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
. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . 336 A3 Matrixexponentials . . . . . . . 251 2 The compound Poisson model .. . . . . . . . . . . . . . . . . .CONTENTS vii IX Ruin probabilities in the presence of heavy tails 251 1 Subexponential distributions . . . . 287 4 Importance sampling for the finite horizon case . . . . . .. . .. . . . . .. . . . . . . . . . . . . . 344 AS Complements on phasetype distributions .. . . . . 292 6 Sensitivity analysis .. . . . . . . . . 261 4 Models with dependent input . . .. . . . 279 X Simulation methodology 281 1 Generalities . . . . . 264 5 Finitehorizon ruin probabilities . . . . . . . . . . . 259 3 The renewal model . . . . . .. . . . . 294 XI Miscellaneous topics 297 1 The ruin problem for Bernoulli random walk and Brownian motion. . . .. . . . . . 326 Appendix 331 Al Renewal theory . . . . . . . . . . . . . . . . . . . .. . . . .. . . . . . . . . . . ... . . . . . . . 297 2 Further applications of martingales . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . .. . . . 323 6 Reinsurance . . . . . . . . . 285 3 Importance sampling via Lundberg conjugation . . . . The twobarrier ruin problem . . . . . . . . . 316 5 Principles for premium calculation . . . .. 350 Bibliography Index 363 383 . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . 306 4 The distribution of the aggregate claims . . . . . . . . 281 2 Simulation via the PollaczeckKhinchine formula . . . . . .. 290 5 Regenerative simulation . . . . . . . . . . . . . . . . . . . 304 3 Large deviations .. . . . .. . . . . . . . . . . . . . . . . . . 331 A2 WienerHopf factorization . . . 340 A4 Some linear algebra . . . . . . . . . . . 271 6 Reservedependent premiums . . . . . . . .
This page is intentionally left blank .
and the series editor Ole BarndorffNielsen for their patience.Preface The most important to say about the history of this book is: it took too long time to write it! In 1991. Apart from these remarks. and has been an active area of research from the days of Lundberg all the way up to today. 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!). 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. it is not by intention. Risk theory in general and ruin probablities in particular is traditionally considered as part of insurance mathematics. and other projects absorbed my interest. and the result is now that the book is much more related to my own research than the initial outline. One reason for writing this book is a feeling that the area has in the recent years achieved a considerable mathematical maturity. In particular. and my belief was that this could be done rather quickly. Let me take this opportunity to thank above all my publisher World Scientific Publishing Co. It has obviously not been possible to cover all subareas. which has in particular removed one of the standard criticisms of the area. Thus. University of Copenhagen. the idea was close to expand these to a short book on the subject. 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. this applies to longrange dependence which is intensely studied in the neighboring ix . I was invited to give a course on ruin probabilities at the Laboratory of Insurance Mathematics. the book is basically mathematical in its flavour. if the formulations occasionally give a different impression. But the pace was much slower than expected. it would not be fair not to say that the practical relevance of the area has been questioned repeatedly. Since I was to produce some handouts for the students anyway. that it can only say something about very simple models and questions. However. The course was never realized.
VII. Resnick & Samorodnitsky [303] and references therein. Here is a suggestion on how to get started with the book.4a. It is obvious that such a system involves a number of inconsistencies and omissions. for the effects on tail probabilities. Hojgaard & Taksar [35] and Paulsen & Gjessing [284]. some papers not cited in the text but judged to be of interest are included in the Bibliography. read Chapter I.2 more properly). for which I apologize to the reader and the authors of the many papers who ought to have been on the list.13 and XI.15. In the classical setting of CramerLundberg models. 111.g.se Lund February 2000 Soren Asmussen . [381]).2.3. IV. One is by model. an area which is becoming increasingly important. The main motivation comes from statistical data for network traffic (e. VIII. In addition.6 (to understand the PollaczeckKhinchine formula in 111. it has not been possible to incorporate more numerical examples than the few there are.13 and IX.x PREFACE field of queueing theory. For a brief orientation. The present book is in between these two possibilities. see also Schmidli [325] and the references in Asmussen & Taksar [52]. The rest is up to your specific interests.45. IV.1.se/matstat / staff/asmus and I am therefore grateful to get relevant material sent by email to asmusfmaths . see e.13.lth. Chapters IIIVII introduce some of the main models and give a first derivation of some of their properties. A book like this can be organized in many ways. the first part of 11. Willinger et al. Hojgaard & Taksar [206]. http:// www. The book does not go into the broader aspects of the interface between insurance mathematics and mathematical finance. IX. I intend to keep a list of misprints and remarks posted on my web page. Another interesting area which is not covered is dynamic control. the standard stochastic control setting of diffusion models has been considered.g. another by method. incorporate 11.lth. 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). For a second reading. Finally.2.g.maths .5. IV. X. VI. Concerning ruin probabilities. I regret that due to time constraints. 111.13.14. some basic discussion can be found in the books by Biihlmann [82] and Gerber [157]. see in particular Michna [259]. Asmussen. Good luck! I have tried to be fairly exhaustive in citing references close to the text. VII. More recently.89. e.
1 is almost identical to Section 2 of Asmussen [26] and reprinted with permission of Blackwell Publishers.3 are reprinted from Asmussen & Rubinstein [46] and parts of VIII. More substantial remarks. IV.1 and X. 5 from Asmussen & Kliippelberg [36] with the permission from Elsevier Science . Fig.2 by Rafal Kulik . Lund September 2001 Soren Asmussen Acknowledgements Many of the figures . Section VIII. . supported by Center for Mathematical Physics and Stochastics (MaPhySto).6 by my 1999 simulation class in Lund. Parts of X.5 from Asmussen [21] with permission from CRC Press. A number of other figures were supplied by Christian Geisler Asmussen . Fig. as well as some additional references continue to be at the web page. not least the more complicated ones.4 from Asmussen. 3 is reprinted from Asmussen & Nielsen [39] and parts of IX. 111 . Schmidli & Schmidt [47] with the permission from Applied Probability Trust . Parts of II. Section VII . of which there are not many at this stage . 5.8 . many of which were pointed out by Hanspeter Schmidli .6 is reprinted from Asmussen & Schmidt [49] and parts of IX. Aarhus.1 by Bjarne Hojgaard and the table in Example 111.6. were produced by Lone Juul Hansen .PREFACE xi The second printing differs from the first only by minor corrections.
This page is intentionally left blank .
1) We also refer to t/) ( u) and 0(u. respectively. results and topics to be studied in the rest of the book. as defined in broad terms . is a model for the time evolution of the reserves of an insurance company. (1. M = (1.2) (O<t<T Ro=ul. T) as ruin probabilities with infinite horizon and finite horizon .3) sup St. t/i(u) = P (infRt < 0) = P (infR t < 0 t>0 t>0 The probability of ruin before time T is t.4) O<t<oo O<t<T 1 . The probability O(u) of ultimate ruin is the probability that the reserve ever drops below zero. For mathematical purposes.Chapter I Introduction 1 The risk process In this chapter . MT = sup St. (1. it is frequently more convenient to work with the claim surplus process {St}t>0 defined by St = u .Rt. (1. we introduce some general notation and terminology. Letting T(u) = inf {t > 0 : Rt < 0} = inf It > 0 : St > u}. and give a very brief summary of some of the models. A risk reserve process { Rt}t>o. They are the main topics of study of the present book. We denote throughout the initial reserve by u = Ro.T) = P inf Rt < 0 I .i(u.
1.pt. . t] is finite.i(u. St = E Uk . (1. (1. That is. Figure 1. and T1 is the time of the first claim.1. we see that Nt Nt Rt = u + pt . Thus. the following setup will cover the vast majority of the book: • There are only finitely many claims in finite time intervals..2 CHAPTER I. per unit time.6) Sofar we have not imposed any assumptions on the risk reserve process. and Nt = min {n > 0 : 0rn+1 > t} = max {n > 0: Un < t}• The size of the nth claim is denoted by Un. say. Putting things together..5) i. (1. the time of arrival of the nth claim is an = T1 + • • • + Tn. We denote the interarrival times of claims by T2.E Uk.b(u) = P (r(u) < oo) = P(M > u). respectively. the ruin probabilities can then alternatively be written as . However.7) k=1 k=1 The sample paths of {Rt} and {St} and the connection between the two processes are illustrated in Fig. T3. INTRODUCTION be the time to ruin and the maxima with infinite and finite horizon. • Premiums flow in at rate p. the number Nt of arrivals in [0.1 .T) = F (MT > u) = P(r(u) < T).
for example. then M < oo a.1. t * oo.s.. then M = oo a. . not discuss whether this actually corresponds to practice. Some main examples of models not incorporated in the above setup are: • Models with a premium depending on the reserve (i. VII. It would appear obvious. We study this case in Ch.e. we shall. For the purpose of studying ruin probabilities this distinction is. on Fig. since any modeling involves some approximative assumptions. and hence . one could well replace Rt by Rtnr(u) or RtA. • Brownian motion or more general diffusions. rl= pP P It is sometimes stated in the theoretical literature that the typical values of the safety loading 77 are relatively small. and the basic ruin probabilities are derived in XI. If 77 < 0.(. one may well argue that Brownian motion in itself could be a reasonable model.b(u) = 1 for all u. We shall discuss Brownian motion somewhat in Chapter IV. The models we consider will typically have the property that there exists a constant p such that Nt a E Uk k=1 p. a basic references is Gerber [127].8) holds.s. however. though many results are straightforward to generalize from the compound Poisson model. and in fact: Proposition 1.1 the slope of {Rt} should depend also on the level). but as an approximation to the risk process rather than as a model of intrinsic merit.1. however.1 Assume that (1. A further basic quantity is the safety loading (or the security loading) n defined as the relative amount by which the premium rate p exceeds p. Thus. 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. If 77 > 0. 1. that the insurance company should try to ensure 77 > 0.8) The interpretation of p is as the average amount of claim per unit time. However. and hence O(u) < 1 for all sufficiently large u. say 10% .) V 0. 1.1. (1. of course. • General Levy processes (defined as continuous time processes with stationary independent increments) where the jump component has infinite Levy measure.. We shall not deal with this case either.20%. immaterial.
. t t p  p' t ^ oo.8) that F N.s.. namely.2 (Cox PROCESSES) Here {Nt} is a Poisson process with random rate /3(t) (say) at time t.10) Again. _ St __ k =1 Uk pt a4. are i.d.11) .10) is a property which we will typically encounter. U2. k=1 (1. and here (1. M < oo a. The simplest concrete example (to be studied in Chapter III) is the compound Poisson model. (1. and independent of {Nt}. rl > 0.T) = i. (1. INTRODUCTION Proof It follows from (1.b(u) < 1 for all u when rl > 0. However. The simplest example is 3(t) = V where V is a r . Thus p may well be random for such processes. zP(u . Here it is easy to see that p = . with the most notable special case being V having a Gamma distribution. If u oo.6EU (on the average. Then the connection between the ruin probabilities for the given risk process {Rt} and those ^(u).i(u. if {(3(t)} is nonergodic.Q (say) and U1. not all models considered in the literature have this feature: Example 1.d. 0(u. where {Nt} is a Poisson process with rate . we obtain typically a somewhat stronger conclusion.8). 0 We shall only encounter a few instances of a Cox process. (1. If 77 < 0. tb(u) = 1 for all u holds also when rl = 0. in connection with risk processes in a Markovian or periodic environment (Chapter VI).10) hold with p constant. Nt)}. U2. and independent of {(0(t). it is not too difficult to show that p as defined by (1.T) for {Rt} is given by V)(u) = t/i (u).3 Assume p 54 1 and define Rt = Rt1p.i. then this limit is > 0 which implies St a$ oo and hence M = oo a.. namely that M = oo a. corresponding to the Pdlya process. .. and that .8) is given by ^t p = EU • lim it (3(s) ds t.i. St In concrete models.s.s.4 CHAPTER I. this needs to be verified in each separate case.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. If U1. . Proposition 1. . This case is referred to as the mixed Poisson process. However.Tp). are i.. then similarly limSt/t < 0.v.oo t 0 J (provided the limit exists).
g.. Insurance: Mathematics and Economics.g. Taylor [364]. Besides in standard journals in probability and applied probability. and in fact p < 1 is the fundamental assumption of queueing theory ensuring steadystate behaviour (existence of a limiting stationary distribution). Schmidli. and we do not get near to the topic anywhere in this book. Mitteilungen der Verein der Schweizerischen Versicherungsmathematiker and the Scandinavian Actuarial Journal. Rolski. see e . Notes and references The study of ruin probabilities. Segerdahl [334] and Philipson [289]. Daykin et al. Sundt [354]. often referred to as collective risk theory or just risk theory. We roughly classify these into two groups . Cox processes are treated extensively in Grandell [171]. CLAIM SIZE DISTRIBUTIONS 5 The proof is trivial. in a number of models. Pentikainen & Pesonen [101]. was largely initiated in Sweden in the first half of the century. U2. 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. Schmidt & Teugels [307] and Seal [326]. which is feasible since in most cases the process { Rt } has a similar structure as {Rt} (for example. in particular.. Buhlmann [82]. Straub [353]. Heilmann [191]. see also Chapter XI. [101]. many results and methods in random walk theory originate from there and the area was ahead of related ones like queueing theory. For mixed Poisson processes and Polya processes. we shall be able to identify p with the traffic intensity of an associated queue. Hipp & Michel [198]. another important early Swedish work is Tacklind [373]. [76]. Daykin.. De Vylder [110]. Note that life insurance (e. Some main later textbooks are (in alphabetical order) Buhlmann [82]. but in probability and applied probability as a whole. The Swedish school was pioneering not only in risk theory. Gerber [159]) has a rather different flavour.2. [330]. Embrechts et al. An idea of the additional topics and problems one may incorporate under risk theory can be obtained from the survey paper [273] by Norberg. lighttailed distributions (sometimes the term . the research literature is often published in journals like Astin Bulletin . Some early surveys are given in Cramer [91]. the role of the result is to justify to take p = 1. The term risk theory is often interpreted in a broader sense than as just to comprise the study of ruin probabilities. while the first mathematically substantial results were given in Lundberg [251] and Cramer [91]. the claim arrivals are Poisson or renewal at the same time). some main texts (typically incorporating some ruin theory but emphasizing the topic to a varying degree) are Bowers et al.. Note that when p = 1. Some of the main general ideas were laid down by Lundberg [250]. the recent survey by Grandell [173] and references therein. Grandell [171]. In the even more general area of nonlife insurance mathematics. [134]. Since { Rt } has premium rate 1. Gerber [157]. the assumption > 0 is equivalent to p < 1.
1 (THE EXPONENTIAL DISTRIBUTION) Here the density is b(x) = beax (2. s<8.1) The parameter 6 is referred to as the rate or the intensity. then the conditional distribution of X . 6 has density r(p)xPleax b(x) P and m.e. Equivalently. if 1 °O AB Jbos x B(dx) > 0. a simple stopping time argument shows that this implies that the conditional distribution of the overshoot ST(u) .f.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'.2 (THE GAMMA DISTRIBUTION) The gamma distribution with parameters p. and can also be interpreted as the (constant) failure rate b(x)/B(x). for the compound Poisson model with exponential claim sizes the ruin probability . (2. a fact which turns out to contain considerable information. The crucial feature is the lack of memory: if X is exponential with rate 6. but different more restrictive definitions are often used: subexponential. Example 2 . In contrast.O(u) can be found in closed form.u at the time of ruin given r(u) is again exponential u with rate 8. As in a number of other applied probability areas.x given X > x is again exponential with rate b (this is essentially equivalent to the failure rate being constant).2) = 0.B(x) satisfies B(x) = O(e8x) for some s > 0. B is heavytailed if b[s] = oo for all s > 0.g. the exponential distribution is by far the simplest to deal with in risk theory as well. Here lighttailed means that the tail B(x) = 1 . For example in the compound Poisson model. i.f.6 CHAPTER I. regularly varying (see below) or even regularly varying with infinite variance. 2a Lighttailed distributions Example 2. the m. where B(bo. INTRODUCTION 'Cramertype conditions' is used).g.3) .2 and /LB is the mean of B. On the more heuristical side. P B[s]= (8Is ) . B[s] is finite for some s > 0. In particular. and heavytailed distributions.
The exact form of the tail B(x) is given by the incomplete Gamma function r(x.). .2..c.1 Poisson events in [0. the Gamma density (2. one has r(bx. p). B(x) = r(p) Asymptotically.. 0 < ai < 1. X2. are i. An appealing feature is its simple connection to the Poisson process: B(x) = P(Xi + • • • + XP > x) is the probability of at most p . 2. > 1 for p < 1 and = 1 for p = 1 (the exponential case).. p) °° where r (x... or just the Erlang(p) distribution.v.ate (b2 ): L• i=o In the present text. Ruin probabilities for the general case has been studied.i.v.1) (or the 1/pth root if p < 1). This special case is referred to as the Erlang distribution with p stages.y i=1 where >i ai = 1. u Example 2 .3 (THE HYPEREXPONENTIAL DISTRIBUTION) This is defined as a finite mixture of exponential distributions..2) can be considered as the pth power of the exponential density (2. In particular. . we develop computationally tractable results mainly for the Erlang case (p = 1. i = 1. if p is integer and X has the gamma distribution p. x] so that B(x) = r` e. . and exponential with rate d. 0. where X1. is > 1. JP 1 B(x) r(p ) XP ie ax In the sense of the theory of infinitely divisible distributions..c. p. P b(x) = r` aibiea.d. CLAIM SIZE DISTRIBUTIONS 7 The mean EX is p/b and the variance Var X is p/b2. among others. then X v Xl + • • • + X. by Grandell & Segerdahl [175] and Thorin [369].) VarX1 (EX )2 p is < 1 for p > 1. An important property of the hyperexponential distribution is that its s. In particular. the squared coefficient of variation (s. u . p) = J tPletdt.
of which one is absorbing and the rest transient.d.e.5 (DISTRIBUTIONS WITH RATIONAL TRANSFORMS) A distribution B has a rational m. See XI. We give a more comprehensive treatment in VIII.4 (PHASETYPE DISTRIBUTIONS) A phasetype distribution is the distribution of the absorption time in a Markov process with finitely many states. which is slightly smaller but more amenable to probabilistic reasoning.6. INTRODUCTION Example 2 . . T) or sometimes the triple (E. it is notable from a practical point of view because of reinsurance: if excessofloss reinsurance has been arranged with retention level xo. the restriction T of the intensity matrix of the Markov process to E and the row vector a = (ai)iEE of initial probabilities.1 and defer further details to u Chapter VIII.8) are realvalued. B(x) = aeTxe where t = Te and e = (1 . a.8 CHAPTER I.6. This class of distributions plays a major role in this book as the one within computationally tractable exact forms of the ruin probability z/)(u) can be obtained.7) q1 b(x) = cjxieWWx + djxi cos(ajx)ea'x + > ejxi sin(bjx)e`ix . Important special cases are the exponential. but the current trend in applied probability is to restrict attention to the class of phasetype distributions. B(x) > 0 for x < xo) is of course a trivial instance of a lighttailed distribution. Equivalent characterizations are that the density b(x) has one of the forms q b(x) j=1 = cjxienbx. We give some theory for matrixu exponential distribution in VIII.8) j=1 j=1 j=1 where the parameters in (2.xo)+ is covered by the reinsurer). q2 q3 (2.g. the Erlang and the hyperexponential distributions. This class of distributions is popular in older literature on both risk theory and queues. However. Example 2 . The density and c. The parameters of a phasetype distribution is the set E of transient states.(2. there exists a xo < oo such that B(x) = 0 for x > xo. a rational Laplace transform) if B[s] _ p(s)/q(s) with p(s) and q(s) being polynomials of finite degree. (or. are b(x) = aeTxt. resp. Example 2 .f..f. The couple (a.. then the claim size which is relevant from the point of view of the insurance company itself is U A xo rather than U u (the excess (U .6 (DISTRIBUTIONS WITH BOUNDED SUPPORT) This example (i. equivalently. 1)' is the column vector with 1 at all entries. T) is called the representation.7) are possibly complexvalued but the parameters in (2.
(2.1. In particular. b(x) = crx''le`xr. 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). Writing c = d/r. Here failure rates b(x) = b(x)/B(x) play an important role. one being B(x) (1 + X)b(x) (1 + x)a+1' x > 0. or equivalently as the distribution of a°U+µ where U .9) which is heavytailed when 0 < r < I.p a 1 (2.12) Sometimes also a location parameter a > 0 and a scale parameter A > 0 is allowed.11) ex log logx 2r p 1 1 2 ( a ) f 1 (lox_P)2} (2. (2.pl = 1 W (logx . b(x) _ A(1 + (x a The pth moment is finite if and only if p < a . a2). It follows that the density is 't (1ogX .N(p. u Example 2 .u l b(x) = d dx or J ax lor 1 exp Asymptotically. the tail is B (x ) 2 x. Example 2 .2. CLAIM SIZE DISTRIBUTIONS 9 2b Heavytailed distributions Example 2. x < a.7 (THE WEIBULL DISTRIBUTION) This distribution originates from reliability theory. All moments are finite. p is defined as the distribution of ev where V . However.10) The loinormal distribution has moments of all orders.8 (THE LOGNORMAL DISTRIBUTION) The lognormal distribution with parameters a2.N(0. There are various variants of the definition around.13) u .9 (THE PARETO DISTRIBUTION) Here the essence is that the tail B(x) decreases like a power of x. we obtain the Weibull distribution B(x) = eCx'. a)/A)a+1' x > a. the exponential distribution representing the simplest example since here b(x) is constant. the mean u is eµ+a /2 and the second moment is e2µ+2o2.1). and then b(x) = 0. (2.
v. x + 00.'s of the form YX. in particular. B(x) = O(xP). the loggamma distribution (with exponent 5) and a Pareto mixture of exponentials.L( x ).15) x2 + 16x3 ) a3x/2) 3 (1 .13). where Y is Pareto distributed with a = (p .10 (THE LOGGAMMA DISTRIBUTION) The loggamma distribution with parameters p. INTRODUCTION Example 2.12) (here L (x) * 1) and ( 2. satisfies L(xt)/L(x) 4 1. { s () 1s+3s29s3log(1+2s I p=3. (2.(1 + 2x + 2x2)e2x) p = 2 (2.12 (DISTRIBUTIONS WITH REGULARLY VARYING TAILS) The tail B(x) of a distribution B is said to be regularly varying with exponent a if B(x) . the loggamma distribution is a Pareto distribution. Thus. examples of distributions with regularly varying tails are the Pareto distribution (2.1)/p. x 4 oo (any L having a limit in (0. another standard example is (log x)'). the density is { 3 (1 . 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). (2. In general. For p = 1.16) 11 Example 2. A = 1 and X is standard exponential. Choudhury & Whitt [1] as the class of distributions of r. oo) is slowly varying .17) where L (x) is slowly varying. The density is 8p(log x)pi b(x) . u Example 2 . u .x6+lr(p) (2. i. The simplest examples correspond to p small and integervalued.11 (PARETO MIXTURES OF EXPONENTIALS) This class was introduced by Abate.10 CHAPTER I.14) The pth moment is finite if p < 5 and infinite if p > 5.(1 + Zx + $ p = 3.e.2). 6 is defined as the distribution of et' where V has the gamma density (2.
it will be seen that we obtain completely different results depending on whether the claim size distribution is exponentially bounded or heavytailed. By far the most prominent case is the compound Poisson (CramerLundberg) model where {Nt} is Poisson and independent of the claim sizes U1. and so is the Weibull distribution with 0 < r < 1. and based upon such information it seems questionable to extrapolate to tail behaviour. the claim size distribution represents of course only one aspect (though a major one). U2.. We return to a closer study in IX. for example the lognormal distribution is subexponential (but not regularly varying). However.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. we may know that such a process (with a covariance function estimated from data) is a reasonable description of the behaviour of the system under study in typical conditions. We give some discussion on standard methods to distinguish between light and heavy tails in Section 4f. From a practical point of view. (2. Namely.18) B(x) It can be proved (see IX. Also. When studying ruin probabilities. but the model also admits a natural interpretation : a large portfolio of insurance holders . though the proof of this is nontrivial... The reason is in part mathematical since this model is the easiest to analyze. THE ARRIVAL PROCESS 11 Example 2.1. Similar discussion applies to the distribution of the accumulated claims (XI. one may argue that this difficulty is not resticted to ruin probability theory alone. this phenomenon represents one of the true controversies of the area. 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. 3 The arrival process For the purpose of modeling a risk process . but can never be sure whether this is also so for atypical levels for which far less detailed statistical information is available.13 (THE SUBEXPONENTIAL CLASS OF DISTRIBUTIONS) We say that a distribution B is subexponential if xroo lim B `2^ = 2. Thus. the knowledge of the claim size distribution will typically be based upon statistical data.1) that any distribution with a regularly varying tail is subexponential.3.. the subexponential class of distributions provide a convenient framework for studying large classes of heavyu tailed distributions. which each have a ( timehomogeneous) small rate of experiencing a .
Some of them have concentrated on the marginal distribution of NT (say T = one year ). and also that the ruin problem may be hard to analyze . The one we focus on (Chapter VI) is a Markovian environment : the environmental conditions are described by a finite Markov process {Jt }too. has some mathematically appealing random walk features . To the author 's knowledge . IV (and. it is more questionable whether it provides a model with a similar intuitive content as the Poisson model. getting away from the simple Poisson process seems a crucial step in making the model more realistic. In order to prove reasonably substantial and interesting results . the first extension to be studied in detail was {Nt } to be renewal (the interarrival times T1 . radioactive decay (a huge number of atoms each splitting with a tiny rate ) and many other applications. with the extension to premiums depending on the reserve. gives rise to an arrival process which is very close to a Poisson process. too general and one neeed to specialize to more concrete assumptions . Another one is Cox processes. but with a general not necessarily exponential distribution ).(3. Cox processes are.. so that . when Jt = i. The point of view we take here is Markov dependent random walks in continuous time (Markov additive processes ). INTRODUCTION claim .d.8 (t) is a periodic function of t. This model . Mathematically. This model can be intuitively understood in some simple cases like { Jt} describing weather conditions in car insurance . A more appealing way to allow for inhomogeneity is by means of an intensity .. the negative binomial distribution. we study this case in VI . however. such that 8(t) = . to be studied in Chapter V. However . In others. the periodic and the Markov modulated models also have attractive features .3(t) fluctuating over time. where {/3 (t)}too is an arbitrary stochastic process . An obvious example is 3(t) depending on the time of the year (the season). The compound Poisson model is studied in detail in Chapters III. not many detailed studies of the goodnessoffit of the Poisson model in insurance are available ... are i.g.12 CHAPTER I. in just the same way as the Poisson process arises in telephone traffic (a large number of subscribers each calling with a small rate). found the Poisson distribution to be inadequate and suggested various other univariate distributions as alternatives .e. T2. its basic feature is to allow more variation (bursty arrivals ) than inherent in the simple Poisson process. e.i. I. Historically. it may be used in a purely descriptive way when it is empirically observed that the claim arrivals are more bursty than allowed for by the simple Poisson process. in particular to allow for certain inhomogeneities. with a common term {Nt} is a Markovmodulated Poisson process . 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 . see 11. Nevertheless . epidemics in life insurance etc. This applies also to the case where the claim size distribution depends on the time of the year or . which facilitate the analysis.6. 5. in Chapter VII).
point processes and so on. however. Mathematically.1) holds as well provided the risk process has a premium rule depending on the reserve. others are quite different.T) = P(VT > u). stochastic differential equations. 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. it is a recurrent theme of this book to stress this connection which is often neglected in the specialized literature on risk theory. The study of the steady state is by far the most dominant topic of queueing and storage theory. and the limit t 4 oo is the steadystate limit. 0(u) = P(V > u). time series and Gaussian processes. this amounts to Vo having the stationary distribution of {Vt}). Some of these have a certain resemblance in flavour and methodology. Similarly. reliability. genetics models. queueing theory. It should be noted. interacting particle systems.1). it is desirable to have a set of formulas like (4. methods or modeling ideas developed in one area often has relevance for the other one as well. More generally. R = p(R) in between jumps. that quite often the emphasis is on computing expected values like EV. and a lot of information on steadystate r. A general release rule p(x) means that {Vt} decreases according to the differential equation V = p(V) in between jumps. In the setting of (4.v. In fact. this gives only f0 O°i (u)du which is of limited . Thus. A stochastic process {Vt } is said to be in the steady state if it is strictly stationary (in the Markov case.6) . The M/G/1 workload process { Vt } may also be seen as one of the simplest storage models. extreme value theory. 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 . ruin probabilities for risk processes with an input process which is renewal.1) permitting to translate freely between risk theory and the queueing/storage setting.0 (u.1) where V is the limit in distribution of Vt as t + oo. (4. The ones which appear most related to risk theory are queueing theory and dam/storage processes. and here (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. dam/storage processes.4. stochastic geometry.'s like V is available. Markovmodulated or periodic can be related to queues with similar characteristics. others being branching processes. with Poisson arrivals and constant release rule p(x) = 1.
1 .2). Similarly.T). much of the study of finite horizon problems (often referred to as transient behaviour) in queueing theory deals with busy period analysis which has no interpretation in risk theory at all. the ideal is to be able to come up with closed form solutions for the ruin probabilities 0(u). • The compound Poisson model with some rather special heavytailed claim size distributions. The infinite horizon (steady state ) case is covered by letting T oo. have to some extent a different flavour. the two areas. B(x) = ebx. 4b Exact solutions Of course . see Boxma & Cohen [74] and Abate & Whitt [3].3. • The compound Poisson model with a claim size distribution degenerate at one point. The fact that Theorem H. Here Vi(u) is given in terms of a matrixexponential function ( Corollary VIII.1 is a sample path relation should be stressed : in this way the approach also applies to models having supplementary r.p(y) y^ Jo p(x) can be written in closed form. though overlapping.'s like the environmental process {Jt} in a Markovmodulated setting. Example VIII. A prototype of the duality results in this book is Theorem 11. • The compound Poisson model with constant premium rate p = 1 and B being phasetype with a just few phases .6.14 CHAPTER I.8. Vi(u. p = 0/8 and y = 8 . see Corollary VII.1. see Corollary III.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 ).3. which can be expanded into a sum of exponential terms by diagonalization (see.3.3. Here ?P(u) is explicit provided that . The cases where this is possible are basically the following for the infinite horizon ruin probability 0(u): • The compound Poisson model with constant premium rate p = 1 and exponential claim size distribution B. INTRODUCTION intrinsic interest . . • The compound Poisson model with premium rate p(x) depending on the reserve and exponential claim size distribution B. the functions w x f d 1 exdx () . 3.g. 3. Here O(u) = peryu where 3 is the arrival intensity. The qualifier 'with just a few phases ' refers to the fact that the diagonalization has to be carried out numerically in higher dimensions.. which gives a sample path version of (4. as is typically the case. e . Thus .v.1).
A notable fact ( see again XI. O(u. Ab(u). [s. a2 (x): Ip (u) = where S(u) = f °O exp {. • An astable Levy process with drift . esuTb( u. Also Brownian models or certain skip free random walks lead to explicit solutions (see XI .f f 2µ(y)/a2(y) dy} dx  (4. Given this can be done. T). (u.Lef$er function. 1). T) can then be calculated numerically by some method for transform inversion.ff 2µ(y)/a2(y) dy} dx . see VIII. T) du dT 0 TO 00 in closed form than the ruin probabilities z/'(u). the second best alternative is a numerical procedure which allows to calculate the exact values of the ruin probabilities.4.7. Embrechts. . say the fast Fourier transform (FFT) as implemented in Grubel [179] for infinite horizon ruin probabilities for the renewal model.u(y)/a2(y) dy} 4c Numerical methods Next to a closedform solution. However. Here are some of the main approaches: Laplace transform inversion Often.1) are so complicated that they should rather be viewed as basis for numerical methods than as closedform solutions. Grubel & Pitts [132] and Grubel & Hermesmeier [180] (see also the Bibliographical Notes in [307] p. We don't discuss Laplace transform inversion much. where Furrer [150] recently computed ii(u) as an infinite series involving the Mittag. Abate & Whitt [2]. relevant references are Grubel [179].b(u)du . 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 . it is easier to find the Laplace transforms = f e8 . the formulas ( IV. f {eXp U LX 2.2) is the natural scale. A SUMMARY OF MAIN RESULTS AND METHODS 15 • The compound Poisson model with a two step premium rule p(x) and B being phasetype with just a few phases. T) themselves.S(u) 1S(oo) f °D exp {. 191).1) is the explicit form of the ruin probability when {Rt} is a diffusion with infinitesimal drift and variance µ(x). For the finite horizon ruin probability 0(u. but are somewhat out of the mainstream of the area .
as the solution of linear differential equations or by some series expansion (not necessarily the straightforward Eo U'u/n! one!). and carry out the solution by some standard numerical method.and integral equations The idea is here to express 'O(u) or '(u. see VIII. whereas for the renewal arrival model and the Markovian environment model U has to be calculated numerically. either as the iterative solution of a fixpoint problem or by finding the diagonal form in terms of the complex roots to certain transcendental equations. 4d Approximations The CramdrLundberg approximation This is one of the most celebrated result of risk theory (and probability theory as a whole). Differential.3) in the compound Poisson model which is an integral equation of Volterra type.3. 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. For the compound Poisson model with p = 1 and claim size distribution B with moment generating function (m. (4. most often it is more difficult to come up with reasonably simple equations than one may believe at a first sight. and in particular the naive idea of conditioning upon process behaviour in [0.f. 0(u) is then given in terms of a matrixexponential function euu (here U is some suitable matrix) which can be computed by diagonalization.1) and y > 0 is the solution of the Lundberg equation (4. In the compound Poisson model with p = 1.7. However.16 CHAPTER L INTRODUCTION Matrixanalytic methods This approach is relevant when the claim size distribution is of phasetype (or matrixexponential).4) 00['Y]1)'Y = 0. One example where this is feasible is the renewal equation for tl'(u) (Corollary III. U is explicit in terms of the model parameters. and in quite a few cases (Chapter VIII).3) where C = (1 .p)/(13B'[ry] . it states that i/i(u) . T) as the solution to a differential. dt] most often leads to equations involving both differential and integral terms.g. .or integral equation. u * oo.) B[s]. which can equivalently be written as f3 [7] = 1 +13 .Ce"u.
In particular. J B dx. often for all u > 0 and not just for large u.2.6) 4e Bounds and inequalities The outstanding result in the area is Lundberg's inequality (u) < e"lu. and use the fact that first passage probabilities are more readily calculated for diffusions than for the risk process itself. However. when the claim size distribution is of phasetype. . (4. In the case of heavytailed distributions. It has generalizations to the models with renewal arrivals. Diffusion approximations are easy to calculate. but typically not very precise in their first naive implementation. 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. incorporating correction terms may change the picture dramatically. u > oo. corrected diffusion approximations (see IV.7 and IV. For example. in such cases the evaluation of C is more cumbersome. The CramerLundberg approximation is renowned not only for its mathematical beauty but also for being very precise. the claim size distribution should have an exponentially decreasing tail B(x). 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.6) are by far the best one can do in terms of finite horizon ruin probabilities '(u.4. some further possibilities are surveyed in 111 . a Markovian environment or periodically varying parameters. in some cases the results are even more complete than for light tails. In fact. for the compound Poisson model ^(u) p pu In fact . However. Approximations for O(u) as well as for 1(u. See Chapter IX. Large claims approximations In order for the CramerLundberg approximation to be valid. T). other approaches are thus required. 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.
empirical evidence shows that the general principle holds in a broad variety of settings. they have however to be estimated from data. UNT are i. though not too many precise mathematical results have been obtained. . fitting a parametric model to U1. and to plot the empirical mean residual life 1 N .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. How do we produce a confidence interval? And.3).k (U(`) . However. . . say degenerate at m.8. For example. 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. to have smaller ruin probabilities than when B is nondegenerate with the same mean m. INTRODUCTION Compared to the CramerLundberg approximation (4. obtained say by observing the risk process in [0. T]. . the difficulty comes in when drawing inference about the ruin probabilities. which is a standard statistical problem since the claim sizes Ui.. e. This procedure in itself is fairly straightforward. This is proved for the compound Poisson model in 111.. We return to various extensions and sharpenings of Lundberg's inequality (finite horizon versions. UNT may be viewed as an interpolation in or smoothing of the histogram). 4f Statistical methods Any of the approaches and results above assume that the parameters of the model are completely known.d. In practice.18 CHAPTER I.g. in the compound Poisson model. as a general rule. it is a general principle that adding random variation to a model increases the risk. However. 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. . one may question whether it is possible to distinguish between claim size distributions which are heavytailed or have an exponentially decaying tail. given NT. of being somewhat easier to generalize beyond the compound Poisson setting. lower bounds etc.U(k)) i =k+ i . The standard suggestion is to observe that the mean residual life E[U .. it splits up into the estimation of the Poisson intensity (the estimator is /l3 = NT/T) and of the parameter(s) of the claim size distribution. more importantly.i. When comparing different risk models. For example. .) at various places and in various settings. one expects a model with a deterministic claim size distribution B.
4g Simulation The development of modern computers have made simulation a popular experimental tool in all branches of applied probability and statistics. UN. in all other chapters than VI where we just write . where U(1) < .d. Klnppelberg & Mikosch [134].v's) which can be generated by simulation. 5 Conventions Numbering and reference system The basic principles are just as in the author's earlier book Applied Probability and Queues (Wiley 1987. (or a functional of the expectation of a set of r. . and also discuss how to develop methods which are efficient in terms of producing a small variance for a fixed simulation budget. CONVENTIONS 19 as function of U(k).. Simulation may be used just to get some vague insight in the process under study: simulate one or several sample paths. Thus Proposition 4.5. 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.4.2. formula (5.. < U(N) are the order statistics based upon N i.. but is not very satisfying.2. and of course the method is relevant in risk theory as well. and in fact methods from that area can often be used in risk theory as well . The chapter number is specified only when it is not the current one. reference [14].. .i. The problem is entirely analogous to estimating steadystate characteristics by simulation in queueing/storage theory. in this book referred to as [APQ]).. the more typical situation is to perform a Monte Carlo experiment to estimate probabilities (or expectations or distributions) which are not analytically available. respectively. and look at them to see whether they exhibit the expected behaviour or some surprises come up.v. See further Embrechts. For example. having small probability) and that therefore naive simulation is expensive or even infeasible in terms of computer time. this is a straightforward way to estimate finite horizon ruin probabilities. However. A main problem is that ruin is typically a rare event (i.3) and Section VI.e. Still.3) or Section 3 of Chapter VI are referred to as Proposition VI. We look at a variety of such methods in Chapter X. The infinite horizon case presents a difficulty. claims U1. Truncation to a finite horizon has been used.3). formula VI.(5.3 (or just VI. because it appears to require an infinitely long simulation.
v. In particular. n i oo. B is concentrated on [0. i. B[s] the m.f. IIGII the total mass (variation ) of a (signed ) measure G . oo).4. for a probability distribution IIGII = 1.3) or Section 3. i.e. with respect to w.d.f. A different type of asymptotics: less precise.p. or a more precise one like eh .h. then for 1s < 5). squared coefficient of variation.s. w. left hand side (of equation) m.Used in asymptotic relations to indicate that the ratio between two expressions is 1 in the limit.ceax. E expectation.c.f. with probability Mathematical notation P probability. cumulant generating function. . b[s] is defined always if Rs < 0 and sometimes in a larger strip (for example. (A.d. moment generating function. right hand side (of equation) r. B(dy).f.The same symbol B is used for a probability measure B(dx) = P(X E dx) and its c. B(x) = P(X < x) = fx.h. .r. cumulative distribution function P(X < x) c.t.i. independent identically distributed i.s.g.g. as for typical claim size distributions. Abbreviations c.g. E. B(x) the tail 1 . mation. References like Proposition A.2. EX2/(EX)2.29) refer to the Appendix. h + 0. say a heuristic approxi1 + h + h2/2.20 CHAPTER L INTRODUCTION Proposition 4. (moment generating function) fm e82B(dx) of the distribution B.f. log E[s] where b[s] is the m. If. infinitely often l.v.g. random variable s.f.g.B(x) = P(X > x) of B. if B(x) .o. r.d. n! 27r nn+1/2en. formula (5. The Laplace transform is b[s]. and for a defective probability distribution IIGII < 1. see under b[s] below.
of numbers.oi denotes the column vector with the xi as components Special notation for risk processes /3 the arrival intensity (when the arrival process is Poisson). . I(A) the indicator function of the event A. matrices have uppercase Roman or Greek letters like T. 21 D [0. (the dimension is usually clear from the context and left unspecified in the notation). Usually. oo) the space of Rvalued functions which are rightcontionuous and have left limits. (xi)diag denotes the diagonal matrix with the xi on the diagonal (xi)row denotes the row vector with the xi as components (xi). E[X. Matrices and vectors are denoted by bold letters. Unless otherwise stated. and column vectors have lowercase Roman letters like t. i. intensity interpretation.e.e. Usually. xa. 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. R(s) the real part of a complex number s. i. an example or a remark. 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. F o r a given set x1. CONVENTIONS {6B the mean EX = f xB(dx) of B ABA' the nth moment EXn = f x"B(dx) of B. though slightly more complicated.. 7r. Notation like f3i and 3(t) in Chapter VI has a similar . In the Frenchinspired literature. Xt_ the left limit limstt X8f i. only have finitely many jumps in each finite interval.e.. all stochastic processes considered in this book are assumed to have sample paths in this space. Thus.A] means E[XI(A)]. . N(it. a2) the normal distribution with mean p and variance oa2. the ith unit row vector is e'i. the ith entry is 1 and all other 0. a.5. the value just before t. . 0 marks the end of a proof. A. often the term 'cadlag' (continues a droite avec limites a gauche) is used for the Dproperty. the processes we consider are piecewise continuous. row vectors have lowercase Greek letters like a.
111.5. interpretation. ry The adjustment coefficient.1.g. though slightly more complicated. cf. . FL. or quantities with a similar time average interpretation.5. I. J the rate parameter of B for the exponential case B(x) = eby. cf. 'q the safety loading . Notation like BE and B(t) in Chapter VI has a similar.22 CHAPTER L INTRODUCTION B the claim size distribution. p the net amount /3pB of claims per unit time. EL the probability measure and its corresponding expectation corresponding to the exponential change of measure given by Lundberg conjugation. e. cf. I. VI.1.
23 . Sections 4. however. in part. The reader should therefore observe that it is possible to skip most of the chapter. All results are proved elsewhere . in most cases via likelihood ratio arguments. Due to the generality of the theory. the relevance for the mainstream of exposition is the following: The martingale approach in Section 1 is essentially only used here. 5) are. not crucial for the rest of the book. however. somewhat more advanced than in the rest of the book. strictly speaking. The likelihood ratio approach in Section 2 is basic for most of the models under study. The duality results in Section 3 (and.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. in particular at a first reading of the book. 5 on random walks and Markov additive processes can be skipped until reading Chapter VI on the Markovian environment model. Sections 4. More precisely. however. fundamental ( at least in the author' s opinion) and the probability involved is rather simple and intuitive. used in Chapter VI on risk processes in a Markovian (or periodic) environment. the level of the exposition is. When encountered for the first time in connection with the compound Poisson model in Chapter III. The topic is. The general theory is. a parallel selfcontained treatment is given of the facts needed there.
V) (u.u denote the overshoot. the time to ruin r(u) is inf It > 0 : St > u}..QµB < 1. As usual.1 Assume that (a) for some ry > 0.24 CHAPTER II. and in the limit (1. Our first result is a representation formula for O(u) obtained by using the martingale optional stopping theorem .2) takes the form 1 = E [e'ys(). T(u) > T] . T(u) < oo] + 0 = eryuE [e7Vu). (b) St a$ oo on {T(u) = oo}. Then e7u (u) = E[e74(u)j7(u) < oo] Proof We shall use optional stopping at time r(u)AT (we cannot use the stopping time T(u) directly because P(T(u) = oo) > 0 and also because the conditions of the optional stopping time theorem present a problem.2 Consider the compound Poisson model with Poisson arrival rate . The more general Theorem 6.1 is basic for the study of the compound Poisson model in Chapter III. and the ruin probabilities are ip(u) = P (T(u) < oo). T) = P(T(u) < T). (1. Thus N.T(u) < cc] = e7uE {e7f(u) I T(u) < cc] z/.(u). T(u) < T] + E [eryST . We get 1 = Ee7So = E e'Y S(. claim size distribution B and p = . using r(u) A T invokes no problems because r(u) A T is bounded by T).2) As T > oo. Example 1 .)AT = E [e7ST(°). Let e(u) = ST(u) .0. SOME GENERAL TOOLS AND RESULTS The ladder height formula in Theorem 6. however. 1 Martingales We consider the claim surplus process {St} of a general risk process.5 can be skipped. the second term converges to 0 by (b) and dominated convergence (e7ST < eryu on {r(u) > T}). Proposition 1. f1 . {e'YS° }t>0 is a martingale. StUit.
f. u Corollary 1. Eeas° = e"(°) where n(a) = a2a2/2 .3 Assume that {Rt} is Brownian motion with variance constant o. the conditional distribution of the overshoot e(u) = U . Now at the time r(u) of ruin {St} upcrosses level u by making a jump . it follows that E [e7st+v I J] = e"rstE [e7(st+vSt) I Ft] = e7StEe"rs° = elst where . Then {St } is Brownian motion with variance constant o2 and drift p < 0.1.i.a = a .g. u Corollary 1.ap.Q and the U. Thus.Q(B[a] . 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.1) . and thus Ee7s° = 1.r" where y = S .u + x is again just exponential with rate S.d. are i.1 is satisfied. Since {St} has stationary independent increments.2.x. From this it is readily seen (see III. Proof Since c(a) = /3 (B[a] . A simple calculation (see Proposition III.. the conditions of Proposition 1.(„)_ = x is that of a claim size U given U > u . Example 1 .6a for details) that typically a solution to the Lundberg equation K(y) = 0 exists. By standard formulas for the m.3/6 < 1. and (b) follows from p < 1 and the law of large numbers (see Proposition III.= e"(') where K(a) = ./3.1.5 For the compound Poisson model with B exponential. of the normal distribution.1 are satisfied. B(x) _ edx.1) .4 (LUNDBERG ' S INEQUALITY ) tion 1 .Q. O(u ) < e7".2 and drift p > 0.a.1) shows that Eels. Thus. and p =.1.2(c)). and thus by the memoryless property of the exponential distribution . the ruin probability is O(u) = pe. condition (a) of Proposition 1. MARTINGALES 25 where {Nt} is a Poisson process with rate . Since {St} has stationary independent increments. with common distribution B (and independent of {Nt}).Ft = a(S" : v < t). From this it is immediately seen that the solution to the Lundberg equation ic(y) = 0 is y = 2p/a2.a it is immediately seen that y = S . 1. and thus Ee'rs° = 1. Thus 00 E [e'rt (") I T(u) < oo] = I e5e  dx = f 5edx . Under the conditions of Proposi Proof Just note that C(u) > 0.
