovmodulated input We consider a risk process {St } in a Markovian environment in the notation of Chapter VI.exponential form of the distribution was found by Sengupta [335] and the phasetype form by the author [18]. T(').
Diagonalization Consider a process {(It. We start in Section 5a with an algorithm involving roots in a similar manner as Corollary 4.6. The stationary distribution is obtained by finding the maximum of the Vcomponent of the version of {(It. •. However. the phase space E(°) for B. Section 5b then gives a representation along the lines of Theorem 4.5.1. 5a Calculations via fluid models. The two environmental states are denoted o.1 In Fig. has states o. and the one E(•) for B. The connection between the two models is a fluid representation of the Markovmodulated risk process given in Fig. 5. p = ql = Q2 = 2.Vt)} obtained by time reversing the I component. (a) 0 0 ♦ o ° tl ♦ • 0 0 o } o o (b) 0 } ♦ • 0 o f o Figure 5. This calculation in a special case gives also the ruin probabilities for the Markovmodulated risk process with phasetype claims. The key unknown is the matrix K.4. the analysis involves new features like an equivalence with first passage problems for Markovian fluids and the use of martingales (these ideas also apply to phasetype renewal models though we have not given the details). states .1. Vt)}t>o such that {It} is a Markov process with a finite state space F and {Vt} has piecewiese linear paths. O. for which the relevant fixpoint problem and iteration scheme has already been studied in VI.2. 5. 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. MARKOVMODULATED INPUT 235 some parameters like the ones T or a+ for the renewal model which need to be determined by similar algorithms.
resp. V. If s is such a number. Bi[s] = a(i)(T(i) + sI)it('). a E E(i) } . The intensity matrix for { It} is (taking p = 3 for simplicity) I A .236 CHAPTER VIII. corresponding to the partitioning + Epp).1))diag + sII = 0 (5. 5. Thus F = {o. i E E. F = E U { (i.1(a).92a(2) 0 0 T(2) 0 0 0 f33a(3) 0 0 T(3) with the four blocks denoted by Ei„ i. In the general formulation . '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.1 A complex number s satisfies 'A+ (f3i(Bi[s] . Eli) + Proposition 5. a) : i E E. This implies that in the fluid context. A claim in state i can then be represented by an E()valued Markov process as on Fig. whereas Ee8s' = oo for all t and all s > so where so < oo. r(i) _ 1. of E into components indexed by E.Vt)} is then obtained by changing the vertical jumps to segments with slope 1. Recall that in the phasetype case.1))diag ) a = sa and the eigenvector b = . 4. MATRIXANALYTIC METHODS 4. a) of {It}. Let E denote the matrix . 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. 5.1) if and only if s is an eigenvalue of E. j = 1.1(b) {(It . o. First. in the fluid model Eel'.(Ni)diag r(i. t. 4}. •. < oo for all s. consider the vector a satisfying (A + (13i(Bi[ s] . 4. a) = 1. Second. F is the disjoint union of E and the Eli).A 0 Or 1A/ _ t(i) 0 t(2) 0 0 0 0 0 t(3) 0 T1 0 0 0 . we have more martingales at our disposal. 2. The fluid model on Fig .31a(l) (/3i)diag .
sI.sI)1t)) iag I = 0 which is the same as (5.sI + E12 (sI .sI) (sI . d = (sI E22)1E21a = E ai(sI .T('))1t(i) . with Eii replaced by Eii . Then (up to a constant) c = a.E22)1 E21a.E21a + sd = sd.1). assume that a is chosen as asserted which means (Ell . where c. resp . and let d = (sI . d correspond to the partitioning of b into components Proof Using the wellknown determinant identity Ell E12 E21 E22 E22 I ' I Ell .E12E22 E21 I .E22)1 E21) a = 0. MARKOVMODULATED INPUT 237 indexed by E.sI+ ((3ia(i)(T(i) .32a(2) (/3i)diag . For the assertions on the eigenvectors. Noting that E11c + E12d = se by definition.E22 . 0 .sI ()3i)diag . c = a.A .sI 0 0 0 T(2) . it follows that Ell E12 ( E 21 E22) (d) = s 1 d I . E(1) + + E(P).5.E22)1 E21a E21a .A . iEE (a> of 0* 1 AI. t(1) 0 0 then also 0 t(2) 0 .sI 0 0 0 T(3) .(sI . it follows that if Qla(1) 0 0 .sI 0 0 t(3) 0 0 = 0.Nla(1) 0 0 T 1. Then E21c+E22d = E21a .
v) yields C{V) = e8 .. a )d(a + e8 °vpi (u . . .4 Assume that E has two states and that B1.. < 0 and let b(v) = I d(„)) be the right eigenvector corresponding to s.2 Assume that E = Or 'Al has q = ql + + qp distinct eigenvalues si.. j.. .Q. e89uc(e)) (d(1) .j)c v . s2 are the negative eigenvalues of Al +01 A1 E _ A 2 b1 0 52 A2 +32 0 . v. Proof Writing Or'Alb( v) = svb( v) as (AI .. v = 1.v) = v) I.j. .. a).v) = = p i( u ./' u = e' (esiuc ( 1) . Example 5 . v.a Solving for the pi(u. j) pi( u . v. we first look for the negative eigenvalue s of E = I 0 I which is s = ry with yy = b .( u.5.v}.. Here E has one state only. Letting v ^ oo and using Rsv < 0 yields e8'u = Epi(u. v > 0. the result u follows. . I' i( V P2 (w (u) < oo..v) = j). B2 are both exponential with rates 51 i b2..3 Consider the Poisson model with exponential claims with rate 5. a) and noting that i1 (u) = >I j. pi(u. .v. MATRIXANALYTIC METHODS Theorem 5. c j. Then . a)). a)). it follows by Proposition II.sv)b(v) = 0. w(u)=inf{t >O:Vtu}.238 CHAPTER VIII... a) = Pi (Vw(u. j. Example 5 . a) = (j. define w(u. We can take a = c = 1 and get d = (s + b)16 = 5/(3 = 1/p. Thus 0(u) = esu/d = pe7 ° as u should be. q.a)d^ ). w(u.4 that {e"1b(v) is a martingale ... d("))1 e. j. For u.upi(u. j..v) = (j.v)=inf{t >0:Vtu orVt=. To determine 0 (u). j. sq with $2s.O.v) = Optional stopping at time w (u. Then we get V)i (u) as sum of two exponential terms where the rates s1.pi(u. u) Iw(u. Iw(u.
( 2.3) .y = to B k7 j # k In particular. dx)Bj(y . •) is phasetype with representation (E(i). 0 Theorem 5 . 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. i.b (u) = Pi(M > u) = 9(i)euue. (') a T( However . according to VI. j. 9('). (5.k. j.2. (y. In terms of K.5 G+(i. Proof We must show that G+ (i.6 For i E E.3j eye.s.33(e = 0 a(j))(K ®T ( j))(ej (9 I). oo)) j)ye.x) 00 f ° (') (j) eT (yy)edx .h. is 0 /3 f R(i . 8^')IT(j)) where e 3^') =. U) where t(j) + t(j)O(j j = k uja.xxej • a 00 oo el .2) the l.5. j.Qj eie 0 f e (j) T(') x T(j)y ej a e dx e e 00 00 eKx ® e T(')' dx (ej (& I)e T(')ye eKa®T(')x dx (ej (9 I)eT(') Ye e(i)eT(')ye. MARKOVMODULATED INPUT 239 5b Computations via K Recall the definition of the matrix K from VI. the Pidistribution of M is phasetype with representation (E(1) + + E(P).
and it just remains to check that U has the asserted form. 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.. a) is obviously chosen according to e(`). the current ladder step of type j must terminate. i. Associated with each ladder step is a phase process.5). +bn0i1 0n +a10n1 +. which occurs w. This yields the asserted form of uja. oo) and b* [0] = f °O eBxb(x) dx the Laplace transform. u Notes and references Section 5a is based upon Asmussen [21] and Section 5b upon Asmussen [17]. An alternative characterization is that such a distribution is matrixexponential..f. T.1 Let b(x) be an integrable function on [0. For a transition from (j. with phase space EU> whenever the corresponding arrival occurs in environmental state j (the ladder step is of type j). which occurs at rate t(i). we have sofar concentrated on a claim size distribution B of phasetype. the ratio between two polynomials (for the form of the density. bn1 bn).. +aii10+anI then a matrixexponential representation is given by b(x) = aeTxt where a = (b1 b2 . equivalently.. intensity matrix U. Bk7 . t) is the representation of the matrixexponential distribution/density): Proposition 6. say. i. see Example 1.y) to occur when j # k.. that the density b(x) can be written as aeTxt for some row vector a.. t = (0 0 .e. 0 1)'. we have the additional possibility of a phase change from a to ry within the ladder step.g. and lifelength M. Piecing together these phase processes yields a terminating Markov process with state space EiEE E(').e. Then b*[0] is rational if and only b(x) is matrixexponential.. which occurs at rate t^^7. a) to (k.) which is rational. some square matrix T and some column vector t (the triple (a. Furthermore.240 CHAPTER VIII. if b* [0] = b1 +b20+b302 +.p. a m. and a new ladder step of type k must start in phase y.k y. However. Numerical illustrations are given in Asmussen & Rolski [43]. MATRIXANALYTIC METHODS Proof We decompose M in the familiar way as sum of ladder steps .2.. 6 Matrixexponential distributions When deriving explicit or algorithmically tractable expressions for the ruin probability. Starting from Jo = i. (6. . the initial value of (i.2) . For j = k.
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.s. but as follows from Proposition 6. 0 0 . 0 1 an an1 an _2 . s = c/ 2 . see Asmussen & Bladt [29] (the representation (6. T= 0 0 1 . personal communication). shows that the distribution B with density b(x) = c(1 cos(21r x))ex. Thus.. a2 a1 Proof If b(x) = aeTxt.6..e(tai1)x/2 + e'T) it follows that a matrixexponential representation ()3.1 0 0 )3 = (111). Namely.3) that we can take 0 1 0 0 a= (1 + 47r2 0 0). we can always obtain a real one (a..1.3 A set of necessary and sufficient conditions for a distribution to be phasetype are given in O'Cinneide [276]. . One of his elementary criteria.3) was suggested by Colm O'Cinneide. b(x) > 0 for x > 0. MATRIXEXPONENTIAL DISTRIBUTIONS 241 T = 0 1 0 0 0 .4) 0 0 1 c This representation is complex. Writing b(x) = c(e( 2ni1 ) y/2 . t). matrixexponentiality implies a rational transform. bn of the denominator to be distinct and expand the r. . (6.(6. namely to asssume the roots 6l. (6. .3) 0 0 0 0 0 .2).47r2 3 . S.. then b*[0] = a(0I T)1t which is rational since each element of (01 .47x2 3 1 0 . where c = 1 + 1/47r 2. . (6. u giving b(x) = E 1 ciebiz/bY./(0 + bi)..T)1 is so. u Remark 6. t= 0 .. Example 6 .. since 1 + 4ir2 03 + 302 + (3 + 47x2)0 + 1 + 47r2 it follows by (6. T. 1 .an_3 an _ 4 . . s) is given by 27r i .h. of (6. The converse follows from the last statement of the theorem.1) as E 1 c.. S = f c/2 0 21ri . For a proof. 0 0 0 0 1 0 0 .2). cannot be phasetype..2 A remarkable feature of Proposition 6.1 0 .
recall that t = Te) that if B is phasetype and (a. and present two algorithms for calculating '(u) in that setting. then: Proposition 6. .6) in Section 3 seems to use the probabilistic interpretation of phasetype distribution in an essential way. For the first. Consider the distribution with density = 15 ((2e2x . q are polynomials without common roots. we have represented ti* [0] as ratio between polynomials (note that 0 must necessarily be a root of the numerator and cancels).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)) . Corollary 111. MATRIXANALYTIC METHODS Example 6 . we use a representation (a. (6.5 (6. t) a phasetype representation with a the initial vector.3. 0.1 to get i (u) = f3esus. we shall only consider the compound Poisson model with arrival rate 0 and a matrixexponential claim size distribution B. T. T the phase generator and t = Te. we take as starting point a representation of b* [0] as p( O)/q(9) where p. and can use this to invert by the method of Proposition 6. As for the role of matrixexponential distributions in ruin probability calculations. t) of b(x). 7 + 155ex b(x) Then it is known from O'Cinneide [276] that b is phasetype when 6 > 0. T.6) holds true also in the matrixexponential case.6) The remarkable fact is. that despite that the proof of (6.5) Thus. For the second algorithm.1 shows that a matrixexponential representation can always be u obtained in dimension only 3 independently of J. Then (cf. leading to matrix calculus in high dimensions when b is small. and that the minimal number of phases in a phasetype representation increases to 0o as 5 . (6. then 5(u) = a+e(T+t+)uTle where a+ = /3aT1.4 This example shows why it is sometimes useful to work with matrixexponential distributions instead of phasetype distributions: for dimension reasons . We recall (see Section 3. But since 15(1 +6)02 + 1205 0 + 2255 + 105 b* [9] _ (7 + 155)03 + (1355 + 63)92 + (161 + 3455)9 + 2256 + 105 Proposition 6.242 CHAPTER VIII.1)2 + 6).
1t = f3a (0I T)1T1t .T)1t ( l .1 + 82 (9I . 519) (A + UBV ).1t du = .T)1T 2 = and 1 = AB IT2 + 82T .1BVA1.b* (6. Now.6b* . (6. we get b+ = 0aT1(9I T). Then in Laplace transform formulation .B=land V=a+.T)1 J0 00 b(x) dx = f aT1t.T)1 + (6I . b+ = a+(9I . U =. xb(x) dx = aT2t. we get (91.7) 9( cf.t.T)1 (91.to+)1T .a+(9I .T .T)1t.T)1ta+(OI .1 = ^(T1 + ( 91T)1). b+ = a +(BI .1 + b+ = b++ 1 .T)1 so that b* b** b** a+(9I .6). this can be verified by analytic continuation from the phasetype domain to the matrixexponential domain .T)1 + 1 ib* (91.T)1t)1a +(9I . but we shall give an algebraic proof.T . (6.5 ). MATRIXEXPONENTIAL DISTRIBUTIONS 243 Proof Write b* = a(9I .to+)1 = (BI .T).6. with A = 91T.1t = b* .T . From the general matrix identity ([331] p.A .1UB(B + BVA1UB).to+)1T . the assertion is equivalent to a+(BI . Presumably.1 = A1 .T)1T1t.'t. (91. since (91T)1T . .
8 a(T1 + (01. of (6. some key early references using distributions with a rational transform for applied probability calculations are Tacklind [373] (ruin probabilities) and Smith [350] (queueing theory). 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.1. From this it is straightforward to check that b**/(b+ . MATRIXANALYTIC METHODS . 7a Computing O(u) via differential equations The representation we use is essentially the same as the ones used in Sections 3 and 4.b*). see Asmussen & Bladt [29]. cf. 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. In Corollary VII.1t = /3a (9I .3a (1 0 T 2 + 1 T 102 (9I + 02 1 T)1) t P + 7.8. 0 Notes and references As noted in the references to section 4.s. (for some remarkable explicit formulas due to Paulsen & Gjessing [286].1. The proof of Proposition 6. A key tool is identifying poles and zeroes of transforms via WienerHopf factorization. 7 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate 0. Much of the flavor of this classical approach and many examples are in Cohen [88].1.T)1)t = 8 (1 . /3aT1(0I . See Fig.1) is the same as the r. to piece together the phases at downcrossing times of {Rt} (upcrossing times of {St}) to a Markov process {mx} with state space E. 3. but the argument of [286] does not apply in any reasonable generality).1.h. VII. see the Notes to VII. Lipsky [247] and Asmussen & O'Cinneide [41]. .5 is similar to arguments used in [29] for formulas in renewal theory. T).82b*. a key early paper is Cox [90] (from where the distribution in Example 6.T)1T2t .3 is taken). 7.la. which is selfexplanatory given Fig. a.7).T)1T. For expositions on the general theory of matrixexponential distributions.244 CHAPTER VIII. the ruin probability(u) was found in explicit form for the case of B being exponential.
t + s) . Since v(u) = (vi(u))iEE is the (defective) initial probability vector for {m8}. Given the v(t) have been computed. the A(t) and hence Vi(u) is available by solving differential equations: Proposition 7. though still Markov. Figure 7. we obtain V)(u) = P(m„ E E) = v(u)P(0.1z I.tl < t2 < u.t)). is no longer timehomogeneous.1 The difference from the case p(r) = p is that {m2}. Ai(t) = P(mt = i). In fact. Proof The first statement is clear by definition. t2) = exp where Q(t) = ds [P(t. RESERVEDEPENDENT PREMIUMS 245 Rt l0 u .1 A(0) = v(u) and A'(t) = A(t)(T + tv(u . 0 < t < u.u)e = A(u)e (7.I] I 80 { tq f Q(v) dvl t1 1 . Let P(tl.e. By general results on timeinhomogeneous Markov processes. the definition of {m8} depends on the initial reserve u = Ro. >iEE Vi (U) is the ruin probability for a risk process with initial reserve 0 and premium function p(u + •). P(tl. in contrast to Section 3. Define further vi(u) as the probability that the risk process starting from RD = u downcrosses level u for the first time in phase i. t) is the vector of state probabilities for mt.t2) be the matrix with ijth element P (mt2 =j I mtl = i). i. Note that in general >iEE Vi (U) < 1. O<.7.1) where A(t) = v(u)P(0. Also.
from a computational point of view the remaining problem is to evaluate the v(t). those corresponding to state changes in the underlying phase process and those corresponding to the present jump of {Rt} being terminated at level u . the probability of which is 1 . whereas in the second case the probability is p(u)dt • tjvi(u).t). 0 Thus. jEE . {mx} has jumps of two types. or it stops between level u + p(u)dt and u. Thus.2 For i E E.t)). Hence Q(t) _ T + tv(u . Proposition 7.(tai + vi(u) E vj(u)tjp (u)  Q + vj (u)tjip ( u). 0 < t < u. dt]. the interpretation of Q(t) as the intensity matrix of {my} at time t shows that Q(t) is made up of two terms: obviously. the probability of downcrossing level u in phase i is 8ji(1 + p(u)dt • tii) + (1 . vi. A'(t) = A(t)Q(t) = A(t)(T + tv(u .3dt. the probability that level u is downcrossed for the first time in phase i is ai.(u) p ( u) = .246 CHAPTER VIII. given A.4) jEE jEE Proof Consider the event A that there are no arrivals in the interval [0. In the first case. (7. Given this occurs.Sj i)p(u)dt • tji = Sji + p(u)tji dt. Given A'. 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. MATRIXANALYTIC METHODS However. two things can happen: either the current jump continues from u + p(u)dt to u. Given A.t and being followed by a downcrossing. the probability that level u + p(u)dt is downcrossed for the first time in phase j is vj (u + p(u)dt).Qdt) vi(u) + vi'(u)p(u)dt + p(u) dt E{tji+tjvi(u)}.t) for the second. The intensity of a jump from i to j is tij for jumps of the first type and tivj(u . we get vi(u) = aidt + (1 .
(v) is given by the r. P u which implies that v.)P"(AnB. Since the processes Rt and Rt coincide under level B. we have p(r) = p = vi (u) 0aTe.4) backwards for {va (t)}v>t>o.2.) + 0 as v + oo. consider a modification of the original process {Rt} by linearizing the process with some rate p. <oo. . Thus.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. Now since both P(A n Bv) 3 0 and P"(A n Bv) . we can first for a given v solve (7. starting from v"(v) = .3 For any fixed u > 0. (u) for any values of u and v such that u < v. say. of (7... Then P(B..0 as v + 00 we have P(A) P"(A) = P(AnBv)+P(AnBv) P"(AnB. supRt>v l t<7 I where o.^ 0. Rt ..h. (u) on both side and dividing by dt yields the asserted differential u equation. When solving the differential equation in Proposition 7. To deal with this. vi (U) = lim v= (u).) is the tail of a (defective) random variable so that P(Bv) + 0 as v 4 oo. refer to the modified process. V . Then pv(r) p(r) r < v p r>v ' and (no matter how p is chosen) we have: Lemma 7. F" etc. after a certain level v. then P(A n Bv) _ P"(A n BV'). and similarly P"(Bv) .5). we face the difficulty that no boundary conditions is immediately available. say.7. denotes the time of downcrossing level u .i7rT1/p.s.) P"(AnBv) = P(AnB. Let p" (t). RESERVEDEPENDENT PREMIUMS 247 Subtracting v. From Section 3. This yields v.
assuming u > v for the moment.6) We may think of process Rt as pieced together of two standard risk processes RI and Rte with constant premiums p1. The f iin in (7. (u)}. 2u.. the evaluation of Vi(u) requires q(u) = 1 . for u > v the distribution of v . the probability of ruin between a and the next upcrossing of v. The precision depends on the particular quadrature rule being employed. 7b Twostep premium rules We now assume the premium function to be constant in two levels as in VII.v v(u)eTat 1 1 . Thus we obtain a convergent sequence of solutions that converges to {vi(t)}u>t>o• Notes and references The exposition is based upon Asmussen & Bladt [30] which also contains numerical illustrations. let v(u) = a+2ieiT +ta+>)(uv). v = u. Corollary 3. 3u etc. MATRIXANALYTIC METHODS Next consider a sequence of solutions obtained from a sequence of initial values {v.1. say. as well p1(u). 0 < u < v.. (7. numerically implemented in Schock Petersen [288]) and the present one based upon differential equations require both discretization along a discrete grid 0.1.7) equals 01 (v .V" M 0 .1.. p(r) P. r<v r > v.zp1(u)/(1 . i.) are so. where v = inf It > 0 : Rt < v}.e. Therefore u pl(vvueTa t 1. typically the complexity in n is 0(n2) for integral equations but 0(n) for integral equations. which is available since the z/i'(.x) dx f v(u)eT xt dx . while the fourthorder RungeKutta method implemented in [30] gives 0(n5). where. p2..RQ (defined for or < oo only) is defective phasetype with representation (v(u). The algorithm based upon numerical solution of a Volterra integral equation (Remark VII. Then v(u) is the initial distribution of the undershoot when downcrossing level v given that the process starts at u. The trapezoidal rule used in [288] gives a precision of 0(n 3). such that Rt coincide with RI under level v and with Rt above level v.q(v dx +( ) ) = ( ) ( q( )) vueTva (7. However.9. 2/n.1a. cf. < 0}).248 CHAPTER VIII.7) f o (the integral is the contribution from {R.10 that in addition to the O'(•). To evaluate p1(u). We recall from Propositon VII. Let ii'( u) = a+'ie(T+ta +^)"e denote the ruin probability for R't where a+ = a+i) = laT1/pi. Recall that q(w) is the probability of upcrossing level v before ruin given the process starts at w < v. 1/n.. > 0} and the last term the contribu tion from {R. T).z51(v)).
I.^1(v) 1 .7.v) + (2^ + 3v2 ea'(u " . (7. Then one gets X20 20 21 f 1ea1(u v) + 1 3 3 ^ A 2(u e . Since µB = 5/21.be the eigenvalues of T + to( 2 ).v(u)eTVe . From Example 3.x) dx} V 1 1(v) f V v(u) eTxt.21 = ? yields 0(u) = 1.u .e6v Let Al = 3 + 2V'2.v(u ) eTV e J v(u)eTxtz/)l (v .. p2 < 3. RESERVEDEPENDENT PREMIUMS 1 . so we consider the nontrivial case example p2 = 4 and p1 = 1. Example 7.2.8) equals v v(u)eTxta+2) e(T+ta +))( vx)edx which using Kronecker calculus (see A.(7 The arrival rate is (i = 3.and A2 = 3 . B is hyperexponential corresponding to 3 0 3 a(2 2)' T= ( 0 7 t.2.8) The integral in (7.v(u)eTve).4 Let {Rt } be as in Example 3.jl (t ®e) Thus.to+))11 {e{T®(Ttoy+ ))}„ . 01(u) _ 24 u + 35 e6u 1 35 e 4(u) _ 35 .24ev .24e.e.v) 1eai(u v) + 7 7 1 e\2(u v) 1 3 ^') eA2 (u.1 from which we see that pl (u) = 1 + 1 249  1 .x) dx 1 ^(v) ( 1 . all quantities involved in the computation of b(u) have been found in matrix form.01 (v .2V"2.4) can be written as (Y(u) ®a+)e(T+t°+>)°1 (T ® (T .e6u 35 .
21(35e6v .24e5v .1)' ?.7) we see that we can write pi (u) = v(u)V2 where V2 depends only on v.24es" .250 CHAPTER VIII.24es" .b(v) = 192esv +8 35e6v + 168esv + 7* Thus all terms involved in the formulae for the ruin probability have been exu plicitly derived. Notes and references [30].24e5v .1 Thus. 192esv + 8 P1 .+ it (3 4'I 1 ea2(uv e1\2(u") 7 + ( 32 +4. The analysis and the example are from Asmussen & Bladt .1 V2 = 4e5"+6 35e6v . MATRIXANALYTIC METHODS From (7. and one gets 12e5" . pi (u) = p12(u)/p1 l(u) where p1i(u) p12(u) 35e6v .1. 21 3 In particular./2) ea 1(u .v)esv + 7 4_ 2../2 ea1(u") .2 35e6v . ) e sv + ( 2v/2.
4. for all t > 0. For example. The definition b[s] = oo for all s > 0 of heavy tails is too general to allow for a general nontrivial results on ruin probabilities. x 2iror2 (c) the Weibull distribution with decreasing failure rate . oo ) and say then that B is subexponential (B E S) if 251 . x 4 oo. we require that B is concentrated on (0. see I.N(µ. III. a2)) with density 1 e(logyFh) 2/2az . For the definition . B(x) = ex0 with 0<0<1.46 and at numerous later occasions require a light tail. B(x) = L(x)/x" where a > 0 and L(x) is slowly varying. and instead we shall work within the class S of subexponential distributions . (b) the lognormal distribution (the distribution of eu where U .f.B(x).g. Some main cases where this lighttail criterion are violated are (a) distributions with a regularly varying tail. 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.2b. For further examples. L(tx)/L(x) 4 1.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. the exponential change of measure techniques discussed in II. A rough distinction between light and heavy tails is that the m.
oo). (b) liminf BB(() ) > 2. B(x) Here B*2 is the convolution square.1 Let B be any distribution on (0.2. the behaviour in the lighttailed case is illustrated in the following example: Example 1. X2) > u x)/B(x) = 2.252 CHAPTER IX. that is. The proof shows that the condition for B E S is that the probability of the set {X1 + X2 > x} is asymptotically the same as the probability of its subset {max(Xi.2 If B E S.3 Consider the standard exponential distribution. To capture the intuition behind this definition. Proof By the inclusionexclusion formula.B(x)2 . if X1 + X2 is large . Thus . As contrast to Proposition 1. x 3 00. X2 with distribution B. X2) > x}.'s.v. Since B is concentrated on (0. X2 but none of them exceeds x. the distribution of independent r. given X1 + X2 > x. proving (a). 1).p. we have {max(Xi. That is. and thus the lim inf in (b) is at least lim inf P(max(Xi.'s X1. the r.2B(x). X2) > x) is P(X1 > x) + P(X2 > x) . 1/2 it has the distribution of X1I X1 > x. oo). P(Xi <yI Xi+X2>x) 1B(y). HEAVY TAILS B*2\ 2. then P(X1>xI X1+X2>x)* 2. X2) > x} C {X1 + X2 > x}. Then: (a) P(max(Xi.1) then means P(X1 +X2 > x) 2P(Xi > x). We later show: Proposition 1.F(X1 > x. P(max(Xi. That is. 1/2 'typical' (with distribution B) and w. B(x) ax. X1 is w. (1.p. In contrast. note first the following fact: Proposition 1.v. X2 > x) = 2B(x) . Thus the liminf in Proposition 1. X2) > x) ^' 2B(x). .1(b) is oo.v. In terms of r. Then X1 +X2 has an Erlang(2) distribution with density yeY so that B*2(x) xex. one can check that x x where U is uniform on (0. in the subexponential case the only way X1 + X2 can get large is by one of the Xi becoming large. then (with high probau bility) so are both of X1.
If lim sup B(x . [In terms of r.6)x)/((1 . Let 0 < 5 < 1/2. y] and (y. Finally lim inf B(x . B( 0 .] Proof Consider first a fixed y.S)x + B(Sx)2 < lim sup B(x) xaoo B(x) lim sup 2L((1 x^oo . The uniformity now follows from what has been shown for y = yo and the obvious inequality y E [0. x].B*n(x .1. then B(B(x)y) * 1 uniformly in y E [0. SUBEXPONENTIAL DISTRIBUTIONS Here is the simplest example of subexponentiality: Proposition 1. a contradiction.z B(x) . and combining with Proposition u 1. then the overshoot X . we therefore get lim sup B*2(x)/B(x) > 1+B(y)+ 1 .5 If B E S. Using the identity B*(n+1)(x) = 1+ + 1)(x) 1+ 2 1 .y)/B(x) > 1 since y > 0. We now turn to the mathematical theory of subexponential distributions.xIX > x converges in distribution tooo.'s: if X .yo].4 Any B with a regularly varying tail is subexponential.1(b) we get B*2(x)/B(x) * 2. Hence lim sup a+oo B*2(x) 2B((1 . we get BZ(x)) > 1 + B(y) + B(B()y) (B(x) .2) B(x) B(x ) B(x) Jo with n = 1 and splitting the integral into two corresponding to the intervals [0.5)x)' + 0 _ 2 L(x)l xa (16) Letting S 10. 1 < B(x ) B( x) Y) < B( 0). Proposition 1. we get limsupB*2(x)/B(x) < 2. If X1 + X2 > x. This follows since the probability of the overshoot to exceed y is B (x + y)/B(x ) which has limit 1. or they both exceed Sx. then either one of the Xi exceeds (1 . 253 Proof Assume B(x) = L(x)/xa with L slowly varying and a > 0.B*(n ) B(dz) (1.B E S.S)x.B(y) = 2. yo] as X + 00.B(y)) .v.y)/B(x) > 1.
.7 If B E S. HEAVY TAILS Corollary 1.254 CHAPTER IX. so assume the proposition has been shown for n.z) B(dz) _y B(x) 111 Lx B . Then by (1. then for any n B*n(x)/B(x) * n.y) sup v>o B(v) B(x) which converges to 0 by Proposition 1. x oo.z) B(dz) 2B(x) o rv 2 0 2 using Proposition 1.6 If B E 8.z) B(dz) (n + O(e)) ^x JO B(x) (n + 0(0) I B (x) . The case n = 2 is just the definition.z) B(dz).z ) (x ) = 1 + (^ B(x . B*(n+1) (x I xy + Jxx y) W. and this immediately yields the desired conclusions. then e"R(x) * oo.1) for all large n so that B(n) > cle6n for all n. choose y such that IB*n(x)/B(x) . This implies B(x) > c2e5x for all x. O The following result is extremely important and is often taken as definition of the class S. Proof We use induction. The first integral is y B(x .2. B(x) \Jo _ B(x . b[c] = oo for all e > 0.nI < e for x > y. Proposition 1.5 and the induction hypothesis. Proof For 0 < 5 < e. 0 Proof of Proposition 1.5 that B(n) > e6B(n . we have by Proposition 1.z) B(x) Here the second integral can be bounded by B*n(y) B(x) .2).B(x .B*2 (x) B(x) (x . 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 . its intuitive content is the same as discussed in the case n = 2 above.5 and dominated convergence. Given e > 0.
z) B(dz) < 1 + A + an(1 + d) . 0 Proposition 1.'s such that Xi has distribution Ai.8 If B E S. Then Al * A2(x) = P(X1 + X2 > x).X1 > xv.z) B(x) < 1 + A + an sup f x B(x .X2 > xv) < A1(xv)A2(x v) .y)Ai(dy) v) f o . 0 Lemma 1.ala2B(x)2 which can be neglected.z) B(x . Proposition 1.(al + a2)B(x). Then Al * A2 (x) .5 easily yields P(X1 + X2 > x. choose T such that (B(x)B*2(x))/B(x) < 1 + b for x > T and let A = 1/B(T). For any fixed v. e > 0.4) .2).B(x .y)Ai(dy) = (x)o(1) (1. an+1 fX B*n( *n(x . A2 be distributions on (0.ajB(x)Ai(v) = ajB( x)(1+o„(1)) (j = 3 . a2 with a1 + a2 > 0.y)B(dy) = B(x)ov (1)• v (1. then there exists a constant K = KE such that B*n(x) < K(1 + e)nB(x) for all n and x. x>T o B(x) The truth of this for all n together with al = 1 implies an < K(1 + 5)2n where K = (1 + A)/e. Since P(X1+X2 > x.v. Proof Define 5 > 0 by (1+5)2 = 1+e.z) B(dz) x . SUBEXPONENTIAL DISTRIBUTIONS 255 Here the first term in {•} converges to 1 (by the definition of B E S) and the second to 0 since it is bounded by (B(x) . Proof Let X1. Then by (1.3) Using the necessity part in the case Al = A2 = B yields f xv B(x . X2 be independent r. Combining these estimates and letting a 4.0 completes the proof. an = supx>o B*n(x)/B(x). Xi <= v Ai (x .1.i).z) B(dz ) + sup < 1 + sup f x<T B ( x) x>T 0 B(x .9 Let A1. oo) such that Ai (x) _ aiB(x) for some B E S and some constants al. it follows that it is necessary and sufficient for the assertion to be true that JX_VA (x .y))/B(x).
V (1. whereas the two first yield B(x)(Ai(v) .aiB(v)) = B(x)o„(1). A2 = B so that a1 = 0. it should hold that B1 * B2 E S and B1 * B2 (x) . Proof Taking Al = A2 = A. In the regularly varying case.h. i = 1.13 Let B have density b and failure rate A(x) such that . of (1.2aB(x) .4).v)B(v) + _'U Aq(x .3) follows if CHAPTER LX. f " By a change of variables. then A E S. it is easy to see that if L1. if q(x) aB(x) for some B E S and some constant a > 0.(x) is decreasing for x > x0 with limit 0 at oo. a2 = 1. u It is tempting to conjecture that S is closed under convolution. Then A * B E S and A * B(x) . B1 * B2 E S does not hold in full generality (but once B1 * B2 E S has been shown.Bl (x) + B2 (x) follows precisely as in the proof of Proposition 1.2. That is.v)Ai(v) . L2 slowly varying. HEAVY TAILS 'VV B(x .Ai(x .10 The class S is closed under tailequivalence. B1 * B2 (x) .y)Ai(dy) = B(x)o„(1). A(x) = o(B(x)). u Corollary 1.12 Assume that Bi(x) = Li(x)lxa. the l. That is. a1 = a2 = a yields A*2(x) .Bl (x) + B2 (x) when B1.2A(x).5) Here approximately the last term is B(x)o„(1) by ( 1. Then L = L1 + L2 is slowly varying and B1 * B2(x) sim L(x)/x«.B(x) Proof Take Al = A. However. . B2 E S.11 Let B E S and let A be any distribution with a ligther tail.5) becomes x B(x .y)B(dy). Then B E S provided fo "O exA(x) b(x) dx < oo. L2 are slowly varying.s.9). u Corollary 1. then so is L = L1 + L2. with a > 0 and L1. Recall that the failure rate A(x) of a distribution B with density b is A(x) = b(x)/B(x) Proposition 1.256 Now (1. Hence Corollary 1. We next give a classical sufficient (and close to necessary) condition for subexponentiality due to Pitman [290].
By (1.y) < A (y) for y < x/2. Thus A(x) is everywhere decreasing. f ' L(y) dy . an integrable function by assumption. In the regularly varying case.(x . Example 1. a(x) = ax01. we can use the same domination for the second integral but now the integrand has limit 0 . Thus by dominated convergence . Thus. Example 1.13 works in this setting.A(xy)A ( y).15 In the lognormal distribution. the first integral has limit 1 .A(y)\(y) dy + fox/ 2 eA(x ). x .y) y\(y)• The rightmost bound shows that the integrand in the first integral is bounded by ey"(v). Goldie & Teugels [66]): Proposition 1. Further.(y) dy. Thus B*2(x )/ B(x) .y) < yA(x . Jo For y < x/2.y) * 0. and exa(x)b(x) = (3x01e(10)x9 is integrable.U) /or) v 2x This yields easily that ex.y ) b(y)dy = B (x) o ox _ J = ox/2 eA( x)A(xy ). proving B E S. Define A(x) = fo .2). Then B(x) = eA(x). To illustrate how Proposition 1.3 < 1. Then b(x) = Ox0lexp.1 B(x) eA( x)A(xv )A(y)A(y) dy f B(x . The middle bound shows that it converges to b(y) for any fixed y since \ (x ...y) dy.12. L(x) y° (a . the DFR Weibull distriu bution is subexponential.A(x .e009xv)2/2a2/(x 2irv2) logx ( ) 't ((logx . subexponentiality has alrady been proved in Corollary 1.16 For L(x) slowly varying and a > 1. the u lognormal distribution is subexponential.1 has limit 1 + 0.`(x)b(x) is integrable.1)xcl1 . B*2(x) .14 Consider the DFR Weibull case B(x) = ex0 with 0 <. Thus. 0 A(x) . replace B by a tail equivalent distribution with a failure rate which is everywhere decreasing). Since ) (x . SUBEXPONENTIAL DISTRIBUTIONS 257 Proof We may assume that A(x) is everywhere decreasing (otherwise. elementary but tedious calculations (which we omit) show that A(x) is ultimately decreasing. we first quote Karamata's theorem (Bingham.A(y)a(y ) = ev'(y) b(y).1.
(c) Under the assumptions of either ( a) or (b). let X W = X .258 From this we get CHAPTER IX. More precisely.y(x)B(x).13 may present a problem in some cases so that the direct proof in Proposition 1.6) EX(x) .L(x)/x" and )t(x) .1))a . Proof ( a): Using Karamata's theorem.1)X(x)/x > y) = P(X > x[1 + y/(a .1) and P(X (.x)+ _ 1 °° P(X > x) P(X>x )J L PX >y)dy 1 x L(y)/ydy L(x)/((a1)x'1) x )l ° J °° ( ()l a x a1 Further P ((a . HEAVY TAILS Proposition 1.ea b(x) is integrable. .4 is necessary in full generality.1/A(x) and P(X ixil'Y (x) > y) * e'. yo] .a/x. Then: Proposition 1.xjX > x. we get (1. Thus exa(x)b(x) . We conclude with a property of subexponential distributions which is often extremely important: under some mild smoothness assumptions. However.1))^ ' (b) Assume that for any yo )t(x + y/A(x)) 1 A(x) uniformly for y E (0. (1 + y/(a .1)])a 1 1 . the monotonicity condition in Proposition 1. Then 7(x) .18 (a) If B has a density of the form b(x) = aL(x)/xa with L(x) slowly varying and a > 1. 'y(x) = EXix>. then 7(x) x/(a .1)] I X > x) L(x[1 + y/(a . then B(x) .E(X .17 If B has a density of the form b(x) = aL(x)/x°+1 with L(x) slowly varying and a > 1. f O B(y) dy .)/Y(x) > y) (1 + y/(a .1)]) xa L(x) (x[1 + y/(a . the overshoot properly normalized has a limit which is Pareto if B is regularly varying and exponential for distributions like the lognormal or Weibull.
. d. Recall that B0 denotes the stationary excess distribution. nG(u). cf.7) is referred to as 1/A(x) being selfneglecting. Kliippelberg & Mikosch [134]. Y2.. St > u}..A(x + y/A(x))} =a(x) a(x + x) dx = ex p ex P . Examples 1.15. Let St = Ei ` Ui .+YK> u) = ^•P(K = n)G* n(u ) . Theorem 2 . It is trivially verified to hold for the Weibull.nn.EK G(u). .(x).1/.v. We get p(yl+. be i. 0 G(u) L G(u) .and lognormal distributions . 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. Then P(Y1 + • • • + YK > u) . P The proof is based upon the following lemma (stated slightly more generally than needed at present).f yl 0 0 = exp {y (1 + 0(1))} 0 fY A( x + u /A( x)) a(x) du } The property (1.2 Let Y1.2. with common distribution G E S and let K be an independent integervalued r.n0 1•P(K= n)•n = EK. THE COMPOUND POISSON MODEL 259 We omit the proof of (c) and that EX (x) . Lemma 2. Bo(x) = f0 B(y) dy / µB. We assume p = /3µB < 1 and are interested in the ruin probability V)(u) = P(M > u) = P(r(u) < oo)..8) in (b) then follows from P (A(x)X (x) > y) = F(X > x + y/.t be the claim surplus at time t and M = sups>0 St. then Vi(u) P Bo(u). with EzK < oo for some z > 1. The remaining statement (1. . r(u) = inf it > 0.A(x) I X > x) = exp {A(x) . Notes and references A good general reference for subexponential distribution is Embrechts.1 If Bo E S. 1. 2 The compound Poisson model Consider the compound Poisson model with arrival intensity /3 and claim size distribution B. u a oo.14. i.
Weibull) one has Bo(x ( B(x) . Note that in these examples . the result follows immediately from Lemma 2. Proof Since B(x + y)/B(x) * 1 uniformly in y E [0.1 is notoriously not very accurate.x400 PBB(x) PB Leta+oo.x^ ) B(x) _ f or ( lox . See also Embrechts & Veraverbeeke [136]. The approximation in Theorem 2. see Abate. r(1/Q) xlQexp B(x) = ex' From this . Bo E S.µ J ) .?(xµ 8 (x). For some numerical studies. The PollaczeckKhinchine formula states that (in the setup of Lemma 2.2. HEAVY TAILS u using dominated convergence with >2 P(K = n) Dz" as majorant. . u The condition Bo E S is for all practical purposes equivalent to B E S. lognormal . u x+a Notes and references Theorem 2. (2..2) M = Yl + • • • +YK where the Yt have distribution Bo and K is geometric with parameter p. Bo is more heavytailed than B .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 ) .p) and EzK < oo whenever pz < 1. _ B(x^sx Bo(x) µ8 I aoB(y )dy = (^) .1 is essentially due to von Bahr [56]. The tail of Bo is easily expressed in terms of the tail of B and the function y(x) in Proposition 1. However. mathematically one must note that there exist (quite intricate) examples where B E S.. P(K = k) = (1. x 4 00. In general: Proposition 2.1.p)p'. in our three main examples (regular variation . Bo E S is immediate in the regularly varying case. Proof of Theorem 2.3 If B E S. we have fx B(y)dy = a B0 (x) > lim inf lim inf x+oo B(x) . and for the lognormal and Weibull cases it can be verified using Pitman 's criterion (Proposition 1. The problem is a very slow rate of convergence as u ^ oo. then Bo(x)/B(x) + 00. a].µB(01 . Borovkov [73] and Pakes [280]. Bo ¢ S.1) In particular .260 CHAPTER IX.18.13). as well as examples where B ¢ S. Since EK = p/(1.
1 Assume that (a) the stationary excess distribution Bo of B is subexponential and that (b) B itself satisfies B(x .y)/B (x) > 1 uniformly on compact y internals. This shows that even the approximation is asymptotically correct in the tail. one may have to go out to values of 1/'(u) which are unrealistically small before the fit is reasonable.1 when u is small or moderately large. M = sup s$ .1 gives 1010. 3 The renewal model We consider the renewal model with claim size distribution B and interarrival distribution A as in Chapter V. Based upon ideas of Hogan [200].+ E A. T1 the ith interarrival time and Xi = U. {n= 0. i=1 . p = iB /µA < 1.1. To Bo(u) u + 00.. . Thus G+ is the ascending ladder height distribution (which is defective because of PB < PA). let t9+ = i9(0) be the first ascending ladder epoch of {Snd> }. E. Define further 0 = IIG+II = P(r9+ < oo).. Kalashnikov [219] and Asmussen & Binswanger [27]. (3. T+ < oo) where r+ = T1 + • • • + T. also a second order term is introduced but unfortunately it does not present a great improvement. Let U= be the ith claim . there is a representation of M similar to the PollaczeckKhinchine formula. We assume positive safety loading.. + Xn..3. THE RENEWAL MODEL 261 Choudhury & Whitt [1]. [279]. Somewhat related work is in Omey & Willekens [278]. Snd) = Xl +. In [1]...9+ < oo) = P(S.1) this end . The main result is: Theorem 3 .e.] The proof is based upon the observation that also in the renewal setting.} Then ik(u) = F ( M > u) = P(i9 (u) < oo).y + as usual denotes the first ascending ladder epoch of the continuous time claim surplus process {St}. 195 there are numerical examples where tp(u) is of order 105 but Theorem 2. Then l/i(u) 1 P P [Note that (b) in particular holds if B E S. in [219] p. Asmussen & Binswanger [27] suggested an approximation which is substantially better than Theorem 2.g.Ti. i. t9(u) = inf {n : Snd> > u} .. G+ (A) = P(Sq+ E A. Then K M=EY.
oo. 0 The lemma implies that (3. Proof By dominated convergence and (b).1 IPG_ I / F(x . (b) and does not rely on the structure Xi = Ui . with distribution G+/9 (the distribution of S.3 G+ (x) . (3.1) holds for a general random walk satisfying the analogues of (a). x + oo.Ti). Lemma 3 . the contribution from the interval (.3) and we will prove it in this form (in the next Section.B(x). As for the compound Poisson model.y) R+(dy ) _ j (x_y)U_(dY) G+ (x) = J 00 00 (the first identity is obvious and the second follows since an easy time reversion argument shows that R+ = U_.d)) E A) denote the pre19+ occupation measure and let and U_ = Eo G'_" be the renewal measure corresponding to G_.9)9'' and Y1. A(dy) = 1. Proof Let R+(A) = E E'+ ' I(S. Write G+( x) = G+ ( x. d+ < oo).Y2.1) is equivalent to P(M > u) " .2 F(x) .d.(. 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. G_(A) = P(S. The heuristics is now that because of (b). B(x) _ J O° B(B(x)y) A(dy) f 1 . oo) = F(S. Then 0 0 F( x . U_ (dy) is close to Lebesgue measure on (.FI(x) /IPG_I.. and hence FI(x) .i.y) dy = 1 Pi (X) oo IPG_ I . Let further 19_ _ inf {n > 0: S^d^ < 0} be the first descending ladder epoch. HEAVY TAILS where K is geometric with parameter 9. cf.PBBo(x). x * oo. we will use the fact that the proof of (3. Lemma 3 . FI (x) _ fz ° F(y) dy. whereas for large y . Let F denote the distribution of the Xi and F1 the integrated tail. 0] to the integral is O(F(x)) = o(FI(x)). 0] normalized by IPG_ I so that we should have to G+(x) .262 CHAPTER IX.y_ E A) the descending ladder height distribution (IIG II = 1 because of PB < P A) and let PG_ be the mean of G_.g+ > x. x > 0. u a 00.. A. are independent of K and i.y+ given r+ < oo).N. P(K = k) = (1 ..FI(u).2).
We then get lim sup G+(x) xro0 Fj(x) < lim sup X)00 o F(x . (3.e) z lim inf G+(x)  FI (x) Ip G_ I Letting a 10.9) 1 . By Lemma 3. n] + 1/I µG_ I.1 I .3.2).G+[s]) .y) U.=1 BIp G_ I (1. and that U_(n . n] F1 ( n=N _1 1+e E F(x+n) 0 + limsup xr00 FI(x) FAG. In the lattice case.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 in the last that FI is asymptotically proportional to Bo E S.1. If G_ is nonlattice. choose N such that F(n . 0] x+00 FI(x) 00 + lim up 1 x) E F(x + n) U_ (n .1.1.I n=N (1 E)2 r00 F(x + y) dy + e) lim sup .3. Hence using dominated convergence precisely as for the compound Poisson model. n] < (1 + e)/1µc_ I for n > N. the proof is complete.y) U_ (dy) 00 FI (x) < lim sup F(x) U(N.O[s])(1 . Given e. THE RENEWAL MODEL 263 We now make this precise.2) yields 00 F F I (u) P(M > u) _ E(1 .UG_ I x. > (1 . F(Y= > x) FI(x)/(OIp _ 1). u Proof of Theorem 3.0)0k k I(u) A.1.9)IpG_ I Differentiating the WienerHopf factorization identity (A. we can assume that the span is 1 and then the same conclusion holds since then U(n .(dy) fN FI ( x) + lim sup ZY00 N F(x . then by Blackwell 's renewal theorem U_ (n . Similarly.1)/F(n) < 1 + e for n > N (this is possible by (b) and Lemma 3.F[s] = (1 .1. n] is just the probability of a renewal at n.
4 Models with dependent input We now generalize one step further and consider risk processes with dependent interclaim times. Mn < u}.(1 ..(u)+n . u)) > P(w(u) < oo)(i lp (0))• On the other hand. Then P(M E (u . must attain a maximum > 0 so that P(M > u.yiui_1.So( u)_1 < a) < P (w(u) < oo)j/i(0) < 0(0) P(M E (u .1)6+[0] .AB iP We conclude by a lemma needed in the next section: Lemma 3 .l.a. S+9(u) . Notes and references Theorem 3.Sty(u)_I < a} we have w(u) < oo. S+q(u) .264 and letting s = 0 yields CHAPTER IX.IIG+II)µc_ = (1 . HEAVY TAILS µF = (1 . Sty(u) .2. with roots in von Bahr [56] and Pakes [280].a) N P(M > u). Note that substantially sharper statements than Lemma 3. 10(0) But since P(M > u .4 on the joint distribution of (S. . on the set {M > u. P(M > u. FJ(u) UBBO(U) PBo(u) N = (10)Ipc_I JUA . see Asmussen & Kliippelberg [36].4 For any a < oo.So(u)) are available..a.a. u)..a. u)).u)) = o(P (M > u)) = o(FI(u)). 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.0)ua_ . Proof Let w(u) = inf {n : Sid) E (u . Therefore by Lemma 3. and {Su.SS(u)}n=o. we have P(M E (u .1 is due to Embrechts & Veraverbeke [136].Se(u)_1 < a) = o(Fj(u)).
2. The idea is now to observe that in the zerodelayed case. Figure 4... Define S. 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 <. (viewed as random elements of the space of Dfunctions with finite lifelengths) are i.1 Note that no specific sample path structure of {St} (like in Fig. Thus the assumption .X2 < .1 except for the first one) is a random walk.. {SX1+t .Sxk}o<t<xk+1xk is the same for all k = 1. t>0 S. Schmidli & Schmidt [47]. 0o(u) etc. and apply it to the Markovmodulated model of Chapter VI.1.. The zerodelayed case corresponds to Xo = Xl = 0 and we write then F0.. 4. For further approaches....X1 is the generic cycle.. assume pp. 4. See Fig....4. E0.Sxi}0<t<x2Xl . and the distribution of {Sxk+t .d.1 = max k=0.1.1 where the filled circles symbolize a regeneration in the path... M. We let F* denote the Podistribution of Si. . M* = max S. We give here one of them. = Sx... (corresponding to the filled circles on Fig. G(x) (4.. MODELS WITH DEPENDENT INPUT 265 Various criteria for this to be true were recently given by Asmussen. < 0 and EoX < oo where X = X2 ..1) is assumed.1 based upon a regenerative assumption.n n=0. such that {SXo+t  SXo}0<t< X 1Xo .1) .. +1. {Sn}n=o.i. Theorem 4.. M = sup St.. see [47]. We return to this point in Example 4.F*(X) = P0(Si > x) .4 below. 4. examples and counterexamples.
1) and (4.266 CHAPTER IX. Proof Since M > M*.. (4.3) applicable so that F(M* > u) 141 F*(u)..Sxn = sup Sxn+t .2. HEAVY TAILS for some G such that both G E S and Go E S makes (3. 4..* i o<t<xn+1x. Fo(Si > X). (4.3) where Mnx) = sup o<t<xn +1 X. Sxn +t . the assumption means that Mix) and Sl are not too far away. See Fig. it suffices by (4. The one we focus on is Fo (Mix) > x) . Then '00 (u) = Fo(M > u) .2) Imposing suitable conditions on the behaviour of {St} within a cycle will then ensure that M and M* are sufficiently close to be tail equivalent.S.3) hold. Since clearly M(x) > Sl . jF11 F* (U).2 Theorem 4..4) liminf u>oo F(M > u) . u p 00. N N Xi=0 N Figure 4. (4.1 Assume that (4.2) to show F(M* > u) > 1.
Let a > 0 be fixed.4.5) which follows since Po (M > u.4). M^xu)+l > a) .+Mn+1>u} 267 (note that {M> u} = {3(u) < oo})... S.. Letting first u + oo and next e . Under suitable conditions . choose a such that Po(Si > x ) > (1 . MW O(u)+1 < a) IN ( U n=1 A1. Theorem 4.1 = limti00 St/t.Sn 0<t<x„+j ( 1 . assume the path structure Nt St = EUit+Zt i=1 . ..Mn +1 >aV(u n=1 00 S. Po(M* > u) .: Sn > u} ..a. E (u .e)Po (M > U).e)Po (M > u. /3(u) = inf{n=1.a. MODELS WITH DEPENDENT INPUT Define 79* (u) = inf {n = 1 . u))/P(M* = 0) = o(Po(M* > u)). 0 yields (4.(u) .S.: S.6) 1 p pBo(u) u where B is the Palm distribution of claims and p .e)Po (MMX> > x). To this end. u)} < P(M* E (u . 2. Mn+l > a V (u . x > a.2.Po (M* > u. We shall use the estimate Po(M > u) Miu^+ 1 < a) = o(Po (M > u)) (4. Given e > 0. )) > (1 .Sn+1Sn>aV(uSn*)) n=1 00 > (1E)EPo(Mn<u.1 can be rewritten as 00 (U) (4.(1 ..E) Po ( n max St u.( u)1 > a) 00 1: Po(Mn<u.4.. Then by Lemma 3.
are Fmeasurable and NX Po J:U=>x i=1 (iv) Po sup Zt > x / (0:5t<x o(B(x)) Then (4.v.7). and ENX Ui . a4' 0. The same is true for Sl. cf. Mix) < > UE + i=1 o<t<x Thus Theorem 4. we get 00 (u) 1 IPF.p) Ju P Bo(u) 1p 0 .X both have tails of sup Zt. i=1 (4.I u J Po(Sl > x) dx 1 EoNxB(x) dx EoX(1 . Assume further that (i) both B and Bo are subexponential. Then the Palm distribution of claims is B(x) = E N Eo 0 I( U1 < x) . HEAVY TAILS N` U.1 is in force. oX (see Proposition A1.2 Assume that {St} is regenerative and satisfies (4. and the rest is just rewriting of constants: since p = 1+tlim St = 1+ . since the tail of Zx is lighter than B(x) by (iv).8) x Write .'s order EoNx • B(x).6 below. the proof of Lemma 4. Corollary 4.268 with {Zt} continuous. independent of {> CHAPTER IX. and also for Mix) since Nx FNX U. X and N.Q = EoNx/EoX. Proof It is easily seen that the r.} and satisfying Zt/t N.4).3PB.6) holds with p = . (ii) EozNX < oo for some z > 1. (iii) For some o field Y.
. and taking F = o. The number N. In particular. Taking again Xo = Xi = 0.3 As a first quick application. . (iii) is obvious.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).t} is standard compound Poisson and {Zt} an independent Brownian motion with mean zero and variance constant a2.4 Assume that St = Zt . The key step of the proof is the following lemma. We consider the case where one or more of the claim size distributions Bi are heavytailed. Thus we conclude that (4. of claims arriving in [0.. (i) holds. .t + EN'I Ui where {>N`1 Ui . in particular lighttailed. The regenerative assumption is satisfied if we take Xo = Xi = 0. X3 = 2. 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. Example 4 .4. i.. 3 The average arrival rate / and the Palm distribution B of the claim sizes are given by P P Q = ir i/i. consider the periodic model of VI. 1) is Poisson with rate /3 = fo /3(s) ds so that (ii) holds. Bo E S.(NX). More precisely. Theorem 4. we conclude just as in Example 4.. .5 Consider the Markovmodulated risk model with claim size distributions satisfying (4.6) holds.6) u holds. X2 = 1.9). we assume that B E S. Again . The arrival rate is /3i and the claim size distribution Bi when Jt = i. > 0. i=1 B = >2 7riaiBi i=1 and we assume p = 014 B = Ep ri/3ipB. We now return to the Markovmodulated risk model of Chapter VI with background Markov process {Jt} with p < oo states and stationary distribution 7r. Assume that B E S. MODELS WITH DEPENDENT INPUT 269 Example 4 . < 1. Then (4. X2 = 1. then (iv) holds since the distribution of supo<t<i Z(t) is the same as that of I Zl 1. Zt .3 that (4. and for some constants ci < oo such that cl + • • • + c. 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 .6) holds.0 (thus (iv) is trivial).. Bo E S. .e. X3 = 2.
v. NP ) and X are . An easy conditioning argument then yields the result when Jo is u random.F) = P(Yo > X+x I •^) G (x +x)>2ciNi i=1 . 2 . i1 = E\ G(x) In the general case. oo) and define p Ni Yx = EEX'i . The same dominated convergence argument completes the proof.Fmeasurable. NP ) be a random vector in {0.270 CHAPTER IX. and the rest of the argument is then just as the proof of Corollary 4. i=1 P(Yx > x ^) < P(Y0 > x I. we can define the regenerations points as the times of returns to i.X i=1 j=1 where conditionally upon F the Xi. 1.. If Jo = i. Then P P(Yx > x) ... . cp with cl + > 0 it holds that Fi(x) .2. P P P(YX and > x I."+Np . i=1 Proof Consider first the case X = 0. i =1 P(Yo > x I ^ ) < CG(x)zN1+ +Np for some C = C(z) < oo. .. and F a aalgebra such that (N1. Markovmodulation typically decreases the adjustment coefficient y and thereby changes the order of magnitude of the ruin .G( x ) > ciNi .. . and that for some + cp distribution G on [0. 6 Let (N1. are independent with distribution Fi for Xij. .. X > 0 a r. For lighttailed distributions. . . Assume EzN1+"'+Np < oo for some z > 1 and all i.. u Proof of Theorem 4. Thus dominated convergence yields ( P(Yo>x P(Yo>x . . as x a oo. oo) such that G E S and some c1.}P.5.c'(x) where c = ciENi . Let {Fi}t=1 P be a family of distributions on [0. HEAVY TAILS Lemma 4 .. It follows by a slight extension of results from Section 1 that P P(Yo > x I Y) G( x) ci Ni.ciG(x).^•) G(x) P ^ E ciNi = C.F) < CG(x)zn'1+.
Then O(u) .2 and Example 4.4. we let PN"N = P(. and independent of (T1. m is a (orfinite) .T2. Within the class of risk processes in a Markovian environment. The main result of this section.5 shows that basically only the tail dominant claim size distributions (those with c. An improvement was given in Asmussen & Hojgaard [33]. as well as a condition for (4. and the final reduction by Jelenkovic & Lazar [213]. I T(u) < oo). i. IV. ) form a general stationary sequence and the U.4. see Schlegel [316].. 5 Finitehorizon ruin probabilities We consider the compound Poisson model with p = /3pB < 1 and the stationary excess distribution Bo subexponential. Theorem 4. FINITEHORIZON RUIN PROBABILITIES 271 probabilities for large u.4.5. > 0) matter for determining the order of magnitude of the ruin probabilities in the heavytailed case. states that under mild additional conditions.e. 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).7).6) to hold in a situation where the interclaim times (T1. It follows from Theorem 4. Theorem 5..4. cf.3.p)Bo(u). Notes and references Theorem 4.. The present approach via Theorem 4. We start by reviewing some general facts which are fundamental for the analysis. this then easily yields approximations for the finite horizon ruin probabilities (Corollary 5. That paper also contains further criteria for regenerative input (in particular also a treatment of the delayed case which we have omitted here). this is applied for example to risk processes with Poisson cluster arrivals.1 is from Asmussen.i.d. i. ). For further studies of perturbations like in Corollary 4.1. As usual.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. the discussion provides an alternative point of view to some results in Chapter IV. VI. Floe Henriksen & Kliippelberg [31] by a lengthy argument which did not provide the constant in front of Bo(u) in final form... this should be compared with the normal limit for the lighttailed case. cf.T2. Combined with the approximation for O(u).7. in particular Proposition 2. 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. In contrast.. Schmidli & Schmidt [47]. r(u) is the time of ruin and as in IV. Theorem 2. for lighttailed distributions the value of the adjustment coefficient y is given by a delicate interaction between all B. Essentially. 5 was first proved by Asmussen. cf.pl(1 .
y = Qx (. {Rt}. to consider only the case Px(w(F`) = 0) 0. For the present purposes it suffices . y = 0). {St} and {Rt} are in classical duality w.1 A compound Poisson risk process {Rt} and its associated claim surplus process {St} are in classical duality w .). The equality of the l. x = 0+ and F = (0. 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} .272 CHAPTER IX. resp. The simplest example is a discrete time discrete state space chain. for states i. .z. Then there is a Markov process {Rt} on E such that fE m(dx)h(x)Exk(Rt) = Lm(dy)k(y)Eyh(St) (5. k on E. and starting from So = y. however . (note that we allow x.2) with t = 1 means m. and (5.s. t. u For F C E. a main difficulty is to make sense to such excursions also when Px(w(F°) = 0) = 1. say. Proof Starting from Ro = x.t. w(Fc) < oo ) 'In general Markov process theory. Let G denote the distribution of ENt U. m. .2) means ffh(a. Sw(F. Lebesgue measure.h. follows by the substitution y = x . k as indicator functions. Thus. oo). where we can take h.rij = mjsji where r13.)k(x . the whole of R and not as usual impose the restrictions x > 0.z) dx G(dz) = ffh(y + z) k(y)dy G(dz). a familiar case is time reversion (here m is the stationary distribution).s=j are the transition probabilities for {St}. j.t.00).h. but the example of relevance for us is the following: Proposition 5.= y. r.s. Rt is distributed as x + t . y to vary in. Say {St} is reflected Brownian motion on [0. We let QS be the corresponding distribution and Qx. Then (5.t + EI U. to the r. St is distributed as y . in the terminology of general Markov process theory.r.2) for all bounded measurable functions h. HEAVY TAILS measure on E such that L for all measurable A C E and all t > 0.>N` Ui.
S.y(2p21 . i1. this simply means the distribution of the path of {Rt} starting from y and stopped when 0 is hit.1 The sample path in (a) is the excursion of {St} conditioned to start in x = 0 and to end in y > 0. io = x.y = Qy Q. . QR and QRy are defined similarly. We consider the discrete time discrete state space case only (wellbehaved cases such as the risk process example can then easily be handled by discrete approximations). w(0. 5.. We can then view Qy. and we let Qy y refer to the time reversed excursion . in with i0. when p = . The theorem is illustrated in Fig .5. Sn+1 E Fc) nx.(0)_ = y < 0 is the same as the distribution of w(y) where w(z) = inf It > 0 : Rt = z}... oo) = r(0) x= St y (a) Figure 5... In particular: Corollary 5. in E F.. Qx.2 Qy. [note that w(z) < oo a. FINITEHORIZON RUIN PROBABILITIES 273 y E F (in discrete time.2.).. Sn = in = y.13AB < 1] Proof of Theorem 5. Sw(F)1 = y) . Sw(Fo)_ should be interpreted as Sw(F^)_1). the one in (b) is the time reversed path. . That is. . But in the risk theory example (corresponding to which the sample paths are drawn).y as a measure on all strings of the form i0i1 .itt) = P Px(w(Fc) < 00..3 The distribution of r(0) given r(0) < oo.= y) Theorem 5 .1 for the case F = (oo.s. in = y. 0]. /^s x (S1 = Z1. z > 0. The theorem states that the path in (b) has the same distribution as an excursion of {Rt} conditioned to start in y < 0 and to end in x = 0. x = 0. . Thus.SS(F. Qx y is the distribution of an excursion of {St} conditioned to start in x E F and terminate in y E F.y() = P ({SW(F`)t} 0<t<w(F °) E So = x.
.ik_1EF Sxin_1 ..in E F. in = x..J (i... HEAVY TAILS E E Px (Si = 21i .. Rn = in = x. .... Si1y k=1 i1 .. Si11 S 1 .... . .TI( 2n2n _1 ..ii ..ik1EF Similarly but easier Sxin_1 .. ... in) = oo jEF^ Sxin1 . i0 = y.in1 .. Sn+1 E Fc) n=1 i1.. (Fc)1 = y) 00 CHAPTER IX. Silt' E Sxik_1 . R Qy x(2p21 .. Rn+1 E FC) TioilTili2.ik_1EF .. 2p)..(F<)1 = Y) S S and Qx y( ipil . . in)  Pt' (R1 = ii. 2n) = Qx.. i0) Q x.. .. 2p) when 20..i„_iEF Similarly. 21 . 20 = y. R .. To show Q y x (i0 i 1 .......... in) = Qx... = in = y...rin_1in E Txj jEFC m21 s2120 m2252221 m in Ssn n1 mjSjx Mx m2p mil min1 jEF` 1 Sinin _ 1 .. Silt' E SO k=1 i1.. t' y and Qy x are measures on all strings of the form ipi l . Rn+1 E F`) F (w(Fc) < 00.. in = x.274 note that Fx(w(Fc) < 00. MY Thus Qx(ioii .. ... S.. in E F. . in with 20.. Si l io E mjSjx. . Rn = in = x. Si1y 00 jEF° E E 5xik_ 1 .. note first that Pt' (R l = il..y(inin _ E SYj jEF` 00 Sxik _1 ..gilt' k=1 ii .. S.......