P correspond to the claim surplus process of two compound Poisson risk processes with Poisson rates /3. which we equip with the natural filtration {. hence so is Nt = limfyo N2`i. Proof Just note that ^(u) = 0 by continuity of Brownian motion. The interesting concept is therefore to look for absolute continuity only on finite time intervals (possibly random. But if a $ ^ . oo). and is further exploited in his book [157]. . However. and in analogy with the theory of measures on finite dimensional spaces one could study conditions for the RadonNikodym derivative dP/dP to exist. Grandell [171]. A]. and by the law of large numbers for the Poisson process . and F. cf. Embrechts. Theorem 2. 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. as shown by the following example this setup is too restrictive: typically'. the parameters of the two processes can be reconstructed from a single infinite path.26 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Corollary 1.v. Two such processes may be represented by probability measures F. I. [172]. The number Nt F) of jumps > e before time t is a (measurable) r.3 below). we look for a process {Lt} (the likelihood ratio process) such that P(A) = E[Lt.e. Delbaen & Haezendonck [103] and Schmidli [320].1) 'though not always: it is not difficult to construct a counterexample say in terms of transient Markov processes. B. A somewhat similar u argument gives singularity when B $ B. F(S) = P(S) = 1. Grandell & Schmidli [131]. P on (DE. then S and S are disjoint .6 If {Rt} is Brownian motion with variance constant a2 and drift p > 0. Example 2 . Thus the sets S = I tlim +oot t =. (2. A E Ft. then z/'(u) = e7" where 'y = 21A/a2. u Notes and references The first use of martingales in risk theory is due to Gerber [156].6 N S = { lim Nt I t +00 t gJ are both in F. More recent references are Dassios & Embrechts [98]..Ft}too and the Borel afield F. on (DE.F). F). 0 and claim size distributions B. P are then singular (concentrated on two disjoint measurable sets).1 Let F.
t. ({Ft} .1) and nonnegativity by letting A = {Lt < 0}.r.r.Pt)) The following result gives the connection to martingales. define P by Pt (A) = E[Lt. P be as in Proposition 2.2) Proof Assume first G C {T < T} for some fixed deterministic T < oo. Hence E [_ . The truth of this for all A E Y. ({. 1 J (2. . Hence the family {Pt} is t>o consistent and hence extendable to a probability measure F on (DE.1) holds.2(i). under the assumptions of (ii) we have for A E rg and s < t that A E Ft as well and hence E[L8.2 Let {Ft}t>o be the natural filtration on DE. A E F8. Ft). LIKELIHOOD RATIOS AND CHANGE OF MEASURE 27 (i. _. ELt = 1 follows by taking A = DE in (2. G l ] = E [_I(G)E[LTIFT] ] = E { _I(G)Lr ] = P(G).. (i) If {Lt}t> o is a nonnegative martingale w. the restriction of P to (DE. Lets < t. By the martingale property. A E F. A] = E[Lt.1) holds. 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 . A]. If r is a stopping time and G E PT. G C {T < oo}.Tt) is absolutely continuous w.r. (ii) Conversely. Then P(A) = E[Lt.Y) such that P(A) = Pt(A). F the Borel o•field and P a given probability measure on (DE.F8] = L8 and the martingale property.e. Then Ft (A) = E[Lt. then {Lt} is a nonnegative martingale w. Finally. . we have E [ LTIFT]1 = LT on {T < T}.3 Let {Lt}.F). Then Lt > 0 and ELt = 1 ensure that Pt is a probability measure on (DE. then { 1 P(G) = EG .Pt}adapted process {Lt}t>o (2.Ft}. G ] = E [LT .t. using the martingale property in the fourth step. that the restriction of P to (DE. P) such that LLt = 1. Conversely. Proof Under the assumptions of (i). This proves (i). implies that E[LtI.A] = EE[LtI(A)IF8] = EI(A)E[LtIFB] = EI(A)L8 = PS(A). A E Ft . Lt < 0] can only be nonnegative if P(A) = 0. Proposition 2. then there exists a unique probability measure Pon . if for some probability measure P and some {. A].F such that (2. F) such that ELt = 1.t.2.
2) follows by monotone convergence. say. T(u) < oo].Ft} . where {Rt} is the risk reserve process.3) to G of{r < T} we get 1111 F(Gn {r <T}) = E[ 1 . u From Theorem 2. {St} = {u .. the natural filtration {.1. SOME GENERAL TOOLS AND RESULTS In the general case . both sides are increasing in T. First we ask when the Markov property is preserved. each F.r. applying (2. is nonnegative and has Ey Lt = 1 for all x. t. we assume for simplicity that {Xt} has Dpaths. is Markov w.r. For the definition. Now just rewrite the r.1: Corollary 2. 1 Since everything is non negative. and letting T * oo.28 CHAPTER II.4) Proof Letting G = {r(u) < oo}. of (2. Rt) or Xt = (Jt. and this problem will now be studied.1) in Proposition 1. (2. The problem is thus to investigate which characteristics of {Xt} and {Lt} ensure a given set of properties of the changed probability measure. A change of measure is performed by finding a process {Lt} which is a martingale w. one would typically have Xt = Rt. To this end.1).1) is that it seems in general easier to deal with the (unconditional) expectation E[eryVu).h. Xt = St.t.t.3 we obtain a likelihood ratio representation of the ruin probability V) (u) parallel to the martingale representation (1..s. we have F(G) = V )(u). Consider a (timehomogeneous) Markov process {Xt} with state space E.4) compared to (1. 5) for processes with some randomwalklike structure. first in the Markov case and next (Sections 4. The crucial step is to obtain information on the process evolving according to F. we need the concept of a multiplicative functional.O(u) = eryuE[e'YC(u). Xt = (Jt. St). (2. In the context of ruin probabilities. Lr(u) 11 The advantage of (2. in continuous time (the discrete time case is parallel but slightly simpler).2) by noting that 1 = ersr(„) = e1'ue7Ou).Gn {r <T} .Rt} the claim surplus process and {Jt} a process of supplementary variables possibly needed to make the process Markovian. .4 Under condition (a) of Proposition 1. r(u) < oo] occuring there than with the (conditional) expectation E[e'r{(u ) Jr(u) < oo] in (1.
(Xtitl) with all t(i) < t + s. (2.Pt+8measurable r.'s of the form Zt • (Y8 o 0t) comprises all r.5) are Tt+e measurable.8) which is the same as (2. which in turn is the same as Ex[Lt+8Zt • (V8 o Bt)] = Ex[Lt • (L8 o 91)Z1 • (Y8 o et)] (2.F8measurable r.Ft]. nonnegative and Lt+8 = Lt•(Lso9t) (2. since Zt • (Y8 o Ot ) is .Ftmeasurable. Lt has the form Lt = 'Pt ({x }0<u<t) for some mapping cot : DE[O.(A) = Ex [Lt.Ft] = Ex.5) is equivalent to Ex[Lt+8Vt+8] = E8[Lt • (L8 o 91)Vt+8] for any .s. since Ext [L8Y8] = E[(Y8 o et)(L8 o 8t)I.Ft+8measurable. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 29 on DE and define {Lt} to be a multiplicative functional if {Lt} is adapted to {.T9measurable Y8. Vt+e.v. A].'s of the form fl' f. Indeed. this in turn means Ex[Lt+8Zt(V8 oet)] = Ex[LtZtExt[L8Y8]].Ft }. The converse follows since the class of r.Y8f t < s..v.v. 0 . which is the same as Ex[Zt(Y8 o Bt)] = E8[ZtEx. ({Xt+u} 0<u<8) Theorem 2.5) Pxa. The precise meaning of this is the following: being . or. t. Proof Since both sides of (2. Zt.7). and then L.v.7).. (2. the Markov property can be written E. s. Similarly. the natural filtration {Ft} on DE. oo). o 9tI.r. t and let Px be the probability measure given by t.5 Let {Xt} be Markov w. let {Lt} be a nonnegative martingale with Ex Lt = 1 for all x.[Y. o 9t = V.2. Y8.6) for any . for all x. where Ot is the shift operator. Ex[Lt+8Zt(Y8 o et)] = Ex[LtZt(Y8 o 0t)(L8 o Bt)].YB] for any Ftmeasurable r. t] * [0. (2. Then the family {Px}xEE defines a timehomogeneous Markov process if and only if {Lt} is a multiplicative functional.t.6) implies (2.v.7) for any Ftmeasurable Zt and any . (2. By definition of Px.
. A more elementary version along the lines of Theorem 2. and the time to ruin is 7(u) = inf {t > 0: Rt < 0}... we shall establish a general connection between ruin probabilities and certain stochastic processes which occurs for example as models for queueing and storage. 3 Duality with other applied probability models In this section. SOME GENERAL TOOLS AND RESULTS to define a timehomogeneous Markov process. Indeed..30 CHAPTER H. UN. We work on a finite time interval [0. } E[Lt+B I. The result is a sample path relation. . and just after time or* {Vt} makes an upwards jump of size UU = UN _k+l. the premium rate is p(r) when the reserve is r (i... .At where At = k: vk <t U. the arrival epochs are Qi.1) The initial condition is arbitrary. and thus for the moment no parametric assumptions (on say the structure of the arrival process) are needed. u Notes and references The results of the present section are essentially known in a very general Markov process formulation. . aN where or* = T UN_k+l. (using the Markov property in the second step) so that the martingale property is automatic. CN. (3.. 0 < vl < .. . then Remark 2. {Vt} . see Dynkin [128] and Kunita [239]. < aN < T. The formulation has applications to virtually all the risk models studied in this book. with a proof somewhat different from the present one.. The storage process {Vt }o<t<T is essentially defined by time reversion.Ft] = LtE[L8 o 9t I. In between jumps.5 can be found in Kuchler & Sorensen [240].e. T] in the following setup: The risk process {Rt}o<t<T has arrivals at epochs or.. Ro = u (say). More precisely . In between jumps. t] = LtExt L8 = Lt.6 For {u . it xEE suffices to assume that {Lt} is a multiplicative functional with Ex Lt = 1 for all x. The corresponding claim sizes are Ul. reflection at zero and initiar condition Vo = 0. .. R = p(R)). t. Thus R = Ro + f p(R8) ds .
. (3.T) = inf Rt < 0 P (O<t<T P(r(u) < T) be the ruin probability..____•_.___ .. .1) we have Vt = At  f P(Vs)ds where A= U= AT . {Vt} remains at 0 until the next arrival). Theorem 3.x._.. V = p(V)).T) = P(VT > u)._: 1} 0 011 =T01N ^N3 To 0 011 014 01N Figure 3. That is.1.1 Define r(u) = inf It > 0: Rt < 0} (r(u) = oo if Rt > 0 for all t < T) and let ii(u. :.11 4. V)(u. DUALITY WITH OTHER APPLIED PROBABILITY MODELS 31 decreases at rate p(r) when Vt = r (i. 3..3. In particular. (3. Then rt°) > rt°) for all t when u > v.2) k: ok <t and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0.3) (u) Proof Let rt' denote the solution of R = p(R) subject to r0 = u.__.. Note that these definitions make {Rt} rightcontinuous (as standard) and {Vt} leftcontinuous. The sample path relation between these two processes is illustrated in Fig.e..__.1 The events {T(u) < T} and {VT > u} coincide.AT_t. instead of (3...
the distinction between right. If VaN > 0. the connection between risk theory and other applied probability areas appears first to have been noted by Prabhu [293] in a queueing context.d. this represents a model for storage. Suppose next VT < u (this situation corresponds to the broken path of {Rt} in Fig. = r(VT) .Ul < roil  Ul = RQ„ Va1V_1 < RQ2. see Siegmund [344].1 and its proof is from Asmussen & Schock Petersen [50].and left continuity is immaterial because the probability of a Poisson arrival at any fixed time t is zero).3). u A basic example is when {Rt} is the risk reserve process corresponding to claims arriving at Poisson rate .2 Consider the compound Poisson risk model with a general premium rule p(r). and then '0 (u) = P(V > u). say V. The arrival epochs correspond to rainfalls. We get: Corollary 3. with distribution B. Then Vo. Historically. and a general premium rule p(r) when the reserve is r.U1 = Rol.i. and since ruin can only occur at the times of claims. Nevertheless. . Some further relevant more general references are Asmussen [21] and Asmussen & Sigman [51]. we can repeat the argument and get VoN_1 > Ra2 and so on.32 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Suppose first VT > u (this situation corresponds to the solid path of {Rt} in Fig. we have r(u) > T. Historically.T l . if and only if O(u) < 1 for all u.U1 > roil . Hence RQ„ > 0 for all n < N. Hence if n satisfies VVN_n+1 = 0 (such a n exists. Corollary 3.1 with Ro = u = u2). 3.2 from Harrison & Resnick [188].3 and being i. 3.1 with Ro = u = ul). Then the time reversibility of the Poisson process ensures that {At } and {At } have the same distribution (for finitedimensional distributions. u Notes and references Some main reference on storage processes are Harrison & Resnick [187] and Brockwell. one may feel that the interaction between the different areas has been surprisingly limited even up to today. and in between rainfalls water is released at rate p(r) when Vt (the content) is r. Thus we may think of {Vt} as having compound Poisson input and being defined for all t < oo. Resnick & Tweedie [79]. if nothing else n = N). we have RQ„ < 0 so that indeed r(u) < T. Then the storage process {Vt} has a proper limit in distribution. Then similarly VVN = r0. and so on. say of water in a dam though other interpretations like the amount of goods stored are also possible. The results can be viewed as special cases of Siegmund duality. Theorem 3. Proof Let T ^ oo in (3.
Proof By (4.1.YNn+1) n=0. In particular. just verify that the r .. (c) The Lindley process {WN} generated by Zl = Y1..2 The following assertions are equivalent: (a) 0(u) = P(r(u) < oo) < 1 for all u > 0..h. . For discrete time random walks .... ZN = ..w. as long as the random walk only takes nonnegative values...min n=0.1.. there is an analogue of Theorem 3. evolves as a random walk with increments Z1i Z2. where the Yi are i . can be viewed as the reflected version of the random walk with increments Z1. .1.Y1 according to Wo = 0.. .. . with common distribution F (say). i. Let further N be fixed and let Wo. of (4.1)).d.i. has a proper limit W in distribution as n + oo. Z2 . W1.. if Wo = 0 then (Z1+•••+Zn) WN = Zl+•••+ZN.1. Theorem 4.N From this the result immediately follows... {Wn}n=0. .... is defined by assigning Wo some arbitrary value > 0 and letting Wn+1 = (Wn + Zn+1)+• (4. N min (Y1 + + Yn)... Z2. W1.e..N (4... 1} is often referred to as simple random walk or Bernoulli random walk). and is reset to 0 once the r. I. Here F is a general probability distribution on R (the special case of F being concentrated on {1..2) satisfies the same recursion as in (4. For a given i.. (b) 1/i(u) = P(•r(u) < oo) > 0 as u * oo. 0 Corollary 4. W2. ..2) (for a rigorous proof. . Most often.2).d. N min (Y1 + • • • + YNn) n=0. hits (oo..1 in terms of Lindley processes . Xo = 0..s. WN = YN .. generated by Z1.. the Lindley process Wo. R valued sequence Z1. Z2 = YN_1 i .1 Let r(u) = inf In: u + Y1 + • • • + Yn < 0}. 0)... ... n=0. ..Yl min (YN ...1.1) Thus {Wn}n=o.....4. = Xo + Y1 + • • • + Y. . Z2 = Y2...... 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.. Z2...1. WN be the Lindley process generated by Z1 = YN. Then the events {r(u) < N} and {WN > u} coincide..
a Markovian change of measure as in Theorem 2.the result is a sample path relation as is Theorem 3. equivalently.l. . In that case .s . 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.i.1 is equivalent to WN D MN = (Z1 + ..ooa. .d.s. Clearly. Proof Since (YN.3 The i.. .=o. The following result gives the necessary and sufficient condition for {Ln} to define a new random walk: . then the restrictions of Fx.. N. (Yi + • • • + Yn) > oo a. or M < oo a... doubly u infinite (n'= 0. the condition 00 F(YI+•••+ Yn<0)<00 n=1 is known to be necessary and sufficient ((APQ] p. SOME GENERAL TOOLS AND RESULTS (d) m = inf. assumption on the Z1. By Kolmogorov's 01 law. W v m and P(W > u) = P (m > u) = 0(u).. Similarly.. there is a more general version of Corollary 4. (e)Yi+•••+Yn 74 . w. ±1.. Y1) has the same distribution as (Y1. the Lindley processes in Corollary 4. e. YN). In general..) sup n=0...s.l. the Y1.s. The converse follows from general random walk theory since it is standard that lim sup (Y1 + • • + Yn) = oo when Y1 + • • • + Yn 74 oo.1 is actually not necessary . either M = oo a.o. Thus the assertion of Theorem 4.1. ZN or. Combining these facts gives easily the equivalence of (a)(d). .....1..N so that WN _P4 M = supra=0. (d) #. .2 and Theorem 4.. 0 By the law of large numbers.g.. + Z. .1.5 does not necessarily lead to a random walk: if. . Px to Fn are equivalent (have the same null sets) so that the likelihood ratio Ln exists. One then assumes Yn to be a stationary sequence. .... . (e). (Z1 + • • • + Zn) = m and P(W > u) = P(M > u) = i (u ). .2. Next consider change of measure via likelihood ratios... ..) and defines Zn = Yn.1 have the same distribution for n = 0.34 CHAPTER II.1. YN in Theorem 4... a sufficient condition for (e) is that EY is welldefined and > 0.g. ±2. 176) but appears to be rather intractable. For a random walk. Remark 4 .
Breiman [78] p. y ). and so onforn =3.. h(Yn) (4.5 corresponds to a new random walk if and only if Ln = h(Y1) . 100 ). u A particular important example is exponential change of measure (h(y) = e°y'(") where r.3) 1Pxa. and define Ln by (4. In particular.. Y) = h(Y ) a..nrc(a )} (4. the random walk property implies Ex f (Y1) = Eo f (Y1 ). we get L2 = L1 (L1 o91 ) = h(Y1)g(X1. this means E(g(x.3) holds for n = 1. In that case.3) holds. cf. implying g (x. where h (y) = g(0.4. then n n Ex [f f = Ex H fi a( YY) i=1 i_1 ( Y=) h(YY) H Ef=(Y=)h(Y=) = II J fi(Y )P( d) from which the random walk property is immediate with the asserted form of F. Y < x].5 corresponds to a new random walk with changed increment distribution F(x) = e'(a) Jr e"'F(dy) . (a) = log F(a] is the c. . Then the change of measure in Theorem 2.. e. = 1 for all n and x. Since L1 has the form g (Xo.4).g. Y) f (Y)] for all f and x.s.. The corresponding likelihood ratio is Ln = exp {a (Y1 + • • • + Yn) ..4 Let {Ln} be a multiplicative functional of a random walk with E_L. We get: Corollary 4. the changed increment distribution is F(x ) = E[h(Y).. RANDOM WALKS INDISCRETE OR CONTINUOUS TIME 35 Proposition 4.4.Y2) = h(1'i)h(I'a). Y ) f (Y)] = E[g(O. Conversely.g. For n = 2. Proof If (4. Y1). Then the change of measure in Theorem 2. for some function h with Eh(Y) = 1. (4.f.4) ({Ln} is the familiar Wald martingale .5 Consider a random walk and an a such that c(a) = log F[a] = log Ee° ' is finite.s. of F).
the claim surplus process for the compound Poisson risk model . The simplest case is 3 = JhvMM < oo. Roughly. see Chapter V.36 CHAPTER II.). . Xt =Xo+pt+oBt+Mt.t(i . 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.. The appropriate generalization of random walks to continuous time is processes with stationary independent increments (Levy processes).7) for all e > 0. or imbedded into continuous time processes . this description needs some amendments. they arise as models for the reserve or claim surplus at a discrete sequence of instants. v2 = 0 and v = 3B). the tradition in the area is to use continuous time models. (4. In discrete time. An equivalent characterisation is {Xt} being Markov with state space R and E [f (Xt+e . {Xt} is a random walk. with premium rate p. the pure jump process is given by its Levy measure v(dx). corresponds to a process with stationary independent increments and u = p. say the beginning of each month or year .5) Note that the structure of such a process admits a complete description. . we are . but we omit the details ). (4.Ft] = Eof (X. however. a Brownian component {Bt} (scaled by a variance constant) and a pure jump process {Mt}. The traditional formal definition is that {Xt} is Rvalued with the increments Xt(1)_t(o). i.6) More precisely.e. Xt (n)t(n1) being independent whenever t(O) < t(1 ) < . e f x:IxJ>e} v(dx) < oo (4. < t(n) and with Xt( i)_t(i_l) having distribution depending only on t(i) .1). the interpretation is that the rate of a jump of size x is v(dx) (if f of Ixlv (dx) = oo. which corresponds to the compound Poisson case: here jumps of {Mt} occur at rate 0 and have distribution B = v/0 (in particular . a positive measure on R with the properties e J x2v(dx) < oo. {Xt} can be written as the independent sum of a pure drift {pt}. SOME GENERAL TOOLS AND RESULTS Discrete time random walks have classical applications in queueing theory via the Lindley process representation of the waiting time . In continuous time. In risk theory.. given by a the increment distribution F(x) = P(Xn+l . say by recording the reserve or claim surplus just before or just after claims (see Chapter V for some fundamental examples). Xt(2)_t(l)... However.Xn < x).Xt)I.
[The condition for V < oo a.4. Then the storage process {Vt} has constant release rate 1. Corollary 4.Q and distribution B of the service times of the arriving customers. then Ee'(xtxo) = Eoeaxt = etx(a). WT = XT .7 If {Xt} has stationary independent increments as in (4. jxJ v(dx) < oo. 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. 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. Proposition 4.2).1. O(u. and the reflected version is then defined by means of the abstract reflection operator as in (4. defined as a system with a single server working at a unit rate.v.6 In the compound Poisson risk model with constant premium rate p(r) .e.3 and decreases linearly at rate 1 in between jumps. Chapter III. i. A different interpretation is as the workload or virtual waiting time process in an M/G/1 queue.1)v(dx) (4. where c(a) = ap + a2a2/2 + f 00 provided the Levy measure of the jump part {Mt} satisfies f". T) = P(VT > u).min Xt (4.8) O<t<T (assuming Wo = Xo = 0 for simplicity).] Processes with a more complicated path structure like Brownian motion or jump processes with unbounded Levy measure are not covered by Section 3. cf. 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. V E [0. having Poisson arrivals with rate .10) . 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).s. (ex . and b(u) = P(V > u).6). oo]. Now consider reflected versions of processes with stationary independent increments. Furthermore. is easily seen to be f3pB < 1. 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. VT + V for some r. has upwards jumps governed by B at the epochs of a Poisson process with rate .
e. Xt +B . This is of course no coincidence since the distribution of Xl . X8) = Eof (X8)L8 = Eof (X8)• For the second. use the representation as limit of compound Poisson processes.Xt)g(s. we show in the compound Poisson case ( IlvIl < oo) in Proposition III. t.Xt)I'Ftl = E [f(Xt+B . Chung [86]).10) is the LevyKhinchine representation of the c. let e" (a ) = Eoeaxl. Q2 = v2.11) (eax . Eea(µt + QBt) = et{aµ +a2OZ/2}. Xt Xo) with E2Lt = 1 for all x.38 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Proof By standard formulas for the normal distribution. if Lt = e9(xt .xo)tk ( e).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 . In particular. 5 has stationary independent increments as well. then the changed parameters in the representation (4.5) and get E [f(Xt+B . (4. of an infinitely divisible distribution (see. v(dx) = e9xv (dx). . u Note that (4. 8 Assume that {Xt} has stationary independent increments and that {Lt} is a nonnegative multiplicative functional of the form Lt = g(t.. Theorem 4 .1 ) v(dx) .g.1)eexv(dx).g. Then the Markov process given by Theorem 2.1. Proof For the first statement . we use the characterization (4.Xt)I Ftl = Eof (X8)g(s.f. 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) . By explicit calculation .Xt)L8 o 0tIFt] = E [f (Xt+s .1 that E eaMt = exp fmoo In the general case .6) are µ = µ + Oo2 .Xo is necessarily infinitely divisible when {Xt} has stationary independent increments.
. (5.l3 and claim u size distribution B. abbreviated as MAP in this section2.11 For an example of a likelihood ratio not covered by Theorem 4.4)..3B[B]. Then we can write v(dx) _ /3eOxB(dx) = / (dx). B(dx) = B[9] B(dx). 0.9 If X0 = 0. U2. the corresponding claim sizes .. b = a = 0) the changed process is the claim surplus process of another compound Poisson risk process with Poisson rate . Recalling that U1. where .5.3 =.. As for processes with stationary independent increments . Example 4 . Ei. and let the Px refer to the claim surplus process of another compound Poisson risk process with Poisson rate. dB/dB = b/b when B. it is then easily seen that Lt = H dB(Ui) i:o. . b with b(x) > 0 for all x such that b(x) > 0). the structure of MAP's is completely understood when E is finite: 2and only there . Ei instead of P2. 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.. MARKOV ADDITIVE PROCESSES 39 Remark 4.St)g(Jt+s)I.o[f (S8)g(J8)].3 and claim size distribution B # B. is defined as a bivariate Markov process {Xt} = {(Jt.. 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 .(3 = . are the arrival times and U1.1) For shorthand . we write Pi.3 and claim size distribution B. u 5 Markov additive processes A Markov additive processes.8.0 in the following. a = 0. MAP stands for the Markovian arrival process discussed below. Thus (since µ = p = 1. B have densities b.0. corresponding to p = 1.10 Let Xt be the claim surplus process of a compound Poisson risk process with Poisson rate . <t whenever the RadonNikodym derivative dB/dB exists (e. Example 4 . then the martingale {eex(t)tk(e)} is the continuous u time analogue of the Wald martingale (4.(3B(dx).2. one reason is that in parts of the applied probability literature.Ft] = Ejt.g. v(dx) _ .
J1=j)= Fij (dx) Pij In simulation language. v.6) depending on i. Jn = j. which we omit and refer to Neveu [272] or cinlar [87].. Fn[a] = F[a]n where P[a] = P . the distribution of which has some distribution Bij. a MAP is specified by the measurevalued matrix (kernel) F(dx) whose ijth element is the defective probability distribution Fij(dx) = Pi.g..) If E is infinite a MAP may be much more complicated. this means that the MAP can be simulated by first simulating the Markov chain {J„} and next the Y1. oo). {Jt} is specified by its intensity matrix A = (Aij)i. consider the matrix Ft [a] with ijth element least Ei . the converse requires a proof. by generating Yn according to Hij when J„_1 = i. . Then a Markov additive process can be defined by letting t St = lim 1 I(IJB1 < e)ds E1o 2d o be the local time at 0 up to time t.Sr_1. a jump of {Jt} from i to j # i has probability qij of giving rise to a jump of {St} at the same time.o(Ji = j.jEE (here pij = Pi(J1 = j)) and the probability measures Hij(dx)=P(Y1 EdxlJo=i. Y1 E dx) where Y„ = S„ . let {Jt} be standard Brownian motion on the line. {St} evolves like a process with stationary independent increments and the parameters pi.. As a generalization of the m. Y2. Proposition 5. a MAP is the same as a semiMarkov or Markov renewal process.[a) = (Ei[easl..9 EE = (iii&ij[a])i j EE . with the Y„ being interpreted as interarrival times. (That a process with this description is a MAP is obvious..f. If all Fij are concentrated on (0. In continuous time (assuming Dpaths). t+s) where Jt .i.it = A. SOME GENERAL TOOLS AND RESULTS In discrete time. In addition.jEE• On an interval [t. An alternative description is in terms of the transition matrix P = (piA. vi(dx) in (4. As an example. 1 J1 ='^])iJEE = (Fij[a])i .40 CHAPTER H.1 For a MAP in discrete time and with E finite.
which in conjunction with Fo[a] = I implies Ft[a] = etK[a) according to the standard solution formula for systems of linear differential equations. 013 . Proof Let {Stt) } be a process with stationary independent increments and pa rameters pi . where K[a] = A+ (r. MARKOV ADDITIVE PROCESSES Proof Conditioning upon (Jn.qkj + k?^j qkj Bkj [a] } = Ei [east. 00 r(i) (a) = api + a2ot /2 + f (e° . Jt = k] { xk kEE j la] . pi. J1 = A which in matrix formulation is the same as Fn+1 [a] = Fn[a]F[a].5. \ diag Ft[a] = Ft[a]K. Jt = j] Ejesh'^ + E Ak j hEi [ease .ijgij(Bij[a] .(')(a)) diag + (). Then the matrix Pt[a] with ijth element Ei [east. u In the following. up to o( h) terms. Jn+1 = A] = 41 Ei[ e 5„. assume that the Markov chain/process {Jt} is ergodic. Jt = k] { 1 . Bij (i. vi(dx) (i E E). u Proposition 5.4c). we infer that in the discrete time case the . In matrix formulation . a= . kEE Jn = k]Ek[e"Y" . Then. this means that F't+h [a] = Ft[a] II+h(rc(i)(a)) +hA+h(Aijgij(Bij[a]1)) I.1 )v(dx). Sn ) yields Ei[easn+ '. By PerronFrobenius theory (see A. Jt = j] is given by etK[a]. qij.1)) . j E E) and So = 0.2 Let E be finite and consider a continuous time Markov additive process with parameters A.1) } (recall that qjj = 0). aSt h = (1 + Ajjh) Ei [east . vi(dx). Jt = j] (1 + htc (j) (a)) j + Ak j qk j (Bk +h E Ei [east .
tK(a)h(a) J jj it L o is a martingale.r. Corollary 5.4 Eie"sth(a) = h=a)et?("). and we shall take V(a)h(a) = 1.h(a)vva)etw(a). u Let k(a) denote the derivative of h() w.5 EiSt = tK'(0) + ki . a.2) where 7r = v(°) is the stationary distribution. Yrh(a ) = 1. The function ic(a) plays in many respects the same role as the cumulant g. h(") may be chosen with strictly positive components.4c). 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. and write k = k(°).etx It then follows that E feast+^(t+v)K(a)h(a) I ^tl l . Proof By PerronFrobenius theory (see A.7. h(") are only given up to a constants.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). (5.4. Proposition 5. In particular. Jt = j] . as will be seen from the following results. cf.42 CHAPTER II.Eikjt = ttc'(0) + ki . Since v("). Furthermore. just note that [a]h(a) = eietK (a)h(a) = etK(a)h(a). The corresponding left and right eigenvectors v("). Jeast.3 Ei [east. of a random walk. Corollary 5. Corollary 5. Proof For the first assertion.f.t. We also get an analogue of the Wald martingale for random walks: Proposition 5.e=e°tk. and appropriate generalizations of the Wald martingale (and the associated change of measure) can be defined in terms of . Eie"sth^a) = e'Pt[a]h( a) = e. Then h(°) = e. we are free to impose two normalizations.c(a) (and h(")). . its derivatives are 'asymptotic cumulants'. cf.
t im v^"St = '(0) Proof The first assertion is immediate by dividing by tin Corollary 5. Ee"st typically grows asymptotically exponential with a rate ic(a) independent of the initial condition (i. (5.5. there is typically a function h = h(") on E and a ^c(a) such that Ey a"st t" (") * h(x). one obtains a generalization of Wald's identity EST = Er • ES.6 For any stopping time T with finite mean. ] = t2tc (0)2 + 2tK'(0)vk + ttc"(0) + O(1) ..a) + ttc (a)2hia ) Multiplying by v=.. In the same way.3) Let a = 0 and recall that h(°) = e so that 0=°) = h(o) = 1.") }) . Remark 5 . for a random walk: Corollary 5. 8 Also for E being infinite (possibly uncountable ).Eikjr . tam E tSt a (0). 43 Ei [Steast h(a) + east k^a)1 = et"(a) (kia) + tic (a)hia)) .2ttc (0)Evkjt + 0(i). (5.7 No matter the initial distribution v of Jo. . Since it is easily seen by an asymptotic independence argument that E„ [Stkjt] u = trc'(0) E„kjt + O(1).5 yields + W (a)k.+ k. MARKOV ADDITIVE PROCESSES Proof By differentiation in Proposition 5.g. [E.5. the distribution of Jo). we differentiate (5.3) to get Ej [St a " st h i(a ) + 2Ste"st k(a) + e"st k^a) J etI(a) (kia )' + ttc (a)ki") + t {ic"(a)h. E=ST = tc'(0)E7. summing and letting a = 0 yields E„ [St + 2Stkj. subtraction yields Vary St = tic"(0) + O(1).4) . Corollary 5.4. More precisely. Squaring in Corollary 5. the existence of exponential moments is assumed ) but can be made rigorous by passing to characteristic functions.St]2 = t2/c'(0 ) 2 + 2ttc (0)vk . u The argument is slightly heuristic (e.e. For the second . t a oo.
6) We shall not exploit this approach systematically. Remark 5. St)} be a MAP and let 0 be such that h(Jt) OStt. and the family {f LEE given by Theorem 2.(9) {Lt}t>o = . 0 Proposition 5. where {Jt} is deterministic period motion on E = [0. in particular that f is bounded. however. inconvenient due to the unboundedness of ea8 so we shall not aim for complete rigour but interpret C in a broader sense. some extra conditions are imposed. (5. St) } as follows. however.e.f (x) tyo t provided the limit exists.6..s.4 that { h(Jt) easttK(a) L o (5. this leads to h(i) + tcha( i. 0) = h(i )( 1 + ttc(a)). For t small .9 The condition that (5. Usually. 0) = n(a) h(i). we take the martingale property as our basic condition below (though this is automatic in the finite case). s) = ea8h(i).for the present purposes. First.5. this is. In view of this discussion . see.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. We then want to determine h and x(a) such that Ejeasth (Jt) = etK(a)h(i).6.. An example beyond the finite case occurs for periodic risk processes in VI.10 Let {(Jt.44 CHAPTER IL SOME GENERAL TOOLS AND RESULTS for all x E E. 1) (i. Then {Lt } is a multiplicative functional. u forsEE). let ha(i. h(Jo) Lo is a Px martingale for each x E E.3b and Remark VI. Jt = (s+t) mod 1 P8a. From (5. G is defined as Gf (x) = lim Exf (Xt) . gha(i. Given a function h on E.5) is a martingale . xEE .5 defines a new MAP.5) is a martingale can be expressed via the infinitesimal generator G of {Xt } = { (Jt. V.
1) . then also vi (dx) is compound Poisson with e Ox ^i = /3iBi[0].(0)j. vi (dx) = f3 Bi(dx) with . Then the MAP in Proposition 5. .tc(0)e = 0 .10 is given by P = eK(e) Oh e) F[e]Oh('). if vi(dx) is compound Poisson. MARKOV ADDITIVE PROCESSES Proof That { Lt} is a multiplicative functional follows from L8 ogt = h(Jt+s) es(St+ .. u Theorem 5. in the discrete time case.12 The expression for A means h(e) Aij = hie) Aij [1 + gij(Bij[0] i 0 j. In the infinite case . Bi. this gives a direct verification that A is an intensity matrix: the offdiagonal elements are nonnegative because Aij > 0. (5.Qi < oo and Bi a probability measure.St)sl(e) h(Jt) 45 The proof that we have a MAP is contained in the proof of Theorem 5. Bi [0] Remark 5.7(dx) Bij [0] Bij(dx) in the continuous time case . one can directly verify that (5.1) holds for the P.7) In particular. Here Oh(e) is the diagonal matrix with the h=e) on the diagonal.c(0)e = tc(0)e . qij = r.11 below in the finite case. That 0 < qij < 1 follows from the inequality qb <1.5. Ai = µi + 0Q. and by A = Oh(°) K [0]Oh(e ) vi(dx) = e"xvi (dx). In particular.ic(0)e = ic(0)Oh e) h(e) .. That the rows sum to 1 follows from Ae = Oh(e) K[O]h(B) . 0<q<1.11 Consider the irreducible case with E finite. 0<b<oo. Bi(dx) = Bi(dx).1) eft ea' f ij (dx) = Hij (dx) Hij [0] . 1 + q(b . ^i = of qij Bij [0] 1 + qij ( Bij [0] . 0 < qij < 1 and Bij [0] > 0. We omit the details.
. Similarly. v= ... Further Fib (dx ) = P=(YI E dx.tc(') (8)/ d)ag h 7 Aiiii (Bii[a + 0] .t. (dx) of a process with stationary independent increments follows from Theorem 4. SOME GENERAL TOOLS AND RESULTS Proof of Theorem 5. a = 0 in (5. it follows that indeed the normalizing constant is H1 [0]. are probability measures . Jt = j] = hie) . Here the stated formula for P follows immediately by letting t = 1.8) h(. v. since Hij.K [O])Oh(e) (0) l + ( A + (tc(') (a + 0) .8). . H1. Jt = A. In matrix notation .tc (') (0) corresponds to the stated parameters µ. Jl = j] :(Yi E dx.11.r. (dx). in continuous time ( 5. Now we can write K[a] =A+A ) ( K[a + 0] .e) Consider first the discrete time case . Yi E dx. this means that Ft[a] = etw ( e)Ohc) Ft[a + 9]oh (e) (5.Bay [0]) That k(') (a + 0) . Hence the same is true for H=j and H. Jl = j) = Ei[Lt. F:j with a density proportional to eei .8. .tc(0)I.46 CHAPTER II.tc(0)' )Ah() = Oh(o) K[a + 0]Oh() . Letting a = 0 yields the stated expression for A. Ji = j) h(e) eeyK(B)p h(8) h(e) eexK ( h=e) e)Fi. This shows that F. is absolutely continuous w. this implies k[a] = A 1 ) (K[a + 0] .8) yields et'[a] = Ohie )et (K[a +e]K(e)I)Oh(°) By a general formula (A.13) for matrixexponentials . First note that the ijth element of Ft[a] is etK(e)Ej [e(a+B)st E:[east Jt = j] = Ej[Lteas' .
i. For the Wald identity in Corollary 5. Much of the pioneering was done in the sixties in papers like Keilson & Wishart [224]. 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]. which.)Ajjgij(Bij[a+0] . [225]. however. [262] in discrete time. hardly a single comprehensive treatment. 7+ < oo). and is typically defective. Though the literature on MAP's is extensive.Bij[0]) = hjel)ijgijBij[0](Bij[a] .7). Write r+ = T(0) and define the associated ladder height ST+ and ladder height distribution by G+(x) = 11 (S.+ < x..1) = Aij4ij(Bij[a] . however.6. Note that G+ is concentrated on (0. [261]. h. has no mass on (oo.. see also Fuh & Lai [149] and Moustakides [264]. [226] and Miller [260]. an extensive bibliography on aspects of the theory can be found in Asmussen [16]. 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). the literature on the continuous time case tends more to deal with special cases. THE LADDER HEIGHT DISTRIBUTION 47 Finally note that by (5. .6.(u) = inf {t > 0 : St > u} to ruin in the particular case u = 0 .e. Notes and references The earliest paper on treatment of MAP's in the present spirit we know of is Nagaev [265]. oo). there is. 6 The ladder height distribution We consider the claim surplus process {St } of a general risk process and the time 7. Conditions for analogues of Corollary 5.1). < x) = 11 (S. is slightly less general than the present setting.3 for an infinite E are given by Ney & Nummelin [266].
a fact which turns out to be extremely useful.ST+(1) and so on.5 below.00 ). Theorem 6 . by approximation with step functions . On Fig. Here bo(x) _ B(x)/µB.B(x) denotes the tail of B. the sum of all the ladder steps (if rl > 0.T+ > t)dt = E f 0T+I(St E A) dt. o 00 (6. The main result of this section is Theorem 6. In any case.i.1. 0 f T+ (6. 1 For the compound Poisson model with p = 01LB < 1.(3B(x ) = pbo(x) on (0.1. Also.48 CHAPTER K. at present we concentrate on the first ladder height. = ST+(1) Figure 6. The first ladder step is precisely ST+.1) The interpretation of R+(A ) is as the expected time {St} spends in the set A before T+. we shall first consider the compound Poisson model in the notation of Example 1. Recall that B(x) = 1 . SOME GENERAL TOOLS AND RESULTS M ST+(2) Sr. the ladder heights are i.. For the proof of Theorem 6. oo). the second ladder point is ST+(2) where r+(2) is the time of the next relative maximum after r+(1) = r+. 6. it follows that for g > 0 measurable. the dependence structure seems too complicated to be useful). In other cases like the Markovian environment model. has no mass on ( 0. 6.d. there are only finitely many). define the prer+occupation measure R+ by R+(A) = E f o "o I(St E A.2. which gives an explicit expression for G+ in a very general setting. In simple cases like the compound Poisson model. the second ladder height (step) is ST+(2) . R+ is concentrated on (oo. i. and the maximum M is the total height of the ladder.1 The term ladder height is motivated from the shape of the process {Mt} of relative maxima.2) . G+ is given by the defective density g + (x) =.1.e. g(y)R+(dy) = E f g(St)dt.e. they have a semiMarkov structure (but in complete generality. see Fig. Thus. 0]. i. To illustrate the ideas.. where basically only stationarity is assumed.
0 < t < T) P(STEA.6. P(STEA.St<0.ST<St. ST < ST_t.O<t<T) = P(STEA. St S* t a Figure 6.ST<St.O<t<T). has the same distribution as {St}o<t<T. 0 < t < T. see Fig. THE LADDER HEIGHT DISTRIBUTION Lemma 6 .T+>T) = P(STEA. 6. That is. since the distribution of the Poisson process is invariant under time reversion.2 R+ is the restriction of Lebesgue measure to (00. 0].ST_t. .2(a): T+ > t Figure 6. {St }o<t<T is constructed from {St}o<t<T by timereversion and hence.2. 49 Proof Let T be fixed and define St = ST .0<t<T) = F(ST E A.2(b): r+ < t Thus.
it follows that R+ (A) is the expected time when ST is in A and at a minimum at the same time . E A} precisely when r+ > t. for A C (0. G+(A) = Q f 0 B(A .St _)I(r+ > t). U + St_ E A.. But since St 4 oo a. 6..T+ > t] dt 0 T+ _ /3E f g( St) dt = 0 f g(y) R+(dy) 0 00 where g(y) = B(A . oo).3 Lemma 6 . . SOME GENERAL TOOLS AND RESULTS Integrating w. The probability of this given { Su}u<t is B(A .y) (here we used the fact that the probability of a jump at u t is zero in the second step.2) in the last). s. Fig.3 where the bold lines correspond to minimal values. That is. this is just the Lebesgue measure of A.St _). T+ > t] 0 _ /3 f E[B( A .3 G+ is the restriction of /3R+*B to (0.y)R+(dy) 00 Proof A jump of {St} at time t and of size U contributes to the event IS. Figure 6.St).50 CHAPTER II. we get G+ (A) = f 00 /3 dt E[B(A . cf.r. oo). and since the jump rate is /3.t dT. and (6.
.T+ < oo).e.s > 0). 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.z)B(dz) _ f I(x < z)B(dz) _ f (x). Lemma 6. The points in the plane (marked by x on Fig. the first component representing time (the arrival time o. Uk) for those k for which ak . THE LADDER HEIGHT DISTRIBUTION 51 Proof of Theorem 6. 0 Generalizing the setup. The marked point process . we define the arrival rate as E# { k : ak E [0 . Uk) (k = 1. 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.1.1 With r+(y) = I(y < 0) the density of R+. i..) where ak = Ti + • • • + Tk . In the stationary case..e.M o 08 shifted by s is defined the obvious way. this does not depend on h). . We call M * stationary if M* o B8 has the same distribution as M* for all s > 0. oo). we consider the claim surplus process {St }t>o of a risk reserve process in a very general setup. 6 . obviously.S8 )t> o = {St }t>o for all s > 0.3 yields g+ (x) = . this is equivalent to the risk process {St*} being stationary in the sense of (6. . {St+8 .6. cf..s. assuming basically stationarity in time and space. 6 . 2.* ) and the second the mark (the claim size Uk ).Q f r+(x .4).. The traditional representation of the input sequence {(TT. h]} /h (by stationarity.4) are (ak. oo) x (0. as a point process on [0. U k)} k=1 a is as a marked point process M *.:T1 +•••+Tk <t}. Nt St =>Uk k=1 t where Nt = max{k = O. Fig. 4 (the points in the plane are (ak . i.
0.s.. o. we define its Palm version M as a marked point process having the conditional distribution of M* given an arrival at time 0 . See. As above .2. h. Uk) k=1.e.. Oh becomes the approximate probability F(ri < h) of an arrival in [0. Assume {Jt} irreducible so that a stationary distribution 7r = (1i)iGE exists.5) does not depend on h. Section 5) which has pure jump structure corresponding to pi = a = 0.g. e.4 Consider a finite Markov additive process (cf. most often one takes h = 1). letting h J.QiBi(dx). SOME GENERAL TOOLS AND RESULTS M* U. . the r. We represent M by the sequence (Tk.52 CHAPTER II. Note also that (again by stationarity) the Palm distribution also represents the conditional distribution of M* o Ot given an arrival at time t. V(M* o eak ). h] and the sum approximately ^o(M*)I(ul < h). and let T = T2 denote the first proper interarrival time . i. Sigman [348] for these and further aspects of Palm theory. where TI = 0. vi(dx) = . Example 6 . The two fundamental formulas connecting M* and M are Eco(M) = aE E. where T is the first arrival time > 0 of M and h > 0 an arbitrary constant (in the literature.. i 1 U2 Us 1_ 0 or Q2 $ U3 *1 L 0 7 X I 11 1 Figure 6.. This more or less gives a proof that indeed (6. h] Eco(M*) = 1 E f co(M o Bt)dt. k: vk E [0. of (6. = 0 .5) represents the conditional distribution of M* given vi = 0.4 Given a stationary marked point process M*.
e. Jo) w. we get a marked point process generated by Poisson arrivals at rate /3i and mark distribution Bi when Jt = i.oo a. and by some additional arrivals which occur w. First choose (Jo_.O fo "o F(x)dx = .p. v. 5 Consider a general stationary claim surplus process {St }t>o./. we note: Corollary 6.p. qij when {Jt} jumps from i to j and have mark distribution Bij.OEU0.6 Under the assumptions of Theorem 6. THE LADDER HEIGHT DISTRIBUTION 53 Interpreting jump times as arrival times and jump sizes as marks.6. . and that p = 0EU0 < 1. Before giving the proof. having the Palm distribution of the claim size and F (x) = F(Uo < x) its distribution .5.O for i # j. an arrival for M* occurs before time t + dt w. Then the ladder height distribution G+ is given by the (defective) density g+(x) = . the distribution of Ul) is the mixture B = E aii Bi + aij Bij J = j#i !i J.OF(x). It follows that we can describe the Palm version M as follows .s. Assume that St * . dt A + E Aijgij j#i Thus the arrival rate for M* is 1] it A + E Aijgij iEE i#i Given that an arrival occurs at time t .6iBi + Aijgij Bij j#i iEE iEE 0 Theorem 6 . A stationary marked point process M* is obtained by assigning Jo distribution Tr. If Jt_ = i. Jt = j is iri(3i /.p. the probability aij of Jt . Note in particular that the Palm distribution of the mark size (i.*(0) with initial reserve u = 0 is p = /3EU0. j) and let the initial mark Ul have distribution Bi when i = j and Bij otherwise. After that. the ruin probability .= i. This follows by noting that iP*(0) = IIG+JI = J0 "o g+(x)dx = ..O for i = j and iriAijgij/. let the arrivals and their marks be generated by {Jt} starting from Jo = j. let U0 be a r. aij for (i.
.).St<Su. 0<u<t) = P(St EA. 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. are point processes on (oo . 2. It follows that for A C (0. .0<u<t) = P(St EA. A standard argument for stationary processes ([78] p. that M* and M have doubly infinite time (i.Su<0.s. Now conditionally upon At .(left limit) when 0 < it < t and is illustrated on Fig . 6..5).St <. has a very simple interpretation as the average amount of claims received per unit time . SOME GENERAL TOOLS AND RESULTS V` (0) = E E Uk k: ak E [0.. the mark at time Qk is denoted by Uk.Q_k and has size U_ k. CHAPTER H. { Su}0<u<t is distributed as a process {Su} . Proof of Theorem 6. The last property is referred to as insensitivity in the applied probability literature. 0). h.St*_ u.5.0<u<t) = P(StEA.e.Su< 0.A.. .$St_ u.o<u<t where a claim arrives at time t and has size Uo.. oo) p(t) = P(St EA.l. oo)).0<u<tIAt) = P(St EA. Let p(t) be the conditional probability that ST+ E A. . We then represent M by the mark (claim size ) Uo of the arrival at time 0.o.g. oo ) and the arrival times 0 > 0_1 > a_2 > .. (k = St}t>o 1.Mt). and the kth preceding claim arrives at time t . 105) shows that one can assume w.54 By (6. in (0. moves down linearly at a unit rate in between jumps and starts from S0 = U. Then clearly * G+ (A) = P(ST+ E A) = Consider a process { f p(t)f3dt..0<u<tIAt) = P(St EA. the arrival times 0 < 0'1 < Q2 < . oo) x (0 .Su_ <0. The sample path relation between { Su } and { Su } amounts to S„ = St .1] here the r .5.. which makes an upwards jump at time . in (oo. T+ = t given the event At that an arrival at t occurs .o.
.5 where it = { St < Su. cf. Uo].6. 2 therefore immediately shows that L(dy) is Lebesgue measure on (oo. and since by assumption St * oo a. and we let L(dy) be the random measure L(A) = fo°° I(St E A. G' (A) = 3 f P(St E A.5. Thus. 6. 0 < u < t } is the event that { Su } has a relative minimum at t . Since So = U0.5 where the boxes on the time axis correspond to time intervals where {St } is at a minimum belonging to A and split A into pieces corresponding to segments where {Su} is at a relative minimum. In Fig. A sample path inspection just as in the proof of Lemma 6 .s. NIt)dt . the left endpoint of the support is oo. Fig. Mt)dt = i3EL(A) o"o . the support of L has right endpoint U0. t a oo. 6. time instants corresponding to such minimal values have been marked with bold lines in the path of { St}. THE LADDER HEIGHT DISTRIBUTION 55 { A Su}0<u<t U0 U0 \t tt u>0 N U_1 Figure 6.
SOME GENERAL TOOLS AND RESULTS = OE f 0 I(Uo>y)I (yEA)dy = Q f IP (Uo>y)dy A 0o a fA P(y) dy• 0 Notes and references Theorem 6.56 CHAPTER II.1). [147]. [263] (a special case of the result appear in Proposition VI. .6 is Bjork & Grandell [67]. A further relevant reference related to Corollary 6.5 is due to Schmidt & coworkers [48].2.
Thus . being of the form Rt = Rt+Bt + Jt where {Rt } is a compound 57 . . • the premium rate is p = 1. say.Chapter III The compound Poisson model We consider throughout this chapter a risk reserve process {Rt } t>o in the terminology and notation of Chapter I. A common view of the literature is to consider such processes as perturbed compound Poisson risk processes . exact matrixexponential solutions under the assumption that B is phasetype (see further VIII. see Chapter IV. St = uRt = EUi t.4 below . and assume that • { Nt}t>o is a Poisson process with rate j3. with common distribution B. i.6) and simulation methods ( Chapter X). For finite horizon ruin probabilities . {Rt} and the associated claims surplus process {St} are given by Nt Nt Rt = u+t EUi. and independent of {Nt}.. • the claim sizes U1. i=1 i=1 An important omission of the discussion in this chapter is the numerical evaluation of the ruin probability.. 4. 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 . are i. Some possibilities are numerical Laplace transform inversion via Corollary 3. Panjer's recursion ( Corollary XI. 3). U2.d.e.
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).f. of the claim surplus St .s.1) = t(p .1). and that B(k)[0] = Pak).+Uk)P(Nt = k) k=O e8t k=O B[s]k . Write pB^) = EUn' YB = Pali = EU. we get Ee8st = 00 e8t c` Ee8 (U1+.t = t(p . (c) Ee8St = et" (8) where c(s) = f3(B[s] . where K(k) (0) is the kth derivative of is at 0.6pBa).u . Dufresne & Gerber [126].58 CHAPTER III. For (c). 1 Introduction For later reference.1 is the expected claim surplus per unit time.. Schmidli [319].1). and Schlegel [316].. for (d) just note that the kth cumulant of St is tic(k) (0). e .1 (a) ESt = t(13µ$ . cumulants . (b) Var St = t. we shall start by giving the basic formulas for moments. THE COMPOUND POISSON MODEL Poisson risk process. [324].t = E E [ U k k=1 k=1 Nt . P = PAB = 1/(1 + rl) Proposition 1.1) . {Bt} a Brownian motion and {Rt} a pure jump process.t = fltpB . See e. We do not spell out in detail such generalizations.'s etc. Furrer [150]. say stable Levy motion. A more formal proof goes as follows: Nt r Nt ESt = E > U k .t = E[Ntµs] . Proof It was noted in Chapter I that p . and this immediately yields (a). (d) The kth cumulant of St is tf3p(k) for k > 2.g. m.Rt.)3t (fit' k t} = etk(8) exp {st '3t + B[s]f Finally.g. 0 .
i.2 (DRIFT AND OSCILLATION) St/ta3'p1 ast >oo. meaning that the increments are stationary and independent. The right hand inequality in (1. lim supt. then St 4 co.4. Here is one immediate application: Proposition 1. we need the following lemma: Lemma 1. 2. so that {Sok } is a random walk with mean EUET = EU.V. The connections to random walks are in fact fundamental. v > 0. if t = nh + v with 0 < v < h. (d) If 17 = 0.Tk.Sok_l = Uk .3) is proved similarly.. 1...1. then St> SnhV>Snhh. S„+V > S„ . Obviously. we have Sok . INTRODUCTION 59 The linear way the index t enters in the formulas in Proposition 1.d. the Uk . Proof We first note that for u. u + v]. Indeed.1 = . which is often used in the literature for obtaining information about {St} and the ruin probabilities. we get a discrete time random walk imbedded in the claim surplus process {St}.h < St < S(n+1)h + h.S„ attains its minimal value when there are no arrivals in (u. In this way. and there are at least two ways to exploit this: Recalling that ok is the time of the kth claim. The point of view in the present chapter is. where Tk is the time between the kth and the (k . then St 00. however. then lien inft. rather to view {St} directly as a random walk in continuous time.3EU01 = 1µs where rt is the safety loading. and the value is then precisely v. We return to this approach in Chapter V. II. (c) If 77 > 0. . (a) No matter the value of 77. St = oo.3 If nh < t < (n + 1)h.. Sn+0 . cf. St = oo.Tk are i.1)th claim. obviously 0(u) = F(maxk Sok > u). In particular. For the proof. For example. then Snh .1 is the same as if {St} was a random walk indexed by t = 0. (b) If 77 < 0.
2: Proposition 1.s.1.1). (c) are immediate consequences of (a). hence by induction i. Part (d) follows by a (slightly more intricate) general random walk result ([APQ].6 Often it is of interest to consider size fluctuations. Corollary 1.2. 169) stating that lim infra.. is a discrete time random walk. 1 since St 4 oo) and repeating the argument. lim supn_.. . Considering the next downcrossing (which occurs w. Proof The case of 17 < 0 is immediate since then M = oo by Proposition 1. Snh u = 00 (the lemma is not needed for (d)). it suffices to prove 4'(0) = F(M > 0) < 1. at least once. u 307).o. If rl > 0..1(b)) that the assertion holds as t 4 oo through values of the form t = 0. and hence it folz lows from standard central limit theory and the expression Var(St) = tf3pB (Proposition 1.60 CHAPTER III.. {Snh}n=o._. is a discretetime random walk for any h > 0.1. Snh/n a4' ESh = h(p ..1.t . and < 1 for all u when 77 > 0.2.. This contradicts u St400.3. if P(M > 0) = 1....p. 2h.3. Notes and references All material of the present section is standard.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). Thus using Lemma 1. However. {Snh}n=o. then {St} upcrosses level 0 a. and (b). The general case now follows either by another easy application of Lemma 1. h A similar argument for lim sup proves (a). 0 Snh = 00. THE COMPOUND POISSON MODEL Proof of Proposition 1.4 The ruin probability 0(u) is 1 for all u when 77 < 0. p. h. Remark 1 . There is also a central limit version of Proposition 1. Assuming that each risk generates claims at Poisson intensity /3 and pays premium 1 per unit time. 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 . 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. or by a general result on discrete skeletons ([APQ] p. it is seen that upcrossing occurs at least twice. where the size of the portfolio at time t is M(t). For any fixed h. and hence by the strong law of large numbers.5 The limiting distribution of St ..
(2. Note that the distribution B0 with density bo is familiar from renewal theory as the limiting stationary distribution of the overshoot (forwards recurrence time ). Combined with i/i(u) = P ( M > u). the ladder heights are i. THE POLLACZECKKHINCHINE FORMULA 61 2 The PollaczeckKhinchine formula The time to ruin r(u) is defined as in Chapter I as inf It > 0: St > u}. p < 1.1. The following results generalizes the fact that the conditional distribution of the deficit ST(o) just after ruin given that ruin occurs (i.just before ruin is again B0. This follows simply by noting that the process repeats itself after reaching a relative maximum.6. which we henceforth refer to as the PollaczeckKhinchine formula.IIG+II) (the parenthesis gives the probability that there are no further ladder steps after the nth ).IIG +II)EG+ . nevertheless. Fig. It is crucial to note that for the compound Poisson model.1) is not entirely satisfying because of the infinite sum of convolution powers. Summing over n. As a vehicle for computing tIi(u). Note that this . cf. d. We assume throughout rl > 0 or..e. 1 The distribution of M is (1.1 provides a representation formula for 0(u). equivalently. Theorem 2. [APQ] Ch. we may view the ladder heights as a terminating renewal process and M becomes then the lifetime. 1e..34 or A. Here bo(x) _ Proof The probability that M is attained in precisely n ladder steps and does not exceed x is G+ (x)(1 . IV. we can rewrite the PollaczeckKhinchine formula as 00 (u) = P (M > u) = (1 .1. oo ). n=0 (2.1) representing the distribution of M as a geometric compound. The decomposition of M as a sum of ladder heights now yields: 00 Theorem 2 .6. and we shall here exploit the decomposition of the maximum M as sum of ladder heights. 0 Alternatively. that r(0) < oo) is Bo: taking y = 0 shows that the conditional distribution of (minus) the surplus ST(o). where G+ is given n=0 by the defective density g+ (x) = 3B (x) = pbo(x) on (0. The expression for g+ was proved in Theorem 11. Thus . 11. B(x)/aB. the formula for the distribution of M follows . i.P) E PnBon(u) . but we shall be able to extract substantial information from the formula. and we further get information about the joint conditional distribution of the surplus and the deficit.2.
cf. (c) the marginal distribution of ST(o)_ is Bo .62 CHAPTER III. 2 The joint distribution of (ST(o )_. In the risk theory literature.5.6. in this setting there is no decomposition of M as a sum of i.1 is traditionally carried out for the imbedded discrete time random walk. Beekman [61]. The proof of Theorem 11. . [62].i. We assume rt > 0 throughout. there is a general marked point process version. Again.2 and it gives an alternative derivation of the distribution of the deficit ST(o) Notes and references The PollaczeckKhinchine formula is standard in queueing theory. As shown in Theorem 11. and the conditional distribution of ST(o) given ST(o)_ = y is the overshoot distribution B(Y) given by Bov)(z) _ Bo (y + z )/Bo(y). ST(o )) given r (0) < oo is the same as the distribution of (VW.2(a) is from Dufresne & Gerber [125]. see Schmidli [323] and references there. the form of G+ is surprisingly insensitive to the form of {St} and holds in a certain general marked point process setup. ST(o) > y. cf. and the conditional distribution of ST(o)_ given ST(o)_ = z is Bo z) The proof is given in IV. (a) 11 (ST(o)_ > x. For the study of the joint distribution of the surplus ST(u)_ just before ruin and the deficit ST(„). 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. Theorem 2. ladder heights so that the results do not appear not too useful for estimating 0(u) for u>0. cf. ST(o)) is given by the following four equivalent statements: B(z) dz.d.V)W) where V. 7r(0 ) < oo) = Q 3 Special cases of the PollaczeckKhinchine formula The model and notation is the same as in the preceding sections. Theorem A1.5.just after ruin. (d) the marginal distribution of ST(o)_ is B0. However. (1 . Asmussen & Schmidt [49]. 1) and W has distribution Fw given by dFyy/ dB(x) = x/µB. f +b (b) the joint distribution of (ST( o). see for example [APQ]. V is uniform on (0. Theorem 2 . where it requires slightly more calculation. Feller [143] or Wolff [384]. the PollaczeckKhinchine formula is often referred to as Beekman 's convolution formula. W are independent.6.
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. 3b Exponential claims Corollary 3. however .p) = S .1)1 00 ( 1 . Bon is the Erlang distribution with n phases and thus the density of M at x > 0 is (1 ..p)pSe a ( l v)x = p( S . a further relevant reference is Bjork & Grandell [67].6. and hence this overshoot has the same distribution as the claims themselves .3. use Laplace transforms.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. As shown in 11. But claims are exponential . Integrating from u to oo. Alternatively. then.p. Thus . B0 is exponential with rate S and the result can now be proved from the Pollaczeck Khinchine formula by elementary calculations .O)e(b0)x. 1 .p. Let r ( x) be the failure rate of M at x > 0.2 If B is exponential with rate S.1 e ax = n1 (n . For a failure at x.e. The result can. 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. also be seen probabilistically without summing infinite series . 0 .p) E pn S n x n. Thus r(x) = S(1 . the current ladder step must terminate which occurs at rate S and there must be no further ones which occurs w. the result follows . I. hence without memory. 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.
(3.1.y)G+(dy ) = f U V(u . Corollary 3. T+ <00) (3. if 3 = 3 and B is a mixture of two exponential distributions with rates 3 and 7.S. u + oo.g.1) For a heavytailed B. is ?7+ ( u). (3.T+ <oo). (Example VIII.+ <u. and weights 1/2 for each.s. Then the first term on the r.4) is similar (equivalently. 2 is one of the main classical early results in the area.h. We mention in particular the following: (a) check that ip (u) = pe (60)u is solution of the renewal equation (3.+ = y yields P(M>u. u .3. (3. (b) use stopped martingales . the survival probability Z(u) = 1 .p + G+ * Z(u) = 1 . we show that expression for /'(u) which are explicit (up to matrix exponentials) come out in a similar way also when B is phasetype. E. (u) 35eu + 35e6u. 0 Proof Write o (u) as P(M>u) = P(S.64 CHAPTER III.y)f3 (y) dy.1 p pBo(u).y)/3B (y) dy. and conditioning upon S.T+ <oo)+P(M> u.4) zu P(M > u .4) can be derived by elementary algebra from (3.3) Equivalently. just insert the explicit form of G+. (3.3.p + f u Z(u . II. cf. A variety of proofs are available .2) Notes and references Corollary 3.+ >u. we use the PollaczeckKhinchine formula in Chapter IX to show that b(u) .S.i(u) satisfies the defective renewal equation Z(u) = 1 .y)G+(dy) For the last identity in (3.3).3)).3) below. THE COMPOUND POISSON MODEL In VIII. 3c Some classical analytical results Recall the notation G+(u) = f^°° G+(dx).2). The case of (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 .+ <U. then 24 1 V.
7) and Corollary 3. Some relevant references are Abate & Whitt [2].s .3. SPECIAL CASES OF POLLACZECKKHINCHINE Corollary 3.5) Proof We first find the m.8) Proof This can be shown.5). In view of (3.4 The Laplace transform of the ruin probability is 65 fo Hence Ee8M 00 e8uiP(u)du .PPB2) EM2 = PPB) + QZPBl 2(1 . [APQ] pp.p = (1 . eau B(u) du = f PB 3PB SPB 0 o (3. 206207).3 .4) can be derived by elementary but tedious manipulations. Embrechts.1 Bo[s] = f oc.Ee8M) f ao e8' ( u)du = a8uP (M > u)du = 0 o 1 ( 1+ (1 . Corollary 3. see e. 191). g.(3 . .5).. which yields the survival probability as 00 f u }t Z(u) = f f3eRtdt 0 from which (3.P)pB' (3.g.pBo[s] no (1 .p) E p"Bo[s]" = 1 . e. either of these sets of formulas are what many authors call the PollaczeckKhinchine formula. 0 Notes and references Corollary 3./3B[s] which is the same as (3.5 can be found in virtually any queueing book.7) s +. [APQ] pp. Of course. by analytical manipulations (L'Hospital's rule) from (3. 111112 or Feller [143]. Bo of B0 as m e8u B(du) = B[s] . (3.(3B[s] 1 . In fact. it is not surprising that such arguments are more cumbersome since the ladder height representation is not used.f. Griibel & Pitts [132].)3B[s]) (3.Ps s(.3 .p)s .p)s s /3 .5 The first two moments of M are 2 EM .p)2 3(1 . We omit the details (see.3 is standard .g.P)PB 2(1 . Also (3.6) 00 = (I . Griibel [179] and Thorin & Wikstad [370] (see also the Bibliographical Notes in [307] p. numerical inversion of the Laplace transform is one of the classical approaches for computing ruin probabilities.7).s . The approach there is to condition upon the first claim occuring at time t and having size x . for example./3B[s] ..
)3(1 .z/'(u) takes the form Z(u) L^J L. THE COMPOUND POISSON MODEL 3d Deterministic claims Corollary 3. we may assume p = 1 so that the stated formula in terms of the survival probability Z(u) = 1 .9) shown for n .1).9).u/p)]k ko k! Proof By replacing {St} by {Stu/p} if necessary./3Z(u . Z^ =eR(k.u) [N(k .u)]k d 1 u) _ a) n ( du ( k! (1  .u) a)Qea" + (1 .u)]k k! (1 L3) 1: e_O(ku) NIN (k (k . Assume (3.4) for Z( u) means f lhu Z(u) = 1./32(u .Q (k 1 k= n  [O(k .s.u) [p(k .Q) k=0 k! E e0( = /32(u) .h.Q) 3e. . of (3. eO('u) [)3(k .u + 1 )]k = QZ(u) .1 < u < n and let Z(u ) denote the r. differentiation yields Z(u) _ /3Z(u) .3I( 0<y<1)dy Z(y)/3I(0<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 (3.9) follows for 0 < u < 1.3+ 18+ J0 Z(uy).6 If B is degenerate at p.u)]k k! k0 The renewal equation (3.1).66 CHAPTER III.1)! k=1 u1 .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. then p) 1: ep(k u/.u)]k1 ku+1) [/3( k . For n < u < n + 1.
d. B9(dx) = B[9] B(dx). or equivalently BB[a] = B[^+ Repeating for t 54 1.4.3) by t.4.1) .3B = .Qe(Bo[a] .rc(9) = . say t = 1: recall from Proposition 1.g. it follows that Z(u) = 2(u) for n<u<n+1. 00 the standard definition of the exponential family {F9} generated by F is FB(dx) = e°xK(e)F(dx).2) (Here 9 is any such number such that r.g. 0 Notes and references Corollary 3. The answer is yes: inserting in (4.f.1) .1 that c(a) = /3(B[a] .4) .(9). (4.2) shows that the solution is Ox [O]0]. we set up .1) or equivalently. See also Iversen & Staalhagen [208] for a discussion of computational aspects and further references.2). CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 67 Since Z(n) = 2(n) by the induction hypothesis. We could first tentatively consider the claim surplus X = St for a single t. but will now be repeated for the sake of selfcontainedness. (4.f.3B[9]. (4. Formalizing this for the purpose of studying the whole process {St}. 4 Change of measure via exponential families If X is a random variable with c.) The adaptation of this construction to stochastic processes with stationary independent increment as {St} has been carried out in 11. K(a) = logEe'X = 109f 00 eaxF(dx) = logF[a]. and thus (4. F and c. The question then naturally arises whether ie is the c. co(a) = rc(a + 9) .r.a.(9) is welldefined. in terms of the c. we just have to multiply (4.4) works as well. of F9.f.f. and define rce by (4.g. 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) .a.6 is identical to the formula for the M/D/1 waiting time distribution derived by Erlang [139].
. BB by (4. for G E FT.8) follows by discrete exponential family theory.g.7) Proof We must prove that if Z is FTmeasurable.6) F(G) = Po (G) = EB [exp {BST + Ttc(0)} . Xn). (4.t. and define 09.2 For any fixed T. Proposition 4. Then FB denotes the probability measure governing the compound Poisson risk process with arrival intensity.9) Proof We first note that for any fixed t. v(Xi.2. it suffices to consider the case where Z is measurable w. Z is measurable w. (4.t. .1 Let P be the probability measure on D[0. THE COMPOUND POISSON MODEL Definition 4. (4. EeeBSt + tk(B) = 1. (4. replications from Fe (replace x by xi in (4.S(k_1)Tln. But let Xk = SkT/n . with common c.68 CHAPTER III. G]. G C {T < oo}.8) By standard measure theory.10) . The following result (Proposition 4. n) for a given n.3 and claim size distribution B. then EBZ = E [Ze9ST _T"(9)I .f. .0e and claim size distribution Be.d. and PBT) the restriction of PB to FT. t < T. in particular the expression (4.i. and dP(T) dP^T) That is.i. and thus (4.1.d. = exp {BST . G]. Ti(a)/n.. the corresponding expectation operator is E9. with T taking the role of n) is the analogue of the expression exp{8(x1 + • • • + xn) .. Then the Xk are i. .(9)} (4.7) now follows by taking Z = eBST+TK(e)I(G) u Theorem 4 . . Then P(G) = Fo(G) = EB [exp {BST + TK(O)} . the PBT) are mutually equivalent on.4). . .3 Let T be any stopping time and let G E FT.5) for the density. oo) governing a given compound Poisson risk process with arrival intensity.FT.r.1) and multiply from 1 to n). The identity (4. (4. Let FT = o(St : t < T) denote the o•algebra spanned by the St.5) for the density of n i.Tic (0)} .r.nr.FTn) = Q(SkTIn : k = 0.
The behaviour at zero is given by the first order Taylor expansion c(a) r.7 1 Some discussion further supporting this statement is given in the next section. t = T .1) . the typical shape of rc is as in Fig. so that PG = EeE0 [exp { 9ST+Trc(9)}I(G)I FT)] = Ee [exp { BST + rrrc(O)} I(G)EB [ exp {9 (ST .1 It is seen that typically) a ry > 0 satisfying 0 = r. Thus by (4. Now consider a general G. and hence (4. .r is deterministic. according to what has just been proved. c(a) is a convex function of a.r)rc(9)}I . LUNDBERG CONJUGATION 69 Now assume first that G C Jr < T} for some deterministic T. Ee [exp { BST +Trc(9)} I(G) FT)] = 1.1(a).9) holds with G replaced by GT. (0) + rc'(0)a = 0 + ES1 a = a (p .(Y) = 13(B['Y] . (a) rc (a) (b) KL(a) 'Y 'Y Figure 5.g.. Letting T t oo and using monotone convergence then shows u that (4.1) _ 1 + a.5. Given FT.7) holds.FT]] = EB [exp { BST + Trc(9)} I(G)] . GT C_ Jr < T}. (4.ST) + (T . 5. Then GT = G n Jr < T} satisfies GT E FT.f. Thus. 5 Lundberg conjugation Being a c. Then G E FT.10). 77 Thus.9) holds for G as well. subject to the basic assumption ij > 0 of a positive safety loading.
Example 5 .1) . 5.1) is precisely what is needed for one of the terms in the exponent . we write FL instead of F7. (5. and (4.1(b). u It is a crucial fact that when governed by FL.1) is known as the Lundberg equation and plays a fundamental role in risk theory .70 CHAPTER III. .4) ELS1 = #L(0) cf. It is then readily seen that the nonzero solution of (5. e.a = i(a + 7). the claim surplus process has positive drift > 0. Thus B[7] = 6/. b[s] = 5/(b .1(b).3. Fig. 5. we further note that ( 5. THE COMPOUND POISSON MODEL exists .g. An established terminology is to call y the adjustment coefficient but there are various alternatives around. Note that KL (a) = /L (BL [a] .2)) is 7 = 5/3.3.QL instead of /37 and so on in the following .3) cf. (5.1 Consider the case of exponential claims. Thus.s). Equation (5.2) 7 Figure 5. 5. G = {T(u) < oo} in Theorem 4. an equivalent version illustrated in Fig.2 s As support for memory. Fig. the Lundberg exponent.4) yields /3L = b and that BL is again exponential with rate bL =. Lundberg conjugation corresponds to interchanging the rates of the interarrival times and the claim sizes. Taking T = r(u).3.1) (or (5.2 is B(7) = 1 + ^. (5.
Then P(ST+ E A) = EL [exp { 7S?+} .3 (THE CRAMERLUNDBERG APPROXIMATION) i'(u) . Proof Just note that e(u) > 0 in (5.4). ST+ E A] .e7x)G+(dx). take first 0 = ry.8) .6) Proof By renewal theory.r. we therefore have ELe7t(u) + C where C ELe7 (00) = µ+) f e7(1 . (5.(oo) (in the sense of weak convergence w.1p .ascending ladder height distribution and µ+ its mean.(u)} .5) Theorem 5 . (5.7) is the same as (5.6 ). T = T+.G+L)(x)) dx ry^+L) J 00 f 0 (1 .2 (LUNDBERG'S INEQUALITY) For all u > 0.G+ L) (x) G+L) (x) IL(+) µ+L) L) where G+L) is the FL.u be the overshoot and noting that PL(T(u) < oo) = 1 by (5.3 takes a particular simple form.5. 0 Theorem 5 . which shows that G(L) (dx) = e7xG +(dx) = e7x /3 (x) dx.t. To this end. Letting e(u) = ST(u) . (5.1. see A . LUNDBERG CONJUGATION 71 to vanish so that Theorem 4. V) (u) = P(T(u) < oo) = EL [exp {ryS. Since a7' is continuous and bounded.+ E A} in Theorem 4. we can rewrite this as 0(u) = e"ELe7^(u). T(u) < oo] .7) 0 and all that is needed to check is that ( 5. e(u) has a limit i. G = {S.3.1 (5. PL ) with density 1 . where C .P Y j o' xeryxOB (x) dx /3k [Y] .Ce7u as u 4 oo.5). V)(u) < e7u.1e.
THE COMPOUND POISSON MODEL In principle. Using (5.1 = ^7 The accuracy of Lundberg's inequality in the exponential case thus depends on how close p is to one.1 .")G + (dx ) = 1  J0 00 3B(x) dx = 1p. A direct proof of C = p is of course easy: B ['y] d S S (S7 )2 d7S y S 02' C 1p 1p _ 1p /3B' [7] 2 1 P1 p. that 7 = S . From this it follows.1)) and 7µ+L) = 'y/3 [7] 7 1/0 = /3B ['y] .8) yields +L) J0 xel'B ( x) dx (5. or equivalently of how close the safety loading 77 is to zero.11) so that I 7B ['Y](B[7]1) BI [7]Q VP (7) 72 7 (using (5.3 (this was found already in Example 5.4 Consider first the exponential case b(x) = Seax. u .4). (5.72 CHAPTER III.12) Example 5 . we get L where 00 (1 . . but some tedious (though elementary) calculations remain to bring the expressions on a final form.1) (5.10) VW = JI c* e° (x) dx = a (B[a] .1 above) and that C = p. this solves the problem of evaluating (5. of course.e.7). Then 0(u) = pe(a_Q)u where p = /3/S. Noting that SIG(L)II = 1 because of (5.
5. LUNDBERG CONJUGATION Remark 5.5 Noting that PL  1 = ,3LIBL  1 = #ci (0 ) = k (ry) _ ,QB' ['Y]  1 ,
73
we can rewrite the CramerLundberg constant C in the nice symmetrical form G, _'(0)1  1  p K'(7) PL1
(5.13)
In Chapter IV, we shall need the following result which follows by a variant of the calculations in the proof of Theorem 5.3: 1  aB[ry  a]  1 Lemma 5 . 6 For a # ry, ELea^ (°°) = 7 aK'(7) 7  a Proof Replacing 7 by a in (5.7) and using ( 5.8), we obtain 1 (I 1  ^ e('ra) x,3 (x)dx) (L ) ELea^*) = a \\\ f
using integration by parts as in (3.6) in the last step . Inserting (5.12), the result follows. u
Notes and references The results of this section are classical, with Lundberg's inequality being given first in Lundberg [251] and the CramerLundberg approximation in Cramer [91]. Therefore, extensions and generalizations are main topics in the area of ruin probabilities, and in particular numerous such results can be found later in this book; in particular, see Sections IV.4, V.3, VI.3, VI.6.
The mathematical approach we have taken is less standard in risk theory (some of the classical ones can be found in the next subsection). The techniques are basically standard ones from sequential analysis, see for example Wald [376] and Siegmund [346].
5a Alternative proofs
For the sake of completeness, we shall here give some classical proofs, first one of Lundberg's inequality which is slightly longer but maybe also slightly more elementary:
74 CHAPTER III. THE COMPOUND POISSON MODEL
Alternative proof of Lundberg 's inequality Let X the value of {St} just after the first claim , F(x) = P(X < x). Then , since X is the independent difference U  T between an interarrival time T and a claim U, ,3+ry F'[7} = Ee7 ( UT) = Ee7U • Ee7T = B['Y] a = 1' where the last equality follows from c(ry) = 1. Let 0( n) (u) denote the probability of ruin after at most n claims. Conditioning upon the value x of X and considering the cases x > u and x < u separately yields
,0(n +1) (u) = F(u) +
Ju
0 (n) (u  x) F(dx).
We claim that this implies /,(n) (u) < e 7u, which completes the proof since Vi(u) = limniw 1/J(n) (u). Indeed , this is obvious for n = 0 since 00)(u) = 0. Assuming it proved for n, we get
„/, (n+1)(u) <
F(u) + e7u
00
Ju
e7(u=) F(dx)
00
<
f
e7x F(dx)
+ fu
e  7(u z) F(dx)
u
o0
= e 7uE[ 'Y] = e 7u.
Of further proofs of Lundberg's inequality, we mention in particular the martingale approach, see II.1. Next consider the CramerLundberg approximation. Here the most standard proof is via the renewal equation in Corollary 3.3 (however, as will be seen, the calculations needed to identify the constant C are precisely the same as above): Alternative proof of the CramerLundberg's approximation Recall from Corollary
3.3 that
(u) = )3
J OO B(x) dx + J U Vi(u  x)/3 (x) dx.
u 0
Multiplying by e7u and letting Z(u) = e7" O(u), we can rewrite this as
u Z(u) =
z(u) = e7u/
J
B(x)dx, F(dx) = e7x,QB(x)dx,
u
z(u)
f +
J
e7(ux ),Y' 1 • l•(u  x) • e7'/B(x) dx,
0
= z(u) +
J0 u Z(u  x)F(dx),
6. MORE ON THE ADJUSTMENT COEFFICIENT 75
i.e. Z = z+F*Z. Note that by (5.11) and the Lundberg equation, ry is precisely the correct exponent which will ensure that F is a proper distribution (IIFII = 1). It is then a matter of routine to verify the conditions of the key renewal theorem (Proposition A1.1) to conclude that Z (u) has the limit C = f z(x)dx/µF, so that it only remains to check that C reduces to the expression given above. However, µF is immediately seen to be the same as a+ calculated in (5.10), whereas
L
00
z(u) du =
f
J
/3e7udu "o
J "o B(x) dx = J "o B(x)dx J y,0eludu
u 0 0
B(x)^ (e7x  1) dx = ^' (B[7]  1)  As] [0 µs] = l y P^
using the Lundberg equation and the calculations in (5.11). Easy calculus now gives (5.6). u
6 Further topics related to the adjustment coefficient
6a On the existence of y
In order that the adjustment coefficient y exists, it is of course necessary that B is lighttailed in the sense of I.2a, i.e. that b[a] < oo for some a > 0. This excludes heavytailed distributions like the lognormal or Pareto, but may in many other cases not appear all that restrictive, and the following possibilities then occur: 1. B[a] < oo for all a < oo. 2. There exists a* < oo such that b[a] < oo for all a < a* and b[a] = 00 for all a > a*. 3. There exists a* < oo such that fl[a] < oo for all a < a* and b[a] = 00 for all a > a*. In particular , monotone convergence yields b[a] T oo as a T oo in case 1, and B[a] T oo as a f a* in case 2 (in exponential family theory , this is often referred to as the steep case). Thus the existence of y is automatic in cases 1 , 2; standard examples are distributions with finite support or tail satisfying B(x) = o(eax)
76 CHAPTER III. THE COMPOUND POISSON MODEL
for all a in case 1, and phasetype or Gamma distributions in case 2. Case 3 may be felt to be rather atypical, but some nonpathological examples exist, for example the inverse Gaussian distribution (see Example 9.7 below for details). In case 3, y exists provided B[a*] > 1+a*/,3 and not otherwise, that is, dependent on whether 0 is larger or smaller than the threshold value a*/(B[a*]  1). Notes and references Ruin probabilities in case 3 with y nonexistent are studied, e.g., by Borovkov [73] p. 132 and Embrechts & Veraverbeeke [136]. To the present authors mind, this is a somewhat special situation and therefore not treated in this book.
6b Bounds and approximations for 'y
Proposition 6.1 ry <
2(1  aps) 2µs
OMB PB)
Proof From U > 0 it follows that B[a] = Eea' > 1 + µsa + pB2)a2/2. Hence 1 = a(B[7]  1) > Q (YPB +72µs)/2) = 3µs + OYµa2) 2 (6.1) 7 'Y from which the results immediately follows. u
The upper bound in Proposition 6.1 is also an approximation for small safety loadings (heavy traffic, cf. Section 7c): Proposition 6.2 Let B be fixed but assume that 0 = ,3(77) varies with the safety loading such that 0 = 1 Then as 77 .0, µB(1 +rl) 2) Y = Y(77) 277 PB Further, the CramerLundberg constant satisfies C = C(r1)  1. Proof Since O(u) + 1 as r7 , 0, it follows from Lundberg's inequality that y * 0. Hence by Taylor expansion, the inequality in (6.1) is also an approximation so that OAY]  1) N Q (711s + 72µB2) /2) = p + 3,,,(2) B 'y 7 2 2(1  p) _ 271µB
QPB PB)
6. MORE ON THE ADJUSTMENT COEFFICIENT 77
That C 4 1 easily follows from y 4 0 and C = ELe7V°O) (in the limit, b(oo) is distributed as the overshoot corresponding to q = 0 ). For an alternative analytic proof, note that C  1P = rlµB 73B' [7]  1 B' [ry)  1/0 711µB µB +7µB2 )  µB(1 +77 ) 'l = 1. 277q
77
7PBIPB
 77
13 Obviously, the approximation (6.2) is easier to calculate than y itself. However, it needs to be used with caution say in Lundberg's inequality or the CramerLundberg approximation, in particular when u is large.
6c A refinement of Lundberg 's inequality
The following result gives a sharpening of Lundberg 's inequality (because obviously C+ < 1) as well as a supplementary lower bound:
Theorem 6 .3 C_eryu < ,)(u) < C+ eryu where
= B(x) = C_ x>o f °° e7( Yx)B(dy )' C+
B(x) xuo f 0 e'r( vx)B(dy)
Proof Let H(dt, dx ) be the PLdistribution of the time r(u) of ruin and the reserve u  S7(„)_ just before ruin . Given r(u) = t, u  ST (u) = x, a claim occurs at time t and has distribution BL(dy)/BL(x), y > x. Hence ELe7£(u) 0
J
°o
H(dt, dx)
fX
eY(Y x) 00 f°° B(dy) x
BL dy
BL(x)
o
f
f H(dt, dx)
L ^ H(dt, dx) f e7B( x)B(dy) Jo oc, < C+
J0 0 o" H(dt, dx) = C. o" J
The upper bound then follows from ik(u) = e7uELeVu), and the proof of the u lower bound is similar.
78 CHAPTER III. THE COMPOUND POISSON MODEL
Example 6.4 If B(x) = eax, then an explicit calculation shows easily that B(x) _ e6X fz ° e7(Yx)B(dy) f x' e(6,6)(Yx)8esydy = 5 = P. Hence C_ = C+ = p so that the bounds in Theorem 6.3 collapse and yield the exact expression pey" for O(u). u The following concluding example illustrates a variety of the topics discussed above (though from a general point of view the calculations are deceivingly simple: typically, 7 and other quantities will have to be calculated numerically). Example 6.5 Assume as for (3.1) that /3 = 3 and b(x) = 2 .3e3x + 2 .7e7x, and recall that the ruin probability is 24 5su 5eu + 3e *(u) = 3 Since the dominant term is 24/35 • e", it follows immediately that 7 = 1 and C = 24/35 = 0.686 (also, bounding aS" by a" confirms Lundberg's inequality). For a direct verification, note that the Lundberg equation is
7 = /3(B['Y]1)
= 3\
2.337
+2.7771
which after some elementary algebra leads to the cubic equation 273  1472 + 127 = 0 with roots 0, 1, 6. Thus indeed 7 = 1 (6 is not in the domain of convergence of B[7] and therefore excluded). Further, 1P = B [7] 181B = 13 2.3+2.71 = 1 3 1 7 I 7'
_ 17
2 (3 a )2 + 2 (7  a)2 «=7=1 2 1p _ 7 _ 24
36 '
3.171 35* 36 For Theorem 6.3, note that the function QB[Y]1 f°°{L 3e_3x+
u
• 7e7x 1 dx
J
3 + 3e4u
f 0c, ex .
I 2 . 3e3x + 2 . 7e7x l dx
l J
9/2 + 7/2e4u
7. VARIOUS APPROXIMATIONS FOR THE RUIN PROBABILITY 79
attains its minimum C_ = 2/3 = 0.667 for u = oo and its maximum C+ = 3/4 = 0.750 for u = 0, so that 0.667 < C < 0.750 in accordance with C = 0.686.
Notes and references Theorem 6.3 is from Taylor [360]. Closely related results are given in a queueing setting in Kingman [231], Ross [308] and Rossberg & Siegel [309]. Some further references on variants and extensions of Lundberg's inequality are Kaas & Govaaerts [217], Willmot [382], Dickson [114] and Kalashnikov [218], [220], all of which also go into aspects of the heavytailed case.
7 Various approximations for the ruin probabil
ity
7a The BeekmanBowers approximation
The idea is to write i (u) as F(M > u), fit a gamma distribution with parameters A, 6 to the distribution of M by matching the two first moments and use the approximation
0(u)
f
u
Sa
r(A)
xa  leax dx.
According to Corollary 3.5, this means that A, 8 are given by A/S = a1, 2A/52 = a2 (2) PIB3) ^ZP(B)2 __ PPB a2 al 2(1  P)PB 3(1  P)µ8 + 2(1  p)2' i.e. S = 2a1 /a2, A = 2a2 1/a2.
Notes and references The approximation was introduced by Beekman [60], with the present version suggested by Bowers in the discussion of [60].
7b De Vylder's approximation
Given a risk process with parameters ,(3, B, p = 1, the idea is to approximate the ruin probability with the one for a different process with exponential claims, say with rate parameter S, arrival intensity a and premium rate p. In order to make the processes look so much as possible alike, we make the first three cumulants match, which according to Proposition 1.1 means p=AUB1=P1,
2N
(2) 6^= =OP
,
/3,4)
.
3 )1 } _ 1p 1 . the approximating risk process has ruin probability z. 7c The heavy traffic approximation The term heavy traffic comes from queueing theory.1 As /3 f Nmax.3* /S.(3)2 P PB 2µB 2µB Letting /3* = /3/P. the premiums exceed only slightly the expected claims. Letting Bo be the stationary excess life distribution. [174]) shows that it may produce surprisingly good results.g. numerical evidence (e. That is. Mathematically. and hence the ruin probability approximation is b(u) e(bAln)u. but has an obvious interpretation also in risk theory: on the average. Proposition 1. cf./3)PBo PB .PBo [s (/3max . we have according to the PollaczeckKhinchine formula in the form (3. Grandell [171] pp. (/3max .80 CHAPTER III.Ps(/3max .7) that Ee$(Amex /j)M _ 1p _ 1p Eo [s (0max 1 .1.1.P . Notes and references The approximation (7.8µBo  Ss' u where 6 = µB/µBo = 2µa/µB 2) .)3 )PBo µB  .b(u) = p*e. heavy traffic conditions mean that the safety loading q is positive but small. THE COMPOUND POISSON MODEL These three equations have solutions 9/3µB2)3 30µa2)2 3µa2) (3) P+ (3) ' 0 . or equivalently that /3 is only slightly smaller than /3max = 1/µ8.2) was suggested by De Vylder [109].3 and Corollary 3.s(/3max .p . we shall represent this situation with a limit where /3 T fl but B is fixed.p = (/3max 0)µB.(bA*)".p + p { 1 1p ti 1 . p* _ ./3)] 1 ./3)M converges in distribution to the 2a exponential distribution with rate S = B' Proof Note first that 1 . 1924. Proposition 7. Though of course it is based upon purely empirical grounds.
ryu . In the setting of risk theory. light traffic is of some interest as a complement to heavy traffic . VARIOUS APPROXIMATIONS FOR THE RUIN PROBABILITY 81 Corollary 7. then P(u) 4 e6„ Proof Write z'(u) as P((/3max .1 1 . but has an obvious interpretation also in risk theory: on the average . and hence 2µ2B 1 . However .0)u. or equivalently that 0 is small compared to µB .3) is reasonable for g being say 1020% and u being small or moderate./3)u).g.2 If .2. This follows since rl = 1/p . [APQ] Ch.4) suggested by the Cramer Lundberg approximation and Proposition 6. Notes and references Heavy traffic limit theory for queues goes back to Kingman [230]. we shall represent this situation with a limit where 3 10 but B is fixed. 7d The light traffic approximation As for heavy traffic .ze a2unµB laB (7. 2 provides the better mathematical foundation. VIII). the first results of heavy traffic type seem to be due to Hadwiger [184]. while the approximation may be far off for large u. 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 . 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./3)u * v. .l3)M > (/3max .p.Q T /3max. Of course. Numerical evidence shows that the fit of (7.p _ 2rl11B PB p.7. the term light traffic comes from queueing theory.B AB ) 6()3max _'3) = However . light traffic conditions mean that the safety loading rl is positive and large . obviously Corollary 7. the premiums are much larger than the expected claims . u * oo in such a way that (3max . Mathematically. as well as it is needed for the interpolation approximation to be studied in the next subsection. We return to heavy traffic from a different point of view (diffusion approximations) in Chapter IV and give further references there . That is .