's are defined w. 7(0) < oo.2. Let Y = Yl = Sr+( 1) be the value of the claim surplus process just after the first ladder epoch . the case r (O) < oo. y > u. Y > y} .')distribution of Yu is Bo").B(a) +a PBBo(u) .p.2 The distribution of (Y. To clarify the ideas we first consider the case where ruin occurs already in the first ladder segment .')distribution of Z since P(Z>aIY>u) = 1 °° B(y) B(y + a) dy FLBBo(u) B (y) J°° (z) dy . Now the P(u. P(o) ). 1 w . Z follows the excess distribution B(Y) given by B(Y) (x) _ B(y + x)/B(y). the distribution w.(o) > y} = {T(0) < oo.2. 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 .v.2. Z) is described in Theorem 111. We are interested in the conditional distribution of T(u) = T(0) given {T(0) < oo. The formulation relevant for the present purposes states that Y has distribution Bo and that conditionally upon Y = y. Bo") is also the P(u. Z = Zl = ST+( 1)_ the value just before the first ladder epoch (these r.')density of Y is B(y)/[.r. P(") = P(. see Fig.r.t. That is.UBBo(u)]. the P(u.t. S. Y > u). that is. 5. ST(o) > y. U T(O) = T (u) Y Figure 5.5. that is.
p) in F(u) distribution. (Y. Fig.276 CHAPTER IX.. more precisely.3.. Then Corollary 5. .. i. Bo") ).3) where the distribution of W is Pareto with mean one in case ( a) and exponential with mean one in case (b). Y1 + • • • + Yn > u} denote the number of ladder steps leading to ruin and P("'n) = P(• I r(u) < oo.. We let K(u) = inf In = 1. we get the same asymptotics as when n = 1. Z/'y(u) * W in Pi "' ')distribution . That is . Then. We now turn to the general case and will see that this conclusion also is true in P(")distribution: Theorem 5 . .e. HEAVY TAILS Let {w(z)}Z^. a slight rewriting may be more appealing.1. must be large and Z1.p) then yields the final result T(u)/y(u) + W/(1 . . Recall the definition of the auxiliary function y(x) in Section 1. . and distributed as (Y. .T+(2).r+ (n . the duration T+ (n) . z ^ oo. this in principle determines the asymptotic behaviour of r(u). conditionally upon r+ (n) < oo. i. cf. the random vectors (YI.o be defined by w(z) = inf It > 0: Rt = z} where {Rt} is is independent of {St}.. and YI. Zn_1 'typical' which implies that the first n1 ladder segment must be short and the last long. . it therefore follows that T(u)/Z converges in Pi"'')probability to 1/(1 .. Since the conditional distribution of Z is known (viz. Now Bo E S implies that the Bo ")(a) + 0 for any fixed a. denote the ladder epochs and let Yk. Zk be defined similarly as Y = Y1. In the proof.d.e. in particular of Z.3 implies that the P("'1)distribution of T(u) = r(0) is that of w(Z). However. The idea is now to observe that if K(u) = n. Z = ZI but relative to the kth ladder segment.. Then 7(u)/y(u) ^ W/(1 .. 1/(1 .. 5. > u with high probability. ...18(c) Bo")(yY (u)) + P(W > y) ( 5. K(u) = n). 4 Assume that Bo E S and that (5.. then by the subexponential property Yn must be large. Hence Z. 2. Zn).p) in Pi"'')distribution.: r+ (n) < oo. r(u)/Z 4 1/(1 ... P(Z < a I Y > u) 3 0. are i. Since w(z)/z a$. let r+(1) = T(0). It is straightforward that under the conditions of Proposition 1.1) of the last ladder segment can be estimated by the same approach as we used above when n = 1. Z).i..3) holds. and since its dominates the first n . Z1).p). Yn_1 'typical'.p).
u) E •) . I A"(u ))II + 0. +Yn1<u.2. . A"(u) _ {K(u)=n} = {Y1+ P(.5.n) (y1. . > u}. A"(u) are events such that P(A'(u) AA"(u)) = o(F (A'(u)) (A = symmetrical difference of events). then IIP( I A'(u)) Taking A'(u) = {Y..u) II 0.. the condition on A'(u) A A"(u) follows from Bo being subexponential (Proposition 1... II ' II denotes the total variation norm between probability measures and ® product measure. FINITEHORIZON RUIN PROBABILITIES 277 16 Z3 Z1 r+(1) T+(1) T+(1) Figure 5..u) E • I A'(u)) = Bo (n1) ®Bou) .Bo (ri1) ®B( . Yn . . P(. Y„1.5 Ilp(u. Proof We shall use the easily proved fact that if A'(u).Yl+ +Yf1>u}. suitably adapted). Further... Lemma 5. I A'(u)) = P(u.Yn1iYn .3 In the following. P (Yj. .n).
Y1 +.. HEAVY TAILS ((Z1'.P) > y) Corollary 5. . copies of {w(z)}.).. ..6 IIPIu'n ) CHAPTER IX... the density of Yn is B(y)/[IBBO(u)]. Proof Let (Y11.1 and Y„ .. Z11)...u has distribution Bout That is.+y 1 p"F(Yn > u) P)Pn1 P/(1 ... wk(Zk) has a proper limit distribution as u + oo for k < n. be independent random vectors such that the conditional distribution of Zk given Y. .. .. Proof of Theorem 5. Let {wl(z)}. 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 •) .. The same calculation as given above when n = 1 shows then that the marginal distribution of Zn is Bou). Then according to Section 5a..6. Zn are independent. + Y" > u) Flul (K (u ) = n) _ Cu) P"F(1'i +. ..278 Lemma 5 . By Lemma 5. n . whereas wn(Zn) has the same limit behaviour as when n = 1 (cf. . P(u) since by Theorem 2. Similarly (replace u by 0)...2. . and clearly Zi. Notes and references Excursion theory for general Markov processes is a fairly abstract and advanced topic. Z' are arbitrary random vectors. Zn). see Fitzsimmons [144])..P(Y' E •)II * 0. ..P(Z' E •)II > 0 (here Y.. . ..y(u)T) . in our example Y = (Y1.. .t.Bo (n1) ®Bo' 0.1. It therefore suffices to show that the P(u'")distribution of T(u) has the asserted limit. the discussion just before the statement of Theorem 5. in particular his Proposition (2.. 2.p) < y)... {wn(z)} be i.' = y is BM. n. Thus F(u'n)(T(u) /7(u) > y) = F(u'n)((wl (Z1) + .. Y'.d. The first step is to observe that K(u) has a proper limit distribution w. y > u. Y") u etc. n_1 < u.1 P PBo(u) • P(W/(1 .4. Zn) E •) .1).P) Bo(u) for n = 1. +wn(Z n))l7( u ) > 1y) ^' P(u'n)(wn (Zn)/7(u) > y) 4 NW/(1 .. . For Theorem 5. and that Yk has marginal distribution B0 for k = 1.7 O (u.. then 11P(Z E •) . k = 1.r. the F'distribution of r(u) is the same as the P'distribution of w1(Zl) + • • • + wn(Zn).i. Z. (Y..4). the marginal distribution of Zk is Bo for k < n.
(3 u u J B(y) dy . .2 Define M. that MQ becomes large as consequence of one big jump. p(Y) and the result follows. Then 0 (u) Qf "O ^) dy. More precisely.e. Proof of Theorem 6. i. u (6. the results only cover the regularly varying case. one expects the level y form which the big jump occurs to be 0(1). Asmussen & Teugels [53] studied approximations of i (u. 6 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate /3. 3. claim size distribution B. nontrivial and we refer to Asmussen [22]. V. = supo<t<0. The heuristic motivation is the usual in the heavytailed area./3Ea B(u). We will show that the stationary density f (x) of {Vt} satisfies f (x) /B(x) r(x) We then get V. x > oo.B(u). T) when T + oo with u fixed. Then P(MT > u) .6. Assume for simplicity that {Vt} regenerates in state 0 . Theorem 6 .1.. max VB>0I Vo=0^ o<s<t J11JJJ Lemma 6 . Extensions to the Markovmodulated model of Chapter VI are in Asmussen & Hojgaard [33]. cf.1 Assume that B is subexponential and that p(x) > 00. Corollary II. the probability that is exceeds u is then B(u . that fo p(x)1 dx < oo.1) The key step in the proof is the following lemma on the cycle maximum of the associated storage process {Vt}. The rigorous proof is. 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).2.y) . however. and define the cycle as a = inf{t>0: Vt=0. and premium rate p(x) at level x of the reserve. 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 .(u) = P(V > u) = f f (y) dy .
there exist constants c(u) 4 0 such that the limiting distribution of r(u)/c(u) given r(u) < oo is exponential. Hence f (u)r(u) = D(u) = Do(u) . D(u) = DQ(u)/µ. Then D(u) = f(u)p(u) and. u Notes and references The results are from Asmussen [22].q(u) Now just use that p(x) * oo implies q (x) + 0. where also the (easier) case of p(x) having a finite limit is treated . by regenerative process theory. HEAVY TAILS Define D(u) as the steadystate rate of downcrossings of {Vt} of level u and Da (u) as the expected number of downcrossings of level u during a cycle. It is also shown in that paper that typically.P(MT > u) $B(u) Ft µ(1 . .280 CHAPTER IX. 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).q ( u)) 1 .
Fox & Schrage [77]. Hence 1.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. estimating z by the empirical mean (Z1 + • • + ZN)/N and the variance of Z by the empirical variance N s2 = E(Z{  N 2. Rubinstein [310] or Rubinstein & Melamed [311] for more detail . replicates Zl.96s z f (1.d. Ripley [304]. z) 2 = Zit NE ii ii According to standard central limit theory . vrN(z .2) is an asymptotic 95% confidence interval . and this is the form in which the result of the simulation experiment is commonly reported..Chapter X Simulation methodology 1 Generalities This section gives a summary of some basic issues in simulation and Monte Carlo methods . We shall be brief concerning general aspects and refer to standard textbooks like Bratley. where a2 = Var(Z ). ZN.z) 4 N(0.. 281 . . topics of direct relevance for the study of ruin probabilities are treated in more depth.. The crude Monte Carlo ( CMC) method then amounts to simulating i. 4Z).
Letting Z' = E[Z I Y]. and in most cases this modest increase of N is totally unproblematic. we then have EZ = EZ = z. lb Variance reduction techniques The purpose of the techniques we study is to reduce the variance on a CMC estimator Z of z.282 CHAPTER X. it is straightforward to use the CMC method to simulate the finite horizon ruin probability z = i. Z = I inf Rt < 0 (0<t<T = I('r(u) < T). variance reduction is hardly worthwhile. SIMULATION METHODOLOGY In the setting of ruin probabilities. one can argue that unless Var(Z') is considerable smaller than Var(Z). there are others which are widely used in other areas and potentially useful also for ruin probabilities. writing Var(Z) = Var(E [Z I Y]) + E(Var[Z I Y]) . We mention in particular ( regression adjusted) control variates and common random numbers. typically by modifying Z to an alternative estimator Z' with EZ' = EZ = z and (hopefully) Var(Z') < Var(Z). Therefore. Typically variance reduction involves both some theoretical idea (in some cases also a mathematical calculation). an added programming effort. We survey two methods which are used below to study ruin probabilities. v. so that Z' is a candidate for a Monte Carlo estimator of z.b(u. 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. 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. generated at the same time as Z. conditional Monte Carlo and importance sampling. 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). Say that Var(Z') = Var(Z)/2. and many sophisticated ideas have been developed. However. 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'. Further. This is a classical area of the simulation literature. Sections 24 deal with alternative representations of Vi(u) allowing to overcome this difficulty. and a longer CPU time to produce one replication.
it gives a guidance: choose P such that dP/dP is as proportional to Z as possible. but tentatively. Then z Var(LZ) = E(LZ)2 . To this end.[E(LZ)] = E Z2 Zz . even if the optimal change of measure is not practical. .z2 = 0..v. one would try to choose P to make large values of Z more likely. GENERALITIES 283 and ignoring the last term shows that Var(Z') < Var(Z) so that conditional Monte Carlo always leads to variance reduction. the argument cheats because we are simulating since z is not avaliable analytically.. it appears that we have produced an estimator with variance zero. the obvious possibility is to take F and P mutually equivalent and L = dP/dP as the likelihood ratio. Importance sampling The idea is to compute z = EZ by simulating from a probability measure P different from the given probability measure F and having the property that there exists a r. Variance reduction may or may not be obtained: it depends on the choice of the alternative measure P.1. Thus.96 sis v^ N 2 1 where srs = N j(LiZi . L such that z = EZ = E[LZ]. a crucial observation is that there is an optimal choice of P: define P by dP/dP = Z/EZ = Z/z. Thus we cannot compute L = Z/z (further. Nevertheless. However. and the problem is to make an efficient choice. .3).E [Z Z]2 = z2 .zrs. (ZN.zrs) = 2 1 N 2 2 2 i=1 i=1 N > Lt Zi . . This may also be difficult to assess . LN) from P and uses the estimator N zrs = N > L:Zj i=1 and the confidence interval zrs f 1. L1). using the CMC method one generates (Z1.3) Thus. it may often be impossible to describe P in such a way that it is straightforward to simulate from P).e. (1. L = z/Z (the event {Z = 0} is not a concern because P(Z = 0) = 0). i. In order to achieve (1.
The CMC method leads to a variance of oZ = z(1 . However. Again. but if the point estimate z is of the order 105. large sample sizes are required. Another way to illustrate the problem is in terms of the sample size N needed to acquire a given relative precision .1.96 2 z2 z increases like z1 as z . In ruin probability theory.e.0. in terms of the halfwidth of the confidence interval. I.96 2Z ( 1 . To introduce these. 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. a confidence interval of width 10 4 may look small. u + oo.96oz /(zV) = 0.100 . say 10%. Thus. let z(u) = P(A(u)).z) 1001. A = {T(u) < T} or A = {r(u) < oo} and the rare events assumption amount to u being large. assume that the rare event A = A(u) depends on a parameter u (say A = {r(u) < oo}). as is the case of typical interest. This leads to the equation 1. We then . SIMULATION METHODOLOGY 1c Rare events simulation The problem is to estimate z = P(A) when z is small . However.284 CHAPTER X. For each u. and further it is usually not practicable to simulate from P(•IA). assume that the A(u) are rare in the sense that z(u) * 0. An example where this works out nicely is given in Section 3.B = iP(AB) = P(BIA).1. i. Two established efficiency criteria in rare events simulation are bounded relative error and logarithmic efficiency.. z I.5 or even much smaller .e. Z z V5 In other words . the optimal P is the conditional distribution given A. we may try to make P look as much like P(•IA) as possible.e. N .z) 1 > 00.. 10 . it does not help telling whether z is of the magnitude 104. and let Z(u) be a Monte Carlo estimator of z(u). the issue is not so much that the precision is good as that relative precision is bad: oZ z(1 . Z = I(A) and A is a rare event.z) which tends to zero as z ^ 0. say of the order 103 or less. if z is small. The optimal change of measure ( as discussed above) is given by P(B) = E [ Z] i. We shall focuse on importance sampling as a potential (though not the only) way to overcome this problem.
it is appealing to combine with some variance reduction method . let Z +. If M > u.0.log Var(Z(u)) lim inf > 2 u+oo .e. Otherwise. Therefore . with common density bo(x) = B(x)/µB and K is geometric with parameter p. let Z +. the mathematical definition puts certain restrictions on this growth rate. Generate K as geometric. . are i. XK from the density bo(x). it is not efficient for large u . . X2. 2. The algorithm gives a solution to the infinite horizon problem . where M = X1 + • • • + XK.i. see Asmussen & Rubinstein [45] and Heidelberger [190]. 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.1) may be written as V) (u) = P(M > u).p)pk. The term logarithmic comes from the equivalent form . the PollaczeckKhinchine formula III. where Z = I(M > u) may be generated as follows: 1. 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. Let M . O (u) = z = EZ. Thus. which gives a logarithmically efficient estimator .. 2 Simulation via the PollaczeckKhinchine formula For the compound Poisson model. logarithmic efficiency is almost as good as bounded relative error. 3.X1 + + XK. .4) for any e > 0...d. Generate X1.1.log z(u) of (1.. Notes and references For surveys on rare events simulation. This allows Var(Z(u)) to decrease slightly slower than z(u)2. and in practice. so that NE (u) may go to infinity.p)pk. i.2. this means that the sample size N = NE(u) required to obtain a given fixed relative precision (say a =10%) remains bounded.4). but as a CMC method . where X1. However. F(K = k) = (1 . We shall here present an algorithm developed by Asmussen & Binswanger [ 271. Var(Z(u)) hm sup U+00 z (u) 2E < oo (1.(2. According to the above discussion. P(K = k) = (1 .
Then (cf.2. < X(K) throw away the largest one X(K). and considering only the remaining ones.p)Bo(x).p/(l . we generate only X1. Z(1) (u) has a smaller variance than Zl (x).X(2).. form the order statistics X(1) < X(2) < . X1 + + XK_ 1 > x when X1 > x...X1 . This calculation shows that the reason that this algorithm does not work well is that the probability of one single Xi to become large is too big.. and let Z(2)(u) = _ P (SK B0((u > u I X(l). Thus. However... To see this. .X(n_1)) Bo(X(„_l) V X) Bo(X(n1)) . Theorem IX. XK1.. assume in the following that Bo(x) .... So.. we thus generate K and X1i . y < 0).S( K_1)) V X(K1)) / Bo(X(K 1)) where S(K_l) = X(1) + X(2) + • • • + X(K_1).... and that Bo(y) = 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. Z(1)(u) is defined as 0). asymptotically it presents no improvement : the variance is of the same order of magnitude F(x).X(2)..XK1] = EBo(uX1 .X(K1)) .. just note that EZ(1)(u ) 2 > E[Bo (x .. .Xl .L(x)/x`' with a > 0 and L(x) slowly varying... conditional probability.286 CHAPTER X.XK_1).. Xl > x. .. For the simulation.. XK. .. SIMULATION METHODOLOGY when the claim size distribution B (and hence Bo) has a regularly varying tail. compute Y = u . A first obvious idea is to use conditional Monte Carlo: write i. note first that To check the formula for the P(X(n) > x I X(1).SK1)2.+XK > uIXl..b(u) = P (Xl +•••+XK>u) = EF[Xl + .1) V)(u) .XK_1 and let Z( 1)(u) = Bo (Y) (if K = 0.. The idea of [27] is to avoid this problem by discarding the largest X.. As a conditional Monte Carlo estimator ... K > 2] = P2p(Xl > x) = P2Bo(x) (here we used that by positivity of the X.
. l)) BO(X(n1)) Theorem 2 . . using 13L. X . it must be noted that a main restriction of both algorithms is that they are so intimately tied up with the compound Poisson model because the explicit form of the PollaczeckKhinchine formula is crucial (say.1) . X(2). .. and simulate from FL. the algorithm for generating Z = Z(u) is: 1. Asmussen . l)) . X(n1)) P(X(TZ) + S(. IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 287 We then get P(S" > x I X( 1). the continuoustime process {St} is simulated by considering it at the discrete epochs {Qk} corresponding to claim arrivals.S(n_1) I X(1). Then the algorithm given by { Z (2) (u) } is logarithmically efficient. and that paper contains one more logarithmically efficient algorithm for the compound Poisson model using the Pollaczeck. and we refer to [27]. Compute y > 0 as solution of the Lundberg equation 0 = K(y) = )3(B[y] .u is the representation 0(u) = e7sr(u) overshoot (cf. 3 Importance sampling via Lundberg conjugation We consider again the compound Poisson model and assume the conditions of Ce7". X(2).. BL instead of 0. 111. For practical purposes. Binswanger and HOjgaard of [28] give a general survey of rare events simulation for heavy tailed distributions . Thus. 1 Assume that Bo (x) = L(x)/x° with L(x) slowly varying. . for the purpose of recording Z(u) = erysr(u). BL(dx) = e7sB(dx)/B[y].5).S (n1)) V X (. The algorithm is sofar the only efficient one which has been developed for the heavytailed case. B.3. and define )3L. in the renewal or Markov. . Notes and references The proof of Theorem 2. However .modulated model P(r+ < oo) and G+ are not explicit ). X(n1)) Bo((x ...Khinchine formula and importance sampling . BL by I3L = /3B[y]. . Also in other respects the findings of [28] are quite negative: the large deviations ideas which are the main approach to rare events simulation in the lighttailed case do not seem to work for heavy tails.1 is elementary but lengty. use the the CramerLundberg approximation so that z(u) = '(u) = e7"ELe7E(") where ^(u) = ST(") . . X (.y. X(2). that is.._1) > P(X(n) > _ X X(1).
Xi = U.i.3. P'[y] < oo for some ry > 0. SIMULATION METHODOLOGY 3.r(u) < oo) = 1. Let FL (dx) = 'For the renewal model. one must restrict attention to the case 4µB > 1. b different from . let S..2 The estimator (3.1) (simulated with parameters ^3. . = X1 + . and avoid simulating the known part e7". If S > u.. and the change of measure F r FL corresponds to B > BL.. with distribution F. The algorithm generalizes easily to the renewal model . BL). so that changing the measure to FL is close to the optimal scheme for importance sampling . Let Sf0 CHAPTER X. We may expect a small variance since we have used our knowledge of the form of 0(u) to isolate what is really unknown.S+U . Ti.T.2ryu _ z (u)2/C2. . More precisely. to deal with the infinite horizon problem . It resolves the infinite horizon problem since FL(.7 tell that P(. B) is not logarithmically efficient when (/3. and assume that µF < 0 and that F[y] = 1. Otherwise. BL could improve the variance of the estimator . We formulate this in a slightly more general random walk setting '. be i..F. the results of IV. r(u) < oo) and FL (both measures restricted to.. and we have: Theorem 3. u It is tempting to ask whether choosing importance sampling parameters .l3 and U from B.d. In fact: Theorem 3.Q. the discussion at the end of Section 1b. 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. The proof is given below as a corollary to Theorem 3. + X. M(u) = inf {n : S„ > u}. Let X1.. Generate T as being exponential with parameter . The answer is no. cf..288 2.1 The estimator Z(u) = e'rs* "u) (simulated from FL) has bounded relative error. Let S .(u)) are asymptotically coincide on {r(u)} < oo. let Z F e_'s. 4. In detail . X2.QL. namely ELery£("). b) # (/3L. return to 3. Proof Just note that EZ(u)2 < e . There are various intuitive reasons that this should be a good algorithm. A > AL as in Chapter V.
where e' = EL Iog dFL (Xi) > 0 by the information inequality. F(XM(u)). let F be an importance sampling distribution equivalent to F and M(u) dF Z(u) _ I (Xi) . it thus follows that for 0 < e < e'/ELXi. Here ELK. Since K1.. . EFZ(u)2 EFZ(u)2 lim sup z(u)2eeU = lim cop C2e2.d.. By the chain rule for RadonNikodym derivatives.2) dF Theorem 3.2) (simulated with distribution F of the X3 has bounded relative error when .i..3 The estimator (3. is not logarithmically efficient.. Since ELM(u)/u + 1/ELXi.+KM(u)}. 1.2'X1 . K2.3. where Kl og (X) (j) 2 ) = log dFL (Xi) .2 > 0. Jensen's inequality and Wald's identity yield EpZ(u)2 > exp {EL(K1 + .. EFZ(u)2 = EeW2(FIF) = Ep [W2(FIFL)W2(FLIF)] = EL [W2 ( FIFL)w(FLIF)] = ELexp {Kl+.. + KM(u))} = exp {ELM(u)(E . are i.2ryELXi.. When F # FL. More generally. write W(F IF) _ F(XI).2ryELXi)} .yu = G. IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 289 e7yF(dx).P = FL.. For the second.yu+elu u +oo etry' 1 > lim up C2e2. Proof The first statement is proved exactly as Theorem 3 . .. = c'. The importance sampling estimator is then Z( u) = e'rSM( ). (3.
The optimality result Theorem 3. U' .3. The queueing literature on related algorithms is extensive .e.T". /3'eQ'YR'( x + y) dy = .4 indicate that we can expect a major difference according to whether y < 1/r. This immediately yields / = 3". The extension to the Markovian environment model is straightforward and was suggested in Asmussen [ 16]. see e. U". with the present (shorter and more elementary) proof taken from Asmussen & Rubinstein [45]. B" and generic claim sizes U'. from 3' P(U'T'>x) ^ = ^ eQ'zB (z) dz. we write T = yu. CHAPTER X. then the estimator Z(u) = e7Sr(°)I(r(u) < yu) (simulated with parameters /3L.1 is from Lehtonen & Nyrhinen [244].T' has a left exponential tail with rate /3' and 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 . BL) has bounded relative error.B'=B". we conclude by differentiation that Bo(x)=B' (x)forallx > 0. optimality is discussed in a heavy traffic limit y 10 rather than when u + oo. First by the memoryless distribution of the exponential distribution .T".4.g. In fact: Proposition 4. Then according to Theorem 3. so that one would expect the change of measure F 4 FL to produce close to optimal results. yu) is close to zk(u).T' D U" . In [13]. U' . generic interarrival times T' .3"eQ x 0 J eQ zB (z) dz x (x > 0) and /3' = /3".i. the references in Asmussen & Rubinstein [45] and Heidelberger [190]. T) with T < oo. The easy case is y > 1/k'(y) where O(u.2.T' = U" . .1 If y > 1/ic'('y). The results of IV.T" > x) J /3"e0 yB (x + y) dy = . Further discussion is in Lehtonen & Nyrhinen [245].290 which completes the proof. /3". T".T" has a left exponential tail with rate /3'. claim size distributions B'. As in IV. Next. all that needs to be shown is that if U' .'(y). then /3' B' = B".'(y) or y > 1/r. SIMULATION METHODOLOGY u Proof of Theorem 3.3'eO'x f f P (U" . 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'.
1) which is all that is needed here can be showed much easier .2) Since the definition of ay is equivalent to Eay r(u) .(u) * 1 (Theorem IV.2 The estimator (4. Remark 4 . Let Qy2 = .'(7). T(u) < yu] . that ryy = ay . the result follows as in the proof of Theorem 3. Bounding u ELZ(u)2 above by a7u. Further .5) follows.1.O(u.log 4')u) 4 u (Theorem IV. We next consider the case y < 1/r. yu) = eayu Eay Leay^(u)+r(u)K(ay). and in fact.1).log Var(Z(u)) l im of .8).yy> 2 . we have ic(ay ) > 0 and get Eay Z(u)2 = Eay [e  2aySr( u)+2r(u )r. The corresponding estimator is Z(u) = eavS' ( u)+T(u)K (ay)I(T( u) < yu). We recall that ay is defined as the solution of a'(a) = 1/y. yu) in the sense that .log Var(Z(u)) _ . .1) so that z(u) = zP(u. Proof Since ryy > y. lim inf uoo 27yu .4. yu) is of order of magnitude a71.8 has a stronger conclusion than (4.3) (simulated with parameters /gay. T( u) < yu] e2ryyuEay le 2ay^(u). one would expect that the change of measure F Pay is in some sense optimal.yu.(ay). 7y (4.4. T(u) < yu] e Hence by (4.4.3) and we have: Theorem 4.1). 3 Theorem IV. Bay) is logarithmically efficient. (4. IMPORTANCE SAMPLING FOR THE FINITE HORIZON CASE 291 Proof The assumption y > 1/n'(y) ensures that 1fi(u. and that ryy > ry.4.to g x ( u ) u u so that (1. (4. yu)/z.yk(ay) determines the order of magnitude of z'(u.
0 rather than when u 3 oo.3.o .o.4).7y x(u) > hm inf u+Oo U . there exists a dual process { V t} such that i.1) where the identity for Vi(u) requires that Vt has a limit in distribution V. N(0.3: for many risk processes {Rt }.1) may be useful. Then z(u) = Eay Z(u) > Eay avS'(u)+T( u)k(av 1 ). yu .(av)Eav l e. (5.yu)/(uyu1/2) .a vt(u).yu1/2 <1 T(u) < yu l r > e7vu +avul/ 2r.T) = P O<t<T inf Rt < 0 = P(VT > u).1) is used to study Voo by simulating {Rt} (for example. However.uaoo U That lim sup < follows similarly but easier as when estimating En. In most of the simulation literature (say in queueing applications).Qyu1/2 < T(u) C yu e. SIMULATION METHODOLOGY Vara„ (r(u))/u so that (T(u) .ryyu +oy u1/2K'(av)Eo l v 1/2) where the last step follows by Stam's lemma (Proposition IV.292 CHAPTER X.b(u. the algorithm in Section 3 produces simulation estimates for the tail P(W > u) of the GI/G/1 waiting time W). zi(u) = INV.a yu +l/ur' (av)Ei`av reav^(u)+(T(++)(U) yu . yu . Hence lira inf log ryyu + vyu 1/2 tc(ay) . One main example is {Vt} being regenerative (see A.1): then by Proposition A1. Z (u)2 above. related discussion is given in a heavy traffic limit q J. 5 Regenerative simulation Our starting point is the duality representations in 11.1) (see Proposition IV. In Asmussen [13].2). > u) = E f I(VV > u) dt 0 (5.4. and (5. 0 Notes and references The results of the present section are new. we believe that there are examples also in risk theory where (5.2) . the object of interest is {Vt} rather than {Rt}. '%(u) = P I info Rt < 0) = P(VV > u)..4.u1/2 < r(u) < yu Le ] l = e.
. .. 02) (5.. 2. To derive confidence intervals . REGENERATIVE SIMULATION 293 where w is the generic cycle for {Vt}. Then Z(1).. a standard transformation technique (sometimes called the delta method) yields 1 V 2 (h (Zi.. EZ2'i = z2 = E Thus. For the ith cycle.h (zl. Taking h(zl. Z2 = N (X21' + . is the cycle length.. )). i (^(u) .3) . consider first the case of independent cycles . Z1 = (Zl1i +. Z2 . + Z1N>) . (u) ?2 = E fo I(Vt > u) dt = 0( u ) zl Ew as N > oo.5. Z2 a4* z2. Thus..t(u)) 4 N(0. let E denote the 2 x 2 covariance matrix of Z('). j = 1.d. The method of regenerative simulation.. Simulate a zerodelayed version of {V t } until a large number N of cycles have been completed. z2)) > N(O. which we survey below . letting J0 'o I (Vt > u) dt .. oh) for h : R2 ^ R and Ch = VhEVh.+Z(N) z 1. Thus the method provides one answer on to how to avoid to simulate { Rt} for an infinitely long time period.. EZ1'i = z1 = Ew. record Zi'i = (Z1'). Zl the LLN yields Z1 a$' Z(1) +. i. Z2'> the time during the cycle where { Vt} exceeds u and zj = EZJ'). + Z2N)) . Vh = (8h/8z1 8h/ 8z2).. Then (Z1z1i Z2z2 ) 4 N2(0. and Z2'>) where Zi'i = w. . Therefore . provides estimates for F ( V. Z(N) are i . z2) z2/z1 yields Vh = (z2/z2 1/zl). For details .E). the regenerative estimator z%(u) is consistent. > u) (and more general expectations Eg(V.
SIMULATION METHODOLOGY 2 Eli = Z2 z1 + 2 E22 . say risk processes with a complicated structure of the point process of claim arrivals and heavy tailed claims . Z of the form Z = ^p(X) where X is a r . in some situations it may be the only one resolving the infinite horizon problem . asymptotic estimates were derived using the renewal equation for z /i(u).g S12 (5.z^ i=1 so a2 can be estimated by 2 2 = 72 S11+ 12 S22 .9.294 where 01 2 CHAPTER X. However . Before going into the complications of ruin probabilities .5) Z1 Z1 Z1 and the 95% confidence interval is z1 (u) ± 1. In 111.g. see e. There is potential also for combining with some variance reduction method.2. We here consider simulation algorithms which have the potential of applying to substantially more complex situations. Let X have a density f (x. v. to evaluate the sensitivity z/i( (u ) = (d/d() 0(u) where ( is some parameter governing the risk process . with distribution depending on a parameter (. 9. 6 Sensitivity analysis We return to the problem of 111 . the expectation z = EZ of a single r. Rubinstein [310] and Rubinstein & Melamed [311]. Here are the ideas of the two main appfoaches in today 's simulation literature: The score function ( SF) method .96s/v"N.Z) ^Z(=) . Notes and references The literature on regenerative simulation is extensive. () dx f Ax) (dl d()f (x' () f ( z.2 E1 2 z1 z1 Z2 The natural estimator for E is the empirical covariance matrix N S = N 1 12 (ZW .v. () depending on C. Then z(() = f cp(x) f (x. () dx = E[SZ] f(X.C)dx = f w(x) d( f ( x. consider an extremely simple example . () dx so that differentiation yields zS d( fco(x)f(x.0 . The regenerative method is not likely to be efficient for large u but rather a brute force one.
A related difficulty occurs in situations involving the Poisson number Nt of claims: also here the sample path derivative w. SZ is an unbiased Monte Carlo estimator of z(. nonpathological examples where sample path derivatives fail to produce estimators with the correct expectation. I(r(u) . ()) is 0 for C < Co and 1 for C > Co so that the sample path derivative cp'(h(U. () = log U/(2. with density f (x. The following example demonstrates how the SF method handles this situation. So assume that a r.v. ()) is 0 w . ()) h((U. for some Co = (o(U). if f (x. the Poisson rate /3 in the compound Poisson model. For the SF method. 11 /3oeOoT. one. () = (8/8()h (u. Then z(() = Ecp(h(U. C)). r(u) = Tl + • • • +TM(u)). IPA will estimate zS by 0 which is obviously not correct.log U/(.1 Consider the sensitivity tka(u) w. () can be generated as h(U. The likelihood ratio up to r(u) for two Poisson processes with rates /3. /3 is 0. cp(h(U. p. /3o is M(u) Oe (3T: < oo) . For IPA there are. In the setting of ruin probabilities . one can take h (U.6. () is an unbiased Monte Carlo estimator of zS. () _ (eSx. Thus . () where U is uniform(0. For example . () = d log f (X. this phenomenon is particularly unpleasant since indicators occur widely in the CMC estimators . SENSITIVITY ANALYSIS where 295 S = (d/d()f (X. C) f(X. The derivations of these two estimators is heuristic in that both use an interchange of expectation and differentiation that needs to be justified.r. Then . just take cp as an indicator function . To see this. Example 6 . () is increasing in C. Thus. () Thus. = E [`d (h(U. Infinitesimal perturbation analysis (IPA) uses sample path derivatives.() d( is the score function familiar from statistics . ( where h( (u.t. say W(x) = I(x > xo) and assume that h(U.r. this is usually unproblematic and involves some application of dominated convergence . however . zc = E [d co(h(U. ()) d( hc(U. cp' (h(U.1). Let M(u) be the number of claims up to the time r(u) of ruin (thus. () = . C). giving h( (U.t. ()).
t.0(1) so that in fact the estimator Zf(u) has bounded relative error. Thus. To resolve the infinite horizon problem .3 (u) is of the order of magnitude ue7u. However. SIMULATION METHODOLOGY Taking expectation. the risk process should be simulated with parameters .T(u)) I(T(u) < co) ] .1 is from Asmussen & Rubinstein [46] who also work out a number of similar sensitivity estimators. differentiating w. different parameters. change the measure to FL as when simulating tp(u). a relevant reference is VazquezAbad [374]. 4) that V5. There have been much work on resolving the difficulties associated with IPA pointed out above.t.3L. we get 1 M(u) 00(u) = E (_Ti)I(T(U)<) E [(M(u) . j3 and letting flo = 0.r. BL). Example 6.r. since ELZp(u)2 < (M(U) _T(u) \ 1 2 a2ryu = O(u2)e27u. in part for different measures of risk than ruin probabilities. ) we have VarL(ZQ(u)) ZO(u)2 O(u2)e2 u2e2ryu yu . We recall (Proposition 111.9 .296 CHAPTER X. for different models and for the sensitivities w. the estimation of z(ip(u) is subject to the same problem concerning relative precision as in rare events simulation .3 (u) (to generate Zp (u). 0 Notes and references A survey of IPA and references is given by Glasserman [161] (see also Suri [358] for a tutorial). whereas for the SF method we refer to Rubinstein & Shapiro [312]. In the setting of ruin probabilities. We then arrive at the estimator ZZ(u) = (M(u) .T(u)) e7uerVu) for ?P. .