[97]. U > u] = /3iE(U . Another way to understand that the present analysis is much simpler than in these references is the fact that in the queueing setting light traffic theory is much easier for virtual waiting times (the probability of the conditioning event {M > 0} is explicit) than for actual waiting times . Asmussen [19] and references there.e. 0 u Notes and references Light traffic limit theory for queues was initiated by Bloomfield & Cox [69]. The crude idea of interpolating between light and heavy traffic leads to 0 (u) C1 . For a more comprehensive treatment. En'=2 • • • = O(/32) so that only the first terms matters. z/' (u) convergence P(U . THE COMPOUND POISSON MODEL Proposition 7. Light traffic does not appear to have been studied in risk theory. ao n=1 00 n=1 (u) P) anllBBon(U) onPaBon(u) • Asymptotically.82 CHAPTER III. ( 3 J O B dx. and hence 00 (U) /3pBBo (u) = 0 / B(x)dx. Omax max m. Indeed.5) follow by integration by parts. . by monotone time T of the first claim .u. 10 ( u The alternative expressions in (7.u)+. Again. cf.(3 10. (7. i.3 As .= 1 aJ 1 a 0+ 1 = = p.T > u).3 is the same which comes out by saying that basically ruin can only occur at the F(U .Q limIP ( u) + Q lim z/'(u) Amax &0 amax ATAm.5) u Proof According to the PollaczeckKhinchine formula. 0(u) /3 J B(x)dx = /3E[U . Sigman [347]. 7e Interpolating between light and heavy traffic We shall now outline an idea of how the heavy and light traffic approximations can be combined. u Note that heuristically the light traffic approximation in Proposition 7.T > u) = J o" B(x + u)/3eax dx . the Poisson case is much easier than the renewal case. see Daley & Rolski [96].
. Instead.O(E)(u) 1 (1 . ^IE) exist: 1 (B) HT QmsxQ hm J e e6" 2µE/µE2)'" = e(1 6)" =  Q1Qm. where further references can be found .O0 M. 8 Comparing the risks of different claim size distributions Given two claim size distributions B(1).8. (U). Another main queueing paper is Whitt [380]. Al .3n. one may hope that some correction of the heavy traffic approximation has been obtained.3 and use similar notation for %(B) (u) = (u). the idea of interpolating between light and heavy traffic is due to Burman & Smith [83 ]. Thus . Let OLT) (u) denote the light traffic approximation given by Proposition 7. . _(E) (u). The adaptation to risk theory is new./3)) . Substituting v = u(. ^ LT Q maxQ m"^ Qlo V LT) ( CHT(v) (say).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. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS 83 which is clearly useless . 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). (7.6) is . 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. 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. z/i(E) (u) = pe(QmaxQ)u.Wmax f(x ) dx + pee6mQ. Notes and references In the queueing setting .VHT) ( ax QmQ ) h (B) ( . available. we see that the following limits HT) (u'). no empirical study of the fit of (7.3). that is.x . [84]. B(2). however. with rate 1/µB = /3max. "/Qmex Cu) CLT(u ( /3max 0) + O16 CHT( U(Qmaz . ) M. even if the safety loading is not very small.
we can assume that 1) < St 2l for all t. cf. Proof According to the above characterization of stochastical ordering. Rather than measuring difference in size. U(2) such that U(l) has distribution B('). U(2) distribution B(2) and U(1) < U(2) a. In terms of the time to ruin. 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). we shall need various ordering properties of distributions. XI. most often the term stoploss ordering is used instead of increasing convex ordering because for a given distribution B.s.ill(u) < V)(2) (U) for all u. Recall that B(') is said to be stochastically smaller than B(2) (in symbols. In particular (consider the convex functions x and x) the definition implies that B(1) and B(2) must have the same mean. THE COMPOUND POISSON MODEL To this end. B(2)) in the increasing convex order if f BM (y) dy < f 00 Bi2i (y) dy x x for all x. the proof is complete.' 1)(u) < V)(2) (U) for all u. Proposition 8. B(') <i. Bill is said to be convexly smaller than B(2) (in symbols. A weaker concept is increasing convex ordering: B(1) is said to be smaller than B(2) (in symbols. equivalent characterizations are f f dB(') < f f dB (2) for any nondecreasing function f.1 is quite weak.2 If B(') <j. this ordering measures difference in variability. Taking probabilities. or the existence of random variables U(l). then . for more detail and background on which we refer to Stoyan [352] or Shaked & Shantikumar [337]. an equivalent characterization is f f dB(') < f f dB (2) for any nondecreasing convex function f. Proposition 8. this implies St T(l)(u) > r(2)(u) for all u so that 17(I) (U) < oo} C_ {T(2)(u) < oo}. B(2)) if f fdB(1) < f fdB(2) for any convex function f.1 If B(') <d B(2). B(' <. whereas (consider x2) B(2) has the larger variance. B(2) and PB(1) = µB(2).6. B(') <d B(2)) if B(1)(x) < B(2)(x) for all x. then Bill = B(2). In the literature on risk theory. u Of course. . then 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. Finally. Here convex ordering is useful: Proposition 8.84 CHAPTER III. we have the convex ordering.
(1) (. from which the result immediately follows.1 and µB at 1 so that the safety loading 11 is 10%. say to p.4 Let D refer to the distribution degenerate at 'LB . . and consider the following claim size distributions: B1: the standard exponential distribution with density ay. then B(1) <. A general picture that emerges from these results and numerical studies like in Example 8. we have Bol) (x) f ' B(1) (y) dy < ' f' B(2) (y) dy = Bo2) (x)• µ 85 I. This u implies that D <.u) = (1 _ P) E /3npnBo( 1):n(u) n=1 00 < (1. Corollary 8. Proof Consider the light traffic approximation in Proposition 7. larger variance is paramount to larger second moment.5 If '0(1)(u) < p(2) (U) for all u and a. Example 8.2 is the following: Proposition 8. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS Proof Since the means are equal. The problem is to specify what 'variation' means.4) certainly supports this view: noting that.6 Fix /3 at 1/1. B(2). we have by Jensen 's inequality that E f (U) > f ( EU).3 If B(1) <. (D) (u) < O(B) (U ) for all u.3 provides another instance of this. it is seen that asymptotically in heavy traffic larger claim size variance leads to larger ruin probabilities. The heavy traffic approximation (7.6 below is that (in a rough formulation) increased variation in B increases the risk (assuming that we fix the mean).. then /'(')(u) < 0(2)(u) for all u.8.1. Then V. B. A first attempt would of course be to identify 'variation' with variance. Hence by the PollaczeckKhinchine formula ..p ) E /3"µ"Bo2)* n(u) _ V(2) (u) n=1 = Corollary 8.e. Proof If f is convex. with fixed mean. B(2). and here is one more result of the same flavor: Corollary 8. Bo1) <_d Bo2) which implies the same order relation for all convolution powers. 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. A partial converse to Proposition 8.
Let ua denote the a fractile of the ruin function. B4: the Pareto distribution with density 3/(1 + 2x)5/2. all distributions have mean 1. in comparison to B2 the effect on the ua does not show before a = 0. which appears to be smaller than the range of interest in insurance risk (certainly not in queueing applications!). 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). 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.) = a.000. and consider a = 5%. A2 = 3. Kluppelberg [234]. with the hyperexponential distribution being more variable than the exponential distribution and the Erlang distribution less. In terms of variances o2. We return to ordering of ruin probabilities in a special problem in VI. However.1358. Note to make the figures comparable.e.0' U0. and this is presumably a consequence of a heavier tail rather than larger variance.86 CHAPTER III. THE COMPOUND POISSON MODEL B2: the hyperexponential distribution with density 0. 1/)(u.001 u0. 9 Sensitivity estimates In a broad setting. = 0. B. Pellerey [287] and (for the convex ordering) Makowski [ 252].. [166]. van Heerwarden [189]. B3 the comparison is as expected from the intutition concerning the variability of these distributions.4.01%.4142.01%.9A2e'2r where A.e'\1x + 0. 11 Notes and references Further relevant references are Goovaerts et al. i. A standard example from queueing theory is . 0. the behaviour of which is governed by a parameter 9. we have 0r3 = 2 < or2 = 1 < 02 = 10 < 04 = 00 so that in this sense B4 is the most variable.lA. One then obtains the following table: U005 U0. For B1i B2. 0. B3: the Erlang distribution with density 4xe2x. 1%.1%.
Example 9.1 Consider the case of claims which are exponential with rate 8 (the premium rate is one). if = a e(6A)u. u Proposition 9. t]. a/3 0 . and hence a _ e(60)u + u e(60)u = ( i + which is of the order of magnitude uV.. obtained say in the natural way as the empirical arrival rate Nt/t in [0. where the partial derivatives are evaluated at p = 1. it follows that ' is approximatively normal N(0. Then if t is large . i.(u) for large u.e.2 Consider a risk process { Rt} with a general premium rate p. Then a p ao = 00 Qa/. increasing in u. where Q2 = fl ( l2 1113 / _ Ou2v)2.9. with 0 the vector of service rates at different nodes and routing probabilities. In the present setting. In particular . the distribution of %3 0 is approximatively normal N(0„ Q/t). say estimated from data. and hence the effect of changing p from 1 to 1 + Ap corresponds to changing /3 to /3/(1 + Op) /3(1 . while /3 = j3 is an estimate. and s(9) the expected sojourn time of a customer in the network. For example. a2/t). the standard deviation on the normalized estimate ^/1' (the relative error ) is approximatively .19P a/ . or we may be interested in aV)/0/3 as a measure of the uncertainty on '0 if 0 is only approximatively known. a0 as ao 80 19P . Thus. Similar conclusions will be found below. Then ib = Pe(613)u. the premium rate p and the claim size distribution B. Then the arrival rate /3(P) for { R(P) } is )31p. Let R(P) = Rtli.3. Assume for example that 8 is known. Thus at p = 1. 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. Proof This is an easy time transformation argument in a similar way as in Proposition 1. we may be interested in a'/ap for assesing the effects of a small change in the premium.1. SENSITIVITY ESTIMATES 87 a queueing network.01/2u.Ap).
x > 0 (9.2) (see Remark 9.w(O.10) below.4). Proposition 9.6 below for some discussion of this assumption). () Proof According to (9. ()} p(dx) .((dx ) = exp {Ox + (t(x) . The most intuitive approach is to rely on the accuracy of the CramerLundberg approximation . e. 5) (Q+'Y)[we(0+7. /3 yields w e(e + Y.()(0 +'0) ' (9 . mathematically a proof is needed basically to show that two limits (u * oo and the differentiation as limit of finite differences) are interchangeable. ^) .^)] 1(/3+y)we(9+'y. Consider first the case of 8/8/3: . However . and write yp = 8y/8/3 and so on . In the case of the claim size distribution B.3. Of course.w(6. 9. 3) ( 9 . it suffices to fix the premium at p = 1 and consider only the effects of changing . Consider first the adjustment coefficient y as function of 3.r. we can rewrite the Lundberg equation as w(9+ y. and the proofs of (9.uypCe7u urypO. this intuition is indeed correct. (3+'y)PC (0+7. (.()YC = 1 +y/ /3 \ Q2 From this (9.(/3 + y)we(9 + 7. (9.3) follows by straightforward algebra. u Now consider the ruin probability 0 = 0 (u) itself.3.3 70 = 'Ye = = 7 /3(1 we(e +'y. () = log(1 + y//3).g.88 CHAPTER III. but we shall concentrate on a special structure covering a number of important cases. various parametric families of claim size distributions could be considered. so that heuristically we obtain '00 50ryu = Coe"u .6) As will be seen below. namely that of a twoparameter exponential family of the form Bo.3 or/and B.Owe (9.0 = t/'(u) and the CramerLundberg constant C.t. for the ruin probabilities . Viei '0(.()^ 1 . Differentiating w. (9. but must look for approximations for the sensitivities 0.()wC(e.5) are similar. Similar notation for partial derivatived are used below. we cannot expect in general to find explicit expressions like in Example 9. THE COMPOUND POISSON MODEL As a consequence. 4) (9 .1 or Proposition 9.
8) Letting cp = e0/e/3 and differentiating (9. () . we get p(u) = J "O B(x) dx + J U O(u .9) (9. ()} .w(9.x) F(dx ) f u J C F(dx) = C as u 4 oo.([a] = exp {w(9 + a. (9. Further write de = [we (9 +'y.x)B(x)dx. we note the formulas Ee. 11 For the following.we(9 . BarndorffNielsen [58]).12)). F(dx) = e'yy/3B(x)dx.x).3 (see in particular (5. ()] exp {w (O + y. () . Be.3(x) dx.p)/C'y. z2(u) _ 1 ^ e'ri`i7i( u . Z(u)/u a C//3PF where PF is the mean of F.QB(x) dx.4 As u oo.11) Ee.w(9.4t (U)e°`U = which are wellknown and easy to show (see e.10) (9. Hence by a variant of the key renewal theorem (Proposition A1.C). () exp {w(9 + a.9. z2(U) = e7" J u b(u . Combining these estimates .g. ()} . By dominated convergence.x)B(x) dx + J U W(u .(e"U = = wS(O. () . O} (9. and alsoo zl(u) + 0 because of B['y ] < oo.8). w((9 + a. it holds that 89 a ue ryu a/3 Q(1 P) 7C2 Proof We shall use the renewal equation (3. the proof is complete. we multiply by e7" and let Z(u) = elt" cp(u).w(O. u 0 Then Z = z + F * Z and F is a proper probability distribution . PF = (1 . But from the proof of Theorem 5. 0(u) = /3 Ju"O B(x) dx + f 0 0(u . Z= zl + z2 where zl (u) = e7u J m B(x)dx. SENSITIVITY ESTIMATES Proposition 9. u 0 Proceeding in a similar way as in the proof of the CramerLundberg approximation based upon (9.3) for z/'(u).St (U) Ee.2 of the Appendix ). () .x).8) (Section 5).
0 x Multiplying by e7" and letting Z(u) = e"uV(u).12) f exp {O y + (t(y) .9)(9. ^) .11).w (9. oo z2 (u) f C . u Z2(U) = e7° f u ^/i(u .1) B(dy) 'f '[t(y) .e7x/3 f 00 [t(y) .x)f3 f ^[t(y) .8) that cp(u) .w(e.w(0. 8^ ue7u. ())B(dy) dx. Then as u > oo. ()](e7v . 8 8() 8( (9. 01 (i+) do = +'Y.w( (0. this implies Z = z + F * Z. ()]B(dy) dx x 0 0C T ON O .6C do 89 1p 8( 1p Proof By straightforward differentiation. ()]B(dy) dx. F(dx) = e7x. C) . )}B(dy)• Letting cp it thus follows from (9.we (0. THE COMPOUND POISSON MODEL [we(e+7.2) holds. By dominated convergence and (9.90 CHAPTER III.wc (O. z = zl + z2.lB(x) dx = e7uzl(u) + e7°zz(u) + V(u T where zl (u) = . ^)} [wc (0 + 7.5 Assume that (9.w( (0. C)] (1 + 7 ) Proposition 9. 2 z 07P N ue7u (3C de .wc(O.QB(x)dx.6e7u f "o f[t(y) .x). ()]e7vB(dy) 'fCd 7 c . ()} 1z(dy) = f [t(y) .wc(9.
(log r(a) a log S)} • r(a) 1.3ary tog('Finally..Y)a+1 ' (9. < = a. U 7µF from which the second assertion of (9. a /(S . ())B(dy) < oo.18) (05 + 57 _'3_y .Sry a/32 + a/37 + /37 . t(x) = logx.12) takes the form y) alp a .) log(9) = %F(a) logs where %1 = F'/]F is the Digamma function. It follows after some elementary calculus that p = a)3/5 and.1 .2) holds with p(dx) = xldx.12) follows. and the proof of the first one u is similar. Here (9.15) (9.QS 1 . () = C/9 = a/S. that C = a. and also zj (u) 4 0 because of f Hence./35' a/i'y + aryl 625ry.w((9. () = log r(a) . 9 = S. ue_Yu 'C2d( 8a 8( 1 p . () ='I'(t..6 Consider the gamma density b (x ) = Sa xa. Example 9. We get w( (0.C log(9).1edz = 1 exp {Sx + a log x .ry) 5a1 cry (5 .yu/3C2do u86 89 1p' az/) = 8z/. w(e.9.rye) S 5rya. ( 9.. we (9.16) (9. Z(u) /3C 91 o c'o e11(t (y) . SENSITIVITY ESTIMATES as u 3 oo. by inserting in the above formulas.17) (9.13) (9.pa+1 . .14) de = d( 7!3 76 = 7e = log ( \ ( \5a_ / \SSry ) 72 .a log S = log r(c) .a/35a&y' ' (9.
S[a] = exp {w (9 + a.2 log (0.CZ try)} 1 C C2 try . Be.3Ee. 9 = .1 = eXP {c(C .l3 of Section 6a needed for the existence of ry becomes e^Q > 1+62 / 2. C = .log c = 2 In particular.9) (() .2a) } Thus the condition B[a*] > 1 + a* /. C) .3.2) with µ(dx) = 2x3zrdx.1 16 +ry c C22ry 2( = + 70 We (e. Straightforward but tedious calculations .7 Consider the inverse Gaussian density ( b(x) Zx37 exp This has the form (9."62 .22.92 CHAPTER III. further yield . t(x) _ .2 . THE COMPOUND POISSON MODEL Example 9. for a < a* = z (. which we omit in part . ()} = exp {c (C . w(e.([Y] .w(9. C) = B = Yc = de = do = . () = Cc .21og 2.
Notes and references The general area of sensitivity analysis (gradient estimation) is currently receiving considerable interest in queueing theory.8 The specific form of (9. the exponent of the density in an exponential family has the form 01 tl (x) + • • • + 9ktk (x). [379] consider a special problem related to reinsurance. 10 Estimation of the adjustment coefficient We consider a nonparametric setup where /3.. in which case we can just let t(x) = 0. then ryT < 0. Thus. (9.12) takes the form a = a 93 ar.+UNT) > 1. the models there (e.a. .oo. Note that if NT = 0. to our knowledge. and hence explicit or asymptotic estimates are in general not possible. the exponent is either Ox. sj=1 and let YT be defined by IKT('ryT) = 0. Also. In general. Thus. That it is no restriction to assume k < 2 follows since if k > 2. Van Wouve et al. if 1 PT = /3TNT(U1+.2) is motivated as follows. However . Finally if k = 1. However. queueing networks) are typically much more complicated than the one considered here. the results presented here are new. 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). we have assumed k = 2 and ti (x) = x.3C2de 1p' z a = c . kT (a) = /T (BT [a] . then BT and hence ryT is undefined. and we estimate y by means of the empirical solution ryT to the Lundberg equation. let NT 16T = ^T . the main tool is simulation.cue_7u)3C P Remark 9.1) . for which we refer to X. Comparatively less work seems to have been done in risk theory.g. by the LLN both F (NT = 0) and F (PT > 1) converge to 0 as T . thus. we can just fix k . or Ct(x).2 of the parameters.. BT [a]= NT ^` e"U. To this end.. B are assumed to be completely unknown. ae t 1lEY u S _ . ESTIMATION OF THE ADJUSTMENT COEFFICIENT Finally. in u which case the extension just described applies.7 and references there.10.
V2 are independent N (0.1) 'YT . 7T a4' 7.2) follows from NT/T a4' .B[7]) + B [7] . N ( n[7].b[Yp'V21 T { (E[7] . THE COMPOUND POISSON MODEL Theorem 10 . For the proof.: N ()3.)vl+ N CO. B[27] . 16T where V1. 1) r.2) rT(7) N N (0.2 As T * oo.3T .T y . Lemma 10 .. B [7]2 (10.1 As T 4 oo.(27)/K'(7)2. vfoVFB[2y].B[7]2 V2 . it is easy to see that we can write \ V1 1 l _ . Hence KT(7) = (F' + (OT a(B[7l 0))((BT [7] . (10.1) . since NT /T . a2 where a2 = /3r. . If furthermore B[27] < oo.'Y . then (10.3/ T).1)2 + E[27] .i3)(B[7] 1) + (3(BT[7]  .B[7]2 }) ( T 0 .94 CHAPTER III.If . we need a lemma.7 + (.1) .B[7]2 n Hence ( 10.Q and Anscombe 's theorem.a BT[7] I B[7] I + .B[7]) 0+ Iv/o(b[y].: N 0.'s.3) Proof Since Var(eryU) = we have B[7]. B[2'Y]  /3T ) . More generally./^ B[27] .v.
Proof of Theorem 10. 7T E (y .'(y). °7IT) . NT i =1 n'(a) for all a so that for all sufficiently large T K7 . Now write KT(7T)  kT(7) = 4T(7T)( 7T 7). Let 0 < E < ry. OT a 95 u 4 /3.1 By the law of large numbers.KT(7) kT(7) K'(7) . first note that e7TU N (e7U u2e27Uo'2/T) 7 . If ryT E (7  we have KT(7 .10. y + E) eventually. lcT(a ) 4 /c(a).e) < 0 < r.Q. Combining ( 10.3)..e. 0 are estimated from data . NT BT [a] Hence r. To this end . By the law of large numbers.E) < 4T(7T) < 4T(7 + E). Then r.e.(ry . and the truth of this for all e > 0 implies ryT at 'y. where ryT is some point between ryT and ry. BT[a] 3 B[a].'T(a) = 1 E Uie°U' a$' EUe "u = B'[a].4) and Lemma 10.2. it follows that 7T7 KT(7T) .c'(7) N (0' T (2(7) / N (0.E ) < 4T(7T) < (7 +0' which implies 'T(ry4) a$' r.4) + E). 6"Y (10. I. Theorem 10. ESTIMATION OF THE ADJUSTMENT COEFFICIENT which is the same as (10.E) < 0 < kT(7 + E) for all sufficiently large T .(ry + e) and hence KT(7 .1 can be used to obtain error bounds on the ruin probabilities when the parameters .
ft.d. > 0 for some t E [Wn_ 1.3 or equivalently p > 1/2 or 11 < 100%. wn = inf{t > W. V. 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. the nth busy cycle is then [Wn1.. Csorgo & Teugels [95].Wn) are i . Deheuvels & Steinebach [102]. Herkenrath [192]. Vt = St . .96 CHAPTER III. For this reason .. if B is exponential with rate 8 so that ry = 8 ...Q.C1e"a ( see e.e. t]}.96 if a = 2.. Embrechts & Mikosch [133]. This approach in fact applies also for many models more general than the compound Poisson one. satisfies b(.0) < 5.i.T = 3TKT ( 21T)IKT (^T)2 is the empirical estimate of vy and fc.f. For example . Further work on estimation of y with different methods can be found in Csorgo & Steinebach [94]. Notes and references Theorem 10.1 : Vt = 0. Hipp [196].e.g.ueryuU ".5%).T VIT where r7ry. various alternatives have been developed. Griibel & Embrechts [292].g. Asmussen [23]) can then be used to produce an estimate of ry. Wn). [197]. with a tail of the form P(Y > y) . 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. Letting Wo = 0. U2.info< „< t S. and the known fact that the Y„ = max Vt tE[W„1.... 6 < 2. i . Mammitzsch [253] and Pitts. = 1.) = a (e. i. it means 2 (8 .1 is from Grandell [170]. Frees [146]. A major restriction of the approach is the condition B[2ry] < oo which may be quite restrictive.
the premium rate is 1. Further let 'Yo be the unique point in (0. In particular. it is assumed that i > 0 and that the adjustment coefficient (Lundberg exponent) y. See Fig.1 (the role of ryy will be explained in Section 4b). The notation is essentially as in Chapter III. B[•] and mean AB. 'y) where c(a) attains it minimum value. Unless otherwise stated. generalizations to other models are either discussed in the Notes and References or in relevant chapters.1 where p = 13µB.s. defined as solution of c(ry) = 0 where ic(s) _ /3(B[s] . The safety loading is q = 1/p .f.g. T) = P( /r(u) <T) \ = PI inf Rt <OIRo=u1 /\0<t<T PI sup St>ul 0<t<T Only the compound Poisson case is treated. the Poisson intensity is 0 and the claim size distribution is B with m. 0.1) . 97 . exists.Chapter IV The probability of ruin within finite time This chapter is concerned with the finite time ruin probabilities 0(u.
we have for k = 1.(.1) (1.) = e7u ELe'Y^(u) ELT(U)k = e'Yu b ELT(u)k = O(u)ELT(u)k. Var[T(u) I T(u) < 00] = VarL T( U) . In particular. PL = 6/0 = 1/p > 1).9).(4.. EL refer to the exponentially tilted process 3 with arrival intensity S and exponential claims with rate / (thus .1 In the compound Poisson model with exponential claims with rate S and safety loading 77 > 0.r. 1 FL. (u) is exponential with rate 0 w. 1 Exponential claims Proposition 1.98 CHAPTER IV.1 The claims surplus is {St}. By the likelihood identity III. 7.5 .t. PROBABILITY OF RUIN IN FINITE TIME Figure 0.u is the overshoot.2) Proof Let as in Example 111. 2 that E [T(u)k. the conditional mean and variance of the time to ruin are given by E[r(u) I T (u) < oo] Var [T ( u) I T( u) < oo] /3u+1 J )3 _ 2/3Su+/3+S (S)3)3 (1. . E[T(u) I T(u) < 00 ] = ELT (U).(U) < 00] = ELT(u)ke'YS. using that the overshoot l. FL and independent of T(u). the time of ruin is T(u) and ^(u) = ST(t&) .
s.12 Thus the l.h. This means that /3(6/(6 .1)) ELST(u) ELT(u) (PL . is V1rLSr( u) +VarL ((PL .1. which leads to the quadratic 02 + (/3 .s.s. of (1.0) .1)T(u)) = VarLe(u) + (PL .6 + a)0 .I (1.2 In the compound Poisson model with exponential claims with rate 6 and safety loading rl > 0./3 .1)ELT(u). Wald's second moment identity yields 2 EL (Sr(u) .h.h.1) .1)T(u) are independent with QL the same mean . Since Sr (u) and (PL .2). u + ELe(u) _ PL . T(u) < oo] fora > r. the Laplace transform of the time to ruin is given by Eea7( u) = E [eaT (u).2) is aLELT( u) .1 /3u + 1 u + 1 //3 = 6/3 6/01 For (1.(PL . 1).6a = 0 with solution 0 (the .6.1)2VarLT(u) + 2 Ca 1I VarLT(u). . 0 Proposition 1./3) . Let 0 > yo be determined by ^c(0 ) = a.1 (6)3)2 which is the same as the r. the 1. where = eBu I 1 .1//32 (6/)3 1)2 26(/3u + 1)/(6 ."(ry) = 26//32. we have by Wald's identity that (note that ELSt = t(pL .1)T(u))2 = UL where = s.V/ is as asserted.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 . EXPONENTIAL CLAIMS For (1 .(yo) = 2V .B = a.
and M(u)+1 is the index of the ladder segment corresponding to T(u).'s with rate 5.1 where Y1.3) we have E [e«T(u ). PROBABILITY OF RUIN IN FINITE TIME sign of the square root is + because 0 > 0). T(u) < oo] = e.100 CHAPTER IV.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.Y1 Y2 Figure 1. Y2. But by the fundamental likelihood ratio identity ( Theorem 111.v.. are the ladder heights which form a terminating sequence of exponential r.9ST(u) +T(u)!c(0)} . T(u) < oo] = EB [exp {aT(u ) ..3 that we can write EeaT( u) = eeuEe 017(o). More precisely.. Using 5 = 6 . (1.0. T2 . Y(u) belonging to a convolution semigroup . M(u) T(u) = T + E Tk k=1 where T = T(0) is the length of the first ladder segment . St Ti F. Fig.1 . Ti.4.v. .. . Note that it follows from Proposition 1. the result follows...3. Cf..T+ Ti a t U T I 1 a i F.. 1. are the lengths of of the ladder segments 2.4) The interpretation of this that T(u) can be written as the independent sum of T(0) plus a r.
T) 1 I fl(O)h(0) fdO where (1.3 Assume that claims are exponential with rate b = 1. where U1.T the service times of the customers awaiting service . .6.3 sin0 + 29) f3(0) = 1+/32/cos9. Hence 00 F(VT > u ) P(QT = N)P(EN > u) N=1 00 N1 k F(QT = N) eu N=1 k=1 °O u k! k Ee k=0 1t P(QT . Let {Qt} be the queue length process of the queue (number in system..1(u. UN. .. Proof We use the formula .ST). . For j = 0. cf. including the customer being currently served). EN has an Erlang distribution with parameters (N. and exponential with rate S = 1.0.. Corollary 11. T) to be evaluated by numerical integration: Proposition 1. Note that the case 6 # 1 is easily reduced to the case S = 1 via the formula V. density xN lex/(N . i. the conditional distribution of VT given QT = N is that of EN where the r. T. [4]) 00 (x/2)2n+3 Ij (x) OnI(n+j)! .v. 2.4.e.cos (u/.T are conditionally i.i. EXPONENTIAL CLAIMS 101 For numerical purposes . U2..T is the residual service time of the customer being currently served and U2 .T. .1.i (u. the following formula is convenient by allowing t..6(u) = Vfl/j l(Su. UN. If QT = N > 0.6) fl(9) f2(0) = = fexp {2iTcos9(1+/3)T+u(/cos91) cos (uisin9) .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.d.T. Since U1 .1.T.I ex cos B cos j O dB fo " ..1 )!.. . 1).k + 1). Then V(u. let (cf. then VT = U1.T + • • • + UN.
cos((k + 2)9)] d9. (1. similar formulas are in [APQ] pp.1)] L _112 /(k+1)/2 [.102 CHAPTER IV.13(k +l)/2ei(k +1)9 R E . and define tj = e(1+R)Taj/2Ij(2vT T). f3(0) .31 /2eie L 1)] 1 I/31/2eie . Then (see Prabhu [294] pp. k k2 + $k+1 E bj 00 t j .)3k+1 = e(1+0)T e201/2Tcos 7r 0 e )3(k +l)/2 [31/ 2 cos ( kO) .44). in particular equations (1.8 ) yields F(QT > k + 1) .cos (( k + 1)0)] f3(9) Hence the integral expression in (1.1 00 ok+lR 00 j=k1 +1)/2e . PROBABILITY OF RUIN IN FINITE TIME denote the modified Bessel function of order j. 8789) 00 E aj j= 00 = 1. 912.cos((k + 1)0)] f3(0) 00 flk +1 > j=k1 3j/2 COS(jB) l)/2ei(k+1)e )3j/2eije = R)3(k+ (31 /2eie .ie . let I _ j (x) = Ij (x).i(k +1)e R [/3( klal/2e:0 (01 /2 e .1 R [.(31/2eie . 00 E '3j/2 cos(je) j=k+1 00 _ j=k+1 ^j/ zeij = .(31/2 cos (( k + 2)9) .112 l 1( k +1)/2 [ 31/ 2 cos(kO) .)3k +1 tj g'(QT >.k + 1) = 1 k +1 + bj j=00 j=00 00 j=kk+1 j=k1 By Euler 's formulas.38).3(k +1)/ 2ei(k + l)6 (.
Ui < x I / (note that P(St < x ) = F(x + t. We first prove two classical formulas which are remarkable by showing that the ruin probabilities can be reconstructed from the distributions of the St.7) that _ [^ au ak+l (30 k L. . there are several misprints in the formula there. expresses V)(0. equivalently. oo (u)31/2e^e)k = )3k z cos(k9) = R k. or. T). going back to Cramer.. T) in terms of F(.T) which. 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. 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. The rest of the proof is easy algebra. 2 The ruin probability with no initial reserve In this section . Related formulas are in Takacs [359]. however. k! k=O k0 i/z Co Uk ate" o'/z e . 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. t )). from the accumulated claim distribution N. is numerically unstable for large T. E Fk. F(x. Seal [327] gives a different numerical integration fomula for 1 . however.3 was given in Asmussen [12] (as pointed out by BarndorffNielsen & Schmidli [59]. u Notes and references Proposition 1. it follows as in (1. we are concerned with describing the distribution of the ruin time T(0) in the case where the initial reserve is u = 0. the numerical examples in [12] are correct).0(u. t) = P .e = e' COS a cos(uf31/2 sin 0). The first formula.
T T o where the second equality follows from II. and the third from the obvious fact (exchangeability properties of the Poisson process) that has the same distribution as St = { Si0)} so that P(M(v. Proof For any v E [0.T)dx.1 In formulas.b (0.0<w<t} St+v .1. T]. co ). [v. ") } is at a minimum at time t.T]. .(6.104 CHAPTER IV. resp.T)) does not {Stv)} depend on v.t)= {Stv) < SM.T))dv E^T I(M(v. v].T) T F(x. Stv^ _ Define M(v. 1 1 . PROBABILITY OF RUIN IN FINITE TIME Theorem 2 .S„ 0 <t<Tv STS„+St_T+v Tv<t<T as the event that IS. Then 1 . 2. we define a new claim surplus process St StM NJ Figure 2.(0.(. T) = P(Tr(0) > T) = P(M(0.T))dv.i. f T lStv)} 0<t<T by a 'cyclic translation'. meaning that we interchange the two segments of the arrival process of {St}o<t<_T corresponding to the intervals [0.T)) 1 fT P(M(v. See Fig.3) with A = (0.
then M(v. in which case there is a last time o where St downcrosses level u. 0<t<v} = {ST < St . Hence T TE f I( M(v. 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. It is then clear from the cyclical nature of the problem that this holds irrespective of whether M(0. Obviously. there exist v such that M(v. t) denote the density of F(•. we can write M(v. v). Indeed. T)).. 0<t <Tv}n{ST<ST Sv+St T+v. cf. . t). then i fT I(M(v.. T) occurs.2. where the last equality follows from ST < St on M(0.T) occurs. T) = M(0. this integral is 0 if STv) . T)) dv = TEST = T fP(ST < x) dx T T NT 1 f P(ST < x) dx = 1 f P Ui T .xdx.ST on M (0. v). w) for some small E. For example. or it occurs. THE RUIN PROBABILITY WITH NO INITIAL RESERVE 105 Now consider the evaluation of fo I(M(v. letting w = inf It > 0 : St_ = mino<w<T Sw}.v<t<T} = {ST<StSv. v).T) = F(u+T.t)dt.2.T) occurs or not as long as ST < 0.T)) dv f T I(M(0. T. v < t < T} n M(0. T T o i =1 Let f (•. It follows that if M(0 . v<t<T}n{ST<STSv+St. If ST < 0. T].T) and Sv < 0 on M(0. Fig 2. T Theorem 2 . T)) dv.2 10(u. We claim that if M(0. ST > 0. we can take v E (w E. T) as {ST<St+ vS.Sv.Tt))f(u+t. v) = M(0.T)f(I z /)(0. v)) dv = ST T T o (note that the Lebesgue measure of the v for which {St} is at a minimum at v is exactly .
which is independent of St and has the stationary excess distribution B0. ST_ _ z} .T) = . u which is the same as the assertion of the theorem. C*(z.u+dt]). 2. Hence P(ST<u) = 1 . 0 < t < T. 0 < t <T.z. 0 < t <T . For a fixed T > 0. Then P(T(0) E • I T(0) < oo) = P(T_ (Z) E •). z > 0. ST_ _ z}.2. Proof of Theorem 111.T) = C(z.T) = {St < 0.2 Here o. t + dt] occurs if and only if St E [u. E [t. which occurs w. The following representation of T(0) will be used in the next section.2. Let Z be a r. Proposition 2.106 CHAPTER IV.3 Define r_ (z) = inf It > 0 : St = z}.t). u + dt] and there is no upcrossing of level u after time t.v. PROBABILITY OF RUIN IN FINITE TIME u Q II T Figure 2. ST_ _ z}.b(u.(0.ST_ t_ and let A(z.T)+ J0 T (1V.Tt))P(StE[u. define St = ST .2 . {S t > z. {St > .p. O(T . The proof is combined with the proof of Theorem 111.
{St }o<t<T have the same distribution .T). THE RUIN PROBABILITY WITH NO INITIAL RESERVE Then 107 P(r(0) E [T. . T(0) < oo) B(y B(z) + z) f3B(z) dz = 3 f °^ B(y + z) dz = f3 + x v f B(z) dz. and since {St}o<t<T. z + dz].T))f3B(z) dz dT.3 But by sample path inspection (cf. z + dz]. 2. T(0) < oo) = OR(z) dz in (2. u which is the assertion of Theorem 111. z + dz]. Fig. z + dz]) = P(A(z. Thus P(Sr(o)_>x.1) that P(T(0) E [T.2. 7( 0) < oo) = P (C(z)) dT. z + dz].T)) = P(Cx.T) = C*(z. r(0) < oo) = 3R(z) dz JP(C(z. (2. Proof of Proposition 2.1) yields P(ST(o)_ E [z.T))dT = Off(z) dz P(T_ (z ) < oo) = 3B(z) dz.T)). we therefore have P(A(z.ST(o) >y.T + dT]. T + dT] I S7(o)_ E [z. ST(o)_ E [z. It follows by division by P(ST(o)_ E [z.2. A(z.T(0)<oc) = f x F(U > y + z U > z) P(Sr(o)_ E [z.3).2. Figure 2.3.1) z T . Hence integrating (2.
1) .(yo). Tak'ecs [359]. z + dz].6. T(0) < oo) = dTP(T_(Z) E [T.108 Hence CHAPTER IV.5a).5 and one upon excursion theory for Markov processes (see IX. PROBABILITY OF RUIN IN FINITE TIME ]P(7(0) E [T. r(0) < oo. Theorem 2.3. ^(0) E dx] (recall that ^(0) = Sr(o)) and write ga[b] = f OD ebxga(x) dx. Let T_ (y) be defined as Proposition 2.3 was noted by Asmussen & Kl(ippelberg [36]. (3.T + dT] T(0) < oo) dT f ' P(C(z))P(Sr( o)_ E [z. some relevant references are Shtatland [338] and Gusak & Korolyuk [181]. Notes and references For Theorems 2.r(a).1. 3 Laplace transforms Throughout in this section.2. z + dz]. In the setting of general Levy processes. a martingale proof is in Delbaen & Haezendonck [103].1 and the present proof is in the spirit of Ballot theorems. one based upon a result of Asmussen & Schmidt [49] generalizing Theorem 11.s.1 Eear( y) = eyr(a). cf. see in addition to Prabhu [293] also Seal [326]. T(0) < oo) 0 = dT f 0 P(C(z))P(Z E [z.c(r(a)) l = l er( a)se+at } u yields 1 = eyr(a)Eear(y). I L Let ga(x) be the density of the measure E[ear(°).(3(B[r( a)] . Proof Optional stopping of the martingale I er (a) 9 t. 2.T+dT]). Lemma 3 . Note that T_ (y) < oo a. Proposition 2.1) where a > r. r(a) denotes the solution < 'Yo of the equation a = ic(r (a)) = .2 ga(x) = Qexr(a) f "o eyr(a)B(dy) x . who instead of the present direct proof gave two arguments. Lemma 3. because of77>0. [329].
x)(a) B(dy)• Lemma 3 . £(0) E dx) = /3B(x + dy) dx and hence ga(x) = f e r)/3B(x + dy) _ /3 f x e(v. Corollary 3. time T(u): u u Here is a classical result : the double m.3.ga [b] 0 TO Using Lemma 3. E[ear (o) I T(0) < oo .3. the result follows after simple algebra.r(a) b .4 E[eaT (o). y + dy].T(0) < oo] = 20[b] = za[b] (9a[b] 9a[0])/b 1 . Hence eb"du E[eaT(").ic(b)/b x(b) + a eb"E[eaT(" ).ST(o)_ just before ruin . (u . (Laplace transform) of the ruin Corollary 3. T(u) < oo] du = Proof Define Za(u) = E [eaT(" ).(v) = ev''(a).x)ga (x) dx where za(u) = f.1] evr(a)B(dy)[ b .g. u . LAPLACE TRANSFORMS 109 Proof Let Z be the surplus .3.f.°° ga(x)dx.2.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)) .r(a) The result now follows by inserting /3B[s] = ic( s) +/3+ s and ic(r(a)) =a.r(a) = a [B[b] B[r(a)]] . Further by Theorem 111. r(u) < oo). It is then easily seen that Za(u) is the solution of the renewal equation Za (u) = za (u) + fo Z.ga [b] 1 . Z = y] = EeaT.2 P(Z E [y. rr(0) < oo) = 1_ r(a) Proof Let b = 0. b . Then by Proposition 2.5 f 00 o a/r(a) .3 ga[b] = c(b) Proof + b + a .
The first main result of the present section is that the value umL.3). That is.h(u. (u) t. t T(u) T(u) T(u) t m = lim = lim = lim Utioo u + Sr(u) u+oo S.110 CHAPTER IV.1. for any c > 0 P( Further. P = /3µB > 1. By Proposition 111. and take basically the form of approximations and inequalities. Theorem 4 .s. mu ) ( 0 m < ML '(u) 1 m > rL.1) i.1 Assume 77 > 0.1)Er(u) .2.. Then given r(u) < 00. Then as u * oo. u 1 ET(u) 1 p1 u where Pw2 = 311B)m3• 7(u) . note that by Wald's identity u + EC(u) = ES.2 Assume ri < 0.s.mL > E T(u) < 00 ) 40. St/t 1 1/m. PROBABILITY OF RUIN IN FINITE TIME 4 When does ruin occur? For the general compound Poisson model.3LELU 1 1p' is in some appropriate sense critical as the most 'likely' time of ruin (here C is the CramerLundberg constant). This proves the first assertion of (4.00 St = lim . and hence a.r(u) = Er(u) • ES. T(u)/u mL as u + oo.(u ) = o(u) a. For the proof. Proposition A1.UProof The assumption 11 < 0 ensures that P(T(u) < oo) = 1 and r(u) a4' oo..e. For the second . for any m T(u) u . Later results then deal with more precise and refined versions of this statement. i. the known results are even less explicit than for the exponential claims case. (4.w ) v/.mu D 2 4 N(0.6. = (p . we need the following auxiliary result: Proposition 4. uoo u using e. T(u) a. where _ 1 _ 1 1 C ML w(ry) 6B'[7J 1 . cf.
1). Thus.1 is standard.3. Notes and references Theorem 4.s.1).mL >E By Proposition 4. cf.mu (2) '• m3/2 µB 7 .1. According to Anscombe' s theorem (e. of (4.h. For (4 .3). the result comes out not only by the present direct proof but also from any of the results in the following subsections. and (4. apB ) .N(0.6. If Z . 4a Segerdahl's normal approximation We shall now prove a classical result due to Segerdahl. the same conclusion holds with t replaced by r(u).1) is T (u)  U mL P( T (u) < I > E.t/m D (2) 111 . T) for T which are close to the critical value umL).2. note first that ( Proposition 111.6µB2) Z v m (3µB2) Z.mL U > E.4. this can be rewritten as u + 1(u) . again Proposition A1.g.^ N (o. 4).1 The l.7 6  11 Proof of Theorem 4. Theorem 7. T(u) < oo f / 00) e7uE L [e_7 (t1). and as a timedependent version of the CramerLundberg approximation.mu m . 1'r(U) .2 of [86]) and (4. T (u) < 00 J 0(u) e7'PL U \ I T u) . though it is not easy to attribute priority to any particular author.5) St .r(u)/m T(u) ti µB2) Z. which may be viewed both as a refinement of Theorem 4. WHEN DOES RUIN OCCUR? and that Ee(u)/u a 0. PL (•)+ 0. .1). Tu) T( u) .2) follows immediately from u (4. implying T(u) . proving (4.1 (by considering 0(u.
we can replace T(u) by r(u').6). (oo. S( u ) < ul/4] < ET(ul / 4) = O(ul/4). Hence Ef (Vu )) 9 (T(u.a C4'(y )• ( 4. and thus in (4. e'°'/b (u. using that ul/4 .4).VU T.f ( (oo)) . oo ).mul h(oo)Eg(Z).(u. Then h(u) 4 h(oo) = E f (6(oo)).ul/4.w2) r.ST( u') = u1/4 .T ( u')] = E[ T ( ul /4 . one has 9 (r(u)_rnu) Ef (^(u)) * E.5) For the proof.e(u') oo w . E9(Z) (4. we need the following auxiliary result: Proposition 4. oo).r. resp . Let h(u) = E f (^(u)). Then for any y. g are continuous and bounded on [0.))I h(ul /4  ^(u)) I(6 (u') C ) f < ul /4 + f(e(u') .um. Proof Define u' = u .4 (SIAM'S LEMMA) If 71 < 0. PROBABILITY OF RUIN IN FINITE TIME Corollary 4. with w2 as in (4.t. P because of ^(u') .6) whenever f. Then the distribution of T(u) .^(T(u')). Using ( 4.112 CHAPTER IV.v.u1/4)I(S(u') > u1 /4) h(oo) + 0.3 (SEGERDAHL [333]) Let C be the CramerLundberg constant and define wL = f3LELU2mL = f3B"[ry]mL where ML = 1/(pL1) = 1/($B'[ry]1). we get E[ T (u) . O .) is readily seen to be degenerate at zero if ST(u•) > u and otherwise that of T(v) with v = u . letting Z be a N(0.T(u') given F.l:(oo) (recall that rt < 0).L+YWLV'U) .)mu \ h(oo)Eg (r(ul) . then e(u) and r(u) are asymptotically independent in the sense that.3). and similarly as above we get E[f(^(u)) I Fr(u.
in practice one would trust (4. The present proof is basically that of Siegmund [342]. Cf. 3 is due to Segerdahl [333].4. 4b Gerber's time.3 in terms of Edgeworth expansions .5) and solve for y = y(T).T) Ce7"4 (T . see also von Bahr [55 ] and Gut [182]. that for the fit of (4. where we used Stain's lemma in the third step and (4.z/)(u .7) to be good.7) To arrive at this . y > k'(7) . umL + ywL f) = e"P(T (u) < umL + ywL) = EL [e7V "). see Asmussen [12] and Malinovskii [254]. PL(T(u ) < umL + ywL) 113 4 C4(y). For refinements of Corollary 4. however .(ay) = 17 7y = ay .1.7) whenever u is large and ly(T)l moderate or small (numerical evidence presented in [12 ] indicates .8) Note that ay > 7o and that 7y > •y (unless for the critical value y = 1/ML). Notes and references Corollary 4 . u needs to be very large).3 ery"z/i(u . also Hoglund [204]. Thus .7) to be valid is that T varies with u in such a way that y(T) has a limit in (. oo ) as u * oo. y u) < . Segerdahl 's result suggests the approximation b(u. T(u) < umL + ywL f.oo. e7v" y < ^'(7) (4 . Theorem 4.4) in the last.9) ( 4 . define ay. y u) < e 7v" . For practical purposes .dependent version of Lundberg's inequality For y > 0. 0. 10) '5(u) . CL Fig.5 '(u . just substitute T = umL + ywL in (4. .yK(ay)• (4. ELe7E (") . The precise condition for (4. WHEN DOES RUIN OCCUR? Proof of Corollary 4.umL wI V"U u (4. yy by 1 K.
1). For a different proof.114 CHAPTER IV. we arrive at the expression in (4. who used a martingale argument. Then ic(ay) > 0 (see Fig . if y > 1/ic'(y). Hoglund [203] treats the renewal case. yu < T (u) < oo 1 l e ayuEav [eT ( u)K(ay).ay4(u)+ T(u)K(ay ). yu ) = < eayuEay [eay^ ( u)+T(U)K ( ay). 0. yu) < C+(ay)e7a„ where l C+(ay) = sup f 00 eayR(xy)B( .3 yields easily the following sharpening of (4.6. which shows that the correct rate of decay of tp(u. T(u) < yu] < eayu + yUr(ay) Y < eayuEav [ eT(u)K(av )L T(u) < yu} Similarly. yu 11 < T(u) < oo j < eayu +Y UK(ay) Remark 4. f Some urther discussion is given in XI. 5 is due to Gerber [156 ]. An easy combination with the proof of Theorem 111. see MartinLM [257] . 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. Numerical comparisons are in Grandell [172 ]. the bound a7y° turns out to be rather crude .8 below . However. we have rc(ay) < 0 and get (u) .8).8).h(u. and generalizations to more general models are given in Chapter VI. u Differentiating w.2.t.7 i. From the proof it is seen that this amounts to that a should maximize ayic(a).yu ) = eayuEav [e . a. In view of Theorem 4. yu) is e 'Yyu/ .6 It may appear that the proof uses considerably less information on ay than is inherent in the definition (4.Y' (u.r.9): Proposition 4.5. and hence t. .v"U. which may be understood from Theorem 4. dy) Notes and references Theorem 4 . PROBABILITY OF RUIN IN FINITE TIME Proof Consider first the case y < 1/K'(y).b (u.
(4.2 yields EaT(u) u u r. then the solution &y < ay of .ayC() . WHEN DOES RUIN OCCUR? 115 4c Arfwedson's saddlepoint approximation Our next objective is to strengthen the timedependent Lundberg inequalities to approximations. u 4 oo. (0) r1 (a) ' I.. yu ) eaauEaye .: T. Using Lemma 111.. (4. we have ryas = ay . the choice of ay.. and in case of ruin probabilities the approach leads to the following result: Theorem 4 . T(u) < yu] . Proposition 4.ay and get Ea e ayf (00) y _ 'Ya( ayKal lay C 1 . For any a > yo.11) ' If y > 1/ r . yu ) ayay e ryyu ayay 27ry/3B"[ay] u Proof In view of Stam 's lemma.^3 ]1/ Bay [lay .'(y ).c(&) = ic(ay) is < 0.e. it is instructive to reinspect the choice of the change of measure in the proof.ay a.8 If y < 1/ic'(ry). This idea is precisely what characterizes the saddlepoint method. then ay > 0.13) . and ii(u) . We thereby obtain that T is 'in the center' of the Padistribution of T(u). yu) = e.(u.ayuEay f eay^ ( u)+T(u)K(ay). and b(u.ay y 'Yay  ay .e. (4. then the relevant choice is precisely a = ay where y = T/u.yu) c ay . T(u) suggests heuristically that l t/. i.yyu y l ay I 21ry/3B" [ay] V fU_ u + 00. the formula 0(u. not inequalities. The traditional application of the saddlepoint method is to derive approximations.z. [eT(u )K( ay).5.4. if we want EaT(u) .i(u.6 with P replaced by Pay and FL by Pay.12) < yu] Here the first expectation can be estimated similarly as in the proof of the CramerLundberg ' s approximation in Chapter III. As a motivation. Ea .
4).a)2 .(ay) _ y(ay .ay) ay +. T(U) < yu] = eyuk (ay)E''ay (ek(ay )"1/2WV.c(ay)ul/2W p 2ir = eyu(ay) dz 1 rc(ay ) 2.c'(a) _ /3a/(8 .13).11) follows.ay)K(ay) ay ayI&YI For the second term in (4. Example 4. (4. a nr=. Writing r(u) and W2 = I3ay{.(j (1 .1) .1B[ay]1 ) y(ay . where V is normal(0.116 CHAPTER IV. The proof of (4. .1)3 = (jB"[ay]l (Pay . (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 . V < 01 Ir 00 er(ay)"1'2"'x eyur.7ruw2 Inserting these estimates in (4.ay ) r. 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).12) is 0 entirely similar. PROBABILITY OF RUIN IN FINITE TIME ry I i .1) under Pay mation (4. The difficulties in making the proof precise is in part to show (4.l'B)y /(Pay .B[ay] /ay &y y(ay .13).9 Assume that B(x) = eay. we get heuristically that Eay Ler (u)r(ay).a.1)3 = y3/3B"[ay].13) rigorously.ay + ayl /BLay] .a) .I ay &y a ^c'(ay) a (1 +.3(5/(S . and the equation ic'(a) = 1/y is easily seen to have . i B[7ay . it seems tempting to apply the normal approxiyu + ul/2wV.1. Then ic(a) = .
p..5. c a 00.tcp) Lo {Wo ( t)}t>0 . y) a''y" L '3 _ fl ) 51 /4(1 +1IY)3/4 \.3+52 1+/351/y' sy 7 B ii[ay] 25 _ 251/2(1 + y)3/2 (5 ./4 ^y for 1/i (u.8 is from Arfwedson [9].= (s. A related result appears in BarndorffNielsen & Schmidli [59].ay)3 0 3/2 and (4.11) gives the expression '31/4 ( .i )( v s vc ('3 + s _2 / . 0 Notes and references Theorem 4.because the c. 2 = Var(Si ) the variance. in discrete time: if p = ES. is undefined for a > 5). then { __ . DIFFUSION APPROXIMATIONS solution ay=5 117 V 1 (the sign of the square root is .1. .. (5.g. The mathematical result behind is Donsker's theorem for a simple random walk {Sn}n=o.. is the drift and o.1) . It follows that 5^y =5ay = /«y =f3+ay=l3+d 1+1/y' V 1+^1/y /35 1+1/y /3' ay ay =Qay say =. 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. yu) when y < 1/ic'('y) = p/1 .f.
n/c < t < (n + 1)/c.tcpp) y = { WC (Sct) pct) } {Wo( t)}t>o (5.3.t} _ {W_1(t)} .7c). It is fairly straightforward to translate Donsker's theorem into a parallel statement for continuous time random walks (Levy processes).3) takes the form LI S(P) { a2 to2/µ2 + t LI S (P) { a2 ta2/µ2 {W0(t)}. St = EN` U= . + {Wo(t ) .. for the purpose of approximating ruin probabilities the centering around the mean (the tcp term in (5.1) with S. PROBABILITY OF RUIN IN FINITE TIME where {W( (t)} is Brownian motion with drift S and variance (diffusion constant) 1 (here 2 refers to weak convergence in D = D[0. Lemma 111. cf. a2 =/3µB2) Proof The first step is to note that { WC (St P) . where p is the critical premium rate APBTheorem 5 . oo)). such that the claim size distribution B and the Poisson rate a are the same for all p (i.118 CHAPTER IV.z } {W_1(t )}t>o (5. p.1)) is inconvenient.3) whenever c = cp f oo as p 1 p. we shall represent this assumption on 77 by a family {StP) L of claim surplus processes indexed by the premium rate p.1.p/c < St(p) < S((n+l)/ c + Pp/c. 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. Indeed . Letting c = a2/pp..1 As p J.1 below). we have o {i!t s: . Mathematically. this is an easy consequence of (5.p. and consider the limit p j p. (5.a = Snp) and the inequalities Sn )C .e.tp). However. We want an approximation of the claim surplus process itself. and this can be obtained under the assumption that the safety loading rt is small and positive.2) t>o where p = pp = p . of which a particular case is the claim surplus process (see the proof of Theorem 5. 0 .
For practical purposes . is 1/ip (ua2 /IpI. DIFFUSION APPROXIMATIONS Now let Tp(u) = inf{t>0: S?)>u}. see Grandell [ 168]. u). 119 It is wellknown (Corollary XI. 263) that the distribution IG(•.h.6) This is the same as the heavy traffic approximation derived in III. since ti(u) has infinite horizon .8 or [APQ] p.t.2 As p j p.Ta2 /p2). 199.1.(u) ti IG(oo.5) Note that letting T * oo in ( 5. 1. (. Because of the direct argument in Chapter III. is IG(T. ('. and the r. [169] or [APQ] pp. 196. the continuous mapping theorem yields sup W Sz2 to lP 4 sup Wi(t)• O<t<T O<t<T a2 Since the r. TS(u)=inf{t>0: WW(t)>u}. (5.5.h. u) =PIT( (u) < x) = 1 .5).s.7c. w.. has a continuous distribution. 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. Corollary 5 .s. ulpI/a2)..h. we omit the details . the continuity argument above does not generalize immediately.1 .T) IG(Tp2/ a2).2 suggests the approximation u 0(u. .1 I 7= . any probability measure concentrated on the continuous functions. C.e. we obtain formally the approximation V.1.r. ulpl /a2) = e2"1µl / or2. and in fact some additional arguments are needed to justify (5.4) Note that IG(. u) is defective when < 0.u).6) from Theorem 5. (5. ^ p2 Proof Since f 4 SUP0<t<T f (t) is continuous on D a.s.. u) of r( (u) (often referred to as the inverse Gaussian distribution) is given by IG(x.f I \\\ J \ (5. Corollary 5. However. (ua2 To2 op \ IPI > IG ( T .( ^ I + e2( \ I .
All material of this section can be found in these references. . Michna & Weron [152] suggested an approximation by a stable Levy process rather than a Brownian motion.Pe. as an example of such a generalization we mention the paper [129] by Emanuel et al.3 Consider a family {Ste) } oc claim surplus processes indexed by a parameter 9. PROBABILITY OF RUIN IN FINITE TIME Checks of the numerical fits of (5.1 and Section VIII. and which is much more precise. See for example Billingsley [64].120 CHAPTER IV. 0) { 2 StQ2 /µ2 D { W_ i(t)}t>o t>o D 2 where p = pe = pe .6) therefore does not appear to of much practical relevance for the compound Poisson model. Theorem 5. the B9. In view of the excellent fit of the CramerLundberg approximation. However. we have ^A.6 of [APQ]. For claims with infinite variance. In contrast.00µB6 + 0.. the claim size distribution B9 and the premium rate p9 depends on 0.g. The proof is a straightforward combination of the proof of Theorem 5.1. a2 = ae = 00µa6 Notes and references Diffusion approximations of random walks via Donsker's theorem is a classical topic of probability theory. as 0 * 00 and that the U2 are uniformly integrable w. in Asmussen [12]. (5. e. in the next subsection we shall derive a refinement of (5. The first application in risk theory is Iglehart [207]. pt? 4 peo. We conclude this section by giving a more general triangular array version of Theorem 5. Furrer. Assume further that 039µB6 < pe. [169]. such that the Poisson rate Oe. Then as 0 _+ 90. for more general models it may be easier to generalize the diffusion approximation than the CramerLundberg approximation.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. B0 * Boo. Further relevant references in this direction are Furrer [151] and Boxma & Cohen [75].Po = 09µB6 .t.5) for the compound Poisson model which does not require much more computation. The picture which emerges is that the approximations are not terribly precise. on the premium rule involving interest. in particular for large u.r. and two further standard references in the area are Grandell [168]. However.5) and (5. the simplicity of (5. that 00 4090. pe .6) are presented.
9o T 0. let P9 refer to the risk process with parameters Q9 = QoB0[9] = QB[9 9o]. which we have seen to play an important role for example for the CramerLundberg approximation . PB('r(u ) < oo) < 1 for 9 < 0. 0(0) = 0.1 > 0.s and p = /3µB < 1. In terms of the given risk process with Poisson intensity . . In this setup.6. whereas there we let the given 3B. . risk process with safety loading 77 > 0 correspond to 9 = 0 . it is more convenient here to use some value 9o < 0 and let 9 = 0 correspond to n = 0 (zero drift). this means the following: 1. Since Brownian motion is skip free.ao (0) _ /c(s + 9 . and we are studying b(u.90] B(dx).90) and the given risk process corresponds to Poo where 90 = 'yo. P9(r (u) < oo) = 1 for 9 > 0.9(s) = Ico ( s + 9) .c(s) = . Determine yo > 0 by r.90) .1) .T) = Peo(r(u) < T) for 90 < 0.6. The setup is the exponential family of compound risk processes with parameters ( B9 constructed in III. 2. CORRECTED DIFFUSION APPROXIMATIONS 121 6 Corrected diffusion approximations The idea behind the simple diffusion approximation is to replace the risk process by a Brownian motion (by fitting the two first moments ) and use the Brownian first passage probabilities as approximation for the ruin probabilities. B9(dx) =Bale] Bo(dx) e9z keo)z = B[9 ./c(9 . Then EOU' = Boki[0] = Biki[eo]/E[9o] and "(s) = k(sBo)k(9o). 77 is close to zero. However . this idea ignores (among other things) the presence of the overshoot e(u).4. Bo(dx) = B[eo]B(dx). Let PO refer to the risk process with parameters e9oz Qo = QB[90]. For each 9.Q (B[s] . and we want to consider the limit 77 10 corresponding to Oo f 0. 77 = 1/p . Then r. this is because in the regime of the diffusion approximation .'(yo) = 0 and let 90 = 'Yo. 3. claim size distribution B . The objective of the corrected diffusion approximation is to take this and other deficits into consideration.
means up to o(u1) terms): .1) IG(x.2) .. Vargo S. C.7(u)/u2} eh(A.e. The corrected diffusion approximation to be derived is (u. (U. IGu+u2. the solution of r. One has (6.Varo S1 = f30Eo U2 = S1. bl IG(t81. _ ^(u) = ST . C) = 2A + (2 .S. tu2 ) i IG (t. write r = T(u). (.C..1) . .3) this implies (take u = 1) Ego exp { .(y) = 0.T) 1+u2 (6. () where h (A.2' where as ususal ry > 0 is the adjustment coefficient for the given risk process. (. Theorem 5. 1) • Since L eatIG (dt. PROBABILITY OF RUIN IN FINITE TIME Recall that IG(x. for brevity. u) = IG(x/u2.3 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 ]. C . 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..u. The first step in the derivation is to note that µ = k (0) = r0 (00) . 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.() The idea of the proof is to improve upon this by an O (u1) term (in the following.122 CHAPTER IV. i.. 0o to. . (01. u) = euh(a . 9otc0" (0) = 0061 = ul.
distributed as Z .h.4.2 ).f.v. calculated using numerical integration and Proposition 1.v. The solid line represents the exact value . Note. (6. the r. is the c.ry2 .5) Once this is established . and the dotted line the corrected diffusion approximation (6.3).s.2). it holds for any fixed A > 0 that Ego exp { Ab1rr(u)/u2} . we get by formal Laplace transform inversion that C 2 u. But the Laplace transform of such a r.7. A numerical illustration is given in Fig.'yu /2)(1 + b2/u)} + Aug 1I J . that whereas the proof of Proposition 6. in (3) and (4). 1% in (2) and (4). we have p =.h. 9o T 0 in such as way that C = Sou is fixed. that the saddlepoint approximation of BarndorffNielsen & Schmidli [59] is a serious competitor and is in fact preferable if 77 is large] . however . . bl I IG I t +2 . of (6. the formal Laplace transform inversion is heuristic: an additional argument would be required to infer that the remainder term in (6.1 + 629. CORRECTED DIFFUSION APPROXIMATIONS 123 Proposition 6. 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. 6 .6.3 = 0.2) is indeed o(u1). of a (defective) r. u is Eeazead2/++ Eeaz[1 + ab2/u] where the last expression coincides with the r. To arrive at (6. The initial reserve u has been selected such that the infinite horizon ruin probability b(u) is 10% in (1) and (3).5) according to (6.z . .exp { h(A.1 + u2 I Indeed.d. 1.1 As u + oo. In ( 1) and (2).52/u where Z has distribution IG (•.3. which is based upon exponential claims with mean µB = 1.1 below is exact. . p = 0. just replace t by Tb1/u2.s. . however.
it gives the right order of magnitude and the ordinary diffusion approximation hopelessly fails for this value of p.4 may not be outstanding but nevertheless. and all of the numerical studies the author knows of indicate that its fit at p = 0.7 or at values of Vi(u) like 1% is unsatisfying.1 W IU.T) 111 0.EB 0 p ex p ( 7 S h ^)u .2 e.W21 0..19)2 11 20 20 i0 T 1n0 Figure 6.124 0. 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.u2 2u3 (e .114 0.aa1 .T1 00.. PROBABILITY OF RUIN IN FINITE TIME 0. . For further numerical illustrations. (Inc 0s 0.07 0. see Asmussen [12].TI CHAPTER IV.() Lemma 6.1 It is seen that the numerical fit is extraordinary for p = 0.111 W(U.08 0.00 0.199 0. OM 0.^) . BarndorffNielsen & Schmidli [59] and Asmussen & Hojgaard [34].T) 0.OOIi O. Similarly.05{ 0.0 0.7.011 L1 60 T IM 11.(061 0.01 0.02 I 90 120 160 2W A0 Z WT 40 80 120 160 100 240 280 T 111 WI. the fit at p = 0.1 proceeds in several steps.08 a. A51 7(SAT 3 3 h(X.
4 Ea. (6.co ((/u)) } Let 8 = (2a + (2)1/2 = h().61a2T (B3 . CORRECTED DIFFUSION APPROXIMATIONS Proof For a>0.T (co (8/u) ..+ h (A.6.2u (B3 . (6..(3)Eea LauT exp i 3J ..C2 = 2). 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 4.2 behaves like C l Eeo eXp r _ ^81T 1 Sl u2 1 u 2u3 [1+h(AC) S . 3 lim Eof (u) = EoC(oo) = a2 Ep = 3EoU2 uroo Proof By partial integration .1) h(A.7) 2 2 .3 EoU2 + 103OoEoU3 + " 2 6 Using d2 .4) that the r.00)(u +C)  'r (. the result follows.h.. the formulas Po(C(0) > x) Po(C(co) > x) imply 1 °° Po(ST(o) > x) = EIU fIP0 (U>y)dy .r0 (00)) } Replacing B by 8/u and Bo by C/u yields e(B() = E eo exp { (e . () 62 Eeo exp u u2 J . + a1b2 + .(3) J t _ aa1T l + eh(A. 1 = PB(T < oo) = Eo0 exp 125 {(B . exp ue } al 1J 3 exP I [2).C)C/u . 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 .co (e) . () + C and note that 2 KO (0) = 102.s.6) u U3 Lemma 6 . in Lemma 6.
Thus by Taylor expansion around ( = 90u.() I 1 + u2 ) y .6) and 7co (Oo) = ico('y + Bo) to get 0 = 21 (^/2 + 2y90) + 1112 (_Y3 + 3_Y200 + 3y9o) + O(u4). and inserting this and 9o 2 = S/u on the r. [2+ (2 . There are two reasons for this : in this way. 2 and (6.7) and using eh(a.h. 0 The last step is to replace h(A.e h(aS)h (^^ 262 exp {_h(.\+ (2 (3 e 2u [ (2.4. l Lemma 6 .6  d h(A. letting formally T * oo yields 7/)(u) C'e7u where C' = e7a2). yields +90 62 0 + O(u 3) 2u2 +O(u 3).2u [2A+ (2 3 . yu/2) 11+ 62 I} S 1 \\\ u/11 l 62 (3 2u 2A Proof Use first (6.h (A..2) for O(u) (indeed . yu/2).1 (y/2 + Oo)u .S) d e 62 .2 (^/2 + 3y9o + 390) + O(u3). we get h(A.\ + () 1 2 / .s. () . yu/2) h(A. 2 + 00 = . PROBABILITY OF RUIN IN FINITE TIME The last term is approximately (e 3 (3) 27. we get the correct asymptotic exponential decay parameter ^/ in the approximation ( 6.(2A + ()1/21 exp S h(A. 5 exp { _h(A) (1 + / y u J)) exp 1.() . and the correction terms which need to be added cancels conveniently with some of the more complicated expressions in Lemma 6.126 CHAPTER IV.2. () by h(\. Thus a2 y = 290 + O (u2). C) ( 1+ u2 The result follows by combining Lemma 6 .x.
() I 1 + u 2 ) } S 1 .4. that is.1: Just insert Lemma 6. The adaptation to risk theory has not been carried out. 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. In Siegmund's book [346]. HOW DOES RUIN OCCUR? exp { h (x. 7 How does ruin occur? We saw in Section 4 that given that ruin occurs. The answer is similar: the process behaved as if it changed its whole distribution to FL. with the translation to risk processes being carried out by the author [12]. the 'typical' value (say in sense of the conditional mean) was umL. We shall now generalize this question by asking what a sample path of the risk process looks like given it leads to ruin. . the same as for the unconditional Lundberg process. Fuh [148] considers the closely related case of discrete time Markov additive processes.7. The corrected diffusion approximation was extended to the renewal model in Asmussen & Hojgaard [34]. i. the approach to the finite horizon case is in part different and uses local central limit theorems.1 (y/2 + Oo)u )} 1 (i + U ) [2+ C2 2u 62 S Pt^ exP { J 62(2 exp { h(A. Hogan [200] considered a variant of the corrected diffusion approximation which does not require exponential moments. 0 1 Proof of Proposition 6.e. () (i+a ) 2A + (2 .(i+ 62 exP{ h(A.5 in Lemma 6. 'yu/2) 127 ( i+ M pz^ exP { h (A. ()} . His ideas were adapted by Asmussen & Binswanger [27] to derive approximations for the infinite horizon ruin probability 'i(u) when claims are heavytailed. the analogous analysis of finite horizon ruin probabilities O(u.T) has not been carried out and seems nontrivial. ()} 3 h (A. 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. and to the Markovmodulated model of Chapter VI in Asmussen [16].
u * oo. and let M(u) be the index of the claim leading to ruin (thus T(u) = Ti + T2 + . PROBABILITY OF RUIN IN FINITE TIME changed its arrival rate from 0 to /3L and its claim size distribution from B to BL.3 to ..(u)_ and similarly the denominator is exactly equal to Ce7u. Recall that 13L = (3B[ry] and BL(dx) = e'rxB(dx)/B[7]. the Poisson rate changes from .(u). stating roughly that under F(u). {St}0< t<T(u)) Proof Write e'rsr(u ) = e'rue'r£(u). we give a typical application of Theorem 7. F(u)c] P(r(u) < oo) ?P(U) < EL[e7u.(u)_ and ^(u) are independent . r(u) < oo) .3L and the claim size distribution from B to BL.TT(u) _measurable.FT(u) = o' (T(u ).r. ^(u) is exponential with rate 8 w. {ST(u)+t . Theorem 7 . Then also P(u)(F(u)) + 1.F.F. Proof P(u) (F(u)c) = F(flu)c. FL As example. F(u)c] ti e' ru]PL (F(u)`) > 0.ST(u)}t> o is just an independent copy of {St}t>o). the numerator becomes e'ruELe7^ (u)PL(F( u)t) = e7uCFL (F(u)°) when F(u) E . P(u) and rate = aL w.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. Note that basically the difference between FT(u) and .vi(u) Ce'Yu Corollary 7.2 If B is exponential. In fact. then P(u) and FL coincide on .(u)_ is that i.T.1. so in the in the proof.t. 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.128 CHAPTER IV. In the exponential case. + TMOO ). .r. (u) and satifying PL(F(u)) * 1.1 Let {F(u)}u>0 be any family of events with F(u) E F.EL[e7S.t. . .(u) is not . Recall that ..
Proof For the first assertion. who also treated the heavytailed case. the subject treated in this section leads into the area of large deviations theory.3 M(u) pcu) 1 . HOW DOES RUIN OCCUR? Corollary 7. take I(Tk < x) .7. however. the queueing results are of a somewhat different type because of the presence of reflection at 0. From a mathematical point of view. 129 M(u) >2 I(Tk < x) M(tu) p(u) M(u) >2 I(Uk < x) BL(x). Notes and references The results of the present section are part of a more general study carried out by the author [11]. A somewhat similar study was carried out in the queueing setting by Anantharam [6].eaLx.eALx) M(u) k=1 u The proof of the second is similar. see further XI.(1 . This is currently a very active area of research.3. .
This page is intentionally left blank .
..1). We use much of the same notation as in Chapter I.7).Chapter V Renewal arrivals 1 Introduction The basic assumption of this chapter states that the arrival epochs O'1.. Thus the premium rate is 1. and M is the maximum of {St}. Var(St) = 11Ba2A + I4AaB lim t goo t PA 131 .d. with common distribution B. In the socalled zerodelayed case. .1..(1. D'2... are i. Proposition 1. A different important possibility is Al to be the stationary delay distribution A° with density A(x)/µA.1 Define p = !µ. with Nt = # {n: Un <t} the number of arrivals before t.Q.. the distribution Al of T1 is A as well. see A. with the same distribution A (say) for T2.Then no matter the distribution Al of T1i B. the one corresponding to the stationary case by 00)(u).1 (T1 = a1)..i. of the risk process form a renewal process: letting Tn = Qn . Then the arrival process is stationary which could be a reasonable assumption in many cases (for these and further basic facts from renewal theory. U2. AA t*oo lim St = lim ESt t tioo t = p . the Tn are independent. and the one corresponding to T1 = s by 1/i8 (u). {St} is the claim surplus process given by I. . T3. r(u) the time to ruin. The ruin probability corresponding to the zerodelayed case is denoted by 1/'(u). . the claim sizes U1.
However . and (1 . The simplest case is of course the Poisson case where A and Al are both exponential with rate 0. A.0 > 0. the . Here are two special cases of the renewal model with a similar direct interpretation: Example 1.1) ENt/t + 1/µA. Nt = Var(PBNt) + E(4Nt) Q2 2 0`2 A tpB B + o(t). This has a direct physical interpretation (a large portfolio with claims arising with small rates and independently). s + t µA PA 0 Of course. Nt ESt = E E UI Nt t = ENt•pB .1) follows . we get similarly by using known facts about ENt and Var Nt that Nt Var(St) = Var E Nt U. Nt + EVar U. CHAPTER V. 2).a is really the accumulated claims over a period u of length a. From this ( 1. Thus.Nt] * a/PA. Example 1 . RENEWAL ARRIVALS lim E [St+a .1 gives the desired interpretation of the constant p as the expected claims per unit time. but the arrival rate in the ON state is . If the environment is Markovian with transition rate A from on to off and u from OFF to ON. The renewal model is often referedd to as the Sparre Andersen process. the definition 77 = 1/p . say at a. by the elementary renewal theorem (cf. such that no arrivals occur in the off state.132 Furthermore for any a > 0. one could imagine that the claims are recorded only at discrete epochs (say each week or month) and thus each U. For (1 . after E.1 of the safety loading appears reasonable here as well.t. stating that E[Nt+a . Sparre Andersen whose 1959 paper [7] was the first to treat renewal assumptions in risk theory in more depth.St] = a(p .2 (DETERMINISTIC ARRIVALS) If A is degenerate. OFF.3 (SWITCHED POISSON ARRIVALS) Assume that the process has a random environment with two states ON. t 4oo Proof Obviously. Proposition 1.1). 3) follows similarly by Blackwell 's renewal theorem.
} with {S(d)} a discrete time random walk with increments distributed as the independent difference U .T between a claim U and an interarrival time T.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. More precisely..d.. and the present author agrees to a large extent to this criticism.1. The values of the claim surplus process just after claims has the same distri bution as {Snd^ }• Since the claim surplus process {St} decreases in between max St = max ^d^. (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 ..4) w. and for historical reasons.1. arrival times.. (1. and then the whole process repeats itself). 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 . integrate (1. U1 . Proof The essence of the argument is that ruin can only occur at claim times. S o<t<oo n=0.. the first term represents the probability F(U1 . we feel it reasonable to present at least some basic features of the model.s.oFF}.s > u) of ruin at the time s of the first claim whereas the second is P(r(u) < oo. A is phasetype (Example 1.4) with phase space {oN. if for nothing else then for the mathematical elegance of the subject.4) fo Indeed.t.y)B(dy).1. in general the mechanism generating a renewal arrival process appears much harder to understand. the relevance of the model has been questioned repeatedly. we have From this the result immediately follows. INTRODUCTION 133 interarrival times become i. Therefore. Ao. The following representation of the ruin probability will be a basic vehicle for studying the ruin probabilities: Proposition 1. However.2.r. the fundamental connections to the theory of queues and random walks.i.s < u). initial vector (1 0) and phase generator 11 However. For the stationary case.. . u For later use.
3* pB. the claims and the premium rate are negative so that the risk reserve process .1 If. (2.1.0* (u) = P (rr* (u) < oo) where rr* (u) = inf It > 0: Rt < 0} ) . If . Theorem 2 . i. That is . each of which receive a payment at constant rate during the lifetime . The compound Poisson model with negative claims We first consider a variant of the compound Poisson model obtained essentially by signreversion . are independent of {Nt} and i. A typical sample path of {Rt } is illustrated in Fig.1) . Using Lundberg conjugation . then 0 * (u) = 1 for all u > 0.134 CHAPTER V. 00). St = t .d. t. the claim surplus process are given by Nt Nt Rt = u+^U. then 0*(u) = e 'r" where ry > 0 is the unique solution of 0 = k*(ry) = *(B*[ry] . U Figure 2.1) +ry. the remaining part of the prepayment (if any ) is made available to the company. < 1. A simple sample path comparison will then provide us with the ruin probabilities for the renewal model with exponential claim size distribution. RENEWAL ARRIVALS 2 Exponential claims.3* (say ) and the U.Ut.1 r* (u) One situation where this model is argued to be relevant is life annuities. resp . b=1 !=1 where {Nt } is a Poisson process with rate . we shall be able to compute the ruin probabilities i(i* (u) for this model very quickly (. with common distribution B* (say) concentrated on (0. 2. At the time of death .a*PB• > 1. The initial reserve is obtained by prepayments from the policy holders.
of {St} is c(a) = is*(a7).(u ) < oo) = E {e7sr_ (u).1. Let B(dx) = ^e7x B*(dx). 2. then by Proposition 111. (a) is*(a) (b) .f. Since ic'(0) < 0. and thus 1 = P(T. the safety loading of { St} is > 0.2(a).(a) 7 Figure 2.2(b).g.UB. If I3*pB* < 1. cf. Define T_ (u) = inf It > 0 : St = u} .0. > 1 . Proof Define 135 St =u .2 sup St = inf St = 00 t>o t>o and hence 0* (u) = 1 follows. EXPONENTIAL OR NEGATIVE CLAIMS [Note that r.3*.s. St=Rtu=St. Hence T_(u) < oo a. Then { St } is the claim surplus process of a standard compound Poisson risk process with parameters 0 *. B* [7] and let {St} be a compound Poisson risk process with parameters . . T_ ( u) < 001 e7"P(T_ (u) < oo) = e"V)* (u). 0 Now return to the renewal model.Rt. Then the c. B*. 0) and has typically the shape on Fig.2 Assume now . and the Lundberg conjugate of {St} is { St } and vice versa.2. B. Hence y exists and is unique. Then the function k* is defined on the whole of (oo. Fig.. 2.* (a) = log Ee'st I. T_ (u) = inf { t > 0 : St = u 'r* (u).
Ti = U1.1 means that M* is exponentially distributed with rate ry. and (2 .f... and hence the failure rate .•. the failure rate of this process is y. with the probability that a particular jump time is not followed by any later maximum values being 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 ..Y] (2. To + M(d) in the notation of Proposition 1.Tn} n=0.g.. the distribution of M(d) is a mixture of an atom at zero and an exponential distribution with rate parameter ry with weights 1 .'s and noting that V)*(u) = P(M* > u) so that Theorem 2.. Taking m. and 5PA > 1..1 it is seen that ruin is equivalent to one of these values being > u. According to Theorem 2. T2 = U2. respectively. with rate S (say).136 CHAPTER V.e. RENEWAL ARRIVALS Theorem 2 .. To + max {Ul+•••+UnTI.. alternatively termination occurs at a jump time (having rate 8).1.Ui .Un } = max St = t>0 n=0.2 If B is exponential. which has the advantage of avoiding transforms and leading up to the basic ideas of the study of the phasetype case in VIII.+Tn U1 Un. and from Fig .)(u) _ 1r+e7" where ry > 0 is the unique solution of 1 = Ee'Y(uT ) = S 8 A[.2).Tr+.7r+ 7r EeTo b/(Sa) + +. A variant of the last part of the proof. Then B* = A. u Hence P(M(d) > u) _ 1r+e'r".. then . 1) means that 8(A[ry] . 3* = 6.1.. 2.4 goes as follows: define 7r+ = P(M(d) > 0) and consider {St*} only when the process is at a maximum value.Y a I.u+ and lr+...• • • . Hence M* max {To + Ti + • • • + Tn ..1.4.1) + ry = 0 which is easily seen to be the same as (2.a) = 1 . However. we get Ee'M(d) = Ee°M* _ Y/(. Now the value of {St*} just before the nth claim is To +T1* +..
7r+) = ry and hence P(M(d) > u) = P(M(d) > 0)e7u = 7r+e'r". CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 137 is b(1.6. letting P(d) refer to the renewal risk model with these changed parameters .T to F(d)(x) = eK^d^(«) ^x e"vFidi(dy) 00 K(d) (a) = log F(d) [a] = log B[a] + log A[a] . Hence the failure rate of M(') is 6(1 . B^d) where Aad> (dt) = ^[ a] A(dt). hence exponential with rate b. the relevant exponential change of measure corresponds to changing the distribution F(d) of Y = U . 111. However.4. 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^. which states that for a given a.2.7r+) and hence r+ = 1.. Bads (dx) = .3. The probability that the first ladder step is finite is 7r+. a ladder step is the overshoot of a claim size. 3a The imbedded random walk The key steps have already been carried out in Corollary 11. we see that ry = 6(1.. This follow since. the imbedded discrete time random walk and Markov additive processes..2. 0 3 Change of measure via exponential families We shall discuss two points of view.5. Furthermore.3 A[a] B[a] F( d) [a +)3] F(d) [a] = Fad) [^] Letting M(u) = inf in = 1. resp.y/b..7r+).B(dx). we have ] A[a )3] E«d'efl' = Bad> [a] A ad> [Q] = B[a +. consider instead the failure rate of M(d) and decompose M(d) into ladder steps as in II. Thus a ladder step terminates at rate b and is followed by one more with probability 7r+. Putting this equal to y. : S(d) > u} .
just note that F7(d) is nonlattice when F is so .Ce"u where C = limu. let 7 > 0 be the solution of r.e.. provided the distribution F of U . Corollary 3. E(d)e 1' (u). and claim (a) follows immediately from this and e (u) > 0.r.1) is explicit given 7. RENEWAL ARRIVALS be the number of claims leading to ruin and ^(u) u = SM(u) . (d) (7) _ 0.. (b) V)(u) .t. For claim (b).2 p.(u) . Consider now the Lundberg case. the evaluation of C is at the same level of difficulty as the evaluation of i/i(u) in matrixexponential form.C8e7u where Cs = Ce78B[7]. 187) and thereby for ^(u) to be nonlattice w.1 implies Cu) = e«uE ( 7d)e«^(u) . This is known to be sufficient for ^(O) ]p (d) ([APQ] Proposition 3. cf.2 In the zerodelayed case.T is nonlattice. 7µA . O(u) = eauE (d)ea{ (u)+M(u)K (d)(a) . 00)(u) . VIII. Proof Proposition 3.u the overshoot . We have the following versions of Lundberg' s inequality and the CramerLundberg approximation: Theorem 3 .p)/($B'[7] .1).138 CHAPTER V. (a) '(u) < eryu..1 For any a such that k(d)' (a) > 0. in the easiest nonexponential case where B is phasetype. i . we get: Proposition 3. to converge in distribution since p(yd) (r(0) < oo) = 1 because of r (d)' (y) > 0. ik.C(°)eryu where C(O) = C0[7] .4. For the stationary case. In fact.3 For the delayed case Tl = s. 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 .
B(x) = o(e7x) and dominated convergence.(s. h(s) = e(a +x( a))8 (3. where G is the infinitesimal generator of {Xt} = {(Jt.1) (normalizing by h(0) = 1). Let P8f E8 refer to the case Jo = s.(°) ( Ce8B[7] Ao(ds) similar manner. E8h0 (Jdt.0 ) = Eo[ha ( Jdt.a . 0) = tc(a)h(s)..3. Sdt) = h(s ./c. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES Proof Using (1. we invoke the behavior at the 1 = h«(0. According to Remark 11. St)} and h.St)} can be defined by taking Jt as the residual time until the next arrival. (s.1) = C(O).Sdt] = Ee'uh(T) means 1 = f ' e°^B(dy) f ' h( s)A(ds). The underlying Markov process {Jt} for the Markov additive process {Xt} = {(Jt. 0 0 . (u + s .5. 3b Markov additive processes We take the Markov additive point of view of II.9. another use of dominated convergence combined with Ao[s] = (A[s] 1)/SPA yields 00 u) e7u iP8(u) Ao(ds) + f 0 = CB['Y](A[y] .h'(s). The expressions are slightly more complicated and we omit the details. we get r u +8 e"8(u) 139 e7uB(u + s) + 4 0 + L 00 J e7(v8)e7(u+8v). To determine boundary 0.dt ) eadt = h ( s) . 0) = ah (s) . For s > 0. y) = e°yh(s).dt(ah ( s) + h'(s)) so that Gha ( s. IPA 0 Of course.y) B(dy) 0 For the stationary case. Here K.5. we look for a function h(s) and a k (both depending on a) such that Gh..4). Equating this to tch(s) and dividing by h(s) yields h'(s)/h(s) _ . delayed version of Lundberg's inequality can be obtained in a e7u.
1)a in agreement u with Chapter III.S„):0<v< )be Lt = eaSt tK(a) h(Jt) = east .rc(a)] = 1.5.s governing {(Jt.c(a)] B[a] Proof Pa. since JT. = J8 = T2.a . Ba(dx) = B(dx). [a + /3] A[b . An easy extension of the argument shows that U1. .2).c(a)] which shows that U1.4 The probability measure Pa.140 CHAPTER V.2) means 1 = B[a]/3/(/3+a+rc (a)). however. i. [e1U1 + 6T2ea ( U1s)stc ( a)e(a+K(a ))( T2s)I B[a +...s(Jo = s) = 1 follows trivially from Lo = 1. .8. ] = E.. Remark 3 . we can now for each a define a new probability measure Pa.rc(a)] B[a] A[a . Further.13]A[b . St)}too by letting the likelihood ratio Lt restricted to Yt = a((J. (3.tK( a)e. .e. An important exception is. (3.e(«+k(a))t esy A(dt).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. . T2. resp.a . 5 For the compound Poisson case where A is exponential with rate . Aa as asserted . .s and P(d). J n+1 u are independent with distribution Aa for the Tk and Ba for the Uk. T2 are independent with distributions Ba. Proposition 3. Ease AU1+6T2 [ AU1+6T2 = Ea a LT.( a+r' (a))(Jt s) h(s) where c(a) is the solution of (3. Ba where Aa (dt) . A[a . U..rc(a)] = B = Ba[13]Aa[5]. Note that the changed distributions of A and B are in general not the same for Pa..2) As in 11. RENEWAL ARRIVALS B[a]A[a .. the determination of the adjustment coefficient ry where the defining equations rc(d) (ry) = 0 and rc(ry) = 0 are the same. . resp. rc(a) = 0 (B[a] .
see e. the time from he arrives to the queue till he starts service. for the zerodelayed case zp8(u. yu ) < e7yu. Using the Markov additive approach yields for example the following analogue of Theorem IV.y yuAa y [ay + K(ay) .rc( ay)] = e(aa+(r ))sb[a ]e7yu L y1 In particular..(u. The claim for the zerodelayed case follows by integration w. it is easily seen that ic(ay ) > 0.6 Let y < let ay > 0 be the solution of ic'(ay) = 1/y.4. Let M(u) be the number of claims leading to ruin .4. yu) = F'ay.t.g. The actual waiting time Wn of customer n is defined as his time spent in queue (excluding the service time). say finite horizon ruin probabilities where the approach via the imbedded random walk yields results on the probability of ruin after N claims. T(u) < yu h(JT(u)) < eayu+yuk(ay ) ( Eia y Le(a(+k(ay))s v. Then J(rr(u)) = TM(u)+1 and hence Ws(u. 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. not after time T. XII.5. defined as the single server queue with first in first out (FIFO. A(ds). which is the same as the asserted inequality for 0. the amount of .yx(ay).r. 2. see in particular Dassios & Embrechts [98] and Asmussen & Rubinstein [45]. and assume that T„ is the time between the arrivals of customers n .)+1 e J j e(ay+w(ay))8 e .5 Proposition 3.yu ) e7vu A[ay .ay+ray))TM(. Proof As in the proof of Theorem IV.1 and n. or FCFS = first come first served) queueing discipline and renewal interarrival times.s eaysr(")+r(u ) K(ay) h (s) . that is. Label the customers 1. that is. and define yy = ay . . u The approach via the embedded random walk is standard. Then "^ e(ay+w(aY))8 Ys(u. [APQ] Ch. Notes and references 4 The duality with queueing theory We first review some basic facts about the GI/G/1 queue. yu). The virtual waiting time Vt at time t is the residual amount of work at time t. and U„ the service time of customer n. For the approach via Markov additive processes.4.. . .