X2.. '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 ). either this makes no difference (P(R. in the Bernoulli random walk example below.g.P(•r(u. a) = r(u) A T+(a). as e.(u) = 0 ) = 0) or it is trivial to translate from one setup to the other. are i. where X1.. wherel T(u) = inf {t > 0 : Rt < 0} . That is. T+(a) = inf It > 0 : Rt > al.. Y'a(U) = P(T (u) = r+(a)) = 1 . a) = r(u))..(u) is defined as the probability of being ruined (starting from u) before the reserve reaches level a > u... }). defined as Ro = u (with u E {0.. R„ = u+X.. The twobarrier ruin problem The twobarrier ruin probability 0. in most cases .i.+• • •+X. T(u.Chapter XI Miscellaneous topics 1 The ruin problem for Bernoulli random walk and Brownian motion.. 297 . .1}valued .d.. Consider first a Bernoulli random walk. Oa(U ) can also be a useful vehicle for computing t/i(u) by letting a * oo.1. and {1. Besides its intrinsic interest . with P(Xk = 1) = 9.
1.+Xn) F[ a]n n=0.o)'t/1a(a . and insertion shows that ( 1.2) Oa(a . = (1 ..o» = z°P (RT ( u.. and the other more advanced but applicable also in some other settings..1) is solution. = z°Va(u) + za(1  .. then 'Oa(u) _ au a We give two proofs .0)/0. The Lundberg equation becomes 1=F[ry]=(19)+9z..(1B)u oJ 0. the solution of F[.e. tba(2) _ (1 .2).(4.y] = 1. zu = EzRO = EzRT(u.a) = 0) + zap ( R. (1.. i.a(u)). 7/la(a .1 For a Bernoulli random walk with 0 0 1/2.a) Y.298 CHAPTER XI. MISCELLANEOUS TOPICS Proposition 1. z and the solution is z = (1 .(u) I\ e = 1 oa ' ()i a = u. By optional stopping. We choose a = ry where ry is the Lundberg exponent. C1_0\a.o)T/la (1) + 8z/'u(3). The martingale is then {zuzXl+•••+X„ } = {zR° }.. and in view of the discrete nature of a Bernoulli random walk we write z = e7..1) = (19)4/'0(a3)+9ba(a1).r(u. Proof 1.4) by ea(u+Xl+. u Proof 2. Conditioning upon X1 yields immediately the recursion 'a(1) = 19+00a(2). In a general random walk setting . u + 1.. where a is any number such that Ee°X = F[a] <oo.1) o If 0 = 1/ 2. Wald's exponential martingale is defined as in 11. one elementary but difficult to generalize to other models.
u)/u.1 If p = 0.1) for p # 0. thenz1 (u)=1. then Proof Since 'Oa (U)  au a Eea(R°. {R.1.1).2) is trivial (z = 1). pa( u) _ u Corollary 1.2 For a Bernoulli random walk with 9 > 1/2. However. 1h (u) = a el u \1 If 9 < 1/ 2.zu)/(za .0a(u)). BROWNIAN MOTION. then Vi(u) = 1. Proof Let a+ oo in (1.3 Let {Rt} be Brownian motion starting from u and with drift p and unit variance . TWO BARRIERS 299 and solving for 4/la(u) yields t/ia(u) = (za . Applying optional stopping to the exponential martingale {e7R. If p = 0.4 For a Brownian motion with drift u > 0.5) .1 yields 't/la(u) = (a . u Proposition 1. If 9 = 1/2.• a2µa e2µu . (1. } yields e7u = Ee7R° = e°Wa(u) + e7a(1 .} is then itself a martingale and we get in a similar manner u = ER° = ER ra( u) = 0 • Y'a (u) + all  au Y'a( u)). Then for p 0 0.e7u)/(e7° . RANDOM WALK. i1(u) = e211 . and solving for 9/la(u) yields Z/)a(u) = (e 76 .u) = et(a2 /2 +aµ) the Lundberg equation is rye/2'yp = 0 with solution y = 2p.. If p<0. {Rt} is itself a martingale and just the same calculation as in the u proof of Proposition 1. Corollary 1.ba(u) = e2µa . (1.1). .
VIII.5 Consider the compound Poisson model with exponential claims (with rate.vi(a) Proof By the upwards skipfree property.0(a) 0 < u < a.7/la(u)). However. 1 .a ) < 0) + e 7aF ( R (u. Ic 5ry 'pa(u) Using y = 6 ..616).0 (u) (where u p =. 5).3. a) I R(u a ) < 0] P (R(u . Here is one more case where this is feasible: Example 1.5a). 7O(u) = 7/la(u) + (1 . however.e7a Again .4). and thus one encounters the problem of controlling the undershoot under level 0. letting a * oo yields the standard expression pe7u for . valid if p < 1 (otherwise .+^a(u))^(a) If 7k(a) < 1. and hence e7u = Ee7Ro E [e7R(.300 Proof Let a * oo in (1.6 If the paths of {Rt} are upwards skipfree and 7//(a) < 1. passing to even more general cases the method quickly becomes unfeasible (see.a) = a) = 5 y = P (R (u. this immediately yields (1. It may then be easier to first compute the onebarrier ruin probability O(u): Proposition 1. MISCELLANEOUS TOPICS u The reason that the calculations work out so smoothly for Bernoulli random walks and Brownian motion is the skipfree nature of the paths.e7a (u) = 6 /0 .7).a) = a on {r (u. (u) _ O(u) . Here the undershoot under 0 is exponential with rate 5.a) < 0) + e7°P (R(u. the paths are upwards skipfree but not downwards. For most standard risk processes . say.a) = r+ (a)} and similarly for the boundary 0. we obtain 'Oa a7u . 0.a) = a ) + e ' ° ( 1 . implying R(u.7) . CHAPTER XI. 7/'(u) = 1). . (1.
)_ _( u)µ2 /2. TWO BARRIERS 301 Note thas this argument has already been used in VII.Rt}.8 ). For µ # 0.8 Let {Rt} be Brownian motion with drift .r(u).4) I = . the density dPµ / dP0 of St is eµsttµ2/2. in particular symmetric so that from time r(u) (where the level is level u) it is equally likely to go to levels < u and levels > u in time T . = eµuTµ2/2Po (T( u) E dT) 2 eµuTµ2/2 u T3/2 ex p u 27r p 12 T .1. and (1 ..1a for computing ruin probabilities for a twostep premium function.2 . + µ2T) } .T) P(MT > u) where MT = maxo<t<T St. we have ili(u.. For the symmetric (drift 0) case these are easily computable by means of the reflection principle: Proposition 1.ST<u) = P(MT>u. MT > u) = P (ST > u) + P (ST > u.9) = 2P(ST > u). P(MT > u) = P(ST > u) + P(ST < u.11) VIT ) Proof For p = 0. = 1 . (1. MT > U) = P(ST > u) + P(ST > u) (1. and hence Pµ('r(u) E dT) = Eo [e µsr(.f.µ so that {St} is Brownian motion with drift µ .. ( 1. Hence P(MT>u. 0(u. Corollary 1.µ%T (1.10) Pµ (T(u) < T) !..ST>U).7 For Brownian motion with drift 0. BROWNIAN MOTION. Then the density and c. 10) follows then by straightforward differentiation. Here {St } is Brownian motion with drift 0 (starting from 0). We now return to Bernoulli random walk and Brownian motion to consider finite horizon ruin probabilities.. (i). (1. T(u) E dT.11 ) is the same as (1. T ) = P(T(u) < T ) = 241.µ T I + e2µ"4) ( .d.8) Proof In terms of the claim surplus process { St} = {u . of r(u) are ( U2 Pµ (T(u ) E dT) = 2^T 3/2 exp µu . RANDOM WALK.
is zero for all u > 0 but that nevertheless Rt ^4 0 (the problem leads into the complicated area of boundary classification of diffusions. is finite for all x > 0. Thus. close to x {Rt} behaves as Brownian motion with drift µ = u(x) and variance a2 = a2(x).10). The expression for F ( ST = v) is just a standard formula for the u binomial distribution.3 we can define the local adjustment coefficient y(x) as the one 2µ(x)/a2(x) for the locally approximating Brownian motion. 0 0 (1.T (1. as defined above as the probability of actually hitting 0. such that the drift µ(x) and the variance a2(x) are continuous functions of x and that a2(x) > 0 . is (1. Let s(y) = ef0 ry(.12) is the same as ( 1.12) P(ST = v) = 0 otherwise. and in a similar spirit as in VII.9).T)dx.g.. We finally consider a general diffusion {Rt} on [0. We assume that u(x) and a2 (x) are continuous with a2 (x) > 0 for x > 0.10 Consider a diffusion process {Rt} on [0.. e. Vi(u.g.T) = P(ST = u) + 2P (ST > u). T are integervalued and nonnegative. oo).h.9 For Bernoulli random walk with 9 = 1/2.11) then follows by checking that the derivative of the r.302 CHAPTER XI. The same argument as used for Corollary 1. S(x) = f x s(y)dy.s. Proof The argument leading to ( 1. If this assumption fails. that 0(u).T+2. as defined in (1. (1. u Small modifications also apply to Bernoulli random walks: Proposition 1. MISCELLANEOUS TOPICS which is the same as (1. Here {2T( (v}TT)/2) v=T.10) and that the value at 0 is 0. whenever u. see e. 226). S(oo) = f c s(y)dy.9) goes through unchanged.8 also applies to the case 9 54 1/2..T2. and (1. oo) with drift µ(x) and variance a2 (x) at x. but we omit the details. Theorem 1.13) The following results gives a complete solution of the ruin problem for the diffusion subject to the assumption that S(x).13) with 0 as lower limit of integration. Breiman [78] or Karlin & Taylor [222] p. the behaviour at the boundary 0 is more complicated and it may happen..
11 Let 0 < b < u < a and let t&0.b(u) + L.b(u) be the probability that {Rt} hits b before a starting from u. TWO BARRIERS 303 for x > 0. elementary calculus shows that we can rewrite L as Lq(u) d 1a2 (u)s(u)d [ s (u) ? ] .6 to Markovmodulated models .b(Rdt). Letting a T oo and considering the cases S(oo) = oo.17) Hence L. Further references on twobarrier ruin problems include Dickson & Gray [116]. 0 Proof of Theorem 1. then 0 < 2l.S(u) (1. so that Y)n. see in particular pp.b('u) = Eu .1. A good introduction to diffusions is in Karlin & Taylor [222].b = 0 implies that VQ b/s is constant.S(b) Proof Recall that under mild conditions on q.e LVa.16). the function S(x) is . E„ q(Rdt) = q(u)+Lq(u)dt. RANDOM WALK.b(u) = S(a) . [117]. For generalizations of Proposition 1.13) is finite for all x > 0.10.b(b) = 1. If (1.S(u)/S(a). if (1.b (Rdt) = Oa. 0 in (1. Using s'/ s = 2p/a2.b('u) = Eu &0.10. where Lq(u) = 0'22u) q "(u) + p(u)q(u) is the differential operator associated with the diffusion.14) S(oo) < 00.14) fails. Notes and references All material of the present section is standard. Lemma 1. Letting b J. . Assume further that S (x) as defined in (1.16) yields 4b (u) = 1 .ba. BROWNIAN MOTION.0(u) = 1 for all u > 0.e. then. Wa.16) S(a) .b(u)dt. Then YIa. 15) i. The obvious boundary conditions '0a. i. see Asmussen & Perry [42]. 191195 for material related to Theorem 1. we can ignore the possibility of ruin or hitting the upper barrier a before dt.ba. (1 . b = 0. S(oo) < oo separately u completes the proof. In view of (1.b(a) = 0 then yield the result. O.b = a+/3S. (1. If b < u < a.(u) < 1 for all u > 0 and ^ S^ Conversely. A classical reference for further aspects of Bernoulli random walks is Feller [142]. and we get Wo. 1'.
1.5) A martingale proof of (2. yielding eau = Ee.17).1) (2. 111 .6. IV.ytc (ay). (2. See Asmussen [20] and Rogers [305] for some recent treatments and references to the vast literature. Markovmodulated Brownian models .(a) (2.6) .5..4. Another basic quantity is the speed measure M . Lo I. yu) '+/1(u) . They all use the fact that ( tx(a) l ( eaRt = eau + aSttx(a) < e7yu. equivalently. yu) where W (ay) = y. is currently an extremely active area of research. one works instead with a lower limit 5 > 0 of integration in (1.9 ) and optional stopping applied to the stopping time r(u) A T.2) C_e7u < t(u) < C+e _7u. variance 0.2.(. y > . 7y = ay .3.4) I. 1 y < k (y).5): _ z/'(u) < e 7u. much of the literature dels with the pure drift case. (2.)AT .3) < e 7yu.aRo .13)).4. but by duality. 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.(7) .aR. The emphasis is often on stationary distributions . MISCELLANEOUS TOPICS referred to as the natural scale in the general theory of diffusions (in case of integrability problems at 0. Remark 11. with the drift and the variance depending on an underlying Markov process .304 CHAPTER XI. which is motivated from the study of modern ATM (asynchronous transfer mode ) technology in telecommunications. and here are alternative martingale proofs of the rest .1 ) was given already in II. (2. defined by the density 1/va(u)s(u) showing up in (1. Lo is a martingale (cf. (2. correponding to piecewise linear paths or .(T(u)AT) r. where C_ = B(x) _ B(x) sup 2no fy° e7(Y )B(dy)' f2e7(Y2)B(dy)' C+ i/i(u.o•K(a) = Ee . information on ruin probabilities can be obtained .t&(u.
d.E [e.1. . dr JO Zoo ) f e7'B(r + dy) B(r) Jo ^00 ^00 H(dt. u Proof of (2.4): We take a = ay in (2. dr) 1 = 1 I0 /o C+ C+ From this the upper inequality follows. Proof of ( 2. For (2. eyuk (ay) = e7yu e > eyu"(ay ) ij(u.yu) Y Similarly for (2. y > r. we have tc(ay) > 0 and we can bound (2.6)..3).4). A claim leading to ruin at time t has c.2. it follows easily from (2.(u.3).1 .7R. we have ic(ay ) < 0 and use the lower bound E [e7Rr („).T(u)K(ay) I yu < r(u) < T] F(yu < r(u) < T) > e. dr) e 7( yr)B(dy) B(r) f oo o 0 r > H(dt. Equivalently.T)  V. Hence E [e7Rr (u) Jr(u) < ool ^00 H( dt.yuk (ay)(u&(u.)r(u)r.2): As noted in Proposition II. (2. yu))• Letting T + oo yield e_ayu > eyur4ay)(0(u)  Notes and references See II.(ay)I T(u) < yu] P(r(u) < yu) (using RT(u) < 0). 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).f. (B(y) . dr) denote the conditional distribution of (T(u).( u ) I T(U) < 00] .1.B(r))/B(r). Rt has distribution B(r + dy)/B(r). when Rt_ = r. and the proof of the lower inequality is similar. RT(u)_) given r(u) < oo.6) with = 'y that eyu . Let H(dt.6) below by 1 E Le7Rr(. so that i/1(uL yu) < eayu .yu))• b(u.
then P C S. ri. gn 4 0. cle . Thus .306 CHAPTER XI. og For sequences fn.1). we will write fn 1.the correct sharp asymptotics might as well have +. gn with fn + 0 ./n E I) for intervals I C R. large deviations results been. .1) is an example of sharp asymptotics : . and gave sharp asymptotics for probabilities of the form P (S. if x > EX1. Cramer considered a random walk Sn = X1 + .1 We will go into some more detail concerning (3. 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.. its generality. The classical result in the area is Cramer's theorem. e.1) does not capture the \ in (3. large deviations results have usually a weaker form. Example 3. logarithmic asymptotics .2) can be rewritten as F (Sn/n > x) 1g a'fin..2).g. The last decades have seen a boom in the area and a considerable body of applications in queueing theory. in being capable of treating many models beyond simple random walks which are not easily treated by other models . Thus.1) amounts to the weaker statement lim 1 log P I Sn > x I = 17. and that a considerable body of theory has been developed. ..nn or C2e.(B) = log EeOX 1 is defined for sufficiently many 0. the parameter will be u rather than n).^ e nn 1 > x n 0o 2xn (3. However . For example. 1) but only the dominant exponential term . (3.means (as at other places in the book) that the ratio is one in the limit (here n * oo).gn if nioo lim 109 fn = 1 log gn (later in this section. which in the setting of (3. The limit result (3. nroo n n /// Note in particular that (3. not quite so much in insurance risk. Accordingly.3na with a < 1.1) where we return to the values of 0. (3. such that the cumulant generating function r. + X. The advantage of the large deviations approach is. v2 later. however . logarithmic asymptotics is usually much easier to derive than sharp asymptotics but also less informative ..?n typically only give the dominant term in an asymptotic expression .
rc(0) where 0 = 0(x) is the solution of x = rc'(0).9S„+n' ( 9). S rtn > x 1.the mean rc'(0) of the distribution of X1 exponentially tilted with 0.96o /] > 0. of P(X1 E dx) = E[e9X1K.960/) * 0. rc*(x) = sup(Ox .r.r. 2 where o2 = o2(x) = rc"(0). if we replace Sn by nx + o / V where V is N(0. V > 0 e.(0)) e 307 (other names are the entropy. LARGE DEVIATIONS Define rc* as the convex conjugate of rc. More precisely. and hence for large n P(Sn/n > x) > E [e. exponential change of measure is a key tool in large deviations methods.tin f o') o e9o^y 1 1 ey2/2 dy 21r = etin 1 Bo 27rn . the LegendreFenchel transform or just the Legendre transform or the large deviations rate function). Since P nn > x) = E {e_8 ' ( 9).t.sseo f which in conjunction with (3. Most often. replacing Sn in the exponent and ignoring the indicator yields the Chernoff bound P Sn > x 1 < e°n (3. P with mean nx and variance no.2). nx < Sn < nx + 1. (3.e. In fact. since Sn is asymptotically normal w.(e)i XI E dx].4) n Next. we get P(Sn/n > x) E [e9nx +nK(9)9" '.425. Define .4) immediately yields (3.1).4 enn +1. we have P(nx < Sn < nx + 1. i.3.3) is put equal to x. the sup in the definition of rc* can be evaluated by differentiation: rc*(x) = Ox .q = rc* (x). which is a saddlepoint equation .
which is of similar spirit as the dicussion in VII.1). Mogulskii's theorem which gives path asymptotics./^ >7 < zn n for n n0. Xn) and sn = x1 + • • • + xn (note that the r.dxn) = 05nKn(7)Fn(dx1. .. Further main results in large deviations theory are the GartnerEllis theorem. X2. and the WentzellFreidlin theory of slow Markov walks. .. We further write µ = tc'(ry). 1]. Sanov's theorem which give rare events asymptotics for empirical distributions.2 (GLYNN & WHITT [163]) Let X1.. Pn Sn1 .p > 7 < zn. (iii) #c (8) = limn. we shall concentrate on a result which give asymptotics under conditions similar to the GartnerEllis theorem: Theorem 3 ..e < 8 < y + e. to be made rigorous.dxn) where Fn is the distribution of (X1i .. see Jensen u [215] or [APQ] p. e > 0 such that (i) Kn (0) = log Ee°Sn is welldefined and finite for 'y . however.o log Ee9Sn /n..h. commonly denoted as is the saddlepoint approximation.308 CHAPTER XI.. The substitution by V needs. Xn given by Fn(dxl.e < 8 < y + e. Assume that there exists 'y.. which is a version of Cramer's theorem where independence is weakened to the existence of c(O) = limn. (iv) tc(ry) = 0 and r.'s. and write Sn = X1 + • • • + Xn. Ee9X n < oo for e < 0 < e.. 260 for details. For the proof.s. Then i/.v. (ii) lim supn. In the application of large deviations to ruin probabilities. 1) and no such that Sn . there exists z E (0. MISCELLANEOUS TOPICS which is the same as (3.. asymptotics for probabilities of the form P ({S[nti/n}o<t<l E r) for a suitable set r of functions on [0.. we introduce a change of measure for X1.. . .. is differentiable at ry with 0 < K'(y) < 00..3 For each i > 0. . r(u) = inf {n : Sn > u} and o(u) = P('r(u) < oo). We shall need: Lemma 3 .'(u) )Ng a"u. n Icn(0) exists and is finite for ry . integrates to 1 by the definition of Icn)... be a sequence of r..3. that is.
y) .71 < e and jq9j < e. S.m(7).n m µ 1 + rl .2. the r . ( U) P(S.s.3.. This establishes the first claim of the lemma .+r7) < zn for n > no.r (7) n = e. We first show that lim inf„_. 0.91) + o(O ) as 0 J. This proves the existence of z < 1 and no such that Pn (Sn/n > µ. can be chosen strictly negative by taking 9 small enough.ne(p+ 17).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 +.µ?7 .> . P n(Sn/n > {c+77) < e no(µ 309 +n)Enees n +n)elcn(B +7).+r.2. u Proof of Theorem 3. 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. Let r7 > 0 be given and let m = m(77) = [u(1 + 77)/µ] + 1. h.h. log zl'(u)/u > 'y. it is easy to see that the r. mµ Sm > u] km e7Sm+n. The rest of the argument is as before. h.. in particular the r.ne(µ limsup 1 log Pn (Sn/n > µ + 17) < ic(9 + ry) .77) follows by symmetry (note that the argument did not use µ > 0). LARGE DEVIATIONS Proof Let 0 < 9 < e where a is as in Theorem 3.Bµ .n > u ) = [ Em [em Em 1e.W. The corresponding claim for Pn(Sn/n < µ .s. Then V.. > 1 +17] m(7)...n e(µ +o)w"(7) [Eep(B +7)Sn]1 /p [EegoX. For Sn1i we have Fn(Sn 1/n > µ+r7) < ene(µ+ 1?)EneeS„1 = ene ( µ+n)EneeSneX„ eno(µ +n) Ee(e+7)Sn ex„ wn (7) < e.s. Clearly. Since I EeqOX „ ] 1/q is bounded for large n by (ii). for Sn. is of order .077 n^oo n and by Taylor expansion and (iv ). S.YS. can be chosen strictly negative by taking p close enough to 1 and 0 close enough to 0.]1/q = e.
3. this is possible by (iii).. I2 = F(T(u) = n).log z) /2 and Sn Fn\ n >lb+S) <Zn. Pn \ > la+ 8 I < zn (3.(•) goes to 1 by Lemma 3.7) so that n(b) I1 < e'Yu E en.0 log i'(u )/u < 'y. P(T(u) = n) < P(Sn > u) = En [e7S.. MISCELLANEOUS TOPICS (7).I < µl1 1+77 I M 1_ 1+277 S.. Obviously. For lim supu.6) for some z < 1 and all n > n(E).YS +^c CHAPTER XI.310 ]Em I e. we write P(T(u) = n) = Il + I2 + I3 + I4 'i/I(u) _ E00 n=1 where n(b) Lu(10/µJ Ii = 1: F(T(u) = n).. logO(u)/u > ry. n=1 .. I > IL exp `S. 0 yields liminfu __. n=1 n=n(b)+1 00 Lu(1 +6) /µJ 13 F( T (u) = P(T(u) n).n(ry)/u 4 0andm/u* (1 + r7)/µ. Sn > u] < eYu+Kn(7)pn(Sn > u) (3.(Y). (iv) and Lemma 3.n Yµ 1 + m + r ('Y) } U n \ 77 m µ µ7 1 < 1+ 77 ) Here E..+wn(7). 14 = = E Lu(16)/aJ+1 Lu(1+6)/µJ+l = n) and n(S) is chosen such that icn('y )/n < 6 A (. and since Ic. we get lum inf z/i(u) 1 +12r7 >_ ry + 77 Letting r7 J..3.
3. we get lim sup log u/00 O (U) < y + b(1 + b) U Letbl0.10) 00 I4 < E F(Sn_1 < u.11) [u(1+6)/µJ+1 1  Thus an upper bound for z/'(u) is n(6) e'Yu n=1 eKn (7) + 2 + (28U + 1) e6u(1+6)/µ Fi 1 zl /2 and using (i).zl/z en6 [u(1 +6)/µJ 1u (1 +6) /µJ ekn(7) < e' 13 < C" E Yu l u(16)/lij+1 Lu(16)/µJ+l1 < e7U Finally. eryu en logz/2p n nt n. S. u . Sn1 C U. LARGE DEVIATIONS Lu(16)/µJ 311 I2 < e"u n=n(6)+1 e'n(Y)P(Sn > u) < Lu(16)/µJ ^. C 26u `p / +1 I e6u(1+6)/µ (3. 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 . µ n=n(6)+1 \ 1u(16)/µ1 00 1 zn < e7u E Z n/2 < e(U xn/2 E n=n(6)+1 n=0 eYu = 1 . > u) Lu(1+6) /µJ +l 00 )^n 'YSn+kn (7) .' 1 + b) n e7u x 1 /2 1 n x n / 2x (3.
2. Corollary 3. For 12.4 Under the assumptions of Theorem 3.(u) = I1+I2+I3+I4'^ ery( u). 4 there is an aj > 0 and a cj < oo such that Ij < c3e.u(1+b)/rc'(7)). we have rcn (a + 7) < 2n^c(7 + a) < 4narc' (7).4.b)/i(7).('+'Y).Q is so small that w = 1 . For 14. u . it holds for each b > 0 that 0(u) 1' g F(T(u) E (u(1 .z 1/z For I1.. e'. (7 + a) < 2arc'(7). Then for n large.312 CHAPTER XI. I2. this is straightforward since the last inequality in (3. ryue«iu . say n n1. we need to redefine n(b) as L. it suffices to show that for j = 1. For I.3ui where . cf.xl/2 to give the desired conclusion. MISCELLANEOUS TOPICS The following corollary shows that given that ruin occurs. the last steps of (3. IV.4/3rc'(y) > 0..7' a"ju. Letting c11 = maxn<n. > u) < e"' E eIsn = ectueKn (a+'Y)Kn(7) where 0 < a < e and a is so small that r. the typical time is u/rc'(7) just as for the compound Poisson model. 13 = P(T (u) E (u(1 b)l^ (7).8) by P(S.9) can then be sharpened to x LQuJ /2 I2 < e7u 1 . u(1 + b)/i(7)) Proof Since V.11 ) can be sharpened to x 4 [u(1+6)/µJ /2 1 . 2. we replace the bound P(Sn > u ) < 1 used in (3. we get Lou] E exp {( 7 + a)u + Kn(a +7)} n=1 Il Lou] exp {(y + a)u} { 111 + exp {4narc'(7)} n=1 exp {('y + a)u} c1 exp {4/3uarc'(7)} = clewhere a1 = aw.
we shall give two continuous time examples and tacitly assume that this can be done. whether P ( sup St > u ltg a ^" 0<t<oo // (3.g. Let {Nt}t>0 be a possibly inhomogeneous Poisson process with arrival rate .. (iv) becomes existence of a limit tc(9) of tct(9) _ log Ee8S° It and a y > 0 with a(y) = 0...1.'(y) > 0.14) is needed in both examples . Assuming that the further regularity conditions can be verified.v. and in fact. but nevertheless. for the ruin probability z/'h(u) of any discrete skeleton {Skh}k=0.. 2 is in force with y = 2p/wz.2 then immediately yields the estimate log F( sup Skh > u) a7u (3. It is then wellknown and easy to prove that Sn has a normal distribution with mean np and a variance wn satisfying i lim wn = wz = Var(X1 ) + 2 E Cov(Xl. Xk+l) k=1 00 naoo n provided the sum converges absolutely.. The problem is whether this is also the correct logarithmic asymptotics for the (larger) ruin probability O(u) of the whole process.(O) = 9µ+02 for all 9 E R. Theorem 3. and we conclude that Theorem 3 .e.f. An event occuring at time s is rewarded by a r. 09(9).2 shows that the discrete time structure is used in an essential way.. V(s) with m. If {St}t> 0 is the claims surplus process. Obviously many of the most interesting examples have a continuous time scale.LARGE DEVIATIONS 313 Example 3 .12) k=0.. criteria are given in Duffield & O'Connell [124].1.5 Assume the Xn form a stationary Gaussian sequence with mean p < 0. Thus the total reward in the interval [0. Hence z z\ 2 z nrn(9) _ n Cn0p+BZn/ * . i.13) One would expect this to hold in considerable generality.3(s) at time s. t] is Rt = E V (Un) n: o„ <t . the key condition similar to (iii). The reader not satisfied by this gap in the argument can easily construct a discrete time version of the models! The following formula (3.. To verify these in concrete examples may well present considerable difficulties. 11 Inspection of the proof of Theorem 3. r.3.
one would take p(t) = (1 + rt)At/ t. More precisely. it contributes to St by the amount Un(t . we conclude that Cu) log e7 u (cf. Then logEeOR° = J0 /3(s)(^8(9) . Un represents the total payment for the nth claim). if the nth claim arrives at time a. e. the Cramer. We further assume that the processes {U1(s)}8>0 are i.2 are trivial to verify. nondecreasing and with finite limits Un as s T oo ( thus..v. where Ft = a(A8 : 0 < s < t). leading to St = At(1+77) Joo t S8 ds. Most obviously. the best estimator of /3µB based upon Ft..0 and assume there are y. the CramerLundberg model implicitly assumes that the Poisson intensity /3 and the claim size distribution B (or at least its mean µB) are known. Example 3.noise model is the same as the one for the Cramer Lundberg model where a claim is immediately settled by the amount Un.g. i.9t.14) (to see this . then the payments from the company in [on.14). Of course .. this is not realistic . O'n +S] is a r . (3. Thus. We let ic (9) = 3(EeWU° . we have rct (9)/t 4 ic (9). MISCELLANEOUS TOPICS are the event times. a differential equation in t).d. It is interesting and intuitively reasonable to note that the adjustment coefficient ry for the shot .Lundberg model has the larger ruin probability. is At .9t = /3 J t (Ee8U° i8l . 7 Given the safety loading 77. . the above discussion of discrete skeletons).It.1) ds . assuming a continuous premium inflow at unit rate. At = . (9) < oo for 9 < 'y + C.1) .t.15) . Kt (0) t (Ee9U"it8i J0 .Q„) .6 We assume that claims arrive according to a homogeneous Poisson process with intensity 0 . Thus. Thus by (3. but that a claim is not settled immediately. 0 and since EeOUn(8) + Ee°U^ as s * oo. If the nth claim arrives at time Qn = s. we have S.1) ds . Un(s). derive . e > 0 such that ic('y) = 0 and that r..s). 0 Example 3 .'`1 U. Of course. n: o. An apparent solution to this problem is to calculate the premium rate p = p(t) at time t based upon claims statistics . <t which is a shotnoise process.314 where the an CHAPTER XI. Since the remaining conditions of Theorem 3.1) ds rt (3. = U„ ( t .
Thus.14) that rt _ 13 Jo _ (a [1_( i+77)log]) ds_flt = t (a) (3.log Oi are i.d. 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. typically the adaptive premium rule leads to a ruin probability which is asymptotically smaller than for the CramerLundberg model .(1 +i) f > i= 1 s ds = E Ui 1 .16) i=1 o i=1 Let ict (a) = log Eeast .2 hold. LARGE DEVIATIONS With the Qi the arrival times.e.21) This follows from the probabilistic interpretation Si EN '1 Yi where Yi = Ui( 1+(1 +r7)log ©i) = Ui(1(1 +17)Vi) where the Oi are i .d. uniform (0.20) (3. (3.1) .17) K(a) f o 1 O (a[I + (1 + 77) log u]) du )3. To see this . we have Nt t N.3. i. It then follows from (3.i. 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 . Indeed. equivalently.18) Thus (iii) of Theorem 3.i. we conclude that t. Ui Nt / t 01i 315 St = Ui . one has y > y' (3./3.(1 + 17)0µB = 0. rewrite first rc as te(a) _ /3E 1 1 +(1+77)aUJ eau 1 .(1 + r7) log t (3. the solution of /3(Eelu .b(u) IN a'Yu (cf. standard exponential .19) with equality if and only if U is degenerate. the Vi = . and since the remaining conditions are trivial to verify.1) or . (3.
rc*' (0 ) < 0. x > x0.20) is due to Tatyana Turova. assuming that the U.i. are i. Further applications of large deviations idea in risk theory occur in Djehiche [122]. k(0) = 0.2. the function k(x) = e7*x . [245].. at time t.d. much of the analysis carries over to more general cases.xo. k'(0) < 0.316 CHAPTER XI. = P(N = n) = e(3an However.7. though we do not always spell this out. we then take t = 1 so that p. the proof of (3. Further. say one year.19). this in turn yields y > y*. the study is motivated from the formulas in IV. Dembo & Zeitouni [105] and Shwartz & Weiss [339]. [257] and Nyrhinen [275]. The main example is Nt being Poisson with rate fit. For Example 3. Lehtonen & Nyrhinen [244].1 . see Nyrhinen [275] and Asmussen [25]. Further. In particular. and k(x) < 0. MISCELLANEOUS TOPICS Next. In addition to Glynn & Whitt [163]. This implies n(y*) < 0. we are interested in estimating P(A > x) for large x.(1 + ri)y*x is convex with k(oo) = 00. For notational simplicity. 0 < x < x0. 11 Notes and references Some standard textbooks on large deviations are Bucklew [81]. with common distribution B and independent of Nt. This is a topic of practical importance in the insurance business for assessing the probability of a great loss in a period of length t. and since tc(s).2 expressing the finite horizon ruin probabilities in terms of the distribution of A. . 4 The distribution of the aggregate claims We study the distribution of the aggregate claims A = ^N' U. Therefore e7'U _ k(U) E [1+(1+77)y*U] .1 E [1+(1+77)y*U] 0 k (+ *y B(+ 1 + (1(+71)y*y B(dy) L xa 1 + f + (1 + rl) Y* xo jJxo k(y) B(dy ) + f' k(y) B(dy) } = 0. MartinL6f [256]. see also Nyrhinen [275] for Theorem 3. so there exists a unique zero xo = xo(r7) > 0 such that k(x) > 0. y = y* can only occur if U . using that Ek(U) = 0 because of (3. a* (s) are convex with tc'(0) < 0 .