Let the T there be the random time UN.1 and Corollary 11..4.1. and we have P(W > u) = V. p < 1. 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. (4. Thus.4. 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. equivalently. whereas in [On .1 Wn+1 = (Wn + U.4): Proposition 4.+ Un.• • • Tk )."^ Vi(N) (u). If W1 = 0. an+1) = [on. Vt converges in distribution to a random variable V. on + Tn) the residual work decreases linearly until possibly zero is hit. the proposition follows.3) Proof Part (a) is contained in Theorem 11. Also {Zt}o<t<T evolves like the leftcontinuous version of the virtual waiting time process up to just before the Nth arrival. we have Wn = Van(left limit). and we have P(V > u) = ?/iiol(u).3.1) The following result shows that {Wn} is a Lindley process in the sense of II..142 CHAPTER V. 0 Applying Theorem 11.n1 (U1 +• • •+Uk Tl . It then jumps to VQ„ . and obviously z/'(u) = limN.3 Assume rl > 0 or.2) (b) as t * oo. in which case {V} remains at zero until time on+1. since customer n arrives at time on. Then: (a) as n + oo. The traffic intensity of the queue is p = EU/ET.Tn)+. then Wn v M. but interchanging the set . (u).1). .Tn)+ Proof The amount of residual work just before customer n arrives is VQ„ . and combining with (4.4: Proposition 4.2. we get: Corollary 4. but we shall present a slightly different proof via the duality result given in Theorem II. equivalently. the waiting time a customer would have if he arrived at time t.4. Thus Vos}1 _ = (Wn + Un . (4.2 Let Mnd) = maxk=o.1.. (4. Wn converges i n distribution to a random variable W..
cf.le) the same is true for the timereversed point process which is the interarrival process for { Zt}o<t < T• Thus as before . K(x) = P(W < x). x > 0. Hence for x > 0.y)F(dy). u Now return to the Poisson case . T1. but this requires some additional arguments (involving regeneration at 0 but not difficult) that we omit.. It follows that P(WN > u) =.4. resp .e. However.oo in Proposition 4. (4. we obtain: Corollary 4.4) 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.. hence (since the residual lifetime at 0 and the age at T have the same distribution . by an obvious reversibility argument this does not affect the distribution .T* < x) fK(x_y)F(dy) (x > 0 is crucial for the second equality!).(N)(u) has the limit tp(u) for all u. i. we get W = (W + U* .. Then the arrivals of {Rt} in [0. Then K(x) = J x00K(x .. Letting n oo in Corollary 4.T* are independent and distributed as U1. { 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: . A.1. THE DUALITY WITH QUEUEING THEORY 143 (T1. In fact .T* = y yields K(x) = P ((W + U* . TN) with (TN. convergence in distribution hold for arbitrary initial conditions .T*)+. Corollary 4. as WN. T1 ..2. T] form a stationary renewal process with interarrival distribution A.Ao in (b).2)... For part (b). and hence in particular ZT is distributed as the virtual waiting time just before the Nth arrival. conditioning upon U* .. where U*. namely W1 = 0 in (a) and Vo = 0.T*)+ < x) = P(W + U* .5) Proof Letting n . Ti) and similarly for the U. which implies the convergence in distribution and (4... we let T be deterministic .4 The steadystate actual waiting time W has the same distribution as M(d).5 (LINDLEY'S INTEGRAL EQUATION) Let F(x) = P(U1T1 < x). Then the corresponding queue is M/G/1. (4.
5) to hold for all x E R and not just x > 0).5) is in fact a homogeneous WienerHopf equation. Hence '(u) = Ali(°)(u). W v V. Cohen [88] or [APQ] Ch. VIII). . implying P(W > u) = P(V > u) for all u. see e. 0 Notes and references The GI/G/1 queue is a favourite of almost any queueing book (see e . the zerodelayed and the stationary renewal processes are identical.g. That is. the actual and the virtual waiting time have the same distribution in the steady state. Note that (4.g. The equation (4. Some early classical papers are Smith [350] and Lindley [246]. Proof For the Poisson case. RENEWAL ARRIVALS Corollary 4. Asmussen [24] and references there.5) looks like the convolution equation K = F * K but is not the same (one would need (4. 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.144 CHAPTER V.6 For the M/G/1 queue with p < 1.
The ruin probabilities with initial environment i are '+ki(u) = pi(T(u ) < oo) = Pi (M > u). 145 Oj( u. The intensity matrix governing {Jt} is denoted by A = (A. Thus.(3i when Jt = i.T) = Pi (T(u) < T).. {St} denotes the claim surplus process. • Claims arriving when Jt = i have distribution Bi.)iJEE and its stationary limiting distribution by lr. here it exists whenever A is irreducible which is assumed throughout. . • The premium rate when Jt = i is pi. dv.f pi. {Jt} describes the environmental conditions for the risk process. M = supt>o St. t St = E Ui . N. As in Chapter I.Chapter VI Risk theory in a Markovian environment 1 Model and examples We assume that arrivals are not homogeneous in time but determined by a Markov process {Jt}0<t<oo with a finite state space E as follows: • The arrival intensity is . and can be computed as the positive solution of WA = 0. i=1 0 and r(u) = inf It > 0: St > u}. Ire = 1.
2. = iii when j E E(i). cf.T(n)). respectively. i and corresponding arrival intensities Qn.1 implicitly assumes that the sojourn times of the environment in the normal and the icy states are exponential. the intensity matrix is A OW) T(i) T(n) t(n)a(i) where t(n) = T(n)e. say. t(i) = T(')e are the exit rates.1 Consider car insurance. and assume that weather conditions play a major role for the occurence of accidents. /3 = Nn when j E E(n). Proposition 1. with rates Ani and Ain. Assume similarly that the sojourn time in the normal state has distribution A(n) which we approximate with a phasetype distribution with representation (E('). MARKOVIAN ENVIRONMENT where as usual Pi refers to the case Jo = i. we could distinguish between normal and icy road conditions.2 (ALTERNATING RENEWAL ENVIRONMENT) The model of Example 1.14. a(i). For example. meaning that accidents occuring during icy road conditions lead to claim amounts which are different from the normal ones. and p is the overall average amount of claims per unit time. Example 1. Bi. leading to E having two states n.4) with representation (E(i). Then the state space for the environment is the disjoint union of E(n) and E(i). in blockpartitioned form. P = E 7riPi. According to Theorem A5. T(=)). the operational time argument given in Example 1.1) iEE Then pi is the average amount of claims received per unit time when the environment is in state i.11 below.146 CHAPTER VI. r^ = P (1. Unless otherwise stated. one expects that 3i > on and presumably also that Bn # Bi. assume that the sojourn time in the icy state has a more general distribution A(i). we shall assume that pi = 1. We let p Pi = /ji/AB. this is no restriction when studying infinite horizon ruin probabilities.a('). say. u .5 below. Example 1 . Thus.. 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. f3i and claim size distributions Bn. which is clearly unrealistic. and we have f3. we can approximate A(i) with a phasetype distribution (cf. An example of how such a mechanism could be relevant in risk theory follows. cf.
is the probability that a long icy period is followed by a short normal one. . we assume again pi = 1 so that the claim surplus is Nt St = ?Ui_t.Q...p. This amounts to a family (A(")) ?CH Of sojourn time distributions. but assume now that the arrival intensity changes during the icy period. u Example 1.J017. Qi = . w.3i/pi. MODEL AND EXAMPLES 147 Example 1 .. i ) : n E H. Example VIII..a(n). it = Jel(t).. such that a sojourn time of type rt is followed by one of type c w. Then for example wi.1.j = .. In the car insurance example. depending only on 77.. 0 Then (by standard operational time arguments) {St } is a risk process in a Markovian environment with unit premium rate.>. resp. t(n) = T("i)e.n.. . (9) where q = CHI. dt.3 Consider again the alternating renewal model for car insurance in Example 1. such that the icy period is of two types (long and short) each with their sojourn time distribution A('L). One way to model this would be to take A(') to be Coxian (cf. T(9) +wggt(9)0.. n8}. Approximating each A('?) by a phasetype distribution with representation (E('l). where W = (w. let T 9(T) = f pi. and . 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. and similarly for the normal period. and 1/ii(u) = t/ii(u). The simplest model for the arrival intensity amounts to .3i.. one could for example have H = {i1. i8f n1. the state space E for the environment is { ('q..5 (MARKOVMODULATED PREMIUMS) Returning for a short while to the case of general premium rates pi depending on the environment i.4) with states i1.1. 1 . St = SB=(t). the parameters are ^ij = aid/pi.3.2. Indeed.. A('^). u From now on. u Example 1 . i E E(n) }. 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) . say it is larger initially. say.tEH is a transition matrix. iq (visited in that order) and letfOil >.T(n)).
The key property for much of the analysis presented below is the following immediate observation: Proposition 1.6 The claim surplus process {St} of a risk process in a Markovian environment is a Markov additive process corresponding to the parameters µi = pi. o = 0. Pi (Ti E dx.5. 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). Nt Nt a .8 As t oo. one can associate in a natural way a standard Poisson one by averaging over the environment. the empirical distribution of the claims is B*. . iEE iEE )3 These parameters are the ones which the statistician would estimate if he ignored the presence of Markovmodulation: Proposition 1. B* = 1 /^* Bi.7 The Pidistribution of T1 is phasetype with representation (ei.(Qi)diag)• More precisely. More precisely. N > 1(Ul < x) a4 B*(x). MARKOVIAN ENVIRONMENT We now turn to some more mathematically oriented basic discussion. JT1 = j) = Qj • e. 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. . 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. t l=1 Note that the last statement of the proposition just means that in the limit.(3iBi(dx). the Markov additive structure will be used for exponential change of measure and thereby versions of Lundberg's inequality and the CramerLundberg approximation. Next we note a semiMarkov structure of the arrival process: Proposition 1. dx.A .e(A(Oi)d'sg)xe. In particular. qij = 0 in the notation of Chapter 11. )3*. we put )3* = E 7fi/3i. vi(dx) = .148 CHAPTER VI.
i. Nt a' t t iEE Also. Bi.aA .x) = Nt E > I (Uk) X) Nt Bi(x) 1=1 iEE k=1 iEE 1: t5 Bi( x) = B*(x). y Ni) i Nti) t a. cf. Conditioning . A. 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 . However . aA.4. has distribution (7ri)3i //3*)iEE and is independent of Ti. N + oo Hence 1 Nt 1 N`+) Nits Nt E I ( Ut <. Then it is standard that ti lt '4' iri as t > oo.. In particular.. this converges to the exponential distribution with rate 0* as a * oo. denoting the sizes of the claims arriving in state i by U(') 1 standard law of large numbers yields U(') the N 1: I(Ukik < x) k=1 N a$. oo) as a 4 oo.7. In particular. {St} to the compound Poisson model with parameters 0 *. Then {St)} + {St*} in D[0. e. Hence Nt'> a . the Fidistribution of T1 in {St(a ) } is phase type with representation (E.1. Bi. By Proposition A5.9 Consider a Markovmodulated risk process {St} with param eters Ni. Example 11.(/3i)aiag). Proposition 1.6. given {Jt}0<t<0. The next result shows that we can think of the averaged compound Poisson risk model as the limit of the Markovmodulated one obtained by speeding up the Markovmodulation. ^j 7riNi. iEE 13 A different interpretation of B* is as the Palm distribution of the claim size. B*. and furthermore in the limit JT. we may view Nt`i as the number of events in a Poisson process where the accumulated intensity at time t is Niti. and let {St °i} refer to the one with parameters Pi..2. the limiting distribution of the first claim size U1 is B*. Proof According to Proposition 1.. Bi(x).* (u) for all u.. zli( (u) .
we first get that 3 (3* = 2. e.150 CHAPTER VI.1 of [145]. from Theorem 3. state 1 appears as more dangerous than state 2.2 +2 2 = 3. those of type E7 with intensity z s = 5 in state 1 and with intensity z . which also yield O(a) (u. U. 0 Example 1. oo).1 with periods with positive drift alternating with periods with negative drift..2.31µB 2 = 2 5 3 7 70 Thus in state 1 where p. > 1... 5 5 where E5 denotes the exponential distribution with intensity parameter 5 and a > 0 is arbitrary. thick.=1. there are p = 2 background states of {Jt}. 1. the overall drift is negative since it = (2 2) so that p = 71P1 + 112P2 = 7. T) for all u and T. 9 ..FT. B1=3E3+2E7. is as in {St }.. s 5 in state 2. B2=1E3+4E7. 132=2.).l3* and U2 having distribution B*. and (at least when a is small such that state changes of the environment are infrequent).2. the company even suffers an average loss.s = o in state 1 and with intensity 1 . 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 = .g.. Continuing in this manner shows that the limiting distribution of (T. T) + ?P* (u. marked by thin.. the paths of the surplus process will exhibit the type of behaviour in Fig.1. MARKOVIAN ENVIRONMENT upon FT. That is. since E3 is a more dangerous claim size distribution than E7 (the mean is larger and the tail is heavier). On Fig..s = 1o in state 2. shows similarly that in the limit (T2. U2) are independent of .2}. with T2 being exponential with rate . Computing the parameters of the averaged compound Poisson model. The fact that indeed 0(a) (U) 3 0* (u) follows. resp. we may imagine that we have two types of claims such that the claim size distributions are E3 and E7. lines in the path of {St}. 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[0.10 Let E_{1. 1. Claims of type E3 arrive with intensity 2 . Thus..
t + oo. That is.8. t * oo.3* = 3/4 of the claims occur in state 1 and the remaining fraction 1/4 in state 2. Hence B* = 415E3+5E7/ +4 ( 51E3 +5 E7) = 1E 3 +2E7.11 (a ) ESt/t * p . 0 The definition (1. note first that EN Uk')/N a4' µgi.1. Hence (i) Nti) 1 U(i) k' N(i)k=1 E t 4 St + t = iEE Nt t 1: 7ri Qi µs. a fraction r.1 a.s.1).1. 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. For (b).(3. 01 /. the averaged compound Poisson model is the same as in III. MODEL AND EXAMPLES 151 Figure 1.1 Thus. = P. iEE . 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). (b) St/t * p ..
\ i and EiX1 Ei f 13 J. [212]. Statistical aspects are not treated here.s. 136 or A. Proposition 1.4. some early studies are in Janssen & Reinhard [211]. Proof The case 77 < 0 is trivial since then the a.1)Eiw = 0. The mainstream of the present chapter follows [16]. and hence oscillates between 0o and oo so that also here M = oo. and so on. PB.Jt=i}. then M = 00 a. dt . having the Pidistribution of X.152 CHAPTER VI.. Now let r) = 0. [APQ]. [302]. [315]. Theorem II. In risk theory. X2 =SW2 So. . Eiw.. X 1 =Sty.2(a) p. If 77 > 0. There seems still to be more to be done in this area. The case 77 > 0 is similarly easy. also + . 2 The ladder height distribution Our mathematical treatment of the ruin problem follows the model of Chapter III for the simple compound Poisson model.a form a renewal process ... let some state i be fixed and define w=wl=inf{t >0:Jt_#i.1 and the Corollary are standard. 38) Eiw1 = 1/ir. see the Notes to Section 7. with some important improvements being obtained in Asmussen [17] in the queueing setting and being implemented numerically in Asmussen & Rolski [43]...g. and a more comprehensive treatment in Asmussen [16].. then M < oo a.12 If 77 < 0. EiX = 0. and hence 1/ii(u ) = 1 for all i and u.. and hence wn /n a4. s. limit p . and hence M = 00. Since the X„ are independent . and . Now obviously the w. + Xn SWn ](1 a ..1(b) is essentially the same as the proof of the strong law of large numbers for cumulative processes.Jt=i}.Eiw o'o Eiw • E ^ifjµs.0i(u) < 1 for all i and u. and involves a version of the . . see [APQ] p.ld. MARKOVIAN ENVIRONMENT Corollary 1. X3. The proof of Proposition 1.1 of St / t is > 0. with X2. n n Thus {SWn l is a discrete time random walk with mean zero..1 jEE = (p .s. Then by standard Markov process formulas (e. See Meier [258] and Ryden [314]. 0 Notes and references The Markovmodulated Poisson process has become very popular in queueing theory during the last decade. w2=inf {t>w1:Jt_#i.
oo)) = f R(i.6.Jt=j)dt. T+ < oo) and let G+ be the measurevalued matrix with ijth element G+(i.2(a) below ) where the ladder height distribution is evaluated by a time reversion argument. •)• kEE Also.dx). for i. 00 (2. only with the product of real numbers replaced by convolution of measures. G+ is the matrix whose ijth element is E G +(i. •) * G +(k.EA.2) R(dx)S((y . Proposition 2.j.. The form of G+ turns out to be explicit (or at least computable). Let further R denote the preT+ occupation kernel. IIG+ II denotes the matrix with ijth element IIG+(i. cf. we get the same ladder height distribution as for the averaged compound Poisson model. However . see also Example II.x.2 (a) The distribution of M is given by 00 1 . e.5. Define the ladder epoch T+ by T+ = inf It : St > 0} = r(0). For measurevalued matrices. k. and S (dx) the measure valued diagonal matrix with /3 Bj(dx) as ith diagonal element. j.Jr+ =j.i.j E E.A) =ZI(St E.6*. •). THE LADDER HEIGHT DISTRIBUTION 153 PollaczeckKhinchine formula (see Proposition 2. .x). the definition of . by specializing results for general stationary risk processes (Theorem II .j. j. (y. T R(i. j. oo)).4) we obtain the following result .3*B *(y)dy.(u) = Pi(M < u) = e' E G+ (u)(I . which represents a nice simplified form of the ladder height distribution G+ when taking certain averages : starting {Jt} stationary. let G+(i.g.A) = Pt(ST+ E A. j. we define the convolution operation by the same rule as for multiplication of realvalued matrices. 6. That is. Proposition 2.1) 0 (b) G+ (y.2. B* in Section 1. oo) = J ao 0 G+(i.1 irG+(dy)e =. Thus. n=0 (2. dx)/jBj(y . •) II = JG+(i.IIG +II)e. but is substantially more involved than for the compound Poisson case .a/i. j.
St < S* for u < t. mx = j when for some (necessarily unique) t we have St = x. thick. hence uniquely specified by its intensity matrix Q (say). III to bring R and G+ on a more explicit form . resp.1) follows by summing over n and j. lines in the path of {St}. MARKOVIAN ENVIRONMENT Proof The probability that there are n proper ladder steps not exceeding x and (x)ej. From this (2.3.IIG+II)e.3 When q > 0.3) We let {St*} be defined as {St}.1 for an illustration in the case of p = 2 environmental states of {Jt}.154 CHAPTER VI. G+ the probability that there are no further ladder steps starting from environment j is e^ ( I . and that the environment is j at the nth when we start from i is e . 0  x Figure 2. The u proof of (2. the intensity matrix A* has ijth element * 7r ^i3 7ri and we have Pi(JT = j) = 7rj P2(JT = i)7ri (2. {mx} is a non terminating Markov process on E.1 The following observation is immediate: Proposition 2. we need as in Chapters II. JJ = j.2) is just the same as the proof of Lemma 11. To make Proposition 2. That is. and let further {my} be the Evalued process obtained by observing {Jt } only when {St*} is at a minimum value.6. only with {Jt} replaced by {Jt } (the /3i and Bi are the same ). we need to invoke the timereversed version {Jt } of {Jt} . see Figure 2. . To this end .2 useful . marked by thin.
In general.(/3i)diag + T S(dx) eQx. 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). Proof The argument relies on an interpretation in terms of excursions. and S(dx) is the diagonal matrix with the f3iBi(dx) on the diagonal. Q( n+l) _ ^.and a jump (claim arrival) occurs at time t. The definitions are illustrated on Fig.0. 0 mms1   ^O \ T. we recursively define the depth of an excursion as 1 plus the maximal depth of a subexcursion. s]. } is a minimum value at v = t. the sequence {Q(n)} A* defined by Q(O) = .2 .2. {S. Figure 2. 2.*. and the excursion is said to have depth 1 if each of these subexcursions have depth 0.. we say that the excursion has depth 0.4 Q satisfies the nonlinear matrix equation Q = W(Q) where 0 co(Q) = n* .2. An excursion of {St*} above level x starts at time t if St = x. = x}. Furthermore. If there are no jumps in (t. corresponding to two subexcursions of depth 0. ( Q( n)) converges monotonically to Q. For example the excursion of depth 2 has one subexcursion which is of depth 1. THE LADDER HEIGHT DISTRIBUTION 155 Proposition 2. Otherwise each jump at a minimum level during the excursion starts a subexcursion.2 where there are three excursions of depth 1. and the excursion ends at time s = inf {v > t : S.(/3i) diag.
4) To show Q = cp(Q). Theorem 2 . p1^) Define a further kernel U by f U(i. StEA . j. mx+dx = j) occurs in two ways .(01)diag = Q. Q = W(Q) follows. Then a jump to j (i.6) . By considering minimum values within the excursion. Fi(mh =i ) = 1 + =hflh+Qihpii+o(h) implies qii = 'iii /i +)3ipii.St EA.T+>t) _ ^iF 7ri (JJ =i. It is clear that { mini } is a terminating Markov process and that { mio) } has subintensity matrix A* .j. Similarly by induction . A).s. A) = f Pi(mx = j) dx eie4xej dx A u (2. or through an arrival starting an excursion terminating with J.156 CHAPTER VI. 7rE Proof We shall show that Fi(Jt=j. = j. It follows that qij = A. 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. 0)). Suppose mx = i. A) = L' U(j. Now let {m ( n) } be {mx } killed at the first time i7n (say) a subexcursion of depth at least n occurs . Writing out in matrix notation . either due to a jump of {Jt } which occurs with intensity A= j. h. (2..Qi + )%pij) Now just note that t pij and insert (2.j +/3ipij. it becomes clear that pij = r [eQh] 0 ij Bi (dy) • (2. i. we first compute qij for i $ j.u< t).5) A (note that we use A = {x : x E Al on the r. The proof of Q = W(Q) then immediately carries over to show that the subintensity matrix of {mil) } is cp (Q(o)) = Q(l).5 R(i.St <S*. e. the subintensity matrix of {min+i ) } is cp (Q(n)) = Q(n +l) which implies that qgj +1) = \!. Similarly.4).. of the definition to make U be concentrated on (co.
K( n (d) the sequence converges monotonically to K.St EA. 0<u<t).4].0<u<t. (Jo = j.7 It is instructive to see how Proposition 2.S„<0.6 (a) R(dx) = eKxdx.6 is hardly all that explicit in general. Remark 2. {Jt }. G+((z..1 can be rederived using the more detailed form of G+ in Corollary 2. e. St E A. consider stationary versions of {Jt}. we shall see that nevertheless we have enough information to derive. it is readily checked that 7r is a left eigenvector of K corresponding to the eigenvalue 0 (when p < 1).0<u<t) = P. and to obtain a simple solution in the . THE LADDER HEIGHT DISTRIBUTION 157 from which the result immediately follows by integrating from 0 to oo w.Qi)diag. From Qe = 0. `` {K(n)} [the W(•) here is of course not the same as in Proposition 2.z+>t) = P.StEA. Jt = i. u It is convenient at this stage to rewrite the above results in terms of the matrix K = 0'Q'A.(Jt=j.=StSt. St < St U. 0 +1) = cp (K( n)) defined by K(o) = A . St EA. and get irPi(Jt =j..St <Su. x < 0. (b) for z > 0.t. (c) the matrix K satisfies the nonlinear matrix equation K = W(K) where W( K) = A ( i) diag + fi J "O eKx S(dx). 0 < u < t) = 7rjPj(Jt =i.. the CramerLundberg approximation (Section 3). To this end.(. where A is the diagonal matrix with 7r on the diagonal: Corollary 2. and we let k be the corresponding right eigenvector normalized by Irk = 1.2. We may then assume Ju=Jtu.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. oo))dx.. and this immediately yields (2..g. S.Jo=i. oo)) = f o' eIXS((x + z.r. dt.
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
for a vector <p(u) = (cpi (u))iEE of functions . (u.00 su e7( ( 3.i [eT(u)K(av ). yu < r(u) < 00 h 4(u) < h(av)C+o)(y)eavuEav . r(u) < yu] hiay)C+ h=av)C+ o) (y)eayu+yuw(av).V)i(u.(ay)}. 1 Similarly.i [e*(u)K(av).11 Let Bj (x) C_ = min 1 • inf jEE hj x>o f2° e'r( vx)Bj(dy) ' C+ _ mE 1 Bj(x) J Y x)Bj (dy). 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. Chie ryu < Vi(u ) < C+hie 7u. we have ic(ay) < 0 and get 'i(u) . we let G+ * W(u) be the vector with ith component E(G+(i. exp {e() + r(u))} .y)G+(z. we shall need the matrices G+ and R of Section 2. However.9) For the proof. as in the classical case (3. yu) f h(av)e v avuE«v..i I (a) exp {aye(u) + r(u)r. oo)) and. dy)• o iEE jEE . (3.3.j. hj P . if y > 1lk'(ry). CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 165 hiav)e _avuE. 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 (3.8 ) Then for all i E E and all u > 0.7.5) will produce the maximal ryy for which the argument works. We further write G(u) for the vector with ith component Gi(u) = EiEE G+(i. r(u) yu o)(y)eavuEav.5).j) * coj)(u) _ f u ^Pj(u . av 'i [h.
Proof Write UN = EN G+ .j. dx) f e7( vu)Bj (dy . U = U".x) x) jEE 0 E Qj f jEE R(i. jEE u 0 j.12 Assume sup1.x ) R(i. = Eo G+ G.dy). MARKOVIAN ENVIRONMENT Lemma 3 .x ) = Gi(u).166 CHAPTER VI. °O .& (u). Then iterating the defining equation ip(n+1) = G + G+ * V(n) we get W(N+1) = UN * G + G+N+1) * ^(o) However. 0 G+(i. dx ) Bj (u . we have G *(N +1) * ^. _ To see that the ith component of U * G(u) equals ?Pi (u). dx) 100 C .j.(0) ] (u) < sup Jt t. 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). dx). dy) 00 C+ ijhj f R(i. dy) = aj f Bj(dy . n > oo. dy) : 1(u) < C+ > hj u e(1tL)G+(i. Then cpin)(u) sit (u) as n + oo. j. 00 Thus C+ > hj f"o e7(Yu)G +(i. j. j. Lemma 3 .7.x) jEE 00 u 0 //^^ C+E.u Iv 2°)(u)I Pi(rr+(N + 1) < oo) + 0. if r+ (n) is the nth ladder epoch. Hence lim cp(n) exists and equals U * G.u IMP:°) (u) I < oo.13 For all i and u.ery(&u+x)Bj (dy) Bj(u Bj (u .3jhj // f 00 R(i. 00 f C_ hj f e(Y)G+(i. j.
We claim by induction that then cpin) (u) > C_ hie7u for all n. we get Wo n +1) (u) = ? 7 i ( U ) + E J u gyp. jEE estimating the first term in (3. T) < C+(')' o)hi7u)e7ou8T . this is obvious if n = 0.12) Proof We first note that just as in the proof of Theorem 3. and assuming it shown for n. 167 u Proof of Theorem 3.11) C_e7u 57 O+[i. j. (3. Here is an estimate of the rate of convergence of the finite horizon ruin probabilities 'i (u.n) ( u .10 ) by Lemma 3 . 9 for the last equality in (3. and using Lemma 3 .MT<u. T) = Pi (7. and let 8 = e'(70).10: Theorem 3 .3.ST). and the proof of the lower one is similar.T) = Pi(M > u) . it follows that Vi(u) < C_(yo) h=70)e7ou. u The proof of the upper inequality is similar . dy) jEE u U +C_ hje7( yu)G jEE"" +(i. (3. ST < u] < C+(yo)e7ouEi [h^7o)e70ST1 l T J = C h(7o)e7ou8T . y]hj = C_ e7uhi.11.11). letting MT = maxo<t<T St.Pi(MT > u) = Pi(MT < u.13 Let first cp=°)(u) = C_ hie"u in Lemma 3.y)G+(i.M>u) = Ei [VGJT (u .13) Hence. j.tpi(u. +i . 14 Let yo > 0 be the solution of 'c'(yo ) = 0.10) C_ 1 f hje7(y.8) with y replaced by yo and hi by h=7o ). dy) jEE o (3. j. j. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES proving the upper inequality.13. MT < u. we have Vii (u) . Then 0< Vi (u )  0i(u. 13 and the second by the induction hypothesis .(u) < T ) to 0i (u) which is different from Theorem 3. Indeed.u)G+(i. taking cps°) (u) = 0. from which the lower inequality follows by letting n * oo.M > u) = Pi(ST<u. let C+(yo) be as in (3. dy) (3.
.3) to B = Bi does not depend on i.o.. we also assume that there exist i # j such that either /3i <.31:5)32 . 0"(u) = P(M" > u)) Now consider the risk process in a Markovian environment and define i' (u) _ >iEE irioi(u). in part from the folklore principle that any added stochastic variation increases the risk.2) alone just amounts to an ordering of the states. MARKOVIAN ENVIRONMENT Notes and references The results and proofs are from Asmussen and Rolski [44]. B2 <_s. u > 0. we refer to .. where it has been observed repeatedly that Markovmodulation increases waiting times and in fact some partial results had been obtained. this correponds to the usual stochastic ordering of the maxima M'. The conditions which play a role in the following are: . The results to be presented show that quite often this is so. The motivation that such a result should be true came in part from numerical studies.2) (4.4) To avoid trivialities. is the one for the Markovmodulated one in the stationary case (the distribution of J0 is 7r). Bp. (4.1) Obviously. this is not the case for (4. The Markov process {Jt} is stochastically monotone (4. V)" if z/i'(u) <'c"" (u)...33 or Bi 0 Bj.. <s. < .o.3). we define the stochastic ordering by 0' < s. Further related discussion is given in Grigelionis [176]. Occasionally we strengthen (4.o.168 CHAPTER VI. (4.3) Bl <_s. It was long conjectured that 0* Vi.0. For the notion of monotone Markov processes.. 4 Comparisons with the compound Poisson model 4a Ordering of the ruin functions For two risk functions 0'. M" of the corresponding two claim surplus proceses (note that 0'(u) _ P(M' > u).3p.5) Note that whereas (4.o.. but that in general the picture is more diverse. (4. [177]. ". and finally in part from queueing theory. where o*(u) is the ruin probability for the averaged compound Poisson model defined in Section 1 and .
2 If al < . Theorem 4 ...r (JT(o) = i.9 ) below).9) (4.. E 7r i Wi(u .6) 7r= fl*B*(u) + p> s=1 +) fu 0 b (u  x)Bt (x) /pB. Conditions (4.3 for the second) *(u) _ /3 *B* (u) +. Proposition 2.x) dx u o i =1 i=1 (4. Conditioning upon the first ladder epoch. < bp and 7ri > 0 (i = 1.9) follows by considering the increasing functions 3iBi (x) and Oi (u . Section 4. Comparing (4.r (Sr(o) E dx Jr(o) = i. Lemma 4 . T(0) < oo) = Bi(x) dx/tcai ... (b) P. also Proposition 2. then P P P 7rjbj.6. 7(0) < oo) = pirf+). p). The first is a standard result going back to Chebycheff and appearing in a more general form in Esary. Then V. dx (4.2)(4.5 (cf.7) 7ri.7) and Lemma 4.6). Proschan & Walkup [140]. 1:7riaibi > E 7riai i=1 i=1 j=1 The equality holds if and only if a1 = ... b1 < . the second follows from an extension of Theorem I1. Lemma 4 .4.r(u x)dx. we obtain (cf. . . note that (4. 0 Here (4..4) is automatic in some simple examples like birthdeath processes or p = 2 . where 7r2+) = QiµBilri/p.3* f uB(x) z/^. we need two lemmas.2. (4.2)(4.3iBi(x)YPi(u .10) Q*B*(u)+. COMPARISONS WITH THE COMPOUND POISSON MODEL 169 Stoyan [352]. = aP or b1 = ..1..2.10) and (4..13* J0 u 0*(u .* For the proof..4) say basically that if i < j .x) of i and using Lemma 4.8) ^j Tri/iBd(x) . = b. Proof of Theorem 4.4) hold..x)B*(x) dx.1 Assume that conditions (4. ^i 7ri = 1. < a.. it follows by a standard .x)dx _ /3*B*(u) + f u / ^ t=1 > 3 * B* ( ) + f (4. then j is the more risky state .1) which with basically the same proof can be found in Asmussen & Schmidt [49]. 3 (a) P.1 for the first term in (4. and it is in fact easy to show that Vii(u ) < t/j(u) (this is used in the derivation of (4.
of (4. u To see that Proposition 4.170 CHAPTER VI. that P P /^ 1r1NiµBi /^2 /^ ^i/ji pBi < 1il3i i=1 i=1 (4. dominates the solution 0* to the renewal equation (4. µB2 = 104.4 is not vacuous.1 of [145] for a formal proof) that z/ii(u) converges to the ruin probability for the compound Poisson model with parameters . i=1 i=1 7'r(0) _ EFioiwi(0) . Rolski & Schmidt [32].4 Assume that . Notes and references The results are from Asmussen.. = 102. For u = 0. u Here is a counterexample showing that the inequality tp* (u) < V)... it will hold for all sufficiently large u.(0) = V. this ruin probability is /3iPBi.0. Q2 = 1. As is seen.* (0). and from this the claim follows. (u) may fail for some u./3*.h. µB. MARKOVIAN ENVIRONMENT argument from renewal theory that tk. Frey.3µi < 1 for all i. (4. Using (4. except possibly for a very special situation .(0) < b *'(0) for e small enough. 4b Ordering of adjustment coefficients Despite the fact that V)* (u) < *.8) we get P P '*' (0) = 3* + /3*1* (0) _ > lri'3qqi • E 7i/ipBi .. 0.s.h. they are at present not quite complete.. it is sufficient to show that 0'. Bi as e J. let = ( 1/2 1/2 ) . of order 101.s.1 and Proposition 4.0*• i=1 But it is intuitively clear (see Theorem 3.2.11) i=1` and that A has the form eAo for some fixed intensity matrix A0. Then i/i*(u) < . 01 = 103.4) is essential (the present author conjectures it is). What is missing in relation to Theorem 4.6). Recall that . (u) is not in general true: Proposition 4. Proof Since 0.11) is of order 104 and the r.6).r (u ) fails for all sufficiently small e > 0. Then the l.4 is the understanding of whether the stochastic monotonicity condition (4.3i.
which in view of EiEE 1ibi = 0 is only possible if Si = 0 for all i E E.a = E irirci(a).(a) > 0 for all a 0 0.12) iEE Theorem 4. It is clear that the distribution of X.g..14) is nonzero so that A"(0) > 0.6 Let (di)iEE be a given set of constants satisfying EiEE iribi = 0 and define A(a) as the eigenvalue with maximal real part of the matrix A + a(bi)diag• Then )t(a) > 0. e.. Hence if 5i 54 0 for some i E E. Then {(Jt.13) implies A(a) > 0 for all a. Further (see Corollary 11.5. it follows by Proposition A1.Jt=kI A (the return time of k) where k E E is some arbitrary but fixed state. The adjustment coefficient y* for the averaged compound Poisson model is the solution > 0 of rc*(ry*) = 0 where rc*(a) _ 13*(B*[a] . in particular . (4. cf. Xt)} is a Markov additive process (a socalled Markovian fluid model.4. This implies that A is strictly convex.13) (4. with strict inequality unless rci (y*) does not depend on iEE.ld) with generic cycle w = inf{t>0: Jt_54 k.7) )i is convex with A'(0) = lim EXt tioo t = iEE 70i = 0.14) A„(O) iioo varXt t t By convexity.4(b) that the limit in (4.5. Jt = i])' EE = vA+n(6. Now we can view {Xt} as a cumulative process (see A. with strict inequality unless a = 0 or bi = 0 for all i E E. (4. Proof Define X= f &ids.5.1) . COMPARISONS WITH THE COMPOUND POISSON MODEL 171 the adjustment coefficient for the Markovmodulated model is defined as the solution y > 0 of ic(y) = 0 where c(a) is the eigenvalue with maximal real part of the matrix A + (rci(a))diag where rci(a) = ai(Bi[a] . and by Proposition II.5 y < ry*. Lemma 4. Asmussen [20]) as discussed in 11.2 we have (Ei[e"X'.)a. is nondegenerate unless bi does not depend on i E E. 0 .a. (4.1) .g.
15) once more and letting e = 0 we get .Qi and Bi are fixed .e7r)1 (Ici(Y*))diage.6. the basic equation is (A + (rci(y))diag)h = 0. Rolski & Schmidt [32]. Notes and references Theorem 4. h(0) = e. The corresponding adjustment coefficient is denoted by ry(e). Thus y(e) * y* as e 10. (4.12) and rc*(y*) = 0. (4. improving upon more incomplete results from Asmussen. 4c Sensitivity estimates for the adjustment coefficient Now assume that the intensity matrix for the environment is Ae = Ao/ e.5 is from Asmussen & O'Cinneide [40].172 CHAPTER VI. we get rc (y*) > 0 which in a similar manner implies that u y < y*...p yi and compute 8y 8a a=0 In both cases. In the case of e. MARKOVIAN ENVIRONMENT Proof of Theorem 4.. we have 7rh' = 0. h depend on the parameter (e or a). Hence letting e = 0 in (4. whereas the . Here we put a = 1/e. y.15) Normalizing h by 7rh = 0. Hence rc (y*) > 0. note that y(a) + mins=1. Further a(1) = rc(y*) by definition of A(. 0 = ((ri(Y))diag + ery (4{('Y))diag)h + (A0 + e(?i'Y))diag)h'.16) Differentiating (4. multiply the basic equation by a to obtain 0 = (A0 + e(r£i(y))diag)h. Let bi = rci(y*). Then > risi = 0 because of (4.) and rc (•). h'(0) = (Ao . If rci(y* ) is not a constant function of i E E.. Frey.15) yields 0 = (Ii(y*)) diage + Aoh'(0) = (rci('Y*)) diage + (Ao . this implies that the solution y > 0 of K(y) = 0 must satisfy y < y*.eir)h'(0).5. where A. a = 1 in Lemma 4. Since ic is convex with rc'(0) < 0 . and our aim is to compute the sensitivity ay e a E=O A dual result deals with the limit a 4 oo.
which has recently received much attention in the queueing literature. then 8a a=o All rci (0) Notes and references The results are from Asmussen..i(7' *))diagh'(0)..18). THE MARKOVIAN ARRIVAL PROCESS 173 0 = 27'(0)(ri(`Y *)) diage + 2(ci('Y* )) diag h' (0) + Aoh" (0) . .18) 0 = 27'(0)p+27r(rs. and may have some relevance in risk theory as well (though this still remains to be implemented). (4.19) Then 'y ^ ryl as a ^ 0 and we may take h(0) = el (the first unit vector). Frey.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).8 If (4. . The analogue of Proposition 4. and we have proved: Proposition 4. 0 = (Ao + ry'(ii(Y)) diag )h + (aAo + (Ki(7'))diag)h'.17) by 7r to the left to get (4.5.7 8ry aE = 1 7r(ci ('Y*))diag ( Ao e7r)1(Xi(Y*))diage *=0 P Now turn to the case of a. We assume that 0 < y < 7i. Rolski & Schmidt [32].19) holds.17) (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. 5 The Markovian arrival process We shall here briefly survey an extension of the model. (4. Inserting (4. We get 0 = (aAo + ( lc&Y))diag)h.16) yields Proposition 4. p.20) Letting a = 0 in (4. (4.8 when ryi < 0 for some i is open. i = 2. multiplying (4. the intensity for such a transition (referred to as marked in the following) is denoted by Aii l and the remaining intensity .
Jt2)) (2.174 CHAPTER VI. let { Jt 1) }. refer to notation) { Jt k) }. u Example 5 . Note that the case that 0 < qij < 1. Here are some main examples: Example 5 . the definition of Bi is redundant because of f3i = 0.^) etc. then {Nt} is a Markov additive process if and only if it corresponds to an arrival mechanism of the type just considered. we may let {Jt} represent the phase processes of the individual interarrival times glued together (see further VIII. where qij is the probability that a transition i * j is accompanied by a claim with distribution. A(l) = tv.2 (SUPERPOSITIONS) A nice feature of the setup is that it is closed under superposition of independent arrival streams . Bii = Bi .2).2) A(1) = A(' 1) ® A(1. Thus . and that are determined by A = A(l ) +A(2) where A is the intensity matrix the governing {Jt}. Indeed. that Bii = Bi . and thus 1i = 0. j(2) } be two independent environmental processes and let E(k).i. but the point process of arrivals is not Poisson but renewal with interclaim times having common distribution A of phasetype with representation (v.2 for details).d. the definition of Bij is redundant for i i4 j. Again . For i = j. II.4). Bij = B. the claim surplus is a Markov additive process (cf. This is the only way in which arrivals can occur. A(1) = A .1 (PHASETYPE RENEWAL ARRIVALS) Consider a risk process where the claim sizes are i.(13i )diag. is neither 0 or 1 is covered by letting Bij have an atom of size qij at 0. MARKOVIAN ENVIRONMENT f o r a transition i + j by A . Jt = (Jtl). we use the convention that a1i = f3i where 3i is the Poisson rate in state i. A(l) = T. B. In the above setting. . T). The extension of the model can also be motivated via Markov additive processes: if {Nt} is the counting process of a point process. A(1'k) A(2 k1). the Markovmodulated compound Poisson model considered sofar corresponds to A(l) = (.6i ) diag. We then let (see the Appendix for the Kronecker E = E(1) x E(2). A ( 2) = A (2`1 ) ® A. and the marked transitions are then the ones corresponding to arrivals. with common distribution B.
iil. 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. i2i ... MARRIED. and that the policy then expires. or. Example 5 . more recently. where ik = 0 means that the kth policy has not yet expired and ik = 1 that it has expired. e. 11.. Thus... The versatility of the setup is even greater than for the Markovmodulated model.iN C17 AilO. iN.iN = a2. Similarly.iil..4 (A SINGLE LIFE INSURANCE POLICY ) Consider the life insurance of a single policy holder which can be in one of several states. Easy modifications apply to allow for • the time until expiration of the kth policy is general phasetype rather than exponential. possibly having a general phasetype sojourn time.. INVALIDIZED. iN. Bilo. Assume further that the ith policy leads to a claim having distribution Ci after a time which is exponential.. • upon a claim. after which it starts afresh.iN = C27 All other offdiagonal elements of A are zero so that all other Bii are redundant. In this way we can model.kl is redundant). DEAD etc..kj = Bik) B13 4k = Bak) 175  (the definition of the remaining Bij. iN = all BOi2...1i2.3 (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). RETIRED.g... WIDOWED.. with rate ai. superpositions of renewal processes. In fact .iN.. E = { WORKING. Hermann [193 ] and Asmussen & Koole [37] showed that in some appropriate . assume that there is a finite number N of policies.. claims occur only at state transitions for the environment so that AN2.}. u Example 5 . as the Markovian arrival process ( MAP). all Al i2.iN are zero and all Bi are redundant. THE MARKOVIAN ARRIVAL PROCESS Bij.. u Notes and references The point process of arrivals was studied in detail by Neuts [267] and is often referred to in the queueing literature as Neuts ' versatile point process . E 10.1i2 . the idea of arrivals at transition epochs can be found in Hermann [193] and Rudemo [313].5.iN. However . DIVORCED.. say... the kth policy enters a recovering state.
we talk of s as the 'time of the year'. MARKOVIAN ENVIRONMENT sense any arrival stream to a risk process can be approximated by a model of the type studied in this section : any marked point process is the weak limit of a sequence of such models . p(t) and B(t) are defined also for t t [0. where i f00 xB(°) (dx) _ .2) Note that p is the average net claim amount per unit time and µ* = p//3* the average mean claim size. but now exhibiting (deterministic) periodic fluctuations rather than (random ) Markovian ones. one limitation for approximation purposes is the inequality Var Nt > ENt which needs not hold for all arrival streams. let the period be 1. a claim arrives with rate /3(s + t) and is distributed according to B(8+0 . B* = J f B(t) ((*) dt. one needs to assume also (as a minimum) that they are measurable in t. continuity would hold in presumably all reasonable examples. p * = 0 p(t) dt. [248]. • Claims arriving at time t of the year have distribution B(t). Thus at time t the premium rate is p(s + t). Obviously. 1).176 CHAPTER VI. By periodic extension. 6 Risk theory in a periodic environment 6a The model We assume as in the previous part of the chapter that the arrival mechanism has a certain timeinhomogeneity. Lucantoni et at.p)/p. )3 t 1 J (6. 1). Some main queueing references using the MAP are Ramaswami [298]. For the Markovmodulated model.1) Then the average arrival rate is /3* and the safety loading rt is 77 = (p* . 0 < t < 1. from an application point of view. we may assume that the functions /3(t).3*µs • p = f /3(v) dv 0 0 (6. We denote throughout the initial season by s and by P(8) the corresponding governing probability measure for the risk process. Without loss of generality. • The premium rate at time t of the year is p(t). Let 1 1 /3* _ f /3(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). Sengupta [336]. Neuts [271] and Asmussen & Perry [42]. for s E E = [0. . Lucantoni [248].
respectively. let . Thus. 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 .t.3(t) = 3A(1 + sin 27rt). the conditional distribution .3) Note that A enters just as a scaling factor of the time axis. 0 Then (by standard operational time arguments ) {St} is a periodic risk process with unit premium rate and the same infinite horizon ruin probabilities. it turns out that they have the same adjustment coefficient. p* as an averaged version of the periodic model. Thus . The claim surplus process {St } two is defined in the obvious way as St = ^N° Ui . we shall see that for the periodic model increasing A increases the effect of the periodic fluctuations.w(t). since the added variation is deterministic. It is easily seen that . Many of the results given below indicate that the averaged and the periodic model share a number of main features.8. In contrast. B*. u Remark 6 . or. equivalently. the discussion in 111. St = SeI(t). in agreement with the general principle of added variation increasing the risk (cf. In particular.w(t)) dt).9). of the periodic model as arising from the compound Poisson model by adding some extra variability. not random.1) and Example 1. We u assume in the rest of this section that p(t) . Section 4b).6.1 As an example to be used for numerical illustration throughout this section.2 Define T 6(T) = p(t ) dt.10.3* = 3A.(3. RISK THEORY IN A PERIODIC ENVIRONMENT 177 In a similar manner as in Proposition 1. The arrival process {Nt}t>0 is a timeinhomogeneous Poisson process with intensity function {/3(s + t)}t>0 . for Markovmodulated model typically the adjustment coefficient is larger than for the averaged model (cf. Example 6 . The behaviour of the periodic model needs not to be seen as a violation of this principle.1. and thus the averaged standard compound Poisson models have the same risk for all A. 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. one may think of the standard compound Poisson model with parameters 3*. In contrast. the average compound Poisson model is the same as in III. and we recall from there that the ruin probability is 24 1 *(u) _ 3 5eu + 35e6u. (6.
Dassios & Embrechts [98] and Asmussen & Rolski [43]. 3 E(8)eaSt = h(s.1) dv . The exposition of the present chapter is basically an extract from [44]. [44] (the literature in the mathematical equivalent setting of queueing theory is somewhat more extensive.a) Proof Conditioning upon whether a claim occurs in [t. [101] .5. with some variants in the proofs.3(v)(B(vl [a] .f. r(u) _ inf It > 0 : St > u} is the time to ruin .al.. of the averaged compound Poisson model (the last expression is independent of s by periodicity).1) a = J8 . 0 (5)(u. a) is periodic on R. we start by deriving formulas giving the m.1) .(3(s + t)dt)e«St adt + /3(s + t)dt . given that the ith claim occurs at time t is B(8+t).e. Daykin et.4) At a first sight this point of view may appear quite artificial. To this end.g.178 CHAPTER VL MARKOVIAN ENVIRONMENT of U. Jt = (s + t) mod 1 P(8) .^8 [.Q(v) (B(„) [a] .(1 . and the ruin probabilities are 0(8) (U) = P(s )(r(u) < 00). and define h(s.tc* (a)] dv then h (. we obtain E.8).3(s + t)dt[B(8+t)[a] .g.a) = exp { .T) = P(8)(r(u) <T). i. t + dt] or not.. The claim surplus process {St} may be seen as a Markov additive process. Notes and references The model has been studied in risk theory by. a) etw*(a) h(s+t.a.a .1]) .adt +. with the underlying Markov process {Jt} being deterministic period motion on E = [0. e.5 (see in particular Remark 11.east B(8+t) [a] east . (6. 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. 6b Lundberg conjugation Motivated by the discussion in Chapter II. As usual.(8) [eaSt+dt I7t] = = (1 .a be the c. . 1).. of the claim surplus process.f. let f 8+1 tc *(a) _ (B* [a] . J Theorem 6 .g.s . see the Notes to Section 7).
5.log h(s.t}t>o = h(s.3(s + t)[D(8 +t)[a] . a) as well as the fact that rc = k` (a) is the correct exponential growth rate of Eeast can be derived via Remark 11.1)dv  o h(t.6. With g the infinitesimal generator of {Xt} = {(Jt.9) eastt. a). B) eoSt t. a) = exp I f t3(v)(kv)[a) . u Remark 6. it then suffices to note that E(8)Le.0(s + t)dt[B(8+t)[a] . E (8)east (a +.adt +.1]) . a) Corollary 6.t. 0) exist and are finite.(8)east 179 = = = = = E(8)east (1 .5 The formula for h(s) = h(s. so that obviously {Lo.1]) .6 . RISK THEORY IN A PERIODIC ENVIRONMENT E(8)east+ dt d Et.* (a) h(s. a) ..3. According to Remark 11. 9) is a P ( 8)martingale with mean one.t = 1 by Theorem 6.t} is a multiplicative functional for the Markov process { (Jt. h(s + t.4).2.(e) Let = h( h(Jo. + v)(B([a] .1]. we can write Lo Jt. St)} and . Proof In the Markov additive sense of (6. a) Thus E(8)east = h(s + t. dt log E(8)east a + f3(s + t) [B(8+t) [a] . St)} . a) et.s. a) h(s + t. a) = h(s. 0) P(8)a. at + f log h(s + t.9 as follows.c* (e) {Le.1)dv l og E(8) et where atetk•(a) h(t.4 For each 0 such that the integrals in the definition of h(t .
g. That is.'y). (ii) use Markovmodulated approximations (Section 6c).f. it follows by Theorem II.y) = eayh(s). such that for any s and T < oo. Proof (i) Check that m. A further important constant is the value yo (located in (0. the requirement is cha(i. Lemma 6 . (iv) finally.1) . .3(s)B(s) [a]h(s). P(s) (T(u) < oo) = 1 for all u > 0. Sdt) = h(s + dt) eadt (1 . St)} with governing probability measures Fes).60(t) = a(t)B(t)[0].180 CHAPTER VI.2. That is.6 The P(s). 0) = h(s) + dt {ah(s) . Proposition 6. the restrictions of Plsi and Pest to Ft are equivalent with likelihood ratio Le.1. J s [.tc] dv} (normalizing by h(0) = 1). yo is determined by 0 = k* (70) = QB*. 0 < s < 1. of St is as for the asserted periodic risk model. as above E (s) ha(Jdt.3(v)( Bi"i [a] . cf. However. y solves n* (y) = 0.7 When a > yo. MARKOVIAN ENVIRONMENT ha(s. we put for short h(s) = h(s. B(s). see [44] for 11 a formal proof.0) = Kh(s).3(s)h(s) + h'(s) +. [70] . Now define 'y as the positive solution of the Lundberg equation for the averaged model. For each 0 satisfying the conditions of Corollary 6.3(s)ks)[a]h(s)} ah(s) 13(s)h(s) + h'(s) +.5 that we can define a new Markov process {(Jt.6 ( s ) exp { 0( s )&s) [a] + tc . correspond to a new periodic risk model with parameters ex .3(s)dt • B(s)[a]h(s) = gha(s.4. Bet)(dx) = ^ B(t ) (dx).3. Proposition 6. ( iii) use approximations with piecewiese constant /3(s). When a = y.T. Equating this to rch (s) and dividing by h(s) yields h(s ) = h(s) = a + .(3(s)dt) +.a . That rc = is*(a) then follows by noting that h(1) _ u h(0) by periodicity. say. ry)) at which n* (a) attains its minimum.
T) = h(s. xEJ 0 (s)b(8)(x) > 0.9) and noting that weak convergence entails convergence of E f (^(u). considered with governing probability measures { E(8) }E[ . the Markov process {(^(u).2). q) = eryx/h(q)).6(v) dv Jo ' xe«xB (°) (dx) r^ xe«xB'(dx) = Q'B' [ a] = ^' J 0 = ^c"'(a) + 1. The relevant likelihood ratio representation of the ruin probabilities now follows immediately from Corollary 11.8) (6.1. a) e«uE(8 ) e «^ . Corollary 6. B(oo)). 9(u)) for any bounded continuous function (e. a) TI h(9(u). a) a > ry0 (6. the mean number of claims per unit time is p« 181 = Jo 1.4. and we refer to [44]. T(u) < (6. e(cc)) Letting u > oo in (6. s E I.u is the overshoot and 9(u) = (T(u) + s) mod 1 the season at the time of ruin. we get: .9(u))} u>0. a)e«uE (a iP(s) (u) = h( s)e7uE(` ) h(O(u)) To obtain the CramerLundberg approximation from Corollary 3. The proof involves machinery from the ergodic theory of Markov chains on a general state space.1) the distribution of (l: (oo). f (x. say s0.9 Assume that there exist open intervals I C [0. has a unique stationary distribution. u which is > 1 by convexity.9) 0(')(u) = h(s.g. have components with densities b(8)(x) satisfying inf sEI. 1). and no matter what is the initial season s. (6. J C R+ such that the B(8). we need the following auxiliary result . Lemma 6 . ^(u) = ST(u) . 0(u)) * (b(oo).6. which is not used elsewhere in the book. Here and in the following.10) Then for each a. RISK THEORY IN A PERIODIC ENVIRONMENT Proof According to (6.7) h(B(u). Wu).8 The ruin probabilities can be computed as (u)+T(u)k'(a) ^/i(8) (u.2.
it does not seem within the range of our methods to compute C explicitly.) C = E1 h(B(oo)) u + oo.182 CHAPTER VI.1 In contrast to h. 6. this provides an algorithm for computing C as a limit. A=1/4 A=1 A=4 0 Figure 6.6 for the Markovmodulated model: Theorem 6 .16. where e. MARKOVIAN ENVIRONMENT Theorem 6. At this stage . Noting that ^(u) > 0 in ( 6.9). Vi(8) (u) .W.10) of Lemma 3. (6. Among other things. Theorem 6 . For our basic Example 6 .11) Note that ( 6.1. 10 shows that certainly ry is the correct Lundberg exponent.ir) } Plots of h for different values of A are given in Fig.11) gives an interpretation of h(s ) as a measure of how the risks of different initial seasons s vary.Ch(s)ery".1. elementary calculus yields h(s) = exp { A C 2^ cos 2irs  4^ sin 21rs + 11 cos 41rs . where C(o) = 1 + info < t<i h(t) . 11 7/'O (u) < C+°)h(s) ery". illustrating that the effect of seasonality increases with A. 1.10 Under the condition (6. we obtain immediately the following version of Lundberg ' s inequality which is a direct parallel of the result given in Corollary 3. which may provide one among many motivations for the Markovmodulated approximation procedure to be considered in Section 6c.
16 In order to apply Theorem 6. . where ay is the unique solution of W(ay) =y• (6.11 as well as it supplements with a lower bound..w) .6. Theorem 6.(ay) > 0 when y < 1/ic' (7). r. in our basic example with A = 1. we obtain Co) = 1. (ay).7x j dx _7x } _ 6w + 6(1 . e7 ( yx)B(t)(dy) > Then for all $ E [0. yu) 000 (u) .(8) (u. whereas ay < y. RISK THEORY IN A PERIODIC ENVIRONMENT Thus. 1 (6.14) < C+)(y)h(s) e7yu. the proofs are basically the same as in Section 3 and we refer to [44] for details. Theorem 6 .47r sin 27rs + 167r cos 47rs .42 so that 183 tp(8) (u) < 1. Lundberg's inequality can be con siderably sharpened and extended.3x + (1 .42 • exp {J_ cos 27rs .17) (6. C_h(s)e7u < V. we substitute T = yu in 0(u.yr. we first note that the function fu° exu {w • 3e . Just as in IV.4.(s)(u) < C+h(s)e7".7e . We state the results below. Consider first the timedependent version of Lundberg's inequality.g.12) As for the Markovian environment model. (6.167r I Cu. 1 ) and all u > 0.0(8) (u+ yu) (6.13) Elementary convexity arguments show that we always have ryy > Y and ay > ry.w)e4u dx 9w + 7(1 .12 Let 00)(y) 1 Then info < t<i h(t. T) and replace the Lundberg exponent ry by ryy = ay . #c( ay) < 0 when y > 1/tc'('y).w)e4u . e.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.w ) • 7e u{w • 3e3x + ( 1 .13 to our basic example.15) The next result improves upon the constant C+) in front of eryu in Theorem 6. ay) • (6.
and in fact.\ 3 C+ = sup 6 exp { A (. Of course. yo). . completing a cycle .20). This observation motivates to look for a more formal connection between the periodic model and the one evolving in a finite Markovian environment. 1) for the environment).013. Then 7oudT . Thus C_ = 2 inf ex cos 2irs . the nth Markovian environmental process {Jt} moves cyclically on {1.16. but thereby also slightly longer.20 •exp { 2n cos 27rs ..T) < C+('Yo)h( s. for A = 1 (where 3 e0.g. we have the following result: Theorem 6 .\ = 0 . C+ = 1.181 s(u) < 1. and let 8 = er' (Y0).16) with 'y replaced by yo and h(t) by h(t.0.9 3 0<8<1 p 27r 47r 167r 161r 2 _ _e.013.1 sin 2irs + 16_ cos 47rs .I eu. Thus. Finally.66.19 } 0 <8<1 8 + cos 21rs Thus e.'Yo)e (6. 14 Let C+('yo) be as in (6. .1 sin 27rs + 1 cos 47rs .(8)(u. such a deterministic periodic environment may be seen as a special case of a Markovian one (allowing a continuous state space E = [0. MARKOVIAN ENVIRONMENT attains its minimum 2 /3 for u = oo and its maximum 6 /(7 + 2w) for u = 0. .19 I eu.. with s the initial season.66. 1). exp 2^ cos 21rs . 6c Markovmodulated approximations A periodic risk model may be seen as a varying environment model. The idea is basically to approximate the (deterministic) continuous clock by a discrete (random) Markovian one with n 'months'.184 CHAPTER VI. where the environment at time t is (s + t) mod 1 E [0. 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. n}.18) Notes and references The material is from Asmussen & Rolski [44].4^ sin 2irs + 16^ cos 41rs . 0 <'p(8)(u ) .cos 27rs .L sin 27rs + 1 I cos 47rs . 1/i18 1 s (u) > 0.
jEE of the risk process. it is desirable to have formulas permitting freely to translate from one setting into the other.7. Notes and references See Rolski [306]. Bi. 7 Dual queueing models The essence of the results of the present section is that the ruin probabilities i/ (u). since the settings are equivalent from a mathematical point of view. To this end. but others are also possible. . This queue is commonly denoted as the Markovmodulated M/G/1 queue and has received considerable attention in the last decade.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. (6. DUAL QUEUEING MODELS 185 within one unit of time on the average . so that the intensity matrix is A(n) given by n n 0 ••• 0 0 n n ••• 0 A(n) _ (6.20) be the claim surplus process of t>o the nth approximating Markovmodulated model. We want to choose the /3ni and Bni in order to achieve good convergence to the periodic model. A be the parameters defining the risk process in a random environment and consider a queueing system governed by a Markov process {Jt } ('Markovmodulated') as follows: • The intensity matrix for {Jt } is the timereversed intensity matrix At _ A ())i. z/'i (u. Let 0j. M(n) = Supt>o Stn). AE= Aii'r?/7ri• The arrival intensity is /3i when Jt = i. and the ruin probability corresponding to the initial state i of the environment is then Y'yn)(t) = F (M(n) > t). Thus.21) which serves as an approximation to 0(1)(u) whenever n is large and i/n s. one simple choice is Oni = 0( i .1 ((i 1)/n) ) and Bni = B . 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. We let {Stn)} (6.
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. ii (u) = it /3 P(W > u. MARKOVIAN ENVIRONMENT • Customers arriving when Jt = i have service time distribution Bi.1 Assume V0 = 0. Jo = j. Then Pi(T(u) < T. J* = i) = P. Proof Consider stationary versions of {Jt}o<t<T. JJ = i). (7.2). Now let In denote the environment when customer n arrives and I* the steadystate limit. . JT = Z).1) follows..T(V > u I J* = i). Proposition 7.3) 7ri where (V.P(V > u. (7.1) over j. and the virtual waiting time (workload) process {Vt}too are defined exactly as for the renewal model in Chapter V.2) and use that limF (VT > u. JT = i) = 'P. T) = 7ri 1 P. The first conclusion of that result then states that the events {T(u) < T. The actual waiting time process 1W1. JT = i} coincide.oo in u (7. Jo = i. 2 . JT = i) = P(V > u.3.=1 . and for (7. Taking probabilities and using the stationarity yields 7riPi(T(u) < T.. JT = j) = LjPj (VT > u. let T . just sum (7.3).186 CHAPTER VI. For (7. (7. (VT > u I JT = 2). I* = i). and (7.2 The relation between the steadystate distributions of the actual and the virtual waiting time distribution is given by F(W > u. Proposition 7. (7. JT = j} and {VT > u. • The queueing discipline is FIFO.1).1) 7ri In particular.2) Oi(u) = 1. J* = i) for all j. JT = j) = 7rjPj(VT > u. J*) is the steadystate limit of (Vt.n(VT > u.. I* )3i P(V > u. J* = i). {Jt }o<t<T• Then we may assume that Jt = JTt.4) where 0* = >jEE 7rj/3j.0i (u . Jt ). In particular. .
on average 0*T customers arrive in [0. Proposition 7. .I. Lemoine [242]. and further references (to which we add Prabhu & Zhu [296]) can be found there. a paper relying heavily on classical complex plane methods.o. and of these. T].4). DUAL QUEUEING MODELS 187 Proof Identifying the distribution of (W.7) of that paper.g. and Rolski [306].3) improving somewhat upon (2. and one has PI'>(rr(u) < T) = P('_T)(VT > u).3). With {Vt} denoting the workload process of the periodic queue.4) can be found in Regterschot & de Smit [301]. [243]. (7. u Notes and references One of the earliest papers drawing attention to the Markovmodulated M/G/1 queue is Burman & Smith [84]. on average /32TP(V > u. P(. >u. n=1 N However. see in particular Harrison & Lemoine [186]. A more probabilistic treatment was given by Asmussen [17].6) (7.T)(T(u) <T) = P(8)(VT > u).=i) a4.I *=i).7. see Regterschot & van Doorn [123]. In the setting of the periodic model of Section 6. P(W >u. the dual queueing model is a periodic M/G/1 queue with arrival rate 0(t) and service time distribution B(') at time t of the year (assuming w.. N * oo.l. if T is large. I* = i. I*) with the timeaverage .8) For treatments of periodic M/G/1 queue.4) and (7. P(1')(r(u) < oo) = P(')(00) > u).5) follows from (7. Taking the ratio yields (7. p < 1 then ensures that V(*) = limNloo VN+9 exists in distribution. The relation (7. that /3(t). a general formalism allowing this type of conclusion is 'conditional PASTA'. with (7. we have 1: I(W.1 is from Asmussen [16]. B(t) have been periodically extended to negative t). J* = i) see W > u. and (7. The first comprehensive solution of the waiting time problem is Regterschot & de Smit [301].7) (7.
This page is intentionally left blank .
. resp .At + p(R8) ds. U2.T) = F(T(u) < T). Thus in between jumps. {Rt} moves according to the differential equation R = p(R). . are i. and that the claim sizes U1.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 .T) = FloinfTRt< OIRo=u1 denote the ruin probabilities with/initial reserve u and infinite. 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..6. t] are Nt At = Ui (1. Thus. the aggregate claims in [0. with common distribution B and independent of {Nt}. finite horizon. i&(u. 189 .1) (other terms are accumulated claims or total claims). However .i. and the evolution of the reserve may be described by the equation Rt = u . z/i(u) = F IinffRt< 0IRo=u 1 (1.2) tk(u.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).
say at interest rate b.Vi(v) u > e(1 .p2. pi > p2 and p(r) = One reason could be competition. 1 .3 (ABSOLUTE RUIN) Consider the same situation as in Example 1. that {Rt} will reach level u before the first claim arrives. the payout rate of interest is Sx and absolute ruin occurs when this exceeds the premium inflow p. when x > p/S.4 Either i. Assume 0(u) < 1 for some u. That is.2 (INTEREST) If the company charges a constant premium rate p u but invests its money at interest rate e. rather than when the reserve itself becomes negative.190 CHAPTER VII. there is positive probability. If Ro = v < u. but when the reserve comes above v. Hence in terms of survival probabilities. Thus at deficit x > 0 (meaning Rt = x).p/S) r > p/S p5(p/5r) 0<r<p/5 Then the ruin problem for {Rt } is of the type defined above.e. and the probability of absolute ruin with initial reserve u E [p/S. Now return to the general model. No tractable necessary and sufficient condition is known in complete generality of the model.2. but assume now that the company borrows the deficit in the bank when the reserve goes negative. dividends are paid out at rate pi .'(u)) > 0 so that V'(v) < 1. where one would try to attract new customers as soon as the business has become reasonably safe. A basic question is thus which premium rules p(r) ensure that 'O(u) < 1. Proof Obviously '(u) < ilb(v) when u > v. we can put Rt = Rt + p/S. oo) is given by i (u + p/S). we get p(r) = p + er. P(r) _ p + e(r . In this situation.i(u) = 1 for all u. say e. RESERVEDEPENDENT PREMIUMS The following examples provide some main motivation for studying the model: 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. it seems reasonable to assume monotonicity (p(r) is u . Example 1. or o(u) < 1 for all u. i. Proposition 1.1 Assume that the company reduces the premium rate from pi to p2 when the reserve comes above some critical value v. However.
instead of (1. Here {Vt}two is a storage process which has reflection at zero and initial condition Vo = 0. INTRODUCTION 191 decreasing in Example 1. {Vt} decreases at rate p(v) when Vt = v (i. { Vt} remains at 0 until the next arrival). In case (b).2 once more. In between jumps.1 and increasing in Example 1. hence Rt < uo also for a whole sequence of is converging to oo. Hence ik(u) < 1 for all u by Proposition III. T] i n such a way that the events {r(u) <T} and {VT > u} coincide.b(u.3µB for all sufficiently large r. (b) If p(r) > /3µB + e for all sufficiently large r and some e > 0.f p(Vs) ds. B and (constant) premium rate p. Theorem 1.3.6 For any T < oo. . one can couple the risk process and the storage process on [0.2(d)).I3IB requires a more detailed analysis and that µB < oo is not always necessary for O(u) < 1 when p(r) 4 oo. the probability that Rt < uo for some t is at least tp(0) = 1 (cf. that u zPp(u . 296297): Theorem 1.. and P(Rt + oo) > 0.T) = P(VT > u). Proposition I1I. then l/i(u) < 1 for all u.e.5) and the process {Vt} has a proper limit in distribution .+ p(r) exists. We next recall the following results. This is basically covered by the following result (but note that the case p(r) . In particular. Let Op(u) refer to the compound Poisson model with the same 0.1. let uo be chosen such that p(r) > p = 0ILB + e for r > uo.1. cf.6) . (1.4) 0 and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0. Then if u > no.uo) < 1. then ?(u) = 1 for all u.2) we have t Vt = At . In case (a). say V. However.1. if and only if V)(u) < 1 for all u. we have z/i(u) <p(u .1. V = p(V)).o(uo) = 1 so that t/'(u) = 1 for all u by Proposition 1.uo) and. appealing to Proposition 111. Proof This follows by a simple comparison with the compound Poisson model. Starting from Ro = uo. (1.4.5 (a) If p(r) < /. and hence by a geometric trials argument. which was proved in 11. (1. Then 0(u) = P(V > u).2) for r sufficiently large so that p(oo) = limr. obviously infu<uo z/'(u) > 0. let uo be chosen such that p(r) < p = /3µB for r > uo. [APQ] pp.2(d). That is.
192 CHAPTER VII.8) as g(x) = p 1 {yo13e_6x +. say.8) is the rate of downcrossings (the event of an arrival in [t. say. where g(x) = p( ^ exp {. t + dt] can be neglected so that a path of {Vt} corresponds to a downcrossing in [t. It follows in particular that 0(u) = fg(Y)dy.8) as the rate of upcrossings. Corollary 1.Sx}. oo).y. and the other being given by a density g(x) on (0. the l. we thus need to look more into the stationary distribution G.7) Proposition 1. one having an atom at 0 of size 'yo. Then the ruin probability is tp (u) = f' g(y)dy. (1. Considering the cases y = 0 and 0 < y < x separately. this means that the rate of upcrossings of level x must be the same as the rate of downcrossings. say if p(r) goes to 0 at rate 1 /r or faster as r j 0.y)g(y) dy. B(x) = e. x] to (x.h. An attempt of an upcrossing occurs as result of an arrival.8) Proof In stationarity.6w(x) . of (1.6x and that w(x) < oo for all x > 0.Sx} dx.9) Proof We may rewrite (1. Oeax f x e'Yg (y) dy } = p) eaxa(x) . for the storage process {Vt}. Note that it may happen that w (x) = oo for all x > 0. of (1. yo ^ 1 + oo Q exp {. t + dt] if and only if Vt E [x.6 applicable. u Define ^x 1 w(x) Jo p(t) dt. we arrive at the desired interpretation of the r. x + p(x)dt]). In view of the path structure of {V t }. 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.s.s.Qw(x) .8 Assume that B is exponential with rate b. the flow of mass from [0. oo) must be the same as the flow the other way. Now obviously. (1.h. It is intuitively obvious and not too hard to prove that G is a mixture of two components.7 p(x)g(x) = tofB (x) + a f (x . and is succesful if the jump size is larger than x . Jo AX) (1. say when {Vt} is in state y. RESERVEDEPENDENT PREMIUMS In order to make Theorem 1.
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
b P/E dx /' P/ ' (p/  x)p/e 150/f I' (/3/E) (6P1'E.V < x)]0 + f P(V > p/E ) + eby fv (p/E .01'E) + (p/E)al aO l febP/E } IF (0 /0 jF From this (2.V. i. .f. The process {St} corresponds to {Wt} so that c(a) or2a2/2 .a) .pa. r (j3/E) In particular.3.200 CHAPTER VII. /^ u Example 2 .x) dx e. assume that {Wt} is Brownian motion with drift µ and variance v2. it follows that logH[a] = f 1 c(y)dy = 1 f '(pa/(a +y))dy f 0 0 Ey R/E 1 [pa + )3log 8 .g. with density x(3/e1aQ/e fV (x) _ e 6X ' x > 0. and the c.e. where V is Gamma(b. then {Rt} is the diffusion with drift function p+Ex and constant variance a2. 13/E).pa. RT(u) has an exponential distribution with rate (S) and hence E [H(RT(u))I r(u) < oo] L Pe6'r (P/C . From ic(a) = . As an example.3/E) By the memoryless property of the exponential distribution.3 is also valid if {Rt} is obtained by adding interest to a more general process {Wt} with stationary independent increments.2) follows by elementary algebra. Proof 2 We use Theorem 2.5 The analysis leading to Theorem 2.2y +µ ) dy .2) follows by elementary algebra. RESERVEDEPENDENT PREMIUMS u from which (2./3 log(b + a)] = log ePa/f (a + a ) e which shows that Z is distributed as p/E . of Z is IogH[a] = f ytc(y)dy = e fa (0. H(u) = P(Z r < u) = P(V > u + p/E) = (8(p + Eu)/E.13 /E) r (.3a/ (5 .
i.g. THE LOCAL ADJUSTMENT COEFFICIENT _ Q2a2 pa 4e E 201 I. 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). Corollary 2.g. [282]. it follows that the ruin probability is Cu) H(u) H(0) 11 Notes and references Theorem 2. The formula (2. A r.3 is from Harrison [185]. se e. Some of these references also go into a stochastic interest rate.3) was derived by Emanuel et at. [129]. it is also used as basis for a diffusion approximation by these authors. of the form Ei° p"X" with the X„ i. that the analysis does not seem to carry over to general phasetype distributions. [129] and Harrison [185]. and since RT = 0 by the continuity of Brownian motion. Z is normal (p/E.8. It must be noted.1) .d. write y* for the solution of the Lundberg equation f3(B[ry *] . Emanuel et at.v.p*. [283]. Paulsen & Gjessing [286] found some remarkable explicit formulas for 0(u) beyond the exponential case in Corollary 1. 134 (the time scale there is discrete but the argument is easily adapted to the continuous case). Gerber [157] p. [357]. see e.. 3 The local adjustment coefficient. Q2/2E). Paulsen [281]. for a martingale proof.Y*p* W*(u) < ery*u = 0. write Vi* (u) for the ruin probability etc. or to nonlinear premium rules p(•).. Further studies of the model with interest can be found in Boogaert & Crijns [71]. Gerber [155]. Delbaen & Haezendonck [104]. not even Erlang(3) or H3. as in the proof of Proposition 2.2 is a special case of a perpetuity.3. Logarithmic asymptotics For the classical risk model with constant premium rule p(x) . Paulsen & Gjessing [286] and Sundt & Teugels [356]. Goldie & Griibel [167].e. however.4 is classical. and recall Lundberg 's inequality .
let p* be a in (3. choose c(.E). (3. Proof of Theorem 3.'y ( x)) = 0 where r.w (u) J dt > c(3)eeu v 1 p(u+ t) .>o 7(x) > 0. choose uo such that p( x) > p* when x > u0E. Letting first E * 0 and next ry * T 5o yields the first statement of the theorem. and (for simplicity) that inf p(x) > (3µs . i. as will hold under the steepness assumption of Theorem 3. 1) and for a given E > 0. a first step is the following: Theorem 3 . Then lim sup u>oo u and e E''p(r) + 0..2) such that p(x) < c(.202 CHAPTER VII. When u > uo.3) When trying to extend these results to the model of this chapter where p(x) depends on x. The steepness assumption and p(x) + oo ensure 'y(x) * So. B(x) > C(2)e(ao+f)x for all x.ap(x).1 ). log ?i(u) < < 00 JO .*(u) . a) = f3(B[a] .log '(u)/u < ry*(1 .1. For the last asssertion . 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.4) we assume existence of y(x) for all x.e.1) .1 Assume that for some 0 < 5o < oo.5) which implies inf. and that p(x) * oo. i. (x. oo for all E > 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). If 60 s f 6o. e(1o+e)2 (x ) u > 00.. we will use the local adjustment coefficient 'y(x). RESERVEDEPENDENT PREMIUMS and the CramerLundberg approximation V. 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) . then log u (u) In the proof as well as in the remaining part of the section .e. c(. which in turn by Lundberg's inequality can be bounded by ery*(1E)" Hence limsup„. it holds that f3[s] T oo. as solution of the equation n(x.C*ef*". Let y* < So. (3. x * oo.1. obviously O(u) can be bounded with the probability that the Cramer Lundberg compound Poisson model with premium rate p* downcrosses level uE starting from u .i)eex. x>0 (3.
The rest of this section deals with tail estimates involving the local adjustment coefficient.g. . let 0e (u) be evaluated for the process only with 3 replaced by /0/e and U. (3. Then limelog l/ie (u) = I(u). the result is not very informative if bo = oo. Condition 3. The first main result in this direction is the following version of Lundberg's inequality: Theorem 3 .3. noting that in many cases the constant C is close to 1. and in particular. the limit is not u + oo but the slow Markov walk limit in large deviations theory (see e.13 below holds. 3. or (b) Condition 3.(u) > so. which essentially says that an overshoot r.. 2 Assume that p(x) is a nondecreasing function of x and let I(u) = fo ry(x)dx. Bucklew [81]). by cU2. However. The form of the result is superficially similar to the CramerLundberg approximation. Then .2 is also an approximation under appropriate conditions. UJU > x cannot have a much heavier tail than the claim U itself. and hence '(u) > c(4)eeuc( 2)e(do+e)u The truth of this for all e > 0 implies lim inf log V.3 Assume that either (a) p(r) is a non decreasing function of r. 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.7) CIO Remarks: 1. Therefore the probability that a claim arrives ( before the reserve has reached level u + v is at least c(. one can then assume that e = 1 is small enough for Theorem 3.e.0 are the same.13 is a technical condition on the claim size distribution B. (u) = O(u/e).3 to be reasonably precise and use e` (u) as approximation to 0 (u).2). (3. {Rte)} defined as in (1.v.1 only presents a first step. ruin will occur if the claim is at least u + v. If p(x) = pis constant . I. then Rte) = CRtie for all t so that V).4)eE" Given such an arrival. THE LOCAL ADJUSTMENT COEFFICIENT 203 where c.' (u) < eI("). 2.ea°/(ecf1)). Theorem 3 . the asymptotics u * oo and c . For e > 0. u Obviously. Theorem 3. 3) = (1 .6) The second main result to be derived states that the bound in Theorem 3.
3. RESERVEDEPENDENT PREMIUMS 4. Then y(x) = b . J0 ^oo g(x ) dx f AX) lexp IOW (X ) bx + b u 1 exp low(x) .4 Consider again the exponential case B(x) = eax as in Corollary 1. 3a Examples Before giving the proofs of Theorems 3. As typical in large deviations theory.bx} dx 1+0.6). One would expect the behaviour in 2) to be important for the quantitative performance of the Lundberg inequality (3.8 in terms of I(u) when the claims are exponential: Example 3 .1 + b f e. the logaritmic form of (3.204 CHAPTER VII. However.exp {/33w(u) .bu}. and r j 1 'Yo v(x)dx = bu  a J0 p(x)ldx = Integrating by parts. it is formally needed only for Theorem 3.(x) dx.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).3. rather than eI(u)).bx} dx = 1 + J0 dodx(x) exp {. we show how to rewrite the explicit solution for ti(u) in Corollary 1. 3. we consider some simple examples. 5.bx} dx fo 00 1 + [exp {/(3w(x) .8. u .bx} dx .(3/p(x).bx}]o + b /' oo exp low (x) .2. say.bx} dx oo exp low(x) bx dx 70 Ju r oo = b J exp low (x) .(iw(x) . First. we get = 1+ J" AX) exp {(3w (x) .
fa 7(x+u)dx/Edy o The analogue of (3.ev 0 O /E) J0 70 70 Yo This implies lim inf A. IE(u) = I(u)/e. and (3. (3.I ( u) fool..e. in the definition of AE converges to 0.3.fo 7(x)dx/Edy f . 3. (X) = µ(x). Choosing yo. > lime log e = 0 and AE * 0. and (3. oo) with drift µ(x) and variance a2 (x) > 0 at x. ry(x /b I u o e f0 °° e  e. we get r 00 e.fory(x+u)dxdy ( 3. the integral is bounded by 1 eventually and hence lim sup AE < lim sup a log 1 = 0.7) follows. note first that the appropriate slow Markov walk assumption amounts to u..3.3 in the particularly simple case of diffusions: Example 3. Similarly.fo 7(x) dx /E dy > av 'yo /Edy = E (1 . In particular.0. u .1. 191195) that 1P (U) = fu0 eI(v)dy = eI(u) follo e.I ( v )dy fo +u) dxdy . It is well known that (see Theorem XI.1/8 . BE * 0. THE LOCAL ADJUSTMENT COEFFICIENT and hence 205 f°° eI(v )dy . 70 > 0 such that 7(x) < 7o for y < yo.2. (u) = I(u) + AE .8) 7(x)dxdy 11 We next give a direct derivations of Theorems 3. applying the inequality 7(x + u) > 7(x) yields immediately the conclusion of Theorem 3.BE. (3. For Theorem 3.9) yields e log . Be = e log U000 e.2.2(X) = ev2(x) so that 7e(x) = 7(x)/e.f y(x)dxd y If 7(x) is increasing . 0. The appropriate definition of the local adjustment coefficient 7(x) is then as the one 2p(x)la2(x) for the locally approximating Brownian motion.5 Assume that {Rt} is a diffusion on [0.9 ) 11000 eI(v)dy f000 e.5) is infx>o 7(x) > 0 which implies that f °O .10 or Karlin & Taylor [222] pp.10) where AE = e log 000 e.
G.Q/p*.206 CHAPTER VII. 0.6/p* so that u 1 I (U) = bu .5. . E+o e*O By (3..1 3.(u) oo.+1 (u) = o( 1). As in Example 3. the slow Markov walk assumption means 5E = b/c. G. Then the solution of the Lundberg equation is y* = b . _ . Of course.5) and 7* = 5 ./3 1 AX dx.10) holds if we redefine AE as AE = flog (j °° efo 7(x)dx/edy _ E/5 I and similarly for B.6 Assume that B is exponential with rate S.e. we have 5 > 7o and get lim inf AE > lime log e . + Gq(u) + o(G9(u))• Gi (u) It should be noted . G. Thus 7e(x) _7(x)/e and (3.4.5 for risk processes with exponential claims is as follows: Example 3 . .. that the interchange of the slow Markov walk oo is not justified and in fact.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. 7(x) is typically not explicit. the results are suggestive in their form and much more explicit than anything else in the literature. however .6) exactly as in Example 3. Further. ) Note that this expression shows up also in the explicit formula for lk(u) in the form given in Example 3. . (u) representing the first few terms in the asymptotic expansion of I(u) as u + oo. . I. Nevertheless .. I(u ) = G1(u) + . .0. this leads to (3. Ignoring 1/5 in the formula there. lim sup Af < lim sup c log(1 .5. RESERVEDEPENDENT PREMIUMS The analogue of Example 3.0/e. 0 Now (3. so our approach is to determine standard functions Gl (u).0) = 0. the slow Markov limit a * 0 and the limit u walk approximation deteriorates as x becomes large.7o C 15 I I.5. > ..
and hence (3.1) leads to . say 1 is the upper limit and B(x) .:. .y/s)dy sn 1 1 f ' evy'7ldy = cse8r(T7) as s T oc. (3.ry*°p*.s)C' f "o o as s T S.3. THE LOCAL ADJUSTMENT COEFFICIENT Example 3 .c3 logu a= 1 J 0 a + bx 1/ ( c4ul 1/° a > 1 where c3 = c2 /b.clxale5x 207 (3. It follows from (3.3cse7*I7(77) . I(u) Pt. in the phasetype case . e. phasetype distributions (Example 1. the typical case is a = 1 which holds .12) with y > 1. 2. x T 1. Here B[s] is defined for all s and B[s] . . B[s] = 1 + s exB(x)dx = 1 +c1SF(a) ('+o(')) (S .11) with a > 0.8). This covers mixtures or convolutions of exponentials or.1/a). More precisely.C2p* C2 = (3clr( a))11'.g. y = 2 if B is uniform on (0.1/k).Y . 77 = 1 if B is degenerate at 1.1) leads to (S7T N Ocp a.c2 Su a dx ) Su a<1 Su . Hence (3.cs(1 . c4 = c2b 1/'/(1 . if the phase generator is irreducible ( Proposition VIII.8 Assume next that B has bounded support. more generally.x)n1.. ry* loge*+ g7loglogp*. For example. fu I(u) Su . 1) and 17 = k + 1 if B is the convolution of k uniforms on (0.7 Assume that B(x) . u(logu + r7loglogu). 1. u Example 3 .11) that b[s] * co as s f S and hence 7* T S as p* + oo.4) or gamma distributions.1 =$ f cse8 Sn f e"B(x)dx = e8 Jo s eIB ( 1 .
x f oo . ec78)2/2c7 dx C7 . (b) 'y(x) <'Yo(x)• Proof That 7(x) is nondecreasing follows easily by inspection of (3.9 As a case intermediate between (3. of U1 + v . The assumption implies that ru(t) .ru(TI)) . 1 = E. By convexity of the m . 3b Proof of Theorem 3.c8 log .•. 7 * .13) We get b[s] . 1 0 3e.f.1) .g. h 10.r„(Ti).r^. g.3 (B[s] . (3. RESERVEDEPENDENT PREMIUMS Example 3 . one could also have considered the increment ru (T1) .Ul up to the first claim (here ru (•) denotes the solution of i = p (r) starting from ru(0) = u).e7o ( u)(ul+u r. assume that B(x) CO x2/2c7.u is a nondecreasing function of u.12).11) and (3. I (u) c8u log u 0 where c8 = 2/c7.(t))dt. . Then: (a) y(x) and 7o(x) are also nondecreasing functions of x..4).10 Assume that p(x) is a nondecreasing function of x. this is only possible if 7o(v) 2 7o(u)• .15) Proposition 3.14) for the m . of the increment in a small time interval [0.log p*. Hence for u<V. h].css 2%rc7eC782/2. This leads to an alternative local adjustment coefficient 7o(u) defined as solution of 1 = Ee''o(u)(vi+u .(T1)) > Ee7o(u)(ul+vr»(Ti)).sp(u).2 We first remark that the definition (3.4) of the local adjustment coefficient is not the only possible one: whereas the motivation for (3.4) is the formula h logEues ( Rhu) .f. (3.208 CHAPTER VII.B[7o (u)] . (3.u .1 Cgs o"O 0 esxex2/2c7 dx = cgsec782/2 f .Ote7o( u)(u.
u > tp(u). THE LOCAL ADJUSTMENT COEFFICIENT For (b).17) shown for n and let Fu(x) = P(U1 + u . Hence 1 = EeYo(u)(U1+uru(T1)) < E.es'Yo(u)Fu(dx)} o0 e fo yo( x)dx j.16) Proof Define 411(n)(u) = P('r(u) < on) as the ruin probability after at most n claims (on = TI + • • • + Tn). we obtain „I.(n+l) (u) 1 .1) .ru(T1) < x).u[70(u)] fo eyo(x)dx .7o (u)p(u)• Since (3. u We prove Theorem 3. this is only possible if yo(u) > 7(u).2 in terms of 7o.3.Fu(u ) + J ^(n)(u . fa 7o(y)dy < u7o(u) < xyo (u) for x > u. We shall show by induction that (' Y'(n) (u) < e fo 'Yo(x)dx (3.4) considered as function of 7 is convex and 0 for y = 0. note that the assumption implies that ru(t) . The case n = 0 is clear since here To = 0 so that ik(°)(u) = 0. (3.10(b): Theorem 3. it is easily seen that fu x7o(y)dy < xyo (u).11 Assume that p(x) is a nondecreasing function of x. Hence „/. Separating after whether ruin occurs at the first claim or not. the case of 7 then follows immediately by Proposition 3.17) from which the theorem follows by letting n + oo.(n+1) (u) efo Yo(x)dxI^"Q exyo( I u u)Fu(dx )+ J .e70(u)(U1P(u)T1) 209 0 + 7o(u)p(u)' 0 <_ 00['Yo( u)] . Assume (3. Also. Then (u) < efo Yo(x)dx.x)Fu(dx) 00 U efo J = o (y) dYF (dx) )+f I 11 /' / 00 e f oFu fu dx) + of u :7o(Y)dYFu(dx) 00 J u 1 l` Considering the cases x > 0 and x < 0 separately.
Y*u /E. for either of Theorems 3.: y(u).n (starting from u/n) without that 2u/n is upcrossed before ruin.n) must first downcross unl.. resp. (3.2.nbe C*.10(a) for some of the inequalities. P k.3/e and U. C*e where the first equality follows by an easy scaling argument and the approximation by (3.I.E ( u/n) ^•e.10(b ) that the bound provided by Theorem 3.3 The idea of the proof is to bound { R( f) } above and below in a small interval [x .12 lim sup4^o f log O.x/n. {RtE)} (starting from u = un.. (u). 3c Proof of Theorem 3. in . and.n. let Op*. Proof For ruin to occur. we used also Proposition 3. x + x/n] by two classical risk processes with a constant p and appeal to the classical results (3.11 is sharper than the one given by Theorem 3.. ryk..2). pk n = uk_l. y* evaluated for p* = Pk. For these reasons.2.n = ku. op*.E (u/n) Now as e . Let Ck. the probability that ruin occurs in the CramerLundberg model with p* = pn. 3.n = sup p(x). yo(u) appears more difficult to evaluate than y(u). The probability of this is at least n n.E (u) denote the ruin probability for the classical model with 0 replaced by .n AX). RESERVEDEPENDENT PREMIUMS where the last identity immediately follows from (3..n so that n.n) pn niE (u /n) n n_1 n. given downcrossing occurs.E (u/n). and here it is easily seen that yo(u) .n <Z auk}l. (u) < I(u).n.n u k}1.210 CHAPTER VII. To this end. (un2.11 be reasonably tight something like the slow Markov walk conditions in Theorem 3. W O .. the value of {R(E)} at the time of downcrossing is < unl. 0 It follows from Proposition 3.n. by €U=.15).E (u/ n) Y'E (un . define uk.3). Also.n.e (u) = v'.3 is required. we have chosen to work with y(u) as the fundamental local adjustment coefficient.n inf n uk1. 0. Lemma 3. in accordance with the notation i/iE (u). However. 0.n) > k =1 II v ^k n. Further.3).
e. (3. 11 Theorem 3.E (urn) < \ *I.7k. uk_1..nk=1 limsupelogv). k=1 k=1 n u _ nE7 k.E (u/n) OP +^p•. in obvious notation one has tC (x) = y(x)/e. (u) CIO < Letting n 4 oo and using a Riemann sum approximation completes the proof. .3. It follows that n log V'C (u) k =1 log Ypk.19) . In case (b).n. 3 now follows easily in case (a).13 There exists a r.E (u/n) Op•. 0 with n and u fixed. THE LOCAL ADJUSTMENT COEFFICIENT particular. 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. *p•. we need the following condition: Condition 3. so that Theorem 3.n cE (2u/n) Ck ne7k.nu /fn( 1 Ck  e.a( u)z.i. 40 Combining with the upper bound of Lemma 3.n = sup ?'(x)..18) (ii) the family of claim overshoot distributions is stochastically dominated by V.nu /En) o(1)).2 gives 7PE (u) < eIi"i/f = lim inf Clog 0E (u) > I (u).n <X<Uk.n + 0(1).12 completes the proof. ne7k. Indeed . for all x ..! (u/n) n n m 7k. y > 0 it holds that F(U>x +yIU>x) B(x + y) < F (V > y). i..log Ck. B(x) (3.n .nu/en(1 + where o(1) refers to the limit e . also ryk. 211 Clearly. /' (u/n) 'T nk. since ry' is an increasing function of p'.F (2u/n). v.
eV) • P (T(E) (u.. infx>2u /n P(x) . u /n) < oo] l = = < E [OE (u/n . u/n) < oo) EV). where El is the contribution from the event that the process does not reach level 2u/n before ruin and E2 is the rest. (u/n .E (u/n . u/n) < oo) . Write EO.5) and the standard formula for b(0). Then the standard Lundberg inequality yields El < E?. . u/n) < oo] E [OE (u/n ..E(E) (u.. The probability of ruin in between two downcrossings is bounded by Epp .1 n. RESERVEDEPENDENT PREMIUMS To complete the proof. T(E) (u.2y 1 ' . let v < u and define T(E) (u.212 CHAPTER VII.R<) (u v). u/n)) I T(E) (u. u/n) < oo] .18) for the last equality). T() (u.^(E) (u.n V.of:>2 in n(x).E (0) cf.( .nu /EnE [e71.EV) = El + E2. N with EN < 1 = infx>2u/nA(x) = 0(1).V) = e71 nu/Eno(l) (using (3. Ei + E2 < e71. ) (u u /n)) . v ) = v . Then Y'E (u) ^(E) (u.nu/En0(1) .v. u /en 0(i) _n so that E2 < e2ryl nu/En0(1). For E2. u/n)) .EV) = EiI 1 . V < u/En] + P(V > u/En) (u/En . (3. P (T(E) (u.^'' = E [ .QEU 1 . (R.EV) = e.. we first note that the number of downcrossings of 2u/n starting from RoE) = 2u/n is bounded by a geometric r.n < ery1. v ) = inf { t > 0 : R(c ) < v R) = u } . (u/n .E (2u/n .