A E dx] . The exponential family generated by A is given by Pe(A E dx) = E [eeA K(9).[s])3/2 = 0.3B[9] and Be is the distribution given by eox B9(dx) = B [9] B(dx). 818' where s' = sup{s : B[s] < oo}. Proposition 4. Hence P(A > x) = E e [e9A+ ic(9).3B"[9].1 Assume that lim8T8."(0) = .1).1. Then Ee"A = e'(") where x(a) _ 0(B[a] . THE DISTRIBUTION OF THE AGGREGATE CLAIMS 317 4a The saddlepoint approximation We impose the Poisson assumption (4. B"[s] = oo. only with 0 replaced by a9 and B by B9.1).4. For a given x. B"' [s] lim (B". Then as x * oo.9(Ax). i. This shows that the Pedistribution of A has a similar compound Poisson form as the Fdistribution. K'(0) _ ic'(9) = x.e. A > x)] = eex+K( e)E9 [e . (4. 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 . Vare(A) = s.1) where )30 = . The analysis largely follows Example 3.ic(9) = .3e(bo[a] . In particular.x)//3B"[9] is standard normal. e9x+K(°) P(A > x) B 2ir /3 B" [9] Proof Since EBA = x.2) implies that the limiting Pedistribution of (A . no(a) = logE9e'A = rc(a + 9) . we define the saddlepoint 9 = 9(x) by EBA = x.
1). where q(x) is bounded away from 0 and oo and h (x) is convex on an interval of the form [xo. In particular. and (4. A covers the exponential distribution and phasetype distributions. b is logconcave. large x. 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 .(D X .2 If B is subexponential and EzN < oo for some z > 1.EN B(x).l3pB. 2). then P(A > x) . For details. 4b The NP approximation In many cases . either of the following is sufficient: A. b(x) = q(x)eh(z). it holds that EA = . or.Q{AB (4. Y satisfies 9(u) ti eu2/2(1 + ibu3) (4. b is gammalike.1 goes all the way back to Esscher [141]. For example. leading to P(A > x) :. In fact. Furthermore 00 b(x)Sdx < oo for some ( E (1. For a rigorous proof. Thus . just the same dominated convergence argument as in the proof of Theorem 2. The (first order) Edgeworth expansion states that if the characteristic function g(u) = Ee"`}' of a r. see Embrechts et al. under the Poisson assumption (4. Jensen [215] and references therein. B covers distributions with finite support or with a density not too far from ax° with a > 1. [138]. Notes and references Proposition 4. Remark 4 . more generally.3) The result to be surveyed below improve upon this and related approximations by taking into account second order terms from the Edgeworth expansion.e. the distribution of A is approximately normal .318 CHAPTER XI.x') where x' = sup {x : b(x) > 0}. MISCELLANEOUS TOPICS It should be noted that the heavytailed asymptotics is much more straightforward.v.(3µB)/(0µB^))1/2 has a limiting standard normal distribution as Q ^ oo.2) is often referred to as the Esscher approximation. The present proof is somewhat heuristical in the CLT steps.1 yields: Proposition 4. Var(A) _ ^3p.ycix °ie6x B. bounded with b(x) . it is quite questionable to use (4.4) . some regularity of the density b(x) of B is required.2i and that (A .3) and related results u for the case of main interest . For example. i.
are the cumulants . and so as a first approximation we obtain a1_E = EA + yle Var(A) .5) follows by integration.. so that 1(u) 3 exp { . Remark 4. the NP (normal power) approximation deals with the quantile al_E. where Kl . defined as the the solution of P(A < yle) = 1 .f.2X2 .5 (y3 .i 6 r 1 3 so that we should take b = ic3/6 in (4..6(1 .99. Heuristically. the density of Y is 1 °° _ eiuy f(u) du 2x _.4. the CLT for Y = Y6 is usually derived via expanding the ch. THE DISTRIBUTION OF THE AGGREGATE CLAIMS where b is a small parameter. Rather than with the tail probabilities F(A > x). f °o 9(y) = 1 e'uye u2/2(1 + iSu3) du 27r _ cc(y) . Let Y = (A . A particular case is a.. If the distribution of Y is close to N(0. however. K4 . zl_e be the 1 . the standard normal distribution. . which is often denoted VaR (the Value at Risk). . K2 = Var (Y). Var(Y) = 1 as above .EA)/ Var(A) and let yl_E. resp. in particular.2 ^ \1 .e..s.c2i.6) . one needs to show that 163.y2)^P(y)• 319 Note as a further warning that the r. u5. one expects the u3 term to dominate the terms of order u4.5) may be negative and is not necessarily an increasing function of y for jyj large. K3 = E(Y .i 3 K3 } Pt^ exp .2K3 + 4i 64 + .l = EY. Thus if EY = 0.. are small.. then P(Y < y) 4(y) .2 2 .equantile in the distribution of Y. of (4.EY)3. (4. If this holds .3& (y).: EA + zl_E Var(A) . ylE should be close to zl_E (cf..1).5). (4. .h. In concrete examples . and from this (4. as u2 u3 u4 9(u) = Ee'uY = exp {iuci ..5) is obtained by noting that by Fourier inversion.. s.3!).
E )Azl E) 4(z1E) + ( ylE .5) by noting that the 4. that [101] distinguishes between the NP and Edgeworth approximations.1)^ 2) µ'E Notes and references We have followed largely Sundt [354].3ni /3 .k = /3µB^1 / (. We can rewrite (4. k3 is small for large /3 but dominates 1c4. b such that EN 1 U%. 21 .zl E)V(zl_E) ..E + (yl. Note.7) as 1 (3) a1E = Qµa +z1 ..S(1 .EA)3 a1_E = EA + z1_E(Var (A))1/2 + 1 Var(A) Under the Poisson assumption (4.(y) terms dominate the S(1 . this yields the NP approximation 6(Z1 _E . as required . 4c Panjer 's recursion Consider A = constants a. [101].y2)cp( y) term.6pBki) d/2. the kth cumulant of A is /3PBk' and so s. In particular .zlE)W(zlE) 1 .EA ) / Var(A)..1)EY3. however. For example.. MISCELLANEOUS TOPICS A correction term may be computed from (4.S(1 .zl E)^o(zl E) .zlE )w(zl _E) = which combined with S = EY3/6 leads to q^ 1 Y1 ..1). . this holds with a = 0.. b = /3 for the Poisson distribution with rate /3 since Pn = Pn1 n! n (n .yi.E)A1 l E) 1 E 4)(yl E) ^' . n = 1. K5 .zi. and assume that there exist n ) Pn_i .1) E (A . let pn Pn = (a+ = P(N = n).E(/3PB^1 )1^2 + s(z1E .6 (1 .E . Another main reference is Daykin et at. Using Y = (A ..320 CHAPTER XI. This leads to t( yl E) .E = z1E + S(zi_E .5(1 .1)! n ^eQ .
. E[a +bU=I >Ui =j l i=1 J (4. if go = 0. (4. THE DISTRIBUTION OF THE AGGREGATE CLAIMS 321 Proposition 4. j = 1. n = k=n1 9k(n1 )9j k • (4.14) is independent of i = 1. j = 0...4. (4.11) Remark 4. Then fo = >20 9onpn and fi = 1 E In particular. and calculating the gj*n recursively by 9*1 = 9j.. then j (a + b!) 1ag k_1 3 gkfj. n.. . 2.5 The crux of Proposition 4. (4. the complexity (number of arithmetic operations required) is O(j3) for (4. .. .4 is that the algorithm is much faster than the naive method. fj = E (a+ b k =1 )9kfi_k . By symmetry.4.13) Namely. (4..12).14) is therefore a + b/n..1.10) f o = po.. Since the sum over i is na + b.4. fj = P(A = j).. 1. . which would consist in noting that (in the case go = 0) fj = pn9jn n=1 (4.} and write gj = 2 ..4 Assume that B is concentrated on {0. 2. the value of (4. j1 g. u Proof of Proposition 4.9). The expression for fo is obvious.. .k .13) but only O(j2) for Proposition 4.. .12) we get for j > 0 that fj n a b + n p nlgj *n 00 U I n 1 *n = E a+bUi=j pn19j n=1 j i=1 CC) n Ui EE n=1 Ia +b Ul i=1 =j pn_1 . 2. .12) where g*n is the nth convolution power of g. . Hence by (4. j = 1.
322
00 J
CHAPTER XI. MISCELLANEOUS TOPICS
EE (a + bk I gkg3 _ k lieni n=ik=0 (a+bk l gkE g j'`kpn = E (a+b!)9kfi_k n=0 k=0 k=0 ^I 1 E(a+b. agofj+ k Jgkfjk, k=i /
and and (4.9) follows . (4.11) is a trivial special case.
u
If the distribution B of the Ui is nonlattice , it is natural to use a discrete approximation . To this end, let U(;+, U(h) be U; rounded upwards, resp. downwards , to the nearest multiple of h and let A}h) = EN U. An obvious modification of Proposition 4.4 applies to evaluate the distribution F(h) of A(h) letting f( ) = P(A() = jh) and
g(h) gkh+
= P (U(h2 = kh) = B((k + 1)h)  B(kh ), k = 0, 1, 2, ... , = P (U4;+ = kh) = B(kh)  B (( k  1)h) = gk  l,, k = 1, 2, ... .
Then the error on the tail probabilities (which can be taken arbitrarily small by choosing h small enough ) can be evaluated by
00 00
< P(A > x ) f (h) j=Lx/hl j=Lx/hl
Further examples ( and in fact the only ones , cf. Sundt & Jewell [355]) where (4.9) holds are the binomial distribution and the negative binomial (in particular, geometric ) distribution . The geometric case is of particular importance because of the following result which immediately follows from by combining Proposition 4.4 and the PollaczeckKhinchine representation: Corollary 4.6 Consider a compound Poisson risk process with Poisson rate 0 and claim size distribution B. Then for any h > 0, the ruin probability zb(u) satisfies 00 00
f^,h) Cu) < E ff,+, j=Lu/hJ j=Lu/hJ (4.15)
f! h)
5. PRINCIPLES FOR PREMIUM CALCULATION
where f^ +, f^ h) are given by the recursions
(h) 3 (h) (h)
323
fj,+ = P 9k fjk,+ ' I = 17 2, .. .
k=1 3 (h)
(h)
=
P
(h)
f9,  (h) gk,fAk, e 1  ago, k=1
j = 1+2,
starting from fo + = 1  p, f(h) = (1  p)/(1  pgoh) and using 07
g(kh) 1 (k+1)h
=
Bo((k + 1 ) h)  Bo(kh ) =  f
AB
kh
B(x) dx, k = 0, 1, 2, ... , k = 1,2 .....
gkh+
Bo(kh )  Bo((k  1 ) h) = 9kh)1 ,
Notes and references The literature on recursive algorithms related to Panjer's recursion is extensive, see e.g. Dickson [115] and references therein.
5 Principles for premium calculation
The standard setting for discussing premium calculation in the actuarial literature does not involve stochastic processes, but only a single risk X > 0. By this we mean that X is a r.v. representing the random payment to be made (possibly 0). A premium rule is then a [0, oo)valued function H of the distribution of X, often written H(X), such that H(X) is the premium to be paid, i.e. the amount for which the company is willing to insure the given risk. The standard premium rules discussed in the literature (not necessarily the same which are used in practice!) are the following: The net premium principle H(X) = EX (also called the equivalence principle). As follows from the fluctuation theory of r.v.'s with mean, this principle will lead to ruin if many independent risks are insured. This motivates the next principle, The expected value principle H(X) = (1 + 77)EX where 77 is a specified safety loading. For 77 = 0, we are back to the net premium principle. A criticism of the expected value principle is that it does not take into account the variability of X which leads to The variance principle H(X) = EX+77Var(X). A modification (motivated from EX and Var(X) not having the same dimension) is
324
CHAPTER XI. MISCELLANEOUS TOPICS
Var(X).
The standard deviation principle H(X) = EX +rl
The principle of zero utility. Here v(x) is a given utility function, assumed to be concave and increasing with (w.lo.g) v(O) = 0; v(x) represents the utility of a capital of size x . The zero utility principle then means v(0) = Ev (H(X)  X); (5.1)
a generalization v(u) = Ev (u + H(X)  X ) takes into account the initial reserve u of the company. By Jensen 's inequality, v(H(X)  EX) > Ev(H(X)  X) = 0 so that H(X) > EX. For v(x) = x, we have equality and are back to the net premium principle. There is also an approximate argument leading to the variance principle as follows. Assuming that the Taylor approximation
v(H(X)  X) ^ 0 +v'(0)(H (X)  X) + v 0 (H(X)  X)2 ,/2
is reasonable , taking expectations leads to the quadratic v"H(X )2 + H(X) (2v'  2v"EX) + v"EX2  2v'EX = 0 (with v', v" evaluated at 0) with solution
H(X)=EXv^±V( ^ )2Var(X).
Write
( vI ) 2 \
Var(X) v^  2v^Var(X)/ I  (
, Var(X) )2
If v"/v' is small, we can ignore the last term. Taking +f then yields H(X) ,:: EX 
2v'(0) VarX;
since v"(0) < 0 by concavity, this is approximately the variance principle. The most important special case of the principle of zero utility is The exponential principle which corresponds to v(x) = (1  e6x)/a for some a > 0. Here (5.1) is equivalent to 0 = 1  e0H(X)EeaX, and we get
H(X) = 1 log Ee 0X .
a
5. PRINCIPLES FOR PREMIUM CALCULATION
325
Since m.g.f.'s are logconcave, it follows that H,, (X) = H(X) is increasing as function of a. Further, limQyo Ha (X) = EX (the net premium princiHa (X) = b (the premium ple) and, provided b = ess supX < oo, lim,, H(X) = b is called the maximal loss principle but is clearly not principle very realistic). In view of this, a is called the risk aversion The percentile principle Here one chooses a (small ) number a, say 0.05 or 0.01, and determines H(X) by P(X < H(X)) = 1  a (assuming a continuous distribution for simplicity). Some standard criteria for evaluating the merits of premium rules are 1. 77 > 0, i .e. H(X) > EX. 2. H(X) < b when b (the ess sup above ) is finite 3. H(X + c) = H(X) + c for any constant c
4. H(X + Y) = H(X) + H(Y) when X, Y are independent
5. H(X) = H(H(XIY)). For example , if X = EN U= is a random sum with the U; independent of N, this yields
H
C^
U; I = H(H(U)N)
(where, of course, H(U) is a constant). Note that H(cX) = cH(X) is not on the list! Considering the examples above, the net premium principle and the exponential principle can be seen to the only ones satisfying all five properties. The expected value principle fails to satisy, e.g., 3), whereas (at least) 4) is violated for the variance principle, the standard deviation principle, and the zero utility principle (unless it is the exponential or net premium principle). For more detail, see e.g. Gerber [157] or Sundt [354]. Proposition 5.1 Consider the compound Poisson case and assume that the premium p is calculated using the exponential principle with time horizon h > 0. That is,
N,,
Ev I P  E U;
i =1
= 0 where
v(x) = 1(1  e°x
a
Then ry = a, i.e. the adjustment coefficient 'y coincides with the risk aversion a.
326
Proof The assumption means
CHAPTER XI. MISCELLANEOUS TOPICS
0 a (1  eareo (B[a11)
l
i.e. /3(B[a]  1)  ap = 0 which is the same as saying that a solves the Lundberg u equation. Notes and references The theory exposed is standard and can be found in many texts on insurance mathematics, e.g. Gerber [157], Heilman [191] and Sundt [354]. For an extensive treatment, see Goovaerts et al. [165].
6 Reinsurance
Reinsurance means that the company (the cedent) insures a part of the risk at another insurance company (the reinsurer). Again, we start by formulation the basic concepts within the framework of a single risk X _> 0. A reinsurance arrangement is then defined in terms of a function h(x) with the property h(x) < x. Here h(x) is the amount of the claim x to be paid by the reinsurer and x  h(x) by the the amount to be paid by the cedent. The function x  h(x) is referred to as the retention function. The most common examples are the following two: Proportional reinsurance h(x) = Ox for some 0 E (0, 1). Also called quota share reinsurance. Stoploss reinsurance h(x) = (x  b)+ for some b E (0, oo), referred to as the retention limit. Note that the retention function is x A b. Concerning terminology, note that in the actuarial literature the stoploss transform of F(x) = P(X < x) (or, equivalently, of X), is defined as the function
b * E(X  b)+ =
f
(s  b)F(dx) _ f
6 00
(x) dx.
An arrangement closely related to stoploss reinsurance is excessofloss reinsurance, see below.
Stoploss reinsurance and excessofloss reinsurance have a number of nice optimality properties. The first we prove is in terms of maximal utility: Proposition 6.1 Let X be a given risk, v a given concave nondecreasing utility function and h a given retention function. Let further b be determined by E(X b)+ = Eh(X). Then for any x,
Ev(x  {X  h(X)}) < Ev(x  X A b).
6. REINSURANCE
327
Remark 6 .2 Proposition 6.1 can be interpreted as follows. Assume that the cedent charges a premium P > EX for the risk X and is willing to pay P1 < P for reinsurance. If the reinsurer applies the expected value principle with safety loading q, this implies that the cedent is looking for retention functions with Eh(X) = P2 = P1/(1 + 77). The expected utility after settling the risk is thus
Ev(u + P  P1  {X  h(X)})
where u is the initial reserve . Letting x = u + P  P1, Proposition 6.1 shows that the stoploss rule h (X) = (X  b)+ with b chosen such that E(X  b)+ u = P2 maximizes the expected utility. For the proof of Proposition 6.1, we shall need the following lemma: Lemma 6 .3 (OHLIN'S LEMMA) Let X1, X2 be two risks with the same mean, such that Fj(x) < F2 (x), x < b, Fi(x) ? F2(x), x > b for some b where Fi(x) = P(Xi < x). Then Eg(X1) < g(X2) for any convex function g. Proof Let Yi=XiAb, Zi=Xivb.
Then
P(Yl < x) _ Fi(x) <_ F2 (x) = P(Y2 < x) x < b 1=P(Y2<x) x>b so that Y1 is larger than Y2 in the sense of stochastical ordering . Similarly, P(Zl < x) _ 0 = P(Z2 < x) x < b Fi(x) > F2(x) = P(Z2 < x) x > b
so that Z2 is larger than Zl in stochastical ordering. Since by convexity, v(x) = g(x)  g(b)  g'(b)(x  b) is nonincreasing on [0, b] and nondecreasing on [b, oo), it follows that Ev(Y1) < Ev(Y2), Ev(Zi) < Ev(Z2). Using v(Yi) + v(Zi) = v(Xi), it follows that
0 < Ev(X2)  Ev(Xi) = Eg(X2)  Eg(X1),
using EX1 = EX2 in the last step. u
Proof of Proposition 6.1. It is easily seen that the asssumptions of Ohlin' s lemma hold when X1 = X A b, X2 = X  h(X); in particular, the requirement EX1
328
CHAPTER XI. MISCELLANEOUS TOPICS
= EX2 is then equivalent to E(X  b)+ = Eh(X). Now just note that v is convex. u
We now turn to the case where the risk can be written as N
X = Ui
i=1
with the Ui independent; N may be random but should then be independent of the Ui. Typically, N could be the number of claims in a given period, say a year, and the Ui the corresponding claim sizes. A reinsurance arrangement of the form h(X) as above is called global; if instead h is applied to the individual claims so that the reinsurer pays the amount EN h(Ui), the arrangement is called local (more generally, one could consider EN hi(Ui) but we shall not discuss this). The following discussion will focus on maximizing the adjustment coefficient. For a global rule with retention function h* (x) and a given premium P* charged for X  h* (X), the cedents adjustment coefficient y* is determined by
1 = Eexp {ry*[X  h*(X)  P*]},
for a local rule corresponding to h(u) and premium P for X look instead for the ry solving
J _f
(6.2) N 1 h (Ui), we
[ X_P_^
1 = Eexp
[ Ei  h(Ui)] P [U
= Eexp{ry
h(Ui)]
l (6.3) This definition of the adjustment coefficients is motivated by considering ruin at a sequence of equally spaced time points, say consecutive years, such that N is the generic number of claims in a year and P, P* the total premiums charged in a year, and referring to the results of V.3a. The following result shows that if we compare only arrangements with P = P*, a global rule if preferable to a local one. Proposition 6.4 To any local rule with retention function h(u) and any
N
J}
P > E X  N h(Ui)
4 =1
(6.4)
there is a global rule with retention function h* (x) such that
N
Eh*(X) = Eh(U1)
i=1
and 'y* > ry where ry* is evaluated with P* = P in (6.3).
u But since ry > 0.d. as often local as global. Proof As in the proof of Proposition 6. Assuming for simplicity that the Ui are i. The arrangement used in practice is. that 01[ry] < 0[y] where 0[y] = Ee'r(U^') . appealing to (6. This follows by taking Xl = U A b.6 Assume the Ui are i.h(U)].P } < 1 = Eexp E[Ui. . i.h(Ui)P JJJ l:='l {ry ] or.4) and u g(x) = e7x in Ohlin's lemma.h(Ui)] . Remark 6. we get EX = EN • EU.5) reduce quite a lot. expectations like those in (6.P]}.4).5) holds trivially. N E X .4).b)+ = Eh(U) (and the same P) satisfies 71 > ry. (6.h( UU) = EN • E[U .h(Ui)] .P I = EC [7]N.d.i. y = Ei [Ui .P > EexP{7[X .h(u) and any P satisfying (6. Eexp 7 [E [Ui . (6.h(U) (as in the proof of Proposition 6. however.6). Then for any local retention function u .h * (X) . ry* > 0 because of (6. Applying the inequality Ecp(Y ) > EW(E (YIX )) (with W convex ) to W(y ) = eryy.5 Because of the independence assumptions . the excess ofloss rule hl (u) = (u . X2 = U . this implies 7* > 7. ' ii (6. Local reinsurance with h(u) = (u .. then (6.3).6.b)+ is referred to as excessofloss reinsurance and plays a particular role: Proposition 6. and so on.6) u where C[ry] = Ee'r(u4(u)). we get N 1 = Eexp ry E[Ui ii . it suffices to show that Eexp {ry ii 'UiAb.P.b)+ with b determined by E(U .h(Ui)] . REINSURANCE Proof Define N 329 h* (x) = E > h(Ui) X = x .4).4.
The original reference for Ohlin's lemma is Ohlin [277]. Bowers et at. See further Hesselager [194] and Dickson & Waters [120].many texts on insurance mathematics. MISCELLANEOUS TOPICS Notes and references The theory exposed is standard and can be found in. [76]. . see also Sundt [354]. e.330 CHAPTER XI. The present proof is from van Dawen [99].g. Heilman [191] and Sundt [354].
Y1. some condition is needed: that F is nonlattice.T„_1). all have the same distribution. t +a]). note in particular that U({0}) = 1.. The associated renewal measure U is defined by U = u F*" where F*" is the nth convolution power of F. Y. If Yo = 0.... That is. are independent and Y1. The mathematical representation is either the ordered set 0 < To < T1 < . The renewal theorem asserts that U(dt) is close to dt/µ. . . of interarrival times and the time Yo = To of the first arrival (that is. The number max k : Tk_j < t of renewals in [0. stating that U(t+a)U (t) ^ a. denoted by F in the following and referred to as the interarrival distribution. 2h. If F satisfies the stronger condition of being spreadout (F*' is nonsingular w .Appendix Al Renewal theory la Renewal processes and the renewal theorem By a simple point process on the line we understand a random collection of time epochs without accumulation points and without multiple points. The point process is called a renewal process if Yo. t]) so that U(t + a) . t] is denoted by Nt. Lebesgue measure dt normalized by the mean to of F. Y2.1) (here U(t) = U([0..r. of epochs or the set Y1.. when t is large..e. U(A) is the expected number of renewals in A C R in a zerodelayed renewal process.. . Y2. Technically. Then Blackwell 's renewal theorem holds. i. = T„ . + U2 where U1 is a finite measure and U2(dt) = u(t)dt where 331 ..U(t) is the expected number of renewals in (t. t 00 (A. not concentrated on {h. the distribution of Yo is called the delay distribution.} for any h > 0.. then Stone 's decomposition holds : U = U. Lebesgue measure for some n > 1). . the renewal process is called zerodelayed..t.
z(x) = 0. IV). wee shall need the following less standard parallel to the key renewal theorem: Proposition A1. then it suffices for (A.5) 2This condition can be weakened considerably .2).x)F(dx). (A.R. the asymptotic behavior of Z(u) is given by the key renewal theorem: Proposition A1.a. IV).3) Further. u u PF 4 00. U Z(u . (A. the statements being EN(t + a) .332 APPENDIX u(t) has limit 1/µ as t 4 oo. and that F has a bounded density2. (A. 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 .2) has the unique solution Z = U * z.i. µF (A. in convolution notation Z = z + F * Z. A weaker (and much easier to prove) statement than Blackwell's renewal theorem is the elementary renewal theorem.EN(t) . see [APQ] Ch. resp.out.9. stating that U(t)/t > 1/p. that z(u) has a limit z(oo) (say) as u 4 oo.1 if F is nonlattice and z (u) is directly Riemann integrable (d. Under weak regularity conditions (see [APQJ Ch. then Z(u) i f0 z(x)dx . oo). Both result are valid for delayed renewal processes. i. Then Z(u) 4 z(oo). Equivalently.i".4) that z is Lebesgue integrable with limZ.4) If F is spread.2 Assume that Z solves the renewal equation (A. Note in particular that F is spreadout if F has a density f.e. and F(dx) a known probability measure . (A.. In 111.2) Z(u) = J0 u z(x)U(dx). but suffices for the present purposes . z(u) a known function.
. that the existence of y may fail for heavytailed F. However. z(x) = e7xz(x). or many queueing processes.} be a renewal process.. . This program has been carried out in III. Z(u) U = 1 u 1 u f z(u . Here the relevant F does not have mass one (F is defective). Yk ). The property of independent cycles is equivalent to the postTk process {XTk+t}t>0 being independent of To. T1. this expression is to be interpreted as a random element of the space of all Evalued sequences with finite lifelengths.k+t }t>o is independent of To. . equivalently.. results from the case fo F(dx) = 1 can then be used to study Z and thereby Z. T1.2) by e7x to obtain Z = z +P * Z where Z(x) = e'Y'Z(x). 0 PF µF 11 In risk theory.(3. cycles. The distribution F of 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. multiply (A.APPENDIX 333 Proof The condition on F implies that U(dx) has a bounded density u(x) with limit 1/µF as x * oo. and its distribution does not depend on k. of Yo. F(dx) = e7xF(dx). However.. however.t. To this end. 1c Regenerative processes Let {T. Hence by dominated convergence.. Tk (or. that F is a probability measure. • . . Eo etc. The simplest case is when {Xt} has i. Assuming that y can be chosen such that f °° Ox F(dx) = 1.. We let FO. is called the cycle length distribution and as before. we let µ denote its mean. Tk and {Xt }o<t<Tk • For example. i. {Tn} if for any k.r. the present more general definition is needed to deal with say Harris recurrent Markov chains. refer to the zerodelayed case. The kth cycle is defined as {XTk+t}o<t<Yk . .3) satisfied by the ruin probability for the compound Poisson model.e.x)u(x) dx = z(u( 1 . where the Tn are the instants where a customer enters an empty system (then cycles = busy cycles). a basic reason that renewal theory is relevant is the renewal equation II..5a. Note.i. . 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 postTk process {XT. Y1 . A stochastic process {Xt}t>0 with a general state space E is called regenerative w..d. A regenerative process converges in distribution under very mild conditions: . . Y2.t))u(ut) dt 0 0 J f z(oo) • 1 dt = z(OO)..
This is the case considered in [APQ] V. An example is Zt = fo f (X8) ds where {Xt} is regenerative w..0 is called cumulative w.r. just the same proof as there carries over to show: Proposition A1..6) id Cumulative processes Let {Tn} be a renewal process with i. Otherwise . but in fact.e.3. (b) If in addition Var(Ul ) < oo..ZT Then: (a) If E sup I ZTo+t . C). resp .t.t.4 Let {Zt}t^. then (Zt . P(C ( t) < a) 4 0 for any a < oo) and ij (t) * oo. then Xt .r..i. µ 0 If F is spreadout. then e (t) . assume that p < 00 and define Un = ZT}1 . is given by Eg(Xoo) = 1 E0 f Ylg (Xt)dt. r. 0<t<Yi then Zt /t a$• EU1/µ..ZTOI < 00.Tk : t < Tk} as the age..0 be cumulative w. {Tn}. 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.. where the distribution of X.oo (i.v. and q(t) = sup It .d.d.t.i.'s by e. Y1) le Residual and past lifetime Consider a renewal process and define e ( t) as the residual lifetime of the renewal interval straddling t. oo). If p = oo. [0. e(t )) .t : t < Tk}.. fi (t) = inf {Tk . Then Xt Di X. cycles (we allow a different distribution of the first cycle).ZT }0<t<Y„+. 2. {Tn}.334 APPENDIX Proposition A1. (A.r. and we have: holds more generally that (rl(t). {i7(t)} are Markov with state spaces (0.. {Tn} if the processes {ZT +t . Then {Zt}t^.tEU1/µ)/f has a limiting normal distribution with mean 0 and variance Var(Ui) + (!)2Var (Yi)_ 2EU1 Cov(U1. are i. i. We denote the limiting r. for n = 1. in total variation. under the condition of Blackwell's renewal theorem.e.. Then it (ii..+ X. Then {e(t)}.. oo).
d. Hence for t large enough. Proof The number Nt of renewal before t satisfies Nt/t a4' p. Then Eo^(t) satisfies a renewal equation with z(t) _ E[Y1 .4. the joint distribution of (rl. if in addition EYo < oo. ^) is given by the following four equivalent statements: (a) P (77 > x. For the second. the first statement follows. (d) the marginal distribution of ^ is FO. use t E^(t)/t = E[Yo .y) = f U(t . U(x + 1) . Since z ( k) < E[Yi . l:) is the same as the distribution of (VW.y)P(Yo E dy) . Y1i Y2. Then fi(t)/t a4' 0 and.d. In IV. but governed by a Markov chain {Jn} (we . = z is Foz) The proof of (a) is straightforward by viewing {(r.t. and the equivalence of (a) with (b)(d) is an easy exercise.. EC(t)/t + 0.APPENDIX 335 Theorem A1. W are independent..i. we can bound e(t) by M(t) = max {Yk : k < 2t/p}.i.U(x) (c < oo because it is easily seen that U(x + 1) . we used: Proposition A1. 0 If Markov renewal theory By a Markov renewal process we understand a point process where the interarrival times Yo . (b) the joint distribution of (ri. 1) and W has distribution Fw given by dFw/dF(x) = x/pF.v.t. Since the maximum Mn of n i.'s with finite mean satisfies Mn/n a$• 0 (BorelCantelli). Hence t t lt ) = f U(dy)z(t .^(t))} as a regenerative process. ^ > y) = 1 f +Y (z)dz. and the conditional distribution of ri given l. are not i.6 Consider a renewal process with µ < oo.dy )z(y) < c ^ l z(k) Eoe(t 0 0 k=o where c = sup.5 Under the condition of Blackwell's renewal theorem. . (c) the marginal distribution of q is FO. V is uniform on (0. In the general case. Yl > t]. Y1 > t] 4 0. (1 V)W) where V. the sum is o(t) so that Eo£(t)/t + 0 .. assume first the renewal process is zerodelayed. r. Yo > 0] + f Eo^ (t .(t). and the conditional distribution of given 17 = y is the overshoot distribution R0(Y) given by FO(Y) (z) = Fo (y+z)/Fo(y).U(x) < U( 1)).
d. Jn +1=j} where J = a(JO.} is nonlattice (it is easily seen that this definition does not depend on i). .r. namely {Twk } where {Wk } is the sequence of instants w where Jo.and regenerative processes. Further: Proposition A1. is given by Eg(X00) = 1 YO vjEj f g(Xt) dt µ jEE o where p = ujEEViAj. . The semiregenerative process is then regenerative w.. IT. For example. .T_ < oo). A2 WienerHopf factorization Let F be a distribution which is not concentrated on (oo. r+ < oo). G_(x) = P(ST_ < x. . A stochastic process {Xt}t>o is called semiregenerative w. Sn = X1 + • • • + Xn the associated random walk. e. Alsmeyer [5] and Thorisson [372].}. T_=inf{n>0: Sn<0}..i . . oo). . 0] or (0 .+ < x...r.. These facts allow many definitions and results to be reduced to ordinary renewal. Y1..) and (Fij )i.. with common distribution F. . in [APQ]. Yn. < yIJ) = Fij( y) on {Jn= i. .t. = io for some arbitrary but fixed reference state io E E. We call r+ (T_) the strict ascending (weak descending) ladder epoch and G+ (G_) the corresponding ladder height distributions. be i. A Markov renewal process {Tn} contains an imbedded renewal process. Jn_1. oo). the Markov renewal process if for any n. distribution ofjXt}t>o itself where Pi refers to the case Jo = i..336 APPENDIX assume here that /the state space E is// finite) in the sense that P(Y..jEE is a family of distributions on (0. Then Xt 4 Xo. the semiregenerative process is called nonlattice if {T.t. where the distribution of X... X2.. G+(x) = P(S. Notes and references Renewal theory and regenerative processes are treated. Let X1. J1 i .g. Assume that uj = EjYo < oo for all j and that {J„} is irreducible with stationary distribution (v3)jEE. .7 Consider a nonlattice semiregenerative process. . and define r+=inf{n>0: Sn>0}. the conditional distribution of {XT„+t}t>o given Yo.. Jo... Jn = i is the same as the P.
and r_ preoccupation measures T+1 r_1 R+(A) = E E I(Sn E A). In (A.G+ * G_: (b) G_ (A) = f °° F(A .. More rigorously. (A.=EGn.r. 0<j<m.x)R_ (dx). oo) (A.1 . Proof Considering the restrictions of measures to (oc. (d) R+ = U_. (c) G+(A) = f °. . (e) R_ = U+. F(A) + (G+ * G_)(A). oo).7) (A. A C (0. define w as the time where the preT_ path S1. oo). 0]. A C (oo. A C (0.T_=n} = {S.APPENDIX 337 Probabilistic WienerHopf theory deals with the relation between F.x)R+(dx).7). Sr_ _1 is at its minimum . On {T_ > 2}. G+.g.=n w=m i Figure A. S. U. >0. we may rewrite (a) as G_ (A) = G+(A) = F(A) + (G+ * G_)(A). . A C (oo. F(A) is the contribution from the event {T_ = 1} = {X1 < 0}. G_.7) follows since G+(A) = 0 when A C (oo. F(A .>0. u . m<j<n}.8) (e.S. n=0 The basic identities are the following: Theorem A2. 0]). n=0 n=0 00 00 and the T+... we consider the last such time (to make w unique) so that {w=m.S. the renewal measures U+=>G+. n 0 R_(A) = E I(Sn E A). 0] and (0. 0).1 (a) F = G+ + G_ .
Aso.3. S.x)R+(dx).m. Sr_ E Adu) (s ee again Fig . SmEdu) = P(T+=m.. SnEAIS. 0<j<m. r+ = n) n=1 n=1 0  C0 E fF(Sk< 0.8) is similar. m < j <n. A._ = n ._ E A .0<k<ri .. Sn1 E dx) n=1  F(A ..= n.u) f0m m=1 n=m+1 00 J0 OO P(S. It follows that for n > 2 F (7.. .1) that P(Sj Sn.F(r_n_mSrEA_u).7) follows. ST_ E A . (b) follows from 00 G+ (A) _ E F(Sn E A..1).x)P(Sk < 0. S. 0 < k < n. m=1 f S mming over n = 2._ E A) n1 f P(r_=nw=m Sm EduSrEA) m=1 n1 F(r+=mSr+Edu).>0.1.+ E du) E P(S. ST_ E A) P(T+ = m.. and reversing the order of summation yields P(T_ > 2.+ E du)P(S.Sn_1Edx. (A.XnEAx) 00 f 0 f 0 00 00 1: F(A . clearly (Sj Sm>0.3 8 APPENDIX Reversing the time points 0. .du) (G+ * G)(A)• C llecting terms. E du) = P(T_=nm. ST+Edu). m it follows (see Fig.. and the proof of (A. .. A.
f.1(a) is from Kennedy [228].SnEA) = P(SnSn_ k.'s. such developments motivate the approach in Chapter VI on the Markovian environment model.g.SnEA) is the probability that n is a weak descending ladder point with Sn E A.0+[s])(1 . . In continuous time. u Remark A2. G_ are trivial.. For (d). it serves as model and motivation for a number of results and arguments in continuous time. WienerHopf theory is only used at a few places in this book. 6+ [s]. Since G+ is concentrated on (0. and using timereversion as in (d) to obtain the explicit form of R+ (Lebesgue measure). u Notes and references In its above discrete time version. 0]. see for example Bingham [65]. cf. However. there are direct analogues of Theorem A2.0<k<n.P as a product H+H_ of functions with such properties. which is basic for the PollaczeckKhinchine formula. For example. In discrete time. E.Sn_k. the survey [15] by the author and the extensive list of references there. we can rewrite (a) as 1 . Summing over n yields R+ (A) = U_ (A). G_ [s] are defined at the same time. Then for A C (oo.G_ [s] is defined and bounded in the halfplane is : ERs > 01 and nonzero in Is : ERs > 0}.9) whenever F[s].g. there is no direct analogue of Theorem A2. then T+ = inf It > 0 : St = 0} is 0 a.T+> n) = P(Sk < O. the analogue of a random walk is a process with stationary independent increments (a Levy process.g.1). Again.APPENDIX 339 and the proof of (c) is similar. consider a fixed n and let Xk = Xn_k+l. Nevertheless. P(SnEA .2 In terms of m. and sometimes in a larger strip. the derivation of the form of G+ for the compound Poisson model (Theorem 11. Another main extension of the theory deals with Markov dependence. In this generality of. this holds always on the line its = 0.G_[s]) (A. The present proof of Theorem A2.6.SnEA) = P(Sn<Sk. H+ (s) = 1G+[s] is defined and bounded in the halfplane Is : ERs < 0} and nonzero in Is: Rs < 01 (because IIG+lI _< 1). and similarly H_ (s) = 1 .s. see e. Sk = X1 + • • • + Xk = Sn . oo).O<k<n. 11. if {St} is Brownian motion. a number of related identities can be derived. being concentrated at 0. and the proof of (e) is similar. The classical analytical form of the WienerHopf problem is to write 1 .1. and G+.O<k<n.4). is based upon representing G+ as in (b).0<k<n..1.SnEA) = P(Sn<Sk.F[s] = (1 .
one needs to compute matrix inverses Q1 and matrix exponentials eQt ( r just eQ ).5 that when handling phase type distributi ons. It is seen from Theorem VIII.10) d dteAt = AeAt = eAtA (A. Here are. write eQ = (eK)m where = Q/m for some suitable integer m (this is the scaling step). Eo Kn/n! converges rapidly and can be evaluated without p oblems.12) eA'AO = Ale AA (A.11) A f eAtdt = eA. and eQ can then be computed as the mth power (by squaring if = 2). To circumvent this. Some fundamental properties are the following: sp(eA) = {e' : A E sp(A)} (A. _I 0 (A. 0 . JAI = max {Jjt : µ E sp(A)} and sp(A) is the set of all eigenvalues of A (the spectrum). 1. if m is s fficiently large. 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). ere A is the eigenvalue of largest absolute value. Here it is standard to compute matrixinverses by GaussJordan el imination with full pivoting . whereas there is no similar single established a proach in the case of matrix exponentials.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.13) henever A is a diagonal matrix with all diagonal elements nonzero.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. three of the c rrently most widely used ones: xample A3. Thus. however .
However . . 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 = q22.3 (DIFFERENTIAL EQUATIONS) Letting Kt = eQt. i.]t)n (A.e. the procedure consists in choosing some suitable i > 0. p different eigenvalues Aj i .. Let vi..7t) n=0 n! u °O n Pn (to see this.APPENDIX 341 Example A3. Zo = h).e. letting P = I + Q/i and truncating the series in the identity = e17t 00 Pn(. In practice.3 i (A. One then can reduce to p linear differential equations by noting that k = ZQ. The probabilistic reason that (A. 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 . .15) Then it is easily checked that P is a transition matrix . Zo = a (Z = QZ. Here is a further method which appears quite appealing at a first sight: Example A3 .2 (UNIFORMIZATION) Formally. To this end.. condition upon the number n of Poisson events in [Olt])  Example A3. vp be the corresponding left . The approach is in particular convenient if one wants eQt for many different u values of t. some jumps are dummy in the sense that no state transition occurs ). assume that Q is the intensity matrix for {Xt} and choose q with rt > max J%J = max qii• 1. Ap.. construction of {Xt} by realizing the jump times as a thinning of a Poisson process {Nt } with constant intensity 77.14) 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}. what is needed is quite often only Zt = TreQt (or eQth) with it (h) a given row (column) vector.14) holds is therefore that the tstep transition matrix for {fft} is eQt = E ent (. 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.4 (DIAGONALIZATION) Assume that Q has diagonal form..
18) Namely. In view of this phenomenon alone care should be taken when using diagonalization as a general tool for computing matrixexponentials. not all ai are real.. (A. of largest real part is often real (say. however. we can take H as the matrix with columns hl.16) (A. Then P P Q = > Aihivi = E Aihi (9 vi. Complex calculus : Typically. Everything is nice and explicit here: 411+q2+D' )12_g11+q2^^ where (411422z + 4412421. i= 1 i=1 P P (A.. say A = (Ai)diag. this last step is equivalent to finding a matrix H such that H1QH is a diagonal matrix.. Qhi = vihi. Nevertheless.18) contains terms which almost cancel and the loss of digits may be disasterous. (A. and hence A2 is so because of A2 = tr(Q).. say Al. i # j. we have an explicit formula for eQt once the A j. hp the corresponding right (column) eigenvectors. vi. There are. two serious drawbacks of this approach: u Numerical instability : If the A5 are too close. The phenomenon occurs not least when the dimension p is large.5 If Q= ( 411 ( q21 q12 q22 is 2 x 2.. v5Q = Aivi. hi have been computed.17) eQt = E e\`thivi = E ea:thi ® vi... the eigenvalue.g H1. D = ) 2 2 . i=1 i=1 Thus. under the conditions of the PerronFrobenius theorem). Then vihj = 0. and we may adapt some normalization convention ensuring vihi = 1. and writing eQt as eQt = He°tH1 = H (e\it)di. hp.342 APPENDIX (row) eigenvectors and hl. and vihi ¢ 0.. Example A3. some cases remain where diagonalization may still be appealing. and we need to have access to software permitting calculations with complex numbers or to perform the cumbersome translation into real and imaginary parts.
k  C k2 ) =b ( A1 q 1 Q11 / where a . i. Then 7r = (ir1 7r2 ) = a (q21 Al .e.19) Example A3 .7 Let 3 9 2 14 7 11 2 2 .k1). 1) . u Example A3.q. Then Al = 0 and the corresponding left and right eigenvectors are the stationary probability distribution 7r and e.21) Here the first term is the stationary limit and the second term thus describes the rate of convergence to stationarity.6 A particular important case arises when Q = q1 qi ) q2 q2 J is an intensity matrix. replacing ai by A2. However.APPENDIX 343 Write 7r (= v1) for the left eigenvector corresponding to a1 and k (= hl) for the right eigenvector. v2 and h2 can be computed in just the same way. l ab (g12g21 + (A1  411) 2) = 1. The other eigenvalue is A = A2 = q1 .20) ir = q2 ql qi +q 2 9l +q2 (A. Of course. it is easier to note that 7rh2 = 0 and v2k = 1 implies v2 = (k2 . b are any constants ensuring//Irk = 1. where (A. h2 = Thus. eqt = eNlt ( ir1ki i2k1 \ ir1 k2 72 k2 + e azt 7r2k2 i2k1 7ri k2 7r1 k1 (A.Q2i and after some trivial calculus one gets eQt = 7r 1 112 + eat 7r1 7r2 / (7fl 7r2) = ( 7r2 1r2 7r1 IF.
e_6u A4 Some linear algebra 4a Generalized inverses A generalized inverse of a matrix A is defined as any matrix A. APPENDIX x1 3/2 . (A. but only that dimensions match . 2 2 1=ab(142+(1+2)2 ) = tab. (AA+)' = AA+. (A+A)' = A+A.11/2 . (A.. A2 = 3/2 .22) Note that in this generality it is not assumed that A is necessarily square.6.11/2 + 5 1.23) . They are most often constructed by imposing some additional properties . Generalized inverses play an important role in statistics. and a generalized inverse may not unique. A+AA+ = A+.344 Then D= 2+ 11)' 7 T4 2 =52.5 . for example AA+A = 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_.satisfying AAA = A..
24) = te7r .P).g.= (I . if A is a possibly singular covariance matrix (nonnegative definite).1 goes under the name fundamental matrix of the Markov chain).23) is called the MoorePenrose inverse of A.eir)1 = I . = 0 where m < p is the rank of A. Then for some b > 0..e.I) (A.P + e7r ).. lt o eAx dx = te7r + D(eAt .ew. one is also faced with singular matrices .eir )1.. . Here is a typical result on the role of such matrices in applied probability: Proposition A4.eir ). Am > 0.e ® 7r)1.D + O(ebt). (A. ( Q .1 Let A be an irreducible intensity matrix with stationary row vector it. one then works with Q = (Q . . most often either an intensity matrix Q or a matrix of the form IP where P is a transition matrix.. 0 01 In applied probability. E. and can define /ail 0 0 0 0 0 0 A+ = C A' 0 0 0 C' . Assume that a unique stationary distribution w exists . and exists and is unique (see for example Rao [300]). (I ...25) . Am+1 = .P + e7r)1 (here ( I . _ A. then there exists an orthogonal matrix C such that A = CDC' where 0 0 D = AP Here we can assume that the A .g.1Q = Q(Q . and define D = (A . Rather than with generalized inverses .APPENDIX 345 A matrix A+ satisfying (A. These matrices are not generalized inverses but act roughly as inverses except that 7r and e play a particular role . are ordered such that Al > 0.
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. h ® it reduces to hit in standard matrix notation.91a(2) . .26) follows by integration by parts: t f t /' xeAx dx = [x {xe7r + D(eAx .h. B'(t) = e7r + DAeAt = eir + (I . h as 1 x m and k x 1 matrices.2e7r . (A. .346 t APPENDIX 2 xe Ax dx = eir + t(D + e7r) + D(eAt . respectively. I.DZ(ent .s.24). and the columns to h.e.s. Interpreting 7r. of (A. u 4b The Kronecker product ® and the Kronecker sum We recall that if A(1) is a k1 x ml and A(2) a k2 x m2 matrix.I)}. Note that h ® it has rank 1. resp.I)} dx. the rows are proportional to it.. the r.D + D2 + O(ebt). o Finally. it follows that h ® it is the k x m matrix with ijth element hi7rj .J {xe^r + D(e . Equivalently.27) Proof Let A(t). in block notation i2h A®B= ( a11B a21 B a12B a22 B Example A4. Then A(O) _ B(O) = 0.h. B(t) denote the l.I) (A.26) 2 = 2 e7r + tD .2 Let it be a row vector with m components and h a column vector with k components. (A.3 Let 2 A= 4 3 Vf' N7 5 )' B= ( 8 ). the formulas involving O(e6t) follow by PerronFrobenius theory.I) . For example.eir)eAt = eAt = A'(t). and in fact any rank 1 matrix can be written on this form. ()®(6 f 6/ 7f 8^ 7 8 )=! ^)( 6 7 8 )=(6^ 7^ 8^) \ u Example A4. see below.
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.3vV/72f 20.3V8. (A. and the number of such factors is precisely given by the relevant binomial coefficient.4vf. then v1B1h1 and v2B2h2 are real numbers.4 eA® B = eA ®eB.3f 4v/. if Al = vi.31) Indeed.31). (A.29).28) In particular. it follows that e® ® e B An _ 0o oo oo Bn 7 I F n! = ` k! (I . Using (A.(A. (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 .29) If A and B are both square (k1 = ml and k2 = m2). such a factor is Ak (&B 1k according to (A.30) eA+B = eAeB function generalizes to Kronecker notation (note that in contrast typically only holds when A and B commute): Proposition A4.5v'8 5vf9 11 A fundamental formula is (A1B1C1) ®(A2B2C2) = (A1 (9 A2)(B1 (9 B2)(C1®C2). A2 = v2 are row vectors and C1 = h1. C2 = h2 are column vectors.k)! ( n0 n=0 t=0 k=0 J _ ® Ak ®Blk r ^.5v/. and v1B1h1 • v2B2h2 = v1B1h1 ® v2B2h2 = ( v1(&v2 )( B1(&B2 )( h1(&h2 ) . each of which is A ® I or I ® B.A9.50 6 7 6 4f 4. if A ® I occurs k times.3v'6.APPENDIX 347 Then A®B = 2 f 20. then the Kronecker sum is defined by A(1) ®A(2) = A(1) ®Ik2 + k ®A(2).
represents ces Q( 1). Let P8f P(Sl).32). first term on the r . Then 2 0 ire At h • ve Bt kdt = (^®v)(A®B)1(e A®Ba . and the form of the bivariate intensity matrix reflects the fact that Yt(2) } cannot change state in both components at due to independence . P(2). where transition matrix of the bivariate Markov chain {X n1). From what has been said about matrices of {Yt( 1). Let further it. h. Thus .32) is the intensity matrix of the bivariate continuous Markov process {Yt(1). P8 = Pal ) ® P82) exp {Q ( 1) ® Q(2)1 = eXp {Q( 1) } ® exp {Q(2) } Also the following formula is basic: B are both square such that a +. A special case of Proposition A4. P(t) Yt(2) }. P8 = exp {sQ} = exp {s (Q(1) ®Q(2)) } .3 < 0 Lemma A4 . { 1't(1) }. the same time. Q(2). Yt(2) where independent Markov processes with intensity matri{y(2) } are {Y(1) }.4 can easily be obtained by probabilistic be the sstep transition reasoning along the same lines . X ) }.348 APPENDIX Remark A4.6 Suppose that A and of B. v whenever a is an eigenvalue of A and 0 is an eigenvalue be any row vectors and h. n2 n1 ) {X(2) } are independent Markov chains with transition matrices P(1). {Yt(1). in the definition (A. and Q = Q(1) ® Q (2) = Q(1) ® I + I ® Q(2) (A.s. k any column vectors.5 Many of the concepts and results in Kronecker calculus have p(2) is the intuitive illustrations in probabilistic terms. we have P8 = Pal) ® p(2). the {Yt(2) } transitions in the {Yt(1) } component and the second transitions in the component .33) . independent Markov chains. Yt(2 ) }.I)(h ® k). (A. resp . p = P(1) ® {X }. { On the other hand. Ps 1) = exp {sQ ( 1) } > p(2 ) = exp {sQ(2) } can therefore be rewritten as Taking s = 1 for simplicity .
. h such that vh = 1. in such that io = i. Now note that the eigenvalues of A ® B are of the form a +. h = e and v = 7r (the stationary row vector). n.7 Let A be a p x pmatrix with nonnegative elements.. A is called aperiodic if the pattern of zero and nonzero elements is the same as for an aperiodic transition matrix.. f o r each i. we have AO = 1. h can be chosen with strictly positive elements.34) Note that for a transition matrix. Then the eigenvalue Ao with largest real part is simple and real.The PerronFrobenius theorem has an analogue for matrices B with properties similar to intensity matrices: Corollary A4.g. . That is. . which can be found in a great number of books. .3 whenever a is an eigenvalue of A and 3 is an eigenvalue of B. . j = 1. the integrand can be written as ( 7r (9 v)( eAt ® eBt )(h ®k ) = ( 7r ®v)(eA (DBt)(h (& k). Here is the PerronFrobenius theorem. [APQ] X. . . il..APPENDIX 349 Proof According to (A. . (A.8 Let B be an irreducible3 p x pmatrix with nonnegative offdiagonal elements. (b) if in addition A is aperiodic. and the corresponding left and right eigenvectors v. . then An = Aohv+O(µ") = Aoh®v+O(µ") for some u. we mean that the pattern of nonzero offdiagonal elements is the same as for an irreducible intensity matrix. so that by asssumption A ® B is u invertible. i. E (0.12). = j and atk_li. Similarly.... then IN < Ao for all A E sp(A). and the corresponding left and right eigenvectors v.29). and if we normalize v.1 and references there (to which we add Berman & Plemmons [63]): Theorem A4. and appeal to (A. 4c The PerronFrobenius theorem Let A be a p x pmatrix with nonnegative elements.. h can be chosen with 3By this. > 0 for k = 1. see e. Then: (a) The spectral radius Ao = max{JAI : A E sp(A)} is itself a strictly positive and simple eigenvalue of A. ao). p there should exist io. We call A irreducible if the pattern of zero and nonzero elements is the same as for an irreducible transition matrix.
35) for some p E (oo. Proposition A5. I. The next result gives a condition for asymptotical exponentiality. To this end. Note that for an intensity matrix.n t AL n=0 n! (cf.1. if we normalize v. The content is that B is approximately exponential if the exit rates ti are small compared to the feedback intensities tij (i # j).e.. Corollary A4. but is an easy consequence of the PerronFrobenius theorem. not only in the tail but in the whole distribution.350 APPENDIX strictly positive elements. the analogy of this procedure with unformization. h such that vh = 1. relate the eigenvalues of B to those of B via (A.(3. we have A0 = 0.1 Let Q be a proper irreducible intensity matrix with stationary distribution a. then eBt = ea0thv + O(eµt) = eA0th ® v + O(et t) (A. For example. 10) and use the formula me at e Bt = e 00 Antn = e . T(°)) is asymptotically exponential with parameter t* _ r EiEE aiti as a 4 oo. let t = (ti)iEE # 0 have nonnegative entries and define T(°) = aQ . Example A3. h = e and v = 7r (the stationary row vector).8. let {Yti°i } be a Markov process with initial distribution a and intensity . Then for any (3.(ti)ding.2). the condition is that t is small compared to Q. note that we can write the phase generator T as Q . the phasetype distribution B(a) with representation (. it was shown that under mild conditions the tail of a phasetype distribution B is asymptotical exponential. A5 Complements on phasetype distributions 5a Asymptotic exponentiality In Proposition VIII. Furthermore.(ti)diag where Q = T + (ti)diag is a proper intensity matrix (Qe = 0). Ao).8 is most often not stated explicitly in textbooks. Bi° (x) + at*x Proof Let { 4 } be the phase process associated with B(a) and (°) its lifelength. 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 can think of ( ( a) as the first event in an inhomogeneous Poisson process ( Cox process ) with intensity process matrix aQ .x (1 . Proof Assume first ti > 0 for all i and let I.)_ = Y(a) = 1'aS(a) = Ya(av)^ it follows that Pi ((. and write Yt = Yt(1). a . We can assume that Jta) = Yt(°). a'/a + 1.(a) > x . Hence O ((a) aa. a' = a . 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. from which the phase process is terminated .g.Yj(av) = j f .YQ(av) = j) Pi ( ci(a'V) > x.a' + oo (e. dx/ti] or not. t < (a). We shall . has a limit distribution: Proposition A5. Let further V be exponential with intensity V and independent of everything else. 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}.jEE. fo tY dv/t a$' t*. Then a(a'V)/a (aV) a' 1. J^O)_ = j) Pi (v(aaV) > x. and that Yt(a) = Yat for all t. J(()) _ = i) + at•x t tt' . Conditioning upon whether { Yt} changes state in [0.aE where 0 < e < 1). v/ t. {t Y( a) } v>0 . and this easily yields a(x)/x a' 1/t*. it states that the state.bij) Hence the intensity matrix of { Ix} is (qij/ti)i. In addition to the asymptotic exponentiality. prove a somewhat more general result which was used in the proof of Proposition VI. By the law of large numbers for Markov processes .1.9.2 Pi (c(a) > x. = YQ(x). Since JJ(. in fact . Then {Ix} is a Markov process with to = Yo.APPENDIX 351 ((1) etc. we get dx F (Idx = j) = (1 + qij t )Sij + qij dt.
j) and initial distribution a.. 1 k=1 1 0 otherwise.352 rr Ia(a'V) Ei I ( > x) P APPENDIX L at (Yo (aV) .. 5b Discrete phasetype distributions The theory of discrete phasetype distributions is a close parallel of the continuous case.Pe.. a). However. Then: (a) The point probabilities are bk = aPklp. A distribution B on {1.. Keilson [223].. P.. (b) the generating function b[z] _ E' .5 Let B be discrete phasetype with representation (P. See also Korolyuk. Example A5. K}.. these results are in the spirit of rare events theory for regenerative processes (e. Example A5. so is the geometric distribution. . the simplest discrete phasetype distribution: here E has only one element. k = 1. 2. so we shall be brief. an easy modification of the argument yields finally the result for the case where t. . ' pk 0 k>1 11 Theorem A5.. u Notes and references Propositions A5. with point probabilities bk = (1 . Indeed. and thus the parameter p of the geometric distribution u can be identified with the exit probability vector p. zkbk is za(I . Penev & Turbin [238]. Gnedenko & Kovalenko [164] and Glasserman & Kou [162]). k>1. say bk = 0. .g.3 As the exponential distribution is the simplest continuous phasetype distribution. Then P is substochastic and the vector of exit probabilities is p = e .1 and A5. let E and Pkj j=k1. is discrete phasetype..4 Any discrete distribution B with finite support.} is said to be discrete phasetype with representation (E. (c) the nth moment k 1 k"bkis 1)"n!aP"p.x k > K. = 0 for one or more i.+ at*x • a't' L ` at t* t* J Reducing the state space of {Ix } to {i E E : t.zP)'p. 2.. a = b = (bk)k=1. Et II I a(a^V) > x) at' . a) if B is the lifelength of a terminating Markov chain (in discrete time) on E which has transition matrix P = (p..p)k1 p. > 0}.2 do not appear to be in the literature.
T(2)). The discrete counterpart is the negative binomial distribution with point probabilities bk k1) (1 k = r.{ 0. T= ( 0 T(2) ) (A. a' .. resp. 11 Example A5..r + 1. U2. and hence the negative binomial distribution is discrete phaseu type.a(2). A. a. resp.36) in blockpartitioned notation (where we could also write a as (a (1) 0)). as is seen by minor modifications of Example A5. initial distribution a and phase generator T. Jt t > U1 + U2.. (E(2).6 (CONVOLUTIONS) Let B1. and piece the processes together by it = 41) 0<t<U1 U1 < t < U1 + U2 2U.APPENDIX 353 5c Closure properties Example A5. Then {Jt} has lifetime U1 + U2 . r .6. _ i E E(1) T(1) t(1)a(2) i E E(2) . . and a=1). B2 be phasetype with representations (E(1). Then the convolution B = B1 * B2 is phasetype with representation (E. A reduced phase diagram (omitting transitions within the two blocks) is am E(1) t(1) a(2) (2) t(2) Figure A.7 (THE NEGATIVE BINOMIAL DISTRIBUTION) The most trivial special case of Example A5.T(1)). T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2). { Jt 2) } with lifetimes U1 .1 This corresponds to a convolution of r geometric distributions with the same parameter p.a(1).6 is the Erlang distribution Er which is the convolution of r exponential distributions..2 The form of these results is easily recognized if one considers two independent phase processes { Jt 1) }.
T. i E E(1) T 0 I (A.0)a(2) E(2) Figure A.T(1)).37) (1) (1 .p)pn1B*n. Let B(") be the corresponding phasetype distribution. then C is the distribution of Ul + • • • + UN.i. Thus. and o'i Oa.9 (INFINITE MIXTURES WITH T FIXED) Assume that a = a(°) depends on a parameter a E A whereas E and T are the same for all a. a. one obvious interpretation of the claim u size distribution B to be a mixture is several types of claims. Then it is trivial to see that B(") is u phasetype with representation (a(").p)pn1. Example A5.0)a(2))). A reduced phase diagram is 0a(1) E(1) A .354 APPENDIX Example A5..'). (E(2).d. U2.p.4 . B2 be phasetype with representations (E(1). with common distribution and N is independent of the Uk and geometrically distributed with parameter p. a.10 (GEOMETRIC COMPOUNDS) Let B be phasetype with representation (E.8 (FINITE MIXTURES) Let B1.a(2). a reduced phase diagram is f a E t Figure A.3 In exactly the same way.O)B2 (0 < 0 < 1) is phasetype with representation (E. Example A5. are i. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2).a(1). we need to restart the phase process for B w. Equivalently. In risk theory. Then the mixture B = 9B1 + (1 . P(N = n) = (1 . a mixture of more than two phasetype distributions is seen to be phasetype. To obtain a phase process for C. and consider B(") = fA B(a) v(da) where v is a probability measure on A.0)ai2). i E E(2) 0 T(2) =IT (in blockpartitioned notation. if U1. resp.. this means that a = (Oa(1) (1 .. T) and C = EO°_1(1 .E) where a(°) = fAa(a)v(da). p at each termination.T(2)).
U2. if U1. 13 (MINIMA AND MAXIMA ) Let U1. E). Then the minimum U1 A U2 and the maximum U1 V U2 are again phasetype..APPENDIX 355 and C is phasetype with representation (E. Corollary VIII. i.g. of F. it follows by mixing (Example A5. then U1 + • • + UN is zeromodified phasetype with representation (a. are i. { 4 } as exit of {Jt}. be the point probabilities of a discrete phasetype distribution with representation (E. cf. T + ta.aF[T].a(1). T) and C = F.1. Minor modifications of the argument show that 1. we then let the governing phase process be {Jt} _ {(411 Jt2))} 2) interpreting exit of either of {4 M }.v. T(2) ). j E F}. a. Proposition VIII. v. let {Jtl)}. It is zeromodified phasetype with representation (E. a(2). then Jy has distribution aeTx.2. 12 (PHASETYPE COMPOUNDS ) Let fl. If we replace x by a r.9) that (U .2.f. P). let the phase space be E x F = {i j : i E E. a. if {Jt} is a phase process for U. if B is defective and N + 1 is the first n with U„ = oo.. T + pta). say with distribution F. To see this. If U1 has a different initial vector.. Indeed. Example A5 .aeTx. U2 be random variables with distributions B1..7. +UN 2.TWWW). Thus the representation is (E(1) x E(2). let B be a continuous phasetype distribution with representation (F.d. (E(2). then U1 +• is phasetype with representation (E.. Example A5 . say v.T + pta). v. resp. a(1) ® a(2 ). with common distribution B and N is independent of the Uk with P(N = n) = f. { Jt2) } be independent with lifetimes U1. Example A5. T(1) ® T(2)). . f2. B2 of phasetype with representations (E(').x)+. U2.°. cf. a..°_1 f„ B*?l. . let the initial vector be a ® v and u let the phase generator be I ® T + P ® (ta). then C is the distribution of U1 + • • • + UN.11 (OVERSHOOTS) The overshoot of U over x is defined as the distribution of (U . T)... but the same T. resp. Note that this was exactly the structure of the lifetime of a terminating renewal u process. Equivalently.X)+ is zeromodified phasetype with representation (E. To obtain a phase representation for C .T) if U is phasetype with representation (E. For U1 A U2.T) where F[T] = J0 "o eTx F(dx) u is the matrix m. X independent of U.
we can assume that ID.8. Then we must find phasetype distributions Bn with B.xq(n)(n)}. The mean of B„ is n/Sn = b and the variance is n/Sn = b2/n. cf. say degenerate at b. Now we can find first a sequence {Dm} of distributions with finite support such that D. and let Bn be the Erlang distribution E. relies more on matrix algebra than the probabilistic interpretation exploited here).. r # oo. Here are the details at two somewhat different levels of abstraction: (diagonal argument . Thus the state space is E(1 ) x E(2) U E(1) U E( 2).n = I:pi(n)Er v ( __ ) n) ) a= 1 .B(bk) I < 1/n for n > k. however. oo).(n) = D. see Neuts [269] (where the proof. oo) can be approximated 'arbitrarily close' by a phasetype distribution B: Theorem A5.356 APPENDIX For U1 V U2.. Then from above. That is.14 To a given distribution B on (0.} of phasetype distributions such that Bn 3 B as n + oo. q(n) q(n) pi(n)a .(bk) + B(bk) for all k as n * oo. i= 1 C. and the closedness of the class of phasetype distributions under the formation of finite mixtures. there is a sequence {B..(bk)'. the initial vector is (a(1) (& a (2) 0 0). 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 . with weight pi(n) for xi(n). Let the support of Dn be {xl(n). the fact that any distribution B can be approximated arbitrarily close by a distribution with finite support. 5d Phasetype approximation A fundamental property of phasetype distributions is denseness . Example A5.. The general case now follows easily from this..2) } to go on (on E(2)) when { i 1) } exits. we need to allow { Jt... Hence it is immediate that Bn 4 B. elementary) Let {bk} be any dense sequence of continuity points for B(x).(Sn) with Sn = n/b.. By the diagonal argument (subsequent thinnings). and vice versa.(bk) + B(bk) for all k. any distribution B on (0. Proof Assume first that B is a onepoint distribution.
14 is fundamental and can motivate phasetype assumptions. for some a < oo. we can then approximate Bo by a phasetype B..D(bk)I < n. Hence G C PET and L = PIT... u 2 (abstract topological ) The essence of the argument above is that the closure (w. It should be noted. Let E be the class of functions f : [0.n (bk) . then it is immediate that WI(B) = p2(B) for all distributions B on [0. oo). the class CO of all discrete distributions. If Cpl (B) and ^02(B) are weakly continuous. and we can take Bn = Cr(n)..e. oo) and any fl. compute W(B) and use this quantity as an approximation to cp(B0)..t. Since PET is closed under the continuous operation of formation of finite mixtures. oo) * [0. however.APPENDIX 357 Hence we can choose r(n) in such a way that ICr( n).. Then ICr( n ). one would use the B given by some statistical fitting procedure (see below). In particular. E E.. But To is the class G of all distributions on [0. For a general Bo. oo) approximation Assume that we can compute a functional W(B) when B is phasetype. that this procedure should be used with care if ^p(B) is the ruin probability O(u) and u is large. replications). Corollary A5.n( b k ) .d.(x)Bf. k < n. the topology for weak convergence) PET of the class PET of phasetype distributions contains all onepoint distributions.r. f2.15 To a given distribution B on (0 .. PIT contains all finite mixtures of onepoint distributions. i.n.i. x 4 oo. 2. . i = 1. u Theorem A5. k < n.B(bk )I < .( dx) * f r f{(x)B(dx). say on the claim size distribution B in risk theory. oo) such that f (x) = O(e«x). and that cp is known to be continuous. 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. there is a sequence {Bn} of phase type distributions such that Bn Di B as n 4 oo and f ' f. if information on Bo is given in terms of observations (i.
\\ 0 Corollary A5.358 Proof By Fatou' s lemma. n B=az. . n. and hence we may choose r(n) such that L 9l) f (x)Cr(n). i = 1. then cc f (x)Bn ( dx) = (?!c ) e'= . Now returning to the proof of (A. By (A. i = 1.. i=1.n(dx) < 1+. if f (x ) = e°x.. . and hence it is sufficient to show that we can obtain limsup n4oo fi(x)Bn(dx) < Jo 0 f fi( x)B(dx ).. for each i..38) We first show that for each f E E.n(dx) + f 0 fi(x)Dn(dx)...16 To a given distribution B on (0 .(dx) > J fi(x)B(dx).39) Indeed. f00 fi(x)Cr. Bn=En z f f (x)Bn(dx) fof (x)B(dx) = ° (A. oo).14 Dn has been chosen such that 00 1 °° f fi(x)D n(dx ) < 1++ '  o \ n o f fi(x)B(dx). and the case of a general f then follows from the definition of the class E and a uniform integrability argument.f (x)B(dx).38 ). TO (A. . .f (z) = f = 1 1 1 1n/ o ..  APPENDIX B implies that 00 o o 00 n. .. liminf B. there is a sequence {Bn} of phase type distributions such that Bn Di B as n + oo and all moments converge..2 .. n. we may assume that in the proof of Theorem A5. 2.. i = 1..39).f ' f (x)B(dx).. f° xtBn(dx ) * f °° x`B( dx).oo J fi(x)B.
0 as i * oo. However. e ) and ei J.18 In the setting of Corollary A5. from a more conceptual . the loggamma or the Weibull have been argued to provide adequate descriptions of claim size distributions. /3) = ry for all n. Proof Let fi(x) = el'r+E. the adjustment coefficient 'y = 7(B. .} of phasetype distributions such that Bfz + B as n * oo and Yn 4 ry where ryn = y(Bn.> y for some sequence {ei} with ei E (0./3) is defined as the unique solution > 0 of B[y] = l+y/j3. and therefore the following result is highly relevant as support for phasetype assumptions in risk theory: Corollary A5.17 To a given /3 > 0 and a given distribution B on (0. The present section is a survey of some of the available approaches and software for inplementing this..16. O We state without proof the following result: Corollary A5.14 is classical. 5e Phasetype fitting As has been mentioned a number of times already. The adjustment coefficient is a fundamental quantity. If ei > 0. and in part from the fact that many of the algorithms that we describe below have been formulated within the setup of fitting distributions. lim inf > is proved similarly. We shall formulate the problem in the slightly broader setting of fitting a phasetype distribution B to a given set of data (1i . (N. I. . (N or a given distribution Bo. . This is motivated in part from the fact that a number of nonphasetype distributions like the lognormal.. .3). lim sup ryn < 7. .. oo) with B[y +e] < oo for some e > y = 7(B. . For practical purposes. . there is substantial advantage in assuming the claim sizes to be phasetype when one wants to compute ruin probabilities. one can obtain 7(Bn.l3µb < 1.APPENDIX 359 In compound Poisson risk processes with arrival intensity /3 and claim size distribution B satisfying . the problem thus arises of how to fit a phasetype distribution B to a given set of data (1.e. then Bn['Y + ei] * B[y + ei] > 1 + 7 Q implies that 'yn < ry + ei for all sufficiently large n . the remaining results may be slightly stronger than those given in the literature. Notes and references Theorem A5./3). but are certainly not unexpected. there is a sequence {B.
and as fitted distribution we may take B. where more than two Erlangs are allowed and in addition to the exact matching of the first three moments a more general deviation measure is minimized (e. three for a mixture of two Erlangs ). and used a nonlinear programming approach .. [202]. [216] ).} of phasetype distribution such that Bo.d. A number of approaches restrict the phase type distribution to a suitable class of mixtures of Erlang distributions . . B„ The problem is that the constructions of {B„} are not economical : the number of phases grows rapidly.g. The characteristics of all of these methods is that even the number of parameters may be low (e. 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.f. Schmickler (the MEDA package.g. [317] ) has considered an extension of this setup. (N is the empirical distribution Be. g. [70]) restrict attention to acyclic phase type distributions . a program package written in C for the SUN workstation or the PC is available as shareware. In a series of papers (e. the number of phases required for a good fit will typically be much larger. the L1 distance between the c . and in practice this sets a limitation to the usefulness (the curse of dimensionality . A method developed by Bobbio and coworkers (see e. for some suitable large n. 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 earliest such reference is Bux & Herzog [85] who assumed that the Erlang distributions have the same rate parameter.. The likelihood function is maximized by a local linearization method allowing to use linear programming techniques. cf. risk theory. we have constructed a sequence { B.f. d. Asmussen & Nerman [38] implemented maximum likelihood in the full class of phasetype distributions via the EM algorithm . 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.. at a a number of selected points .. we do not not want to perform matrix calculus in hundreds or thousands dimensions). Of course.360 APPENDIX point of view the two sets of problems are hardly different : an equivalent representation of a set of data (1 . Johnson & Taaffe considered a mixture of two Erlangs (with different rates ) and matched (when possible ) the first three moments . .g .'s). giving mass 1 /N to each S=. and this is what matters when using phasetype distributions as computational vehicle in say renewal theory. It seems therefore a key issue to develop methods allowing for a more general phase diagram. defined by the absence of loops in the phase diagram . The observation is that the statistical problem would be straightforward if the whole ( EAvalued) phase process { Jtk)} o<t<( k associated with each observa . e .g.
Thus.g.. .T(n) (Nik IC1. jEEA.x)t(n) 1 and this and similar expressions are then computed by numerical solution of a set of differential equations. (N ) (^ 54 k )+ and similarly for the cn+1) The crux is the computation of the conditional expectations. .T (n)(TiI(1. one is lead to an iterative scheme. the methods of [70] and [38] appear to produce almost identical results.. In practice. (n+1) _ Ea (n). (N) tJk Ea ( n). then the estimators would be of simple occurenceexposure type. since this is parameterdependent. E.APPENDIX 361 tion Sk was available. = j) f k=1 k =1 tE[0. N Ti = I(J= i) dt.. it seems open whether the restriction to the acyclic case is a severe loss of generality..g. Nii = = . EN where ai = N 1 I ((k) = i) tii=i iEE.T(n) k=1 I (Jti) dt o \f a(n)eT(n )(kt(n) N f:i a(n)eT(n)xei ..T(n) (Ti ^^ 1. The general idea of the EM algorithm ([106]) is to replace such unobserved quantities by the conditional expectation given the observations. it is easy to see that N (k Ea(n). ... e... In fact. . .(N) = E Ea(n). eieT(n)((k.(k] (Ti is the total time spent in state i and Nii is the total number of jumps from i to j).
This page is intentionally left blank .
Whitt (1998) Explicit M/G/1 waitingtime distributions for a class of longtail service time distributions . Whitt ( 1992) The Fourierseries method for inverting transforms of probability distributions . Abate. with applications to risk reserve processes and the GI /G/1 queue. 587. The Mathematical Scientist 14. 219229. 101116. pp. 213229. Preprint. Appl. [8] G. 53100. Borokov. Asmussen (1984) Approximations for the probability of ruin within finite time . Stegun ( 1972 ) Handbook of Mathematical Functions (10th ed. Chichester New York. 14. 1985. [4] M. [3] J. [2] J. Abramowitz & I. 37. New York. [14] S. [5] G. Asmussen ( 1982 ) Conditioned limit theorems relating a random walk to its associate. Act.A. Choudhury & W. In: Limit Theorems and Related Problems (A. [15] S. Abate & W. [13] S. Whitt (1994) Waitingtime tail probabilities in queues with longtail service time distributions . [7] E. ibid. 20. Aktuar Tidskr. Dover.). [9] G. Teubner. 253267. 191223. Stoch. 345368. The case of equal risk sums. AT&T. [10] K. [6] V. Adv. Stuttgart. Aktuar Tidskr. New York. Asmussen (1989a) Aspects of matrix WienerHopf factorisation in applied probability. Asmussen (1987) Applied Probability and Queues.G. 1984 . ed. Proc. Sparre Andersen (1957) On the collective theory of risk in the case of contagion between the claims . [11) S. 143170. Probab. Anantharam (1988 ) How large delays build up in a GI/GI11 queue. John Wiley & Sons. Arfwedson ( 1955) Research in collective risk theory. 311338. Skand. Queueing Systems 10. Scand.L. 57. J. Abate & W. II.Bibliography [1] J. Skand. Arndt (1984) On the distribution of the supremum of a random walk on a Markov chain . Transactions XVth International Congress of Actuaries. 3157 . Queueing Systems 16. [12] S. 363 . Appl.). B. Alsmeyer (1991) Erneuerungstheorie. 38. G. Queueing Systems 5 . Optimizations Software. New York. Asmussen (1985) Conjugate processes and the simulation of ruin problems. Arfwedson ( 1954) Research in collective risk theory.
Binswanger (1997) Simulation of ruin probabilities for subexponential claims.).). Probabilistic Analysis of Rare Events (V. Act. MatrixAnalytic Methods in Stochastic Models (A. [20] S. Ann. Marcel Dekker. Asmussen (1995) Stationary distributions via first passage times. Appl. J. 1999 . [17] S. SIAM Review 40. Asmussen (2000) Matrixanalytic models and their analysis. 1989 . 25. 313326. Kliippelberg (1994) Large claims approximations for risk processes in a Markovian environment. Act. Statist. Act. Asmussen (1989b) Risk theory in a Markovian environment. 2. 1996. [18] S. Asmussen (1998a) Subexponential asymptotics for stochastic processes: extremal behaviour. Appl. Chakravarty. 54. 37. [32] S. Probab. T. CRC Press. Proc. Asmussen (1992c) Stationary distributions for fluid flow models and Markovmodulated reflected Brownian motion. J. 96107. [27] S. 4966. Proc. [19] S. eds. Florida. Th. 313341.M. Bladt (1996) Renewal theory and queueing algorithms for matrixexponential distributions. J. Asmussen (1992a) Phasetype representations in random walk and queueing problems. Scand. Probab. Ann. Hojgaard (2000) Rare events simulation for heavytailed distributions. [26] S. Asmussen (1998c) A probabilistic look at the WienerHopf equation. 297318. Boca Raton. eds. Asmussen & M. Random Processes 2. Frey. [31] S. Riga Aviation University. [22] S. 8. Advances in Queueing: Models.). 2943. 555574. [24] S. Appl. Asmussen. 354374. Bladt (1996) Phasetype distributions and risk processes with premiums dependent on the current reserve. Probab. Ann. 1936. A. stationary distributions and first passage probabilities. Asmussen (1991) Ladder heights and the Markovmodulated M/G/1 queue. 315. Schmidt (1995) Does Markovmodulation increase the risk? ASTIN Bull. Appl. Alfa & S. Asmussen & M. Astin Bulletin 27. New York. Asmussen.S. [21] S. Extremes 1 . 189201. [33] S. [23] S. Asmussen & B. 79102. 106119. 20. Methods V Problems (J. [29] S. Floe Henriksen & C. K. Stoch. 69100. Asmussen (1998b) Extreme value theory for queues via cycle maxima.K. Asmussen (1992b) Light traffic equivalence in single server queues. J. Bernoulli 6. Asmussen & B. Dshalalow ed. Asmussen (1999) On the ruin problem for some adapted premium rules. [25] S. Andronov. Stoch. Hojgaard (1996) Finite horizon ruin probabilities for Markovmodulated risk processes with heavy tails. 193226. Scand. Hojgaard (1999) Approximations for finite horizon ruin prob abilities in the renewal model. Asmussen. L. [28] S.364 BIBLIOGRAPHY [16] S. Binswanger & B. [30] S. [34] S. 2149. Stochastic Models 11. Scand. . Rolski & V. 772789. 137168. 303322. Asmussen & K. Kalashnikov & A. Scand. 27.
419441. [52] S. Asmussen & R. Probab. 422447. [40] S. Taksar (2000) Optimal risk control and dividend distribution policies . Asmussen & C. Adv. Appl. Asmussen & C. Opns. Methods 8 Problems (J. 120. Oper. 103125. Asmussen & V. [43] S. Probab. Johnson. 105119. Math. Scand. [39] S. Asmussen & C. 23. Appl. B. (to appear). 58. Wiley. O'Cinneide (2000/2001) Matrixexponential distributions [Distributions with a rational Laplace transform] Encyclopedia of Statistical Sciences. [45] S. Statistica Neerlandica 47. Asmussen & K. H. Adv. [37] S. [46] S. Asmussen & S. 19. 429466. 31. Asmussen . Insurance: Mathematics and Economics 20. 11251141. 410433. Nerman & M. Rolski (1991) Computational methods in risk theory: a matrixalgorithmic approach. Statist.Y. Management Science 45. Fitting phasetype distributions via the EM algorithm. Asmussen & H. Schmidt (1995) Ladder height distributions with marks. Asmussen & M. Stoch. Supplementary Volume (Kotz. [42] S. O'Cinneide (2000/01) On the tail of the waiting time in a Markovmodulated M/G/1 queue. Schmidt (1999) Tail approximations for nonstandard risk and queueing processes with subexponential tails. Asmussen. Sigman (1996) Monotone stochastic recursions and their duals. Appl. Example of excessoffloss reinsurance for an insurance corporation. [50] S. Asmussen & G. Stochastic Models 8.BIBLIOGRAPHY 365 [35] S. Asmussen & V. O. Eng. Advances in Queueing: Models. Schmidt (1993) The ascending ladder height distribution for a class of dependent random walks. Taksar (1997) Controlled diffusion models for optimal dividend payout. Sc. Proc. Asmussen & R. 299324. Olsson (1996). J. Perry (1992) On cycle maxima. [51] S.Y. Read eds. 913916. Probab. Boca Raton. Appl. Asmussen & T. 259274. 20. Finance and Stochastics 4. [36] S. Kliippelberg (1996) Large deviations results for subexponential tails. Asmussen & T. Hojgaard & M. Probab. 365372. first passage problems and extreme value theory for queues. Proc. Asmussen. Schmidli & V. Rubinstein (1999) Sensitivity analysis of insurance risk models. Dshalalow ed. Res.M. Res. Appl.). [38] S. Th. 64. Insurance: Mathematics and Economics 10. . Schock Petersen (1989) Ruin probabilities expressed in terms of storage processes. CRC Press.A. 736755. [49] S. Asmussen & D. [41] S. 421458. Rubinstein (1995) Steadystate rare events simulation in queueing models and its complexity properties. J.). with applications to insurance risk Stoch. 115. Koole (1993). [48] S. 30. Florida. Appl. 10. J. Marked point processes as limits of Markovian arrival streams. Nielsen (1995) Ruin probabilities via local adjustment coefficients . [44] S.A. Inf. Rolski (1994) Risk theory in a periodic environment: Lundberg's inequality and the CramerLundberg approximation. [47] S. Prob 32 .
von Bahr (1974) Ruin probabilities expressed in terms of ladder height distributions. J. 275279. Probab. J.T. Teugels ( 1996) Convergence rates for M /G/1 queues and ruin problems with heavy tails . Goldie & J. BarndorffNielsen (1978) Information and Exponential Families in Statistical Theory. Bjork & J. J. Bobbio & M. C. Act. [66] N. 77111. Insurance: Mathematics and Economics 9. 190204. [56] B. [59] O. Adv. New York.R.B . SpringerVerlag.L. Act. Cambridge University Press.A. 832840. von Bahr (1975) Asymptotic ruin probabilities when exponential moments do not exist. [64] P. Probab. Bingham. Beekman (1974) Two Stochastic Processes.J. 9. 1995. J. [70] A. [57] C.M. [67] T. Grandell (1988) Exponential inequalities for ruin probabilities in the Cox case. [65] N. 610. Appl. Boogaert & A. Teugels (1987) Regular Variation. Borovkov (1976) Asymptotic Methods in Queueing Theory. Athreya & P. 7. Stochastic Models 10. 1985 . Trans. Act. Asmussen & J. Appl. Act. Berman & R. Crijns (1987) Upper bounds on ruin probabilities in case of negative loadings and positive interest rates. New York.H. Ney ( 1972) Branching Processes . Beekman (1969) A ruin function approximation.L. Clarendon Press. New York. Halsted Press. Grandell (1985) An insensitivity property of the ruin probability. de Waegenaere (1990) Simulation of ruin probabilities. Scand. Insurance: Mathematics and Economics 6. Appl. 4148. . BarndorffNielsen & H. Probab. 661667. 1974. Cambridge. [55] B. [58] O. Scand. 1988 . Springer . [63] A.A. Berlin. Telek (1994) A bencmark for PH estimation algorithms: results for acyclic PH. SIAM. [62] J. 129134. Schmidli (1995) Saddlepoint approximations for the probability of ruin in finite time. Soc. Wiley. [69] P. 9599. Act. Baker (1977) The Numerical Solution of Integral Equations. [73] A. 221232.H. J. Bjork & J. Plemmons (1994) Nonnegative Matrices in the Mathematical Sciences. Cox (1972) A low traffic approximation for queues.366 BIBLIOGRAPHY [53] S. [61] J. 148156. [72] P. [68] T. Beekman (1985) A series for infinite time ruin probabilities. Billingsley (1968) Convergence of Probability Measures. Scand. Oxford. J. 169186. Bloomfield & D. Bingham (1975) Fluctuation theory in continuous time. Scand. 705766. 1975. [54] K. Wiley. 33. Actuaries 21. Boogaert & V. 11811190. Scand. Insurance: Mathematics and Economics 4.H. [60] J. J. [71] P.
177204. Nonparametric estimators for the probability of ruin. [83] D.Y . 93121. [84] D. 313319. Springer. 16. [91] H. 62. Z. Cohen (1982) The Single Server Queue (2nd ed. Reading. [75] O.U. [82] H. Cohen (1999) Heavytraffic analysis for the GI/G/1 queue with heavytailed service time distributions . Cramer (1955) Collective risk theory. Cohen (1998) The M/G/1 queue with heavytailed service time distribution. Insurance : Mathematics and Economics 5.). Geb. IEEE Trans. AddisonWesley. Area Commun. J. Appl. J. [93] K. 907920. Chandy & M. 24. [79] P. The Society of Actuaries . H. Itasca. IEEE J. Smith ( 1986) An asymptotic analysis of a queueing system with Markovmodulated arrivals . [90] D.R. [92] H. Sel. B.W. Aut.R.BIBLIOGRAPHY 367 [74] O. Berlin. Tweedie (1982) Storage processes with general release rule and additive inputs . Fort & G. Oper. Broeck . Goovaerts & F. [88] J.. Boxma & J. Academic Press. Boxma & J. New York. New York San Francisco London. NorthHolland. Stockholm. Philos. [76] N. [81] J. [78] L. 2338.I. Malgouyres (1983) Large deviations and rare events in the study of stochastic algorithms . Cox (1955) Use of complex probabilities in the theory of stochastic processes. Queueing Systems 33. 749763.M. Res.R. Smith ( 1983) Asymptotic analysis of a queueing system with bursty traffic. Qinlar (1972) Markov additive processes. Brockwell . Skandia Jubilee Volume. 14. Bratley. Tech. verve. New York. M. Stockholm. 105119. 34. Bowers .J. D. Hickman.W. Cottrell. [87] E. [86] K.L.J.L. Bux & U. Fox & L. Probab. The Jubilee volume of Forsakringsbolaget Skandia. [80] F. 14331453. Insurance : Mathematics and Economics 9. [89] M. Reiser eds. Proc. SpringerVerlag. Chung (1974) A Course in Probability Theory (2nd ed. Burman & D.A. Control AC28. Biihlmann ( 1970) Mathematical Methods in Risk Theory. Jones & C.L. Soc.C. II. Adv.W. Veraverbeke (1990).Y. Amsterdam. Syst. De Vylder (1986) Ordering of risks and ruin probabilities. Herzog (1977) The phase concept : approximations of measured data and performance analysis. Bucklew ( 1990) Large Deviation Techniques in Decision.).A. J. Burman & D . Wiley. Computer Performance (K. Jr. [77] P.C. Nesbitt (1986) Actuarial Mathematics . . Illinois. 51. Cambr.L. Amsterdam. 392433. Croux & N. [85] W. Bell. S. 127130. Schrage (1987) A Guide to Simulation. 3539.) NorthHolland. Simulation and Estimation. Cramer (1930) On the Mathematical Theory of Risk. Breiman ( 1968) Probability. Wahrscheinlichkeitsth. Resnick & R. Gerber.
[111] F. [99] R. Rolski (1984) A light traffic approximation for a single server queue. Chapman & Hall. Csorgo & J.. J. 7083 (1969). Haezendonck (1985) Inversed martingales in risk theory. Steinebach ( 1991 ) On the estimation of the adjustment coefficient in risk theory via intermediate order statistics . [108] F. 583602. Daykin. J. N. [104] F. [107] L. Tidskr. Soc. Daley & T. [106] A.B. 5771. Springer. Jones and Bartlett. Res. [105] A. New York. Dassios & P. Oper. Daley & T. T. 135159. 435436. [100] A. 114119. Editions de 1'Universite de Bruxelles.D. Suppl. Aktuar. Res. [110] F. Steinebach (1990) On some alternative estimators of the adjustment coefficient in risk theory. Math. Laird & D. Devroye (1986) Nonuniform Random Variate Generation.P. Pesonen (1994) Practical Risk Theory for Actuaries. Oper. Insurance: Mathematics and Economics 4. [95] M. 121131. Embrechts (1989). De Vylder & M. Rolski (1991) Light traffic approximations in queues. Goovaerts (1988) Recursive calculation of finitetime ruin probabilities. 227229. Scand. Appl. 85116. 1990 . 3750. Actuarial J. Dembo & O. Insurance: Mathematics and Economics 6. J. Teugels ( 1990) Empirical Laplace transform and approximation of compound distributions . [101] C. Insurance: Mathematics and Economics 10. Math. Probab. Martingales and insurance risk. Rubin (1977) Maximum likelihood from incomplete data via the EM algorithm. Forsiikringsmatematiska Studier Tillagnade Filip Lundberg. Deheuvels & J. Statist. [109] F. [96] D. Csorgo & J. Math. [102] P. De Vylder & M. Stochastic Models 5. Delbaen & J.M. van Dawen (1986) Ein einfacher Beweis fur Ohlins Lemma. Goovaerts (1984) Bounds for classical ruin probabilities. English version published in Skand. Haezendonck (1987) Classical risk theory in an economic environment. De Vylder (1978) A practical solution to the problem of ultimate ruin probability. 9 . Davidson (1946) On the ruin problem of collective risk theory under the assumption of a variable safety loading (in Swedish). . Appl.368 BIBLIOGRAPHY [94] M. Boston. [97] D. [112] F. 88101. Pentikainen & E. 17. 624628. 201206.L. Insurance: Mathematics and Economics 3. 16. Roy. Insurance: Mathematics and Economics 7. [98] A. 27. Comp. 22. Zeitouni (1992) Large Deviations Techniques and Applications. Act. De Vylder (1977) A new proof of a known result in risk theory. Delbaen & J. J. Dempster. 3. [103] F. Blotter der deutschen Gesellschaft fur Versicherungsmathematik XVII. Stockholm.. 181217. De Vylder (1996) Advanced Risk Theory. Scand.
R.C. J. Scand. British Actuarial J. Dickson & J. 174186. Astin Bulletin 21. Insurance : Mathematics and Economics 11. Gray (1984) Exact solutions for ruin probability in the presence of an upper absorbing barrier. 229232. 147167. Act. Emanuel .U. Scand. 1993 . 4259.C.C. 1984 . [122] B. 177192. Gerber (1991) Risk theory for the compound Poisson process that is perturbed by a diffusion. 1741. [117] D.M. Gerber & E. [114] D. [130] P. 251262. Scand.R. [124] N.U. Res.C.W. 105115.M. Springer . 107124. [128] E. 5159. [126] F. J.M. Embrechts ( 1988 ).C. Act. Waters (1999) Ruin probabilities wih compounding.C. [129] D . O'Connell (1995) Large deviations and overflow probabilities for the general singleserver queue. Scand.R. Duffield & N.K. Letters 7. Act. 1. J. 269274.C. Insurance: Mathematics and Economics 7.M. J. [123] E. Waters (1996 ) Reinsurance and ruin . [121] D. Hipp (1998) Ruin probabilities for Erlang (2) risk processes.M. Grandell & H. Act. J. Scand. 3745.C. 1993 . Gerber (1988) The surpluses immediately before and at ruin.R. Scand. Dufresne.B. 193199. Harrison & A. 2000 . Schmidli (1993) Finitetime Lundberg inequalities in the Cox case .M. Act. 4962. Insurance: Mathematics and Economics 19.M. Ruin estimates for large claims . Philos. Soc. 118. Oper. J. 1984 . Taylor ( 1975) A diffusion approximation for the ruin probability with compounding assets. [116] D.C. . 1994 . Insurance : Mathematics and Economics 10. J. Regterschot (1988) Conditional PASTA.C. [120] D.J. [127] F. Dickson (1992) On the distribution of the surplus prior to ruin .G. and the amount of claim causing ruin. [131] P.M. Act.M. Insurance : Mathematics and Economics 22. Hipp (1999) Ruin problems for phasetype(2) risk processes. van Doorn & G. [119] D. Proc. Dynkin (1965) Markov Processes I. with applications. Dickson & C. Berlin Gottingen Heidelberg. 363374. J. Scand. J. Dufresne & H. Dickson (1995) A review of Panjer's recursion formula and its applications. Camb. Act. Dickson & C. Embrechts . Math. Dufresne & H. Dickson & H. Dickson & J. 131138. 6180. 191207. Gray (1984) Approximations to the ruin probability in the presence of an upper absorbing barrier. Dickson (1994) An upper bound for the probability of ultimate ruin. Shiu (1991) Risk theory with the Gamma process. [115] D. Insurance: Mathematics and Economics 25. H.M. [118] D. 1975 . [125] F.S.J. Dickson & H.BIBLIOGRAPHY 369 [113] D. Djehiche (1993) A large deviation estimate for ruin probabilities. Insurance : Mathematics and Economics 7.U.
J. . 623637. 269274. 181190. Probab.K. 38. Wiley. [140] J. A bootstrap procedure for estimating the adjustment coefficient. Heidelberg. C. Math. Probab. Act. Griibel & S. Embrechts & J. Embrechts & N. Astin Bulletin 16. 404422. Teugels (1985).). 14661474. 29. first passage times and moments of ladder variables in Markov random walks. Esary. Esscher (1932) On the probability function in the collective theory of risk. Erlang (1909) Sandsynlighedsregning og telefonsamtaler . 566580. Appl.W. Embrechts & T. Embrechts. F. [144] P.L. with applications. 26. Walkup (1967) Association of random variables. Tidsskr. Schmidli (1994) Ruin estimation for a general insurance risk model. Nyt Tidsskrift for Matematik B20. [142] W.J. Probab. 8190. Probab. 131137. [137] P. Sci. Probab. Franken. [143] W. Fields 75. 5975. 159178. [133] P. [138] P. Mikosch (1991). Jensen. [141] F. Statist. [135] P. New York. Adv. Veraverbeke (1982) Estimates for the probability of ruin with special emphasis on the possibility of large claims . [134] P. Proschan & D. Kliippelberg & T. Furrer (1998) Risk processes perturbed by astable Levy motion. Statistica Neerlandica 47. J. Embrechts. Frees (1986) Nonparametric estimation of the probability of ruin. Approximations for compound Poisson and Polya processes. Wiley. Danish Academy Tech.D. Ann. [149] C.W. 616624. 17. Frenz & V. Feller (1966) An Introduction to Probability Theory and its Applications I (3nd ed. [146] E.K. Konig. Fuh (1997) Corrected ruin probabilities for ruin probabilities in Markov random walks. Villasenor (1988) Ruin estimates for large claims. 5572. 695712. Arndt & V. Insurance : Mathematics and Economics 1. [136] P. J. [145] P.L.370 BIBLIOGRAPHY [132] P. Adv. Mikosch (1997) Modelling Extremal Events for Finance and Insurance.). 175195. Maejima & J. Wiley. 3339. Probab. Embrechts.M. [148] C. Reprinted 1948 as 'The theory of probabilities and telephone conversations' in The Life and work of A. Fitzsimmons (1987). Schmidt (1982) Queues and Point Processes. New York. Erlang 2. Akt.A. 29. Fuh & T.D. Springer. Scand. [139] A. Skand. Insurance: Mathematics and Economics 10. R. U. Pitts (1993) Some applications of the fast Fourier transform in insurance mathematics . D. J. Rel. Feller (1971) An Introduction to Probability Theory and its Applications II (2nd ed. Th. 1998 . Appl. [147] M. Embrechts & H. Insurance: Mathematics and Economics 7. Schmidt (1992) An insensitivity property of ladder height distributions.D. [150] H. 35. Appl. M. Lai (1998) Wald's equations. 5974. Trans. Appl. Appl. Adv. On the excursions of Markov processes in classical duality.
Goovaerts. Amsterdam. [153] H.U.J.S. Kovalenko (1989) Introduction to Queueing Theory (2nd ed. Schweiz. Appl. Huebner Foundation Monographs. [163] P. Z. Suppl. de Vylder & J. Birkhauser . Amsterdam. Galambos & J. Gerber (1970) An extension of the renewal equation and its application in the collective theory of risk. [160] H. Furrer. 105115.G. A.. J. Probab. 205216. Gnedenko & I. University of Pennsylvania. Vers. Mitt. 28. Grandell (1978) A remark on 'A class of approximations of ruin probabilities'. Scand. Mitt. Goldie & R. Gerber (1971) Der Einfluss von Zins auf die Ruinwahrscheinlichkeit. Gerber (1981) On the probability of ruin in the presence of a linear dividend barrier.BIBLIOGRAPHY 371 [151] H. 1978 . Kaas. [165] M. Glynn & W. Scand. [159] H.U. [157] H. Ann. R. [164] B. Insurance : Mathematics and Economics 7. Weron (1997) Stable Levy motion approximation in collective risk theory. [162] P. ETH Zurich.J. Springer. 131156. 205210.V. Ver. Unpublished.). [152] H. Appl. Probab. van Heerwarden & T. Vers. Gerber (1986). 31A. NorthHolland. [168] J. 73. [154] H. . [167] C. Life Insurance Mathematics. 3752. 1977. Kluwer. 6370. Griibel (1996) Perpetuities with thin tails. Glasserman (1991) Gradient Estimation via Perturbation Analysis. Aktuarietidskrift 1970 . Skand.U.E.). Boston.U. [169] J.71. 1523. Basel. contained in the authors PhD thesis. Haezendonck (1984) Insurance Premiums. London. F.M. Appl. Whitt (1994) Logarithmic asymptotics for steadystate tail probabilities in a singleserver queue. Glasserman & S. Gerber (1973) Martingales in risk theory. J. Probab.U. Insurance: Mathematics and Economics 20..N. Furrer & H. 5. 97114. 1981 . Math. Gani. Goovaerts. Michna & A. Act. Gerber (1979) An Introduction to Mathematical Risk Theory. NorthHolland. J. Scand. [156] H. Schweiz. Dordrecht. In Studies in Applied Probability (J. Act. [158] H. [161] P. Insurance: Mathematics and Economics 15. 7778. Schmidli (1994) Exponential inequalities for ruin probabilities of risk processes perturbed by diffusion. J. Act. eds.W. Grandell (1977) A class of approximations of ruin probabilities. Math. 463480. [166] M. Ver. Bauwelinckx (1990) Effective Actuarial Methods. 2336. Kou (1995) Limits of first passage times to rare sets in regenerative processes. Adv.U. Gerber (1988) Mathematical fun with ruin theory. 424445. S. Furrer (1996) A note on the convergence of the infinitetime ruin probabilities when the weak approximation is astable Levy motion.U. [155] H.
A. 400409. Berlin. [180] R. 54. Stoch. Appl. Astin Bulletin 29. 1. Resnick (1976 ) The stationary distribution and first exit probabilities of a storage process with general release rule .V. ed.372 BIBLIOGRAPHY [170] J. Gusak & V. Opns . Appl. Statistical Algorithm 265. SpringerVerlag. Grandell (1999) Simple approximations of ruin functions. Proc .131135. 167176. Informatica 2. Gyllenberg & D.). 14. Harrison & S. SpringerVerlag.243255. Math. Harrison (1977) Ruin problems with compounding assets . Aktuar Tidskr.and Sozialforschung 6. [181] D. [185] J. M. Grandell ( 1992) Finite time ruin probabilities and martingales . [189] A. Segerdahl (1971) A comparison of some approximations of ruin probabilities. [179] R. 5766. [175] J. Grigelionis ( 1992) On Lundberg inequalities in a Markovian environment. Lemoine (1977) Limit theorems for periodic queues. [187] J. 332. Grandell (1997) Mixed Poisson Processes. [178] B. 3041. M. Silvestrov (2000) Cram€rLundberg approximations for nonlinearly perturbed risk processes . [183] M. Th. New York. Harrison & A. Amsterdam.I. Grandell (1979) Empirical bounds for ruin probabilities. [176] B. [171] J. 14.S. M. Rink. Korolyuk (1969) On the joint distribution of a process with stationary increments and its maximum .J. Insurance: Mathematics and Economics 26. [177] B. 355365. Grigelionis ( 1993) Twosided Lundberg inequalities in a Markovian environment. Preprint. 8. 143158. Probability Theory and Mathematical Statistics (I. [174] J. Stoch. J. AkademieVerlag. Probab. . Chapman & Hall. Probab. Grnbel (1991) G/G/1 via FFT. Grandell & C. Applied Statistics 40. Skand. Tinbergen Institute Research Series 20. Griibel & R. [173] J. Proc. 347358. KTH. [172] J. 3 . Res. Ibragimov. Gut (1988 ) Stopped Random Walks . Liet. Theory and Applications. Matem. Appl. van Heerwarden (1991) Ordering of Risks: Theory and Actuarial Applications. Appl. Resnick (1977) The recurrence classification of risk and storage processes . Math. Hermesmeier (1999) Computation of compound distributions I: aliasing errors and exponential tilting. Grigelionis (1996) Lundbergtype stochastic processes.I. Archiv v. 6779.. [182] A. M. [186] J. Winter School on Stochastic Analysis and Appl . E. 7590. Res. Proc. mathematische Wirtschaft . 33. Grandell (1990) Aspects of Risk Theory. [188] J. Opns . Hadwiger (1940) Uber die Wahrscheinlichkeit des Ruins bei einer grossen Zahl von Geschiiften . 5 .O. [184] H . 566576. Harrison & S. 197214.
Scand. Haggstrom.A. [206] B. [202] 0. Heilmann (1987) Grundbegriffe der Risikotheorie. Hoglund (1991). 1998 . [198] C. Iversen & L. 25. Hoglund (1990) An asymptotic expression for the probability of ruin within finite time.L. Hoglund (1974) Central limit theorems and statistical inference for Markov chains. Scand. Hermann (1965) Bin Approximationssatz fiir Verteilungen stationarer zufalliger Punktfolgen.V. Hipp & R. Verlag Versicherungswirtschaft e. S. [205] T.R. eds. Horvath & E.L. Appl. 11631174. 1990 . 285292.V. [203] T. [204] T. Huskova. Act. No. Act. Chalmers University of Technology and the University of Goteborg. 259268. Math. ACM TOMACS 6. Karlsruhe. [197] C. Hesselager (1990) Some results on optimal reinsurance in terms of the adjustment coefficient. Wiley. Heidelberger (1995) Fast simulation of rare events in queueing and reliability models. Foruth Prague Symp. Staalhagen (1999) Waiting time distributions in M/D/1 queueing systems. [207] D. Ann. Iglehart (1969) Diffusion approximations in collective risk theory. Geb. on Asymptotic Statistics (P. Hojgaard & M. Wahrscheinlichkeitsth. New York.). Electronic Letters 35. Mathematical Statistics. [196] C. 4385. Hipp (1989a) Efficient estimators for ruin probabilities. [192] U. 378389. Hipp (1989b) Estimators and bootstrap confidence intervals for ruin probabilities. 123151. 12981310. 6. 377381. Probab.B. J. Willekens (1986) Estimates for the probability of ruin starting with a large initial reserve. Probab. Prob. [195] B. 8996. [208] V. 166180. 18. Studies in Statistical Quality Control and Reliability 1992 :4. 285293. Herkenrath (1986) On the estimation of the adjustment coefficient in risk theory by means of stochastic approximation procedures.M. 23. Asmussen & 0. Versicherungswirtschaft. Hill (1975) A simple general approach to inference about the tail of a distribution. Klugman (1984) Loss Distributions. Mandl & M.a program for fitting phasetype distributions. [191] W. 19. [193] U. Proc. Astin Bulletin 19. Taksar (2000) Optimal proportional reinsurance policies for diffusion models. verve. [200] M. 8095. Michel (1990) Risikotheorie: Stochastische Modelle and Statistische Methoden. . 5770. Hogg & S. [201] L. 305313. Statist. Ann. J. Ann. [194] 0. [199] R. The ruin problem for finite Markov chains. Karlsruhe.BIBLIOGRAPHY 373 [190] P. Probab. Hogan (1986) Comment on corrected diffusion approximations in certain random walk problems. Insurance: Mathematics and Economics 5. Z. J. 3. Insurance: Mathematics and Economics 5. Nerman (1992) EMPHT . J. Nachrichten 30. 29.. Appl.
New York. [222] S. Cambridge Philos. J. Munksgaard. [225] J. Stochastic Models 1. Keilson & D. Reinhard (1985) Probabilities de ruine pour une classe de modeles de risque semiMarkoviens. [227] J. Jagerman (1991) Analytical and numerical solution of Volterra integral equations with applications to queues. 116129. Scand. Jelenkovic & A. 37. 63. 123133. Janssen (1980) Some transient results on the M/SM/1 special semiMarkov model in risk and queueing theories. Jagerman (1985) Certain Volterra integral equations arising in queueing.R. Kalashnikov (1996) Twosided bounds of ruin probabilities.G. [221] E. Proc. 11. ibid.H. [214] A. [210] D. [223] J. [211] J. Lazar (1998) Subexponential asymptotics of a Markovmodulated random walk with a queueing application. 31. Kaas & M.374 BIBLIOGRAPHY [209] D.C.M. Kennedy (1994) Understanding the WienerHopf factorization for the simple random walk. 4151. Soc. Goovaerts (1986) General bound on ruin probabilities. N. Kemperman (1961) The Passage Problem for a Markov Chain. [219] V. 165167. 6. 3. Amer. Stochastic Models 5. Chicago. Johnson & M. 15. [213] P. . [212] J. 547567. J.J. Technometrics 30.L. 1996. Wishart (1964) A central limit theorem for processes defined on a finite Markov chain.M. Appl. University of Chicago Press. Princeton. Appl.A. Keilson (1966) A limit theorem for passage times in ergodic regenerative processes. Jensen (1995) Saddle Point Approximations. Soc. Addenda to for processes defined on a finite Markov chain. J. Copenhagen. Janssen & J. [215] J. Kalashnikov (1997) Geometric Sums: Bounds for Rare Event with Applications. Wishart (1964) Boundary problems for additive processes defined on a finite Markov chain. 6. Oxford. Act. [228] J. 325247. Proc. NEC Research Institute. J.M. Kalashnikov (1999) Bounds for ruin probabilities in the presence of large claims and their comparison. [226] J.P. Cambridge Philos. Proc.B. [216] M.M. 61. Kao (1988) Computing the phasetype renewal and related functions. Kluwer. Keilson & D. [224] J.M. Probab. Manuscript. 866870. Keilson & D. Ann. Probab. Astin Bull. 35. Karlin & H. Math. 60. 711743. 259281. 239256. [220] V. ibid. Cambridge Philos. 187193. Jensen (1953) A Distribution Model Applicable to Economics. Clarendon Press. N. Academic Press. Soc. Statist. 561563.G. [218] V. [217] R. Wishart (1964).G. Insurance: Mathematics and Economics 5. Astin Bull. Taaffe (1989/90) Matching moments to phase distributions. 118. 8793. 173190. Taylor (1981) A Second Course in Stochastic Processes. 283306. Act.J.
J. Philos.BIBLIOGRAPHY 375 [229] J. Oxford 12. [232] C. Penev & A. J. [230] J. Kliippelberg (1988) Subexponential distributions and integrated tails. [233] C. (235] C.C. Proc. Probab. Quart. [236] C. 154168. Kingman (1964) A martingale inequality in the theory of queues. Kingman (1961) A convexity property of positive matrices .. Bernoulli 1 . Turbin (1973) An asymptotic expansion for the absorption time of a Markov chain distribution. [239] H. Seminaire de Probabilties X . [248] D. Kingman (1962) On queues in heavy traffic. Lipsky (1992) Queueing Theory .P.F. 4958. Prob. Appl. Appl. 48. 1998 . Nyrhinen (1992b) On asymptotically efficient simulation of ruin probabilities in a Markovian environment. Scand.J. Mikosch (1995) Explosive Poisson shot noise with applications to risk retention.S. 383392. Macmillan. 259264. Sorensen (1997) Exponential Families of Stochastic Processes. Roy. 279285. [242] A. Ramaswami (1999) Introduction to MatrixAnalytic Methods in Stochastic Modelling.C. [246] D. 359361. Klnppelberg (1993) Asymptotic ordering of risks and ruin probabilities. [245] T. Appl. . Insurance: Mathematics and Economics 8. [234] C. Probab. Scand. 60 . Soc. 133135 (in Russian). Nyrhinen (1992a) Simulating levelcrossing probabilities by importance sampling. Kliippelberg & U. 18. Philos.J. 277289. J. Stochastic Models 7. [247] L. 25. Klnppelberg & T. 132141. Scand. B24. Cambr. Appl. Stadtmiiller (1998) Ruin probabilities in the presence of heavytails and interest rates . [241] G. [244] T. 390397. I. Actuarial J. Soc. Proc. Lecture Notes in Mathematics 511. [238] V. [243] A. Lindley (1952) The theory of queues with a single server. Lemoine (1981) On queues with periodic Poisson input. Latouche & V. 26.F. 858874. Kunita (1976) Absolute continuity of Markov processes. Klnppelberg & T. Korolyuk. 125147. J. Klnppelberg (1989) Estimation of ruin probabilities by means of hazard rates. New York. Statist. Math.F. Probab. SpringerVerlag.C. [240] U. Quart. Camb. J. 4477. Insurance : Mathematics and Economics 12. 889900. 24. 146. SIAM. [231] J. Lemoine (1989) Waiting time and workload in queues with periodic Poisson input. Lehtonen & H. J. Lehtonen & H. Act.a Linear Algebraic Approach. Act. 1995 . Adv.F. Kuchler & M. J. Soc. Mikosch (1995) Modelling delay in claim settlement. 6075. Cybernetica 4. [237] C. SpringerVerlag. 283284. Lucantoni (1991) New results on the single server queue with a batch Markovian arrival process.
Moustakides ( 1999) Extension of Wald 's first lemma to Markov processes. MeierHellstern & M . Meier ( 1984). Michna ( 1998) Selfsimilar processes in collective risk theory. [261] H. 36. K. with an application . Ann. Appl.K. 834841. Probab. 286298. Proc. Ann. Probability and Mathematical Statistics (A.). Appl. 147149. Ann. Ney & E. Cambridge Philos. Adv. .M. Uppsala. Stockholm. [252] A.F. [265] S. [263] M . a useful concept in risk theory. Math. Act. 1261 (0?)1270. 763776. Insurance : Mathematics and Economics 5. Probab . [266] P. Appl. 161174. J. 31. Schmidt ( 1993) On ladder height distributions of general risk processes . Nagaev (1957) Some limit theorems for stationary Markov chains. J. Statist. [264] G. Probab. Gut & J. Miller ( 1962 ) Absorption probabilities for sums of random variables defined on a finite Markov chain . Soc.D.V. J. Makowski ( 1994) On an elementary characterization of the increasing convex order . 676705. MartinL3f (1986) Entropy.D. 3. 4859. [260] H. MartinL6f (1983) Entropy estimates for ruin probabilities . Mammitzsch ( 1986 ) A note on the adjustment coefficient in ruin theory. [254] V . 2939. A fitting algorithm for Markovmodulated Poisson processes having two arrival rates . Appl. Lucantoni . Scand.D. Lundberg (1926) Forsdkringsteknisk Riskutjdmning. 124147. 58 . Eigenvalue properties and limit theorems . J. [257] A. 378406. 561592. Appl. [256] A. 32. Res. 11 . Miller ( 1961 ) A convexity property in the theory of random variables defined on a finite Markov chain.V. 29 . Miller (1962 ) A matrix factorization problem in the theory of random variables defined on a finite Markov chain . II Aterforsdkring av Kollektivrisker .376 BIBLIOGRAPHY [249] D. [253] V. 429448. 223235. Scand. F. Neuts ( 1990) A single server queue with server vacations and a class of nonrenewal arrival processes . J. Cambridge Philos. Math. Anal. J. 58. 268285.S. [262] H . Lundberg (1903) I Approximerad Framstdllning av Sannolikhetsfunktionen. Appl. 15. Soc. [255] V. Th. Malinovskii ( 1996) Approximation and upper bounds on probabilities of large deviations of ruin within finite time. Almqvist & Wiksell. Stoch.K. Englunds Boktryckeri AB. Nummelin (1987) Markov additive processes I. 1994. 1996 . . 2. 22. Proc. Scand. Probab. Malinovskii ( 1994) Corrected normal approximation for the probability of ruin within finite time. Act.S. Act. J. [250] F. Europ. Opns. 1986 . Hoist eds. [259] Z . Probab. [258] K . 370377. Probab. [251] F. Miyazawa & V.
Insurance: Mathematics and Economics 20. [281] J. 339353.F. Willekens (1986) Second order behaviour of the tail of a subordinated probability distributions. Stochastic Models 3 . Paulsen (1998) Ruin theory with compounding assets . Paulsen & H. [275] H. Sem. J. Omey & E. Stoch. Hamburg. [276] C. Appl. Proc. [282] J . 10081026. Pakes (1975) On the tail of waiting time distributions. Proc.F. Statistics 21. 311342. 46. Neveu ( 1961 ) Une generalisation des processus a accroisances independantes. .F. Stoch. 16. Johns Hopkins University Press. Appl. New York. Gjessing (1997a) Optimal choice of dividend barriers for a risk process with stochastic return on investments. Adv.BIBLIOGRAPHY 377 [267] M . Nyrhinen ( 1998 ) Rough descriptions of ruin for a general class of surplus processes. 12551265. 21. Appl. Commun. Appl. 12. London. 123144. 249266. [285] J. J. 36. 273299. Stoch. 30. Norberg ( 1990) Risk theory and its statistics environment . Neuts ( 1989) Structured Stochastic Matrices of the M/G/1 Type and their Applications. Math. Paulsen (1993) Risk theory in a stochastic economic environment. Marcel Dekker. [279] E. 75. [280] A. [271] M. Proc. E75B .K. Naval Research Logistics Quarterly 25. Stoch. [269] M . Paulsen ( 1998) Sharp conditions for certain ruin in a risk process with stochastic return on investments. Astin Bulletin 5 . Neuts (1977) A versatile Markovian point process. 71. 135148. 764779. Abh. Willekens (1987) Second order behaviour of distributions subordinate to a distribution with finite mean . 157. IEICE Trans. Nyrhinen (1999) Large deviations for the time of ruin. 555564.K. O'Cinneide (1990) Characterization of phasetype distributions. Neuts ( 1981 ) MatrixGeometric Solutions in Stochastic Models. Stoch. [283] J. Neuts (1992) Models based on the Markovian arrival process. J. [270] M . Baltimore . 327361. Probab. [272] J . Appl. Appl.A. Gjessing (1997b) Present value distributions with applications to ruin theory and stochastic equations. Neuts ( 1978) Renewal processes of phase type. Appl. [274] H . Appl. [284] J.G. 733746. Omey & E.a survey. Probab.F. [278] E. [273] R. Probab. [268] M . 215223. [277] J. 316. Insurance: Mathematics and Economics 22. 445454. Probab. Models 6. Ohlin (1969) On a class of measures for dispersion with application to optimal insurance.F. Proc. Paulsen & H.
Reinhard (1984) On a class of semiMarkov risk models obtained as classical risk models in a markovian environment. Insurance Risk. [297] R. Prabhu (1961). Math. Ann. Math. Prabhu & Zhu (1989) Markovmodulated queueing systems. Queues. Wiley. 117133. Ann.H. 4. Stat. 32.J.U. Pitts. [294] N. [295] N. 147159. [289] C. 29. J. 215246. Pellerey (1995) On the preservation of some orderings of risks under convolution. R. 61. Astin Bulletin XIV. Res. [293] N. de Smit (1986) The queue M/G/1 with Markovmodulated arrivals and services. New York. [292] S. Gjessing (1997c) Ruin theory with stochastic return on investments. Math. 390413. Appl. Pyke (1959) The supremum and infimum of the Poisson process. [290] E. [305] C.L. Probab. J. [288] S. 222261. [300] C. Griibel & P. Act. Appl. Probab. 1989 . Ramsay (1984) The asymptotic ruin problem when the healthy and sick periods form an alternating renewal process. [302] J.K. Adv. Rogers (1994) Fluid models in queueing theory and WienerHopf factorisation of Markov chains. 28. 2330. Probab. Austr. Probab. Prabhu (1965) Queues and Inventories. 46. [291] S. Res. Samorodnitsky (1997) Performance degradation in a single server exponential queueing model with long range dependence.U. Appl. Opns. 537555. Wiley. 4568. 30. Wiley. Philipson (1968) A review of the collective theory of risk. Adv.. 691707. New York. Queueing Systems 5. Embrechts (1996) Confidence bounds for the adjustment coefficient. 19.R.U. Appl. 965985. Adv.M.J. On the ruin problem of collective risk theory.M. 757764. [298] V. Berlin. and Dams. Math. Inst. Appl. Springer. [287] F. 337347. 11. Aktuar. Rolski (1987) Approximation of periodic queues. 568576. Ann.G. Soc.M. [303] S.G. [299] C. 139143. Rao (1965 ) Linear Statistical Inference and Its Applications. Resnick & G. [296] N. 235243. 12.A. 820827. Statist. Math. Ripley (1987) Stochastic Simulation . 465483. Oper. Insurance: Mathematics and Economics 3. Ramaswami (1980) The N/G/1 queue and its detailed analysis. Prabhu (1980) Stochastic Storage Processes. Pitts (1994) Nonparametric estimation of compound distributions with applications in insurance. Paulsen & H. [304] B. 29A. Ann. Probab . Tidskr. Skand.378 BIBLIOGRAPHY [286] J. Pitman (1980) Subexponential distribution functions. . [306] T.U.K. [301] G. Insurance: Mathematics and Economics 16. J. Heidelberg. Scand. Schock Petersen (1989) Calculation of ruin probabilities when the premium depends on the current reserve. Statist. Regterschot & J. 2343.
11. Stochastic Models 6. Wiley. Th. Operationsforsch. Stoch. Schmidli ( 1997b) An extension to the renewal theorem and an application in risk theory. Insurance: Mathematics and Economics 22. 5. Teugels (1999) Stochastic Processes for Insurance and Finance. Rubinstein & B. J. 365388. Astin Bulletin 29. [315] T.J. Shapiro (1993) Discrete Event Systems: Sensitivity Analysis and Stochastic Optimization via the Score Function Method. . Insurance: Mathematics and Economics 16. 5. Stochastic Models 10. 131156. [308] S. Stochastic Models 10 . [322] H. Ross (1974) Bounds on the delay distribution in GI/G/1 queues. Data Anal. 2001 . [325] H . Wiley. J. Rudemo (1973) Point processes generated by transitions of a Markov chain. Schmidli ( 2001 ) Optimal proportional reinsurance policies in a dynamic setting . Math.BIBLIOGRAPHY 379 [307] T. 145165. Statist. Schmidli. Probab . Comp. Adv. Schmidt & J. Ann. Schmidli ( 1997a) Estimation of the Lundberg coefficient for a Markov modulated risk model . Scand. Scand. 93104. [317] L. Schmidli ( 1996) Martingales and insurance risk. Lecture Notes of the 8th Summer School on Probability and Mathematical Statistics ( Varna). Proc.M.L. [324] H . Rossberg & GSiegel (1974) Die Bedeutung von Kingmans Integralgleichungen bei der Approximation der stationiiren Wartezeitverteilung im Modell GI/C/1 mit and ohne Verzogerung beim Beginn einer Beschiiftigungsperiode. J. 1997.Y. [311] R. [312] R. Act. Ryden (1996) An EM algorithm for estimation in Markovmodulated Poisson processes.Y. [310] R. Appl. 227244. Science Culture Technology Publishing . 687699. Wiley. Statist. Wiley. 7. [323] H. Rolski. [321] H. Schmidli ( 1994) Diffusion approximations for a risk process with the possibility of borrowing and interest. Singapore. 21. 795829. 431447 [316] S. Schmidli ( 1999b) Perturbed risk processes: a review . [326] H.Y. 155188. 4857. V. Schmickler ( 1992 ) MEDA : mixed Erlang distributions as phasetype representations of empirical distribution functions . 121133. Schmidli ( 1999a) On the distribution of the surplus prior and at ruin. 5. Melamed (1998) Modern Simulation and Modelling. Rubinstein & A. Seal ( 1969) The Stochastic Theory of a Risk Business. Wiley. [318] H. Rubinstein (1981) Simulation and the Monte Carlo Method. Ryden (1994) Parameter estimation for Markov modulated Poisson processes. 417421. Probab. Schlegel ( 1998) Ruin probabilities in perturbed risk models. 4068. 135149. Appi. H. Act. [319] H. Probab. [309] H. [314] T. [320] H. 262286. Schmidli ( 1995 ) CramerLundberg approximations for ruin probabilities of risk processes perturbed by a diffusion . [313] M. Appl.
Segerdahl (1959) A survey of results in the collective theory of risk. 1987. 72. 701719. Mitteil. Siegmund (1976) Importance sampling in the Monte Carlo study of sequential tests. New York. [346) D . Stochastic Models 6. Probab. Sengupta ( 1989) Markov processes whose steadystate distribution is matrixgeometric with an application to the GI/PH/1 queue. Siegmund ( 1985) Sequential Analysis . Aktuar Tidsskr. 191197. 429442. 483487. 77100. pp. Shwartz & A. 383413. [342] D. Statist. Scand. Ann. [330] H. Shantikumar (1993) Stochastic Orders and Their Applications.W. Segerdahl (1955) When does ruin occur in the collective theory of risk? Skand. Shaked & J. Seal (1972) Numerical calculcation of the probability of ruin in the Poisson/Exponential case. [343] D. Shin (1987) Convolution of uniform distributions and ruin probability. Grenander).the Harald Cramer volume (ed. 1974. Math. 171178.S. [341] E. Insurance: Mathematics and Economics 8. Segerdahl ( 1942) Uber einige Risikotheoretische Fagestellungen . [345] D. Risk teory and the single server queue . 914924. Appl. J. 4. Versich. Siegmund (1976) The equivalence of absorbing and reflecting barrier problems for stochastically monotone Markov processes . Queueing Systems 11. 121139. 159180. Wiley. [334] C: O. [331] G.W. 11. 2236. Siegmund (1979) Corrected diffusion approximations in certain random walk problems. Seal (1974) The numerical calculation of U(w. Chapman and Hall. 75. Verein Schweiz.380 BIBLIOGRAPHY [327] H. Siegmund (1975) The time until ruin in collective risk theory.A. Sengupta (1990) The semiMarkov queue: theory and applications. Appl. [332] C : O. [340] E. . 61. . 11. 21. J. 1955. Act. Shtatland ( 1966) On the distribution of the maximum of a process with independent increments. the probability of nonruin in an interval (0. Weiss (1995) Large Deviations for Performance Analysis. [335] B . Almqvist & Wiksell. [337] M. Versich. Math. Seber (1984) Multivariate Observations.S. Chapman & Hall. Seal ( 1972).L. Ann. Math. Verein Schweiz. SpringerVerlag. Mitt. [333] C: O. 4. [344] D. [338] E. Sigman ( 1994 ) Stationary Marked Point Processes: An Intuitive Approach. [336] B. t). Skand. t). In Probability and statistics . Seal (1978) Survival Probabilities. Probab. Sigman (1992) Light traffic for workload in queues. Verein Schweiz Versich. 279299. 673684. Aktuar Tidsskr.L. Stockholm.L. Adv. [329] H. Adv. Th. Mitt. 243249.F.S. Shin (1989) Ruin probability by operational calculus.G. Wiley. Appl. 72. [348] K . [347] K. 4383. Scand. Act. [339] A. 157166. Probab . [328] H. L. Academic Press. Probab.
. 235254. Act. Scand. Daley ed.V. 8594. Taylor (1976 ) Use of differential and integral inequalities to bound ruin and queueing probabilities . Straub (1988) NonLife Insurance Mathematics. Stroinski ( 1994) Recursive method for computing finitetime ruin probabilities for phasedistributed claim sizes. Slud & C. Ruin probabilities prepared for numerical calculations. J. 5776. Act. Astin Bulletin 12. J. J. Insurance: Mathematics and Economics 16. [351] D . [364] G. Act.C. Proc. Soc 49. [353] E. New York. Smith ( 1953) Distribution of queueing times.. Actuarial J. [363] G. Taylor (1978) Representation and explicit calculation of finitetime ruin probabilities. 65102. 163175.L. 1980 . 8199. 1982 . Wiley. Astin Bulletin 16 . New York. Act. Thorin (1986) Ruin probabilities when the claim amounts are gamma distributed. Stoyan (1983) Comparison Methods for Queues and Other Stochastic Models (D. Verlag Versicherungswirtschaft e. J. Taylor (1979) Probability of ruin with variable premium rate. Takhcs (1967) Combinatorial Methods in the Theory of Stochastic Processes. Thorin (1982) Probabilities of ruin. [365] J. 722. [357] B. [362] G. Taylor (1979) Probability of ruin under inflationary conditions or under experience ratings .J. Insurance: Mathematics and Economics 19. 2739.BIBLIOGRAPHY 381 [349] E. Teugels (1982) Estimation of ruin probabilities .L. [366] O. Unpublished manuscript. 149162. NorthHolland. Springer. Taylor (1986) Claims reserving in NonLife Insurance. . Scand. 118. 114137. [367] O. [361] G. Thorin (1974) On the asymptotic behaviour of the ruin probability when the epochs of claims form a renewal process. Scand. J. Cambridge Philos. Teugels ( 1995 ) Ruin estimates under interest force . [368] O. Karlsruhe. 1978. Sundt & J. [355] B . 1974.C. [356] B . Sundt & W. Sundt & J. 725741. Scand.C. Insurance: Mathematics and Economics 1. [360] G. [350] W. Appl. 1976. Hoesman ( 1989) Moderate. Stanford & K.J. Teugels (1997) The adjustment coefficient in ruin estimates under interest force . [358] R. Proceeedings of the IEEE 77. [352] D. 449461.L. Adv.C. 65102. Sundt ( 1993) An Introduction to NonLife Insurance Mathematics (3rd ed.). Jewell ( 1981 ) Further results on recursive evaluation of compound distributions .L. Scand. Suri ( 1989) Perturbation analysis : the state of the art and research issues explained via the GI/G/1 queue. Thorin (1977). Act.A. [354] B.). [369] O. John Wiley & Sons. Scand. [359] L. 21.C. 197208. Probab.and largedeviations probabilities in actuarial risk theory. Astin Bulletin 24.S.
[380] W.R. Wallace (1969) The solution of quasi birth and death processes arising from multiple access computer systems. Wilson (1995) Selfsimilarity in highspeed traffic: analysis and modeling of ethernet traffic measurements. Wikstad (1977) Numerical evaluation of ruin probabilities for a finite period. Wald (1947) Sequential Analysis. 231246. Van Wouve. [381] W. [376] A. Wiley. [378] H. De Vylder.E. M. D. Skand. Unpublished Ph. [379] M. Sherman & D. Thorin & N. thesis. [371] 0. F. Goovaerts (1983) The influence of reinsurance limits on infinite time ruin probabilities. Astin Bulletin VII. J. 743756. Haezendonck eds. 137153. SpringerVerlag. Wikstad (1977) Calculation of ruin probabilities when the claim distribution is lognormal. [383] G. Act. [384] R. Insurance : Mathematics and Economics 13.382 BIBLIOGRAPHY [370] 0. Res. Statistical Science 10. [375] N. Willmot (1994) Refinements and distributional generalizations of Lundberg's inequality. Submitted. 936952. 148164. Oper. R.). Lin (1994) Lundberg bounds on the tails of compound distributions. Astin Bulletin IX. Whitt (1989) An interpolation approximation for the mean workload in a GI/G/1 queue. Tidskr. [377] V. Thorin & N. 5762. [382] G. Waters (1983) Probability of ruin for a risk process with claims cost inflation. Aktuar. 1942. 4963. [372] H. VazquezAbad (1999) RPA pathwise derivatives of ruin probabilities. M. Scand. Thorisson (2000) Coupling. In: Premium Calculation in Insurance (F. Probab. Willinger. J. [373] S. . 37. Appl. Veraverbeke (1993) Asymptotic estimates for the probability of ruin in a Poisson model with diffusion. 6785. De Vylder & M. University of Michigan. Taqqu.W. Goovaerts. PrenticeHall. 1983 .E. 142. Reidel. J. Wolff (1990) Stochastic Modeling and the Theory of Queues. Stationarity and Regeneration. [374] F. Tacklind (1942) Sur le risque dans les jeux inequitables. 31. Dordrecht Boston Lancaster. Insurance: Mathematics and Economics 15. Willmot & X.
100.249250 integral equation 16 Lindley 143 renewal 64. 117128. 323 Coxian distribution 147. 79. 7475.185187. 111117. 37.86.292293 Edgeworth expansion 113.346349 383 .287292.6779. 196201 inverse Gaussian distribution 76. 8283 hyperexponential distribution 7. 1819. 12 CramerLundberg approximation 1617.135. 135. 316323 Bessel function 102. 80 81. 14.226.307312 compound Poisson model 4.203.217. 201214. 7079. 217. 5796. 141144. 278 gamma distribution 67. 360 excursion 155156. 162164. 89.251280 heavy traffic 76. 39.228229. 3032.160167. 201 Brownian motion 3 . 122. 283. 97129. 170173.259261. 5. 138139. 207 heavytailed distribution 6.314316.328330. 245248. 3436.200201. 239. 9496.203.272.269.9899. 97. 218 Cox process 4.and sum 221. 227229. 17. 226.178184. 302303 diffusion approximation 17.150. 361 diffusion 3.182. 119. 3839.318320 change of measure 2630.308 CramerLundberg model: see compound Poisson model cumulative process 334 dams: see storage process differential equation 16.121129. 271274.281. 189. 110113. 341. 2526.249. 86. 2425.359 aggregate claims 103106.285292. 15. 205. 1112.137141.301 central limit theorem 60 . 3334. 40.4447. 7179. 4851. 308. 301 Kronecker product. 180182. 9396. 7879. 17. 9293. 332333 Volterra 192194. 1415.Index adjustment coefficient 17.299. 91. 318319 Erlang distribution 7.293294. 248 WienerHopf 144 interest rate 190.242. 117127 corrected 121127 duality 1314.
139141.238.275278. 176185 nonhomogeneous 60 PollaczeckKhinchine formula 6167. 295.180. 25.287.152160. 304305 random walk 3336.161. 134135. 35. 2730.123. 261264. 44. 39. 71. 176185.269271.297299.227230. nonlinear 155. 149. 245 M/G/1 13. 37.201. 271274. 145187.146148. 179 NP approximation 318320 Palm distribution 5253. 133.302. 25. 16. 35. 106108.161164.349 350 perturbation 172173. 227228.348 terminating 215216.178182. 3639. 112113.128129. 3947. 99. 41. 7576. 108109.285287 queue 14 . 96. 251. 6162. 113114. 175 light traffic 8183 Lindley integral equation 143 process 3334. 100. 86 periodicity 12. 178 modulation 12. 162.134135. 38. see also sensitivity analysis phasetype distribution 8.234.161. 5758. 15. 32. 3947. 157. 142 likelihood ratio : see change of measure lognormal distribution 9. 6970. 267269 Panjer's recursion 320323 Pareto distribution 910.160161.304 process 2830. 38. 203 Markov additive process 12.287291 INDEX matrix equation .259261.350361 Poisson process Markovmodulated 12 periodic 12.148.340350 multiplicative functional 2830. 132133. 234 matrixexponential distribution 240244 matrixexponentials 14. 7179.234240. 260 Lundberg conjugation 6979 . 154. 171. 269 PerronFrobenius theory 4142. 203204. 306316 Levy process 3.288290. 134. 213214. 42. 4446. 59. 65.261264. 16. 44. 137139. 108 life insurance 5. 185187 GI/G/1 141144 M/D/1 6667 equation 16. 14. 133.315 inequality 1718. 138.240244.384 ladder heights 4756.298299. 80.108. 9899.339 large deviations 129. 52 53. 144.218221.174.215250. 230. 229 M/M/1 101 Markovmodulated 185187 periodic 187 martingale 2426. 141144.336339 Laplace transform 15.336339 . 257.
292294. 338 utility 324. 257. 31.279280 Rouche roots 158. 307308. 9693. 251.262263. 331336 equation 64. 256258. 222. 160. 60.273274.INDEX 385 waiting time 141. 253. 37. 261264 reservedependent premiums 14. 294296 shotnoise process 314 simulation 19. 186187 renewal process 131.336339 workload 13. 251. 174. 5455. 238 saddlepoint method 115117. 281296 stable process 15. 186187 virtual: see workload rational Laplace transform 8. 131144. 233234. 107. 233. 4950. 229234. 189214. 213. 251280 time change 4. 12. 332333 model 12. 146. 123. 162. 240. 89.186. 244.244250. 141144. 333334 regular variation 10. 317318 semiMarkov 147. 168172 storage process 13. 280. 1819.154157. 177 timereversion 14. 87.314. 223226. 7475. 191192. 279280 subexponential distribution 11. 260 WienerHopf theory 144. 11. 327 . 3032. 8386. 147. 326330 Weibull distribution 9. see also matrixexponential distribution regenerative process 264 268. 260 reinsurance 8. 172173. 120 statistics x. 335336 sensitivity analysis 8693. 152.359361 stochastic control x stochastic ordering 18.
T [Ail i The book is a comprehensive treatment of  I i I \ classical and modern ruin probability theory. Special features of the book are the emphasis on change of measure techniques.com 2779 he 9 "789810ll22293211 .Advanced Series on Statistical Science & Applied Probability . extensions of the classical compound Poisson model to allow f o r reservedependent premiums. the ^W A l \ i l ' ''' CramerLundberg approximation." Short Book Reviews ISBN 9810222939 mi u inn i nun I I I I I I i in u www.Vol. 2 A I 11 JjVb l' i  i Yj . worldscientific. P'i yfliother approximations (e. "This book is a must for anybody working in applied probability.. for heavytailed claim size distributions). Some i (l I JL I J r of the topics are Lundberg's inequality. I 1! Ruin Probabilities . It is a comprehensive treatment of the known results on ruin probabilities. Markovmodulation or periodicity.g.. phasetype distributions as a computational vehicle and the connection to other applied probability areas like queueing theory. y finite horizon ruin probabilities. exact solutions.
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 listening from where you left off, or restart the preview.