(u) 40 213 lim inf e log(Ei +E2) + logP (r(`) (u.) = exp . the approximation (3. Similarly.3.7(x)) (3.3EU) (3.)Ui } .7) for ruin probabilities in the presence of an upper barrier b appears in Cottrell et al. the results are from Asmussen & Nielsen [39]. l o JJJ o . u Notes and references With the exception of Theorem 3.21) (the initial condition is r(0) = u in both cases). u/n) < oo { 40 )I U nryl n+liminfelogP (T(')(u. 0 ) (= p(x) .=1 J An approximation similar to (3. where the key mathematical tool is the deep WentzellFreidlin theory of slow Markov walks (see e . T) is maximized over T by taking T as the time for (3.)ds + Y(R2. one can in fact arrive at the optimal path by showing that the approximation for 0(u. whereas the most probable path leading to ruin is the solution of r(x) _ k (x. it might be possible to show that the limits e . s). Bucklew [81]). 0 and b T 00 are interchangeable in the setting of [89]. Comparing these references with the present work shows that in the slow Markov walk setup. they also discuss simulation based upon 'local exponential change of measure' for which the likelihood ratio is ( /'t /'t Ns Lt = exp S .J y(Rs)dR. s) as in (3.20) (with ic(x. Typically.13. the risk process itself is close to the solution of the differential equation r(x) _ r (x.g.1. Whereas the result of [122] is given in terms of an action integral which does not look very explicit. 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].u/n) < oo) CI  > u n n ryi n' i=1 Another Riemann sum approximation completes the proof. Djehiche [122] gives an approximation for tp(u.T) = P „(info<t <T Rt < 0) via related large deviations techniques. THE LOCAL ADJUSTMENT COEFFICIENT Hence lim inf e log Ali.7) then comes out (at least heuristically) by analytical manipulations with the action integral.4) and the prime meaning differentiation w.t.r.J r(Rs)p(R.21) to pass from u to 0. .
214 CHAPTER VII.g. however.3. e. RESERVEDEPENDENT PREMIUMS the simplest being to require b[s] to be defined for all s > 0 (thus excluding . to point out as a maybe much more important fact that the present approach is far more elementary and selfcontained than that using large deviations theory. the exponential distribution ). For different types of applications of large deviations to ruin probabilities .. We should like. see XI. .
then the problem may admit an algorithmic solution involving a reasonable degree of computational effort if one allows for the more general assumption of phasetype structure. if a problem can be solved explicitly when the relevant distributions are exponentials.Chapter VIII Matrixanalytic methods 1 Definition and basic properties of phasetype distributions Phasetype distributions are the computational vehicle of much of modern applied probability. Note that since (1. This implies in particular that the intensity matrix for { it } can be written in blockpartitioned form as T 0 0 . on Eo = E U {A} where A is some extra state which is absorbing. A distribution B on (0. that is. 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<. if v = (vi)iEE is a probability distribution. A proper knowledge of phasetype distributions seems therefore a must for anyone working in an applied probability area like risk theory. F (Jt = A eventually) = 1 for all i E E 1 and where all states i E E are transient. Typically. P. We often write p for the number of elements of E. and not in other cases. oo) is said to be of phasetype if B is the distribution of the lifetime of a terminating Markov process {Jt}t>o with finitely many states and time homogeneous transition rates. we write Pv for the case where Jo has distribution v so that Pv = KER viPi• 215 . More precisely. refers to the case Jo = i.1) is 'Here as usual .
0 = t11. and the phasetype distribution is the lifetime of a particle with constant failure rate /3. an exponential distribution with rate parameter . j. 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. that is. the ith component ti gives the intensity in state i for leaving E and going to the absorbing state A.3. The initial vector a is written as a row vector. Then a = a1 = 1. (1. B(t) = Fa(^ < t ). and we have t = Te.1 Suppose that p = 1 and write . k}. We now say that B is of phasetype with representation (E. Equivalently. In particular. Here are some important special cases: Example 1 . i.2) The interpretation of the column vector t is as the exit rate vector. A convenient graphical representation is the phase diagram in terms of the entrance probabilities ai.1 The phase diagram of a phasetype distribution with 3 phases.e. the exit rates ti and the transition rates (intensities) tij: tj 3 aj ai i ti tk tjk FkJ ak Figure 1. t1 = /3.T)) if B is the Padistribution of the absorption time C = inf{t > 0 : it = A}. T) (or sometimes just (a.e. i. E = {i. C is the lifetime sup It > 0 : Jt E E} of {Jt}. 0 2this means that tii < 0. T is a subintensity matrix2. tij > 0 for i 54 j and EjEE tij < 0 . Thus the phasetype distributions with p = 1 is exactly the class of exponential distributions. a. MATRIXANALYTIC METHODS the intensity matrix of a nonterminating Markov process.216 CHAPTER VIII.
2 corresponding to E = {1. 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... 0 0 0 0 S 6 .... 0 0 0 T= t= 0 ••• S S 0 0 0 0 0 0 ..1. 0 SP 0 and the phase diagram is (p = 2) . 6. a = (1 0 0 . . .. 0 S 6 Example 1. 0 ••• 0 0 Sp1 0 0 t= 0 0 00 •..1)!e Since this corresponds to a convolution of p exponential densities with the same rate S.. .. so that the density is P E ai6ie6..3 The hyperexponential distribution HP with p parallel channels is defined as a mixture of p exponential distributions with rates 51.. . 0 •. 00)) S s o . ... p}.x i=1 Thus E _ Si 0 T 0 S2 0 0 . PHASETYPE DISTRIBUTIONS 217 Example 1. .
3 0 Example 1 .g.1)"n! aT"e. the backwards equation for {Jt} (e. Theorem 1 . MATRIXANALYTIC METHODS Figure 1.aeTxe. and is defined as the class of phasetype distributions with a phase diagram of the following form: 1 617 ti t2 2 b2.e.g. T). see A. 5 Let B be phasetype with representation (E. Then for i . Proof Let P8 = (p ^) be the sstep EA x EA transition matrix for {Jt } and P8 the sstep E x Etransition matrix for {Jt} . the Erlang distribution is a special case of a Coxian distribution. Recall that the matrixexponential eK is defined by the standard series expansion Eo K"/n! 3. the restriction of P8 to E. [APQ ] p.f is B (x) = 1 . 36) yields s d. a. ds^ = ds' = ttlaj + tikpkj. i. E t ikp kj = kEE kEE 3For a number of additional important properties of matrixexponentials and discussion of computational aspects . Then: (a) the c.f. dp.3 . p:. B[s] = f0°O esxB (dx) is a(sI T)lt (d) the nth moment f0°O xnB(dx) is (.218 CHAPTER VIII. (c) the m.t2 yt bP.d.1 tP1 1 Figure 1. .4 For example. The basic analytical properties of phase type distributions are given by the following result . (b) the density is b(x ) = B'(x) = aeTxt. j E E.4 (COXIAN DISTRIBUTIONS) This class of distributions is popular in much of the applied literature.
T) 1t.g. for n = 1 we may put ki = Ei( and get as in (1.5) as hi(tii + s) = ti  t ij hj. hj .T) 't = (. = aPxe.12) for integrating matrixexponentials yields B[s] = J esxaeTxt dx = a ( f°°e(81+T)dx ) t a(sI . For (c).e.f. of the initial sojourn in state i.1 ) n +l n ! a (s I + T ) .p. tij / . and since obviously P° = I. and since b[s] = ah. d8 P8 = TP8. Alternatively. Since 1 . d" dsn a (.jEE B'(x) _ cx Pxe = aeTxTe = aeTxt (since T and eTx commute). i. the solution is P8 = eT8. B(n)[0] = _ Alternatively. Then h tit ti + ti3 h j .g.1.f.n1t = (1)nn!aTn1Te (1)nn! aTne.p.5) Indeed .6) . PHASETYPE DISTRIBUTIONS 219 That is.B(x) = 1'a (( > x) = P. or w. and (b) then follows from 1: aipF. .f. After that. this means in vector notation that (T + sI)h = t. we i w.g. define hi = Eie8S. in which case the time to absorption is 0 with m . i.tii is the rate of the exponential holding time of state i and hence (tii)/(tii . we arrive once more at the stated expression for B[s]. ti/ .5) ki = 1 + tii L jj:Ai tii (1.tii and have an additional time to absorption either go to state j which has m . Part (d) follows by differentiating the m.n lt ..tii tii .tii we go to A.f. h = (T + sI)1t. Rewriting ( 1.g. the rule (A. 1.. (1) n+1n!aT .s) is the m .s j# tii i (1. j#i jEE tijhj + his = ti. (Jx E E) = this proves (a).s I .
we get the density as 9 9 6 (1 1) 10 7 1 0 10 2 aeTyt = e x .s. T= 2 111 so that 2 2 Then (cf. making the problem trivial. another the case p = 2 where explicit diagonalization formulas are always available. 0 Example 1. Consider for example 3 9 a= (2 2). One obvious instance is the hyperexponential distribution."n! ( ( l 2 2 ) 17 9 0 \ 1 / 10 10 32 n! 35 6" +n!353 Similarly.6 Though typically the evaluation of matrixexponentials is most conveniently carried out on a computer.220 CHAPTER VIII. see the Appendix. MATRIXANALYTIC METHODS which is solved as above to get k = aTle.7) the diagonal form of T is 9 9 1 9 T 10 7 10 70 1 10 6 10 7 0 70 9 1 10 where the two matrices on the r. are idempotent. Example A3. This implies that we can compute the nth moment as (1)"n! aT "e 1"n! 1 1 22 9 9 10 70 7 1 10 10 1 9 +6.h. there are some examples where it is appealing to write T on diagonal form.
g.7 If B is phasetype with representation (v.f. a random variable U having a defective phasetype distribution with representation (a. This is the traditional choice in the literature. and in fact one also most often there allows a to have a component ao at A.e a mixture of a phasetype distribution with representation (a/llall. • The phasetype distribution B is zeromodified.4b for definitions and basic rules): Proposition 1.4. i.T) with weight hall and an atom at zero with weight 1 . or one just lets U be undefined on this additional set.1. i. There are two ways to interpret this: • The phasetype distribution B is defective.T).29) and Proposition A4.hall.11aDD. hail = E=EE a. B[Q] of B is f3[Q] = J e'1zB(dx) _ (v (9 I)(T ® Q)1(t ® I). 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).7) Proof According to (A. 0 Sometimes it is relevant also to consider phasetype distributions. T) is then defined to be oo on a set of probability 1. where the initial vector a is substochastic. . then the matrix m. 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.e 11BIJ = 1laDD < 1. (1. 5 and serves at this stage to introduce Kronecker notation and calculus (see A. < 1.
the conditions of Proposition 1. it is easily seen that the asymptotic form of the tail of a general phasetype distribution has the form B(x) _ Cxkenx. the Erlang case). . Rolski. but the relevant T is not irreducible. the text is essentially identical to Section 2 of Asmussen [26].g. Lipsky [247]. v.222 CHAPTER VIII.8).hve7x. O'Cinneide [276] gave a necessary and sufficient for a distribution B with a rational m.8 Let B be phasetype with representation (a. In older literature. T). and we have eTx . All material of the present section is standard. the result follows (with C = (ah)(ve)). see his book [269] (a historical important intermediate step is Jensen [214]). but todays interest in the topic was largely initiated by M.q be the eigenvalue of largest real part of T. 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. where C. Then the tail B(x) is asymptotically exponential. 2. let v. here k = p1). See in particular the notes to Section 6. 77 > 0 and k = 0..Ce7'. assume that T is irreducible . (or Laplace transform) are often used where one would now work instead with phasetype distributions. (1.f. Other expositions of the basic theory of phasetype distributions can be found in [APQ].F. x * oo. not only in the tail but in the whole distribution.g.8 are far from necessary ( a mixture of phasetype distributions with the respective T(') irreducible has obviously an asymptotically exponential tail. h can be chosen with strictly positive component.. cf. Example A5. B(x) . Using B(x) = aeTxe .1 of the Appendix. Schmidli. Schmidt & Teugels [307] and Wolff [384]. 1. we give a criterion for asymptotical exponentiality of a phasetype distribution B. distributions with a rational m. In Proposition A5. but in many practical cases. Neuts. i is real and positive.4c). MATRIXANALYTIC METHODS la Asymptotic exponentiality Writing T on the Jordan canonical form. Here is a sufficient condition: Proposition 1. Notes and references The idea behind using phasetype distributions goes back to Erlang. let . No satisfying .f. one has k = 0.8) Proof By PerronFrobenius theory (A. 0 Of course. h be the corresponding left and right eigenvectors normalized by vh = 1 and define C = ah • ve . The Erlang distribution gives an example where k > 0 (in fact. cf.
and U(A) is then the expected number of replacements (renewals) in A.. (2.1 Consider a renewal process with interarrivals which are phasetype with representation (cr. or the density is available ) is. The explicit calculation of the renewal density (or the renewal measure) is often thought of as infeasible for other distributions... If B is exponential with rate 0.r.+UnEA). Let U1. known. is Markov and has two types of jumps .: U1 + ..f. + U0 is 0 .. Lebesgue measure. +UnEA} 00 = EEI(U1 +.g. .. however. Then the renewal density exists and is given by u(x) = ae(T+ta)xt. the renewals form a Poisson process and we have u(x) = 0. we denote the density by u(x) and refer to u as the renewal density. the problem has an algorithmically tractable solution if B is phasetype: Theorem 2.. Jt={Jt?ul}. RENEWAL THEORY 223 algorithm for finding a phase representation of a distribution B (which is known to be phasetype and for which the m..d. A related important unsolved problem deals with minimal representations: given a phasetype distribution . we refer to U as the renewal measure.i. but nevertheless.1 of the Appendix. 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 +. if U is absolutely continuous on (0. U2. For this reason.1.. but is in part repeated below. be i.. 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.. ..1) Proof Let {Jtk)} be the governing phase process for Uk and define {Jt} by piecing the { J(k) } together.2.t. JtJt1) Then { 0<t<U1 . oo) w. as the lifetimes of items (say electrical bulbs) which are replaced upon failure. n=O We may think of the U. .. U1<t < U1+U2. with common distribution B and define4 U(A) = E# {n = 0.T).
T). The renewal density at x is now just the rate of jumps of the second type.1 Corollary 2. which is ti in state i. i. fi(t) U2 U1 .T) where vt = ae (T+ta)t . Hence ( 2. Hence the intensity matrix is T + ta. and the jumps of the first type are governed by T. and let µB = aTle be the mean of B.T) where v = aT1 /µB.T). i.3 Consider a renewal process with interarrivals which are phasetype with representation (a. see Fig. the lifetime of the renewal process. IIafl < 1. .1) follows by the law of total probability. Then: (a) the excess life t(t) at time t is phasetype with representation ( vt. Then the lifetime is zeromodified phase type with representation (a. MATRIXANALYTIC METHODS of the last type from i to j occurs at rate tiaj .1) remains valid for that case.2 Consider a terminating renewal process with interarrivals which are defective phasetype with representation (a. is the first k with Uk = 00.. that is. this is welldefined. and hence ( 2. . B is defective . (b) £(t) has a limiting distribution as t * oo. Proof Just note that { it } is a governing phase process for the lifetime. This is defined as U1 + . as the time of the last renewal. which is phase type with representation (v. However. Corollary 2.1.U1 U3 U2 U3 U4 Figure 2.e. the phasetype assumptions also yield the distribution of a further quantity of fundamental importance in later parts of this chapter . define the excess life e(t) at time t as the time until the next renewal following t. 2.T + ta).224 CHAPTER VIII. u The argument goes through without change if the renewal process is terminating. the density is veTxt = B(x)/µB. since Uk = oo with probability 1 .IIBII which is > 0 in the defective case. and the distribution of Jx is ae ( T+t«)x. + Uit_1 where s.e.. u Returning to nonterminating renewal processes . Equivalently.
i.6.T) where vt is the distribution of it which is obviously given by the expression in (a).e. 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 (2.4 Consider a nonterminating renewal process with two phases. we first compute the stationary distribution of Q. Hence in (b) it is immediate that v exists and is the stationary limiting distribution of it.q2. the unique positive solution of ve = 1.2): aT1 e = AB = 1 µB µB a + aT'Tea aT1(T + ta) µB PB a + aea a + a µB µB =0. cf. Al. hence e(t) is phasetype with representation (vt.e. (ii) First check the asserted identity for the density: since T. RENEWAL THEORY 225 Proof Consider again the process { Jt } in the proof of Theorem 2.2.1. The formulas involve the matrixexponential of the intensity matrix Q = T + to = ( tll + tlal t12 + t2al tlz + tlaz _ q1 ql t22 + t2a2 q2 q2 (say).2) v(T + ta) = 0. we get B(x) aeTxe aT1eTxTe µB µB PB = veTxt.) ( t2 ) . The renewal density is then aeQtt = (al a2) ( 7i 7"2. T1 and eTx commute. (2. Next appeal to the standard fact from renewal theory that the limiting distribution of e(x) has density B(x)/µB. = qz ql (x1 xz) = ql + qz ql + q ' and the nonzero eigenvalue A = ql . According to Example A3. u Example 2 .
Then Q= 0 55 )+(1o)=( j ad ).t2) 1 + eat (a17r2 .4 yields the renewal density as u(t) = 2 (1 . The present treatment is somewhat more probabilistic.52) 25152 51x2+5251 51a2+5251 Notes and references Renewal theory for phasetype distributions is treated in Neuts [268] and Kao [221].226 CHAPTER VIII. )t (51 . t1B 0 Example 2 .4 yields the renewal density as u(t) = 5152 e.a27r1) (t1 .a27rl) (tl . Hence 7r = (1/2 1/2).52a1.6 Let B be hyperexponential. MATRIXANALYTIC METHODS e.e2bt) 13 Example 2 . A = 25.(biaz + aza. and Example 2.`t (al a2) + C 11 172 ir12 / \ t 2 ) r1 (7r1 7r2) ( t2 7rltl + J + eAt (al a2) ( 71(t2 . and Example 2.5 Let B be Erlang(2).t2) . Then _ Q Hence 51 0 0 52 + 51 52 _ 5152 51a2 ) (al a2) 52a1 62a1 Slat + 52a1 51a2 51a2+52a1 A = 51a2 .tl) 7r2t2 + eat (a17r2 . .
(b) V. T) where a+ is given by a+ = .) = F(ST(o) E •. T + to+).1 on the next page. The essence is contained in Fig. the Markov processes representing ladder steps can be pieced together to one {my}. Next. 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. The stars represent the ladder points ST+(k). with 0 denoting the Poisson intensity. and M is zeromodified phasetype with representation (a+. Since the results is so basic. Corollary 2. the transitions are governed by T whereas termination of ladder steps may lead to some additional ones: a transition from i to j occurs if the ladder step terminates in state i. however. Then each claim (jump) corresponds to one (finite) sample path of the Markov process.i. Then: (a) G+ is defective phasetype with representation (a+. Here we have taken the terminating Markov process underlying B with two states. T). cf. Proof The result follows immediately by combining the PollaczeckKhinchine formula by general results on phasetype distributions: for (a). represent the maximum M as the lifetime of a terminating renewal process and use Corollary 2. T). {St} the claim surplus process.f3aT1.2.e. B the claim size distribution. Corollary 3. and if there is a subsequent ladder step starting in j whic occurs w. We asssume that B is phasetype with representation (a. r(u) the time of ruin with initial reserve u. Considering the first. itself phasetype with the same phase generator T and the initial vector a+ being the distribution of the upcrossing Markov process at time ST+_. Thus the total rate is tip + tia+. and rewriting in matrix form yields the phase generator of {my} as T + ta+. . For (b). we shall. THE COMPOUND POISSON MODEL 227 3 The compound Poisson model 3a Phasetype claims Consider the compound Poisson (CramerLundberg) model in the notation of Section 1. T(0) < oo) the ladder height distribution and M = supt>o St.3.p. a+j. Now just observe that the initial vector of {mx} is a+ and that the lifelength is M.3. which occurs at rate ti. G+(. i. marked by thin and thick lines on the figure. 3. Within ladder steps. add a more selfcontained explanation of why of the phasetype structure is preserved.1 Assume that the claim size distribution B is phasetype with representation (a.(u) = a+e(T+tQ+)u Note in particular that p = IIG+II = a+e. use the phasetype representation of Bo.
T = (3 .. 3e3x + . see Corollary 2.. This is in fact a simple consequence of the form of the excess distribution B0.228 CHAPTER VIII.Q = 3 and b(x) = . 7e7x 2 2 Thus b is hyperexponential (a mixture of exponential distributions) with a (2 2 ).QaT1.2 Assume that . 0 Example 3. Figure 3.3.1 .7)diag so that a+ = QaT 1 = 3 ( 3 2 2) 0 3 9 2 14 7 2 11 2 T+ta+ = 3 0 07/+( 7I \ 2 14 .1 This derivation is a complete proof except for the identification of a+ with ..M {mx} ST+(2)  S .t t d kkt S. MATRIXANALYTIC METHODS t .
but there the vector a+ is not explicit but needs to be calculated (typically by an iteration). For an attempt. In the next sections.1): Proposition 4. We assume p = PB/µA < 1 and that B is phasetype with representation (a. THE RENEWAL MODEL This is the same matrix as is Example 1. the discussion around Fig.4. (a) G+ is of phasetype with representation (a+. if we define {mz} just as for the Poisson case (cf.6. 4 The renewal model We consider the renewal model in the notation of Chapter V.6).j). For further more or less explicit computations of ruin probabilities. T) for some vector a+ = (a+.4. with A denoting the interarrival distribution and B the service time distribution. We shall derive phasetype representations of the ruin probabilities V) (u). his derivation of +'(u) is different. see Stanford & Stroinski [351] .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+. For the compound Poisson model. T).and Markovmodulated models. It is notable that the phasetype assumption does not seem to simplify the computation of finite horizon ruin probabilities substantially. Fig. the duality result given in Corollary 11. The parameters of Example 3. this was obtained in Section 3. we encounter similar expressions for the ruin probabilities in the renewal. where a+ is the (defective) . but that such a simple and general solution exists does not appear to have been well known to the risk theoretic community. see Section 6. and the argument for the renewal case starts in just the same way (cf. The result carries over to B being matrixexponential. cf.T). 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 . 0(8) (u) (recall that z/i(u) refers to the zerodelayed case and iY(8) (u) to the stationary case). 3.1 In the zerodelayed case. 3.1 can be found in Neuts [269] (in the setting of M/G/1 queues.2 are taken from Gerber [157]. That is. see Shin [340].^(u) = a+e( T+ta+)ue = 24eu + 1 e6u 35 35 0 Notes and references Corollary 3.
T).3 a+ satisfies a+ = V(a+).T).6. B0 is phasetype with representation (aT1/µa. the form in which we derive a+ for the renewal model is as the unique solution of a fixpoint problem a+ = cp(a+). which for numerical purposes can be solved by iteration. Proof Obviously. In fact.1. but with initial distribution a rather than a+. Hence by Theorem 11. We have now almost collected all pieces of the main result of this section: Theorem 4 .3. (4. Since the conditional distribution of my given T1 = y is ae4y.Sy} in the same way as {mx} is defined from {St}. where a(8) = aT1/PA. the calculation of the first ladder height is simple in the stationary case: Proposition 4.4 Consider the renewal model with interarrival distribution A and the claim size distribution B being of phasetype with representation (a. obviously mo = m. The key difference from the Poisson case is that it is more difficult to evaluate a+. (c) {mx } is a (terminating) Markov process on E. MATRIXANALYTIC METHODS (b) The maximum claim surplus M is the lifetime of {mx}. Fig.1).5. the Palm distribution of the claim size is just B. Also. where B0 is the stationary excess life distribution corresponding to B. with intensity matrix Q given by Q = T + to+. it follows by integrating y out that the distribution a+ u of mo is given by the final expression in (4.*'} is Markov with the same transition intensities as {mx}. cf. CHAPTER VIII.230 distribution of mo. Then .1) Proof We condition upon T1 = y and define {m. G(') = pBo.*} from {St+y . where u w(a +) = aA[T + to+) = a J0 e(T+t+)1A(dy).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). Nevertheless. Then {m. 4. But by Corollary 2.T)• Proposition 4.
by a+ = lim a +n) where a+°) . The second follows in a similar way by noting that only the first ladder step has a different distribution in the stationary case. a+ can be computed by iteration of (4.e.M. a+2) = ^p (a+l)) .3) (defined on the domain of subprobability vectors .•.3). and that this is given by Proposition 4. thus . i y ^ T1= y `•r Figure 4. (4.2.1(b). a+) > 0 = a+o) implies a+) _ (a+) > W (a+)) = a+) .2) where a+ satisfies (4.1 by noting that the distribution of mo is a+. THE RENEWAL MODEL 231 . only with initial distribution a(*) for mo. The term tf3 in cp(i3) represents feedback with rate vector t and feedback probability vector (3. (4.0) is an increasing function of /3.0. .^(u) = a+e ( T+ta+)xe.1) and a(8) _ aT..3) Proof The first expression in (4. the maximum claim surplus for the stationary case has a similar representation as in Proposition 4.^(8)(u) = a ( 8)e(T+ta +) xe.. i. It remains to prove convergence of the iteration scheme (4. I {mx} .4. Hence ^p(. .2 ) follows from Proposition 4..1 ..1). In particular . a+l ) = cp (a+°)) . Furthermore .1/pA.
} can contain at most n .1. let F be the distribution of U1 . and let &+". Let Fn = {T1 + • • • + Tn+1 > r+}be the event that {my} has at most n arrivals in [T1.5) yields h = (sI .232 CHAPTER VIII.4).1 arrivals (n arrivals are excluded because of the initial arrival at time T1 ). To this end. a+ ) exists . and hence we may assume that h has been normalized such that ahA[s] = 1.T)1t. F[s] being interpreted in the sense of the analytical continuation of the m. However.2. 7+ ].T)'t • A[s] (4. Thus .T)It. (4. s ¢ sp(T). In that case.f. To prove the converse inequality. the corresponding right eigenvector may be taken as (sI . It follows that n1) so that on Fn the feedback to {mz} after each ladder step cannot exceed &+ a+ n) < a f ^ e(T+ t&+ 1))YA(dy) o < a is e(T+t«+1')YA(dy) _ w (a+1 )) = a+n).g. the normalization is equivalent to F(s) = 1. Thus by (4. 0 = a+) < a+ yields a+) _ (a+0)) (a+) = a+ (n and by induction that a(n) < a+ for all n . MATRIXANALYTIC METHODS and (by induction ) that { a+ n) } is an increasing sequence such that limn. Obviously.. which links together the phasetype setting and the classical complex plane approach to the renewal model (see further the notes). Proof Suppose first Qh = sh.4. Then each subexcursion of {St+Tl .4) makes sense and provides an analytic continuation of F[•] as long as s ¢ sp(T). .5 Let s be some complex number with k(s) > 0.4) whenever EeR(S)U < oo.5) Since s $ sp(T). n) &+n) T a+.P[s] = A[s]B[s]. Fn ). For n = 0. this implies that ahA[s] # 0. both quantities are just 0 .) = P(mTl = i. 0 0 We next give an alternative algorithm. Then (4. limn4oo a ) < a+. Theorem 4. Then e4'h = e82h and hence sh = Qh = (T + taA[Q])h = Th + A[s]tah. Similarly. so to complete the proof it suffices to show that &+ < a+) for all n.ST. Then s is an eigenvalue of Q = T + ta+ if and only if 1 =. Assume the assertion shown for n . Then F[s] = a(sI . with B[s].T1. we use an argument similar to the proof of Proposition VI. (4.
Then G+ is phase.5(c) means that a+(sI T)1t = 1.' that the equation F(s) = 1 has d distinct roots p1. . 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. Corollary 4. W v M(d) in the notation of Chapter V).6) i=1 i=1 Proof Appealing to Theorem 4. the classical algorithm starts by looking for roots in the complex plane of the equation f3[y]A[ry] = 1. Further.T) = 1 ata+ = a+. hd... Let d denote the number of phases. . Pd with corresponding eigenvectors hl.T)lt + t = s(sI . .type with representation (a+. Since R(s) > 0 and G _ is concentrated on (oo. t(ry) > 0.. we have IG_ [s] I < 1 .. in turn. This gives d roots 'y. D that with columns p1 hl...p1i . the matrix Q in Theorem 2. and the solution is ..lt we get Qh = (T + to+)h = T(sI . (4.. pdhd.5. and hence by the WienerHopf factorization identity (A. . and define hi = (piI . Given T has been computed. This immediately implies that Q has the form CD1 and the last assertion on the diagonal form ... hd.6.T)lt = sh.T)It. 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 ). Q = CD1 where C is the matrix with columns hl. 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.6 Suppose u < 0. . T) with a+ = a(QT)/at. Q has diagonal form d d Q = dpivi®hi = dpihivi... 0).4..1 has the d distinct eigenvalues . yd satisfying R(ryi) > 0. . In older literature .9) we have G+[s] = 1 which according to Theorem 1. As in Corollary 4... THE RENEWAL MODEL 233 Suppose next F(s) = 1. . Pd in the domain ER(s) > 0 . we get at a(Q .. Hence with h = (sI T). Notes and references Results like those of the present section have a long history. letting vi be the left eigenvector of Q corresponding to pi and normalised by vihi = 1 ..
The arrival rate in background state i is a. The solutions are based upon iterations schemes like in Theorem 4. It turns out that subject to the phase. the background Markov process with p states is {Jt}. 5 Markovmodulated input We consider a risk process {St } in a Markovian environment in the notation of Chapter VI. E(t)). a pioneering paper in this direction is Tacklind [373]. which contains somewhat stronger results concerning the fixpoint problem and the iteration scheme. with representation say (a(' ). and the distribution of an arrival claim is B.234 then in transform terms CHAPTER VIII. e. is phasetype. [119].4. MATRIXANALYTIC METHODS d F 1 + a J e°" ip(u) du = Ee°w = 11(t. That is . the intensity matrix is A and the stationary row vector is ir .. T(').F. We assume that each B. an alternative approach (the matrixgeometric method ) has been developed largely by M. the fixpoint problems look like R=Ao+RAI+R2A2+ .exponential form of the distribution was found by Sengupta [335] and the phasetype form by the author [18].type assumptions are basic. involving .. the ruin probability can be found in matrixexponential form just as for the renewal model. The exposition here is based upon [18]. and appears already in some early work by Wallace [377]. The matrix. The distribution of W comes out from the approach but in a rather complicated form . where R is an unknown matrix. Asmussen & O'Cinneide [ 41] for a short self. Numerical examples appear in Asmussen & Rolski [43].g. see Neuts [269].. In queueing theory.) d (see. Here phase. The number of elements of El=> is denoted by q. In risk theory. whereas the approach was introduced in queueing theory by Smith [350]. 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]. Neuts and his students. 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. For further explicit computations of ruin probabilities in the phasetype renewal case . For surveys . see Dickson & Hipp [118].contained derivation). starting around in 1975.type assumption . [270] and Latouche & Ramaswami [241].
(a) 0 0 ♦ o ° tl ♦ • 0 0 o } o o (b) 0 } ♦ • 0 o f o Figure 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. 5.Vt)} obtained by time reversing the I component.5. Diagonalization Consider a process {(It.1 In Fig. the phase space E(°) for B.1. •. and the one E(•) for B. 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). The stationary distribution is obtained by finding the maximum of the Vcomponent of the version of {(It. We start in Section 5a with an algorithm involving roots in a similar manner as Corollary 4. has states o. Vt)}t>o such that {It} is a Markov process with a finite state space F and {Vt} has piecewiese linear paths. say with slope r(i) on intervals where It = i.1. states . MARKOVMODULATED INPUT 235 some parameters like the ones T or a+ for the renewal model which need to be determined by similar algorithms. The connection between the two models is a fluid representation of the Markovmodulated risk process given in Fig. O. 5a Calculations via fluid models. However. The two environmental states are denoted o.4. p = ql = Q2 = 2. Section 5b then gives a representation along the lines of Theorem 4.6. This calculation in a special case gives also the ruin probabilities for the Markovmodulated risk process with phasetype claims. for which the relevant fixpoint problem and iteration scheme has already been studied in VI.2. The key unknown is the matrix K. 5.
Thus F = {o. 4. F = E U { (i. Eli) + Proposition 5.31a(l) (/3i)diag . we have more martingales at our disposal. •. of E into components indexed by E. in the fluid model Eel'.1))diag ) a = sa and the eigenvector b = . t.1) if and only if s is an eigenvalue of E.1(b) {(It . Let E denote the matrix .1 A complex number s satisfies 'A+ (f3i(Bi[s] . First. '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. MATRIXANALYTIC METHODS 4.Vt)} is then obtained by changing the vertical jumps to segments with slope 1. Bi[s] = a(i)(T(i) + sI)it('). consider the vector a satisfying (A + (13i(Bi[ s] . < oo for all s. The fluid model on Fig . In the general formulation . F is the disjoint union of E and the Eli). Recall that in the phasetype case. a) : i E E. r(i) _ 1. a) = 1. 5. a E E(i) } .A 0 Or 1A/ _ t(i) 0 t(2) 0 0 0 0 0 t(3) 0 T1 0 0 0 . 2. A claim in state i can then be represented by an E()valued Markov process as on Fig. 4.1))diag + sII = 0 (5. 4}. This implies that in the fluid context. o.1(a).(Ni)diag r(i.236 CHAPTER VIII. whereas Ee8s' = oo for all t and all s > so where so < oo.92a(2) 0 0 T(2) 0 0 0 f33a(3) 0 0 T(3) with the four blocks denoted by Ei„ i. Second. If s is such a number. j = 1. The intensity matrix for { It} is (taking p = 3 for simplicity) I A . 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. V. corresponding to the partitioning + Epp). 5. a) of {It}. resp. i E E.
sI) (sI .E21a + sd = sd.sI ()3i)diag .E22)1 E21a E21a .5.A .sI. it follows that if Qla(1) 0 0 . Then E21c+E22d = E21a .sI 0 0 0 T(3) . d = (sI E22)1E21a = E ai(sI . d correspond to the partitioning of b into components Proof Using the wellknown determinant identity Ell E12 E21 E22 E22 I ' I Ell .A . Then (up to a constant) c = a.sI)1t)) iag I = 0 which is the same as (5. and let d = (sI . iEE (a> of 0* 1 AI. For the assertions on the eigenvectors.(sI . 0 . MARKOVMODULATED INPUT 237 indexed by E. c = a.E22)1 E21a. resp . t(1) 0 0 then also 0 t(2) 0 . with Eii replaced by Eii . assume that a is chosen as asserted which means (Ell . it follows that Ell E12 ( E 21 E22) (d) = s 1 d I .E22)1 E21) a = 0.32a(2) (/3i)diag .E12E22 E21 I .sI 0 0 t(3) 0 0 = 0. E(1) + + E(P).1).sI 0 0 0 T(2) .Nla(1) 0 0 T 1. where c.T('))1t(i) .sI + E12 (sI .sI+ ((3ia(i)(T(i) . Noting that E11c + E12d = se by definition.E22 .
j) pi( u .( u.. Then .v.O. j.upi(u.v) = v) I.. Example 5 .. Then we get V)i (u) as sum of two exponential terms where the rates s1. a) = Pi (Vw(u.. B2 are both exponential with rates 51 i b2. .v}.a Solving for the pi(u. < 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.a)d^ ). define w(u. c j.. We can take a = c = 1 and get d = (s + b)16 = 5/(3 = 1/p. j. . a )d(a + e8 °vpi (u . .v) = = p i( u . To determine 0 (u). pi(u. I' i( V P2 (w (u) < oo.pi(u.j)c v . Here E has one state only...v) = j). j. s2 are the negative eigenvalues of Al +01 A1 E _ A 2 b1 0 52 A2 +32 0 . j.Q. v = 1.3 Consider the Poisson model with exponential claims with rate 5..4 Assume that E has two states and that B1. v > 0.5. a). v.4 that {e"1b(v) is a martingale .. Example 5 . j..v)=inf{t >0:Vtu orVt=./' u = e' (esiuc ( 1) . Thus 0(u) = esu/d = pe7 ° as u should be. sq with $2s. e89uc(e)) (d(1) . a)). w(u. .. it follows by Proposition II. a) = (j. a)).. Proof Writing Or'Alb( v) = svb( v) as (AI . a) and noting that i1 (u) = >I j. Letting v ^ oo and using Rsv < 0 yields e8'u = Epi(u.v) = Optional stopping at time w (u.sv)b(v) = 0. u) Iw(u.. v. w(u)=inf{t >O:Vtu}. v) yields C{V) = e8 .238 CHAPTER VIII.. v. we first look for the negative eigenvalue s of E = I 0 I which is s = ry with yy = b . the result u follows. Iw(u. d("))1 e.j.v) = (j. For u. . MATRIXANALYTIC METHODS Theorem 5. q.
9(').xxej • a 00 oo el .( 2. j. 8^')IT(j)) where e 3^') =. oo)) j)ye. MARKOVMODULATED INPUT 239 5b Computations via K Recall the definition of the matrix K from VI.b (u) = Pi(M > u) = 9(i)euue.2) the l.5 G+(i.2. In terms of K.k.h. j. 0 Theorem 5 . i. is 0 /3 f R(i . (') a T( However .3) .s.3j eye.6 For i E E.y = to B k7 j # k In particular. according to VI. Proof We must show that G+ (i. dx)Bj(y . (5.33(e = 0 a(j))(K ®T ( j))(ej (9 I).Qj eie 0 f e (j) T(') x T(j)y ej a e dx e e 00 00 eKx ® e T(')' dx (ej (& I)e T(')ye eKa®T(')x dx (ej (9 I)eT(') Ye e(i)eT(')ye. (y. the Pidistribution of M is phasetype with representation (E(1) + + E(P). we get the following phasetype representation for the ladder heights (see the Appendix for the definition of the Kronecker product 0 and the Kronecker sum ®): Proposition 5. •) is phasetype with representation (E(i). j. U) where t(j) + t(j)O(j j = k uja.5.x) 00 f ° (') (j) eT (yy)edx .
intensity matrix U. T. we have sofar concentrated on a claim size distribution B of phasetype. and a new ladder step of type k must start in phase y. the ratio between two polynomials (for the form of the density..y) to occur when j # k. MATRIXANALYTIC METHODS Proof We decompose M in the familiar way as sum of ladder steps ..k y. An alternative characterization is that such a distribution is matrixexponential.2. a) to (k. which occurs at rate t^^7. (6. For a transition from (j..e. oo) and b* [0] = f °O eBxb(x) dx the Laplace transform. Then b*[0] is rational if and only b(x) is matrixexponential.. +bn0i1 0n +a10n1 +. equivalently. if b* [0] = b1 +b20+b302 +. a m. the initial value of (i. Bk7 . which occurs w.1 Let b(x) be an integrable function on [0. we have the additional possibility of a phase change from a to ry within the ladder step. 6 Matrixexponential distributions When deriving explicit or algorithmically tractable expressions for the ruin probability. . Furthermore.. t) is the representation of the matrixexponential distribution/density): Proposition 6. For j = k. some square matrix T and some column vector t (the triple (a. Numerical illustrations are given in Asmussen & Rolski [43].. i. Starting from Jo = i. +aii10+anI then a matrixexponential representation is given by b(x) = aeTxt where a = (b1 b2 .e..) which is rational.5).p. that the density b(x) can be written as aeTxt for some row vector a. with phase space EU> whenever the corresponding arrival occurs in environmental state j (the ladder step is of type j). see Example 1. say.f.240 CHAPTER VIII. 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. i. Associated with each ladder step is a phase process.. a) is obviously chosen according to e(`). 0 1)'. and lifelength M.2) . bn1 bn). This yields the asserted form of uja. and it just remains to check that U has the asserted form. the current ladder step of type j must terminate. Piecing together these phase processes yields a terminating Markov process with state space EiEE E(').g. u Notes and references Section 5a is based upon Asmussen [21] and Section 5b upon Asmussen [17]. However. t = (0 0 . which occurs at rate t(i).
3) was suggested by Colm O'Cinneide. bn of the denominator to be distinct and expand the r.. Example 6 . 1 . see Asmussen & Bladt [29] (the representation (6. . u giving b(x) = E 1 ciebiz/bY.4) 0 0 1 c This representation is complex.s.47x2 3 1 0 .(6.47r2 3 . (6. t). s) is given by 27r i . The converse follows from the last statement of the theorem.1 0 0 )3 = (111).2). b(x) > 0 for x > 0.. 0 0 0 0 1 0 0 .1 0 ..3) 0 0 0 0 0 . T. (6.e(tai1)x/2 + e'T) it follows that a matrixexponential representation ()3. then b*[0] = a(0I T)1t which is rational since each element of (01 .1. . One of his elementary criteria. (6. cannot be phasetype. of (6. 0 1 an an1 an _2 . where c = 1 + 1/47r 2.6. t= 0 . .2 A remarkable feature of Proposition 6.3) that we can take 0 1 0 0 a= (1 + 47r2 0 0).T)1 is so. but as follows from Proposition 6. we can always obtain a real one (a.h. personal communication). ... namely to asssume the roots 6l. Namely.. Thus. S = f c/2 0 21ri . u Remark 6. . matrixexponentiality implies a rational transform.. a2 a1 Proof If b(x) = aeTxt.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.1) as E 1 c. s = c/ 2 . For a proof../(0 + bi). MATRIXEXPONENTIAL DISTRIBUTIONS 241 T = 0 1 0 0 0 . T= 0 0 1 . since 1 + 4ir2 03 + 302 + (3 + 47x2)0 + 1 + 47r2 it follows by (6.an_3 an _ 4 .. Writing b(x) = c(e( 2ni1 ) y/2 . S. shows that the distribution B with density b(x) = c(1 cos(21r x))ex.2). 0 0 .3 A set of necessary and sufficient conditions for a distribution to be phasetype are given in O'Cinneide [276].
T. recall that t = Te) that if B is phasetype and (a.5 (6. we shall only consider the compound Poisson model with arrival rate 0 and a matrixexponential claim size distribution B.6) holds true also in the matrixexponential case. t) a phasetype representation with a the initial vector.3.4 This example shows why it is sometimes useful to work with matrixexponential distributions instead of phasetype distributions: for dimension reasons . For the second algorithm. 7 + 155ex b(x) Then it is known from O'Cinneide [276] that b is phasetype when 6 > 0.242 CHAPTER VIII. we have represented ti* [0] as ratio between polynomials (note that 0 must necessarily be a root of the numerator and cancels). and that the minimal number of phases in a phasetype representation increases to 0o as 5 . Then (cf.6) in Section 3 seems to use the probabilistic interpretation of phasetype distribution in an essential way. then: Proposition 6.4) the Laplace transform of the ruin probability is /g(e)PO 0*[e] _ /' eeu^G(u)dU = 0 9(/3a0p(9)ap (9)/q(9)) . T the phase generator and t = Te. that despite that the proof of (6. we use a representation (a. As for the role of matrixexponential distributions in ruin probability calculations. Consider the distribution with density = 15 ((2e2x . q are polynomials without common roots. then 5(u) = a+e(T+t+)uTle where a+ = /3aT1. Corollary 111. and can use this to invert by the method of Proposition 6. MATRIXANALYTIC METHODS Example 6 . For the first. 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 6.1 shows that a matrixexponential representation can always be u obtained in dimension only 3 independently of J. leading to matrix calculus in high dimensions when b is small.5) Thus.1 to get i (u) = f3esus. we take as starting point a representation of b* [0] as p( O)/q(9) where p. 0. (6.6) The remarkable fact is. T.1)2 + 6). and present two algorithms for calculating '(u) in that setting. (6. . t) of b(x).
we get b+ = 0aT1(9I T).6. From the general matrix identity ([331] p.'t. b+ = a +(BI .T)1t.1 + b+ = b++ 1 .to+)1 = (BI .7) 9( cf. we get (91. Presumably.T)1 + (6I .1t = f3a (0I T)1T1t .to+)1T .T)1T 2 = and 1 = AB IT2 + 82T . since (91T)1T .to+)1T .1t = b* .5 ).6). Then in Laplace transform formulation .T)1t)1a +(9I .1UB(B + BVA1UB).T .T)1 (91.T)1 so that b* b** b** a+(9I . xb(x) dx = aT2t. 519) (A + UBV ). . (91.1BVA1.1 = ^(T1 + ( 91T)1).6b* . U =.T)1 + 1 ib* (91. MATRIXEXPONENTIAL DISTRIBUTIONS 243 Proof Write b* = a(9I .A .T . Now. with A = 91T.1 + 82 (9I .T)1T1t.T)1ta+(OI .b* (6.T .T)1t ( l .1 = A1 . this can be verified by analytic continuation from the phasetype domain to the matrixexponential domain . b+ = a+(9I . but we shall give an algebraic proof.T).a+(9I . (6.t.T)1 J0 00 b(x) dx = f aT1t. (6.1t du = .B=land V=a+. the assertion is equivalent to a+(BI .
82b*. 0 Notes and references As noted in the references to section 4. VII.8 a(T1 + (01. 7 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate 0. The proof of Proposition 6.244 CHAPTER VIII.1.1.s. /3aT1(0I . From this it is straightforward to check that b**/(b+ . a key early paper is Cox [90] (from where the distribution in Example 6. Lipsky [247] and Asmussen & O'Cinneide [41]. For expositions on the general theory of matrixexponential distributions.1.T)1T. see the Notes to VII.7). to piece together the phases at downcrossing times of {Rt} (upcrossing times of {St}) to a Markov process {mx} with state space E. of (6. which is selfexplanatory given Fig. a.1. We present here first a computational approach for the general phasetype case (Section 7a) and next (Section 7b) a set of formulas covering the case of a twostep premium rule. but the argument of [286] does not apply in any reasonable generality). some key early references using distributions with a rational transform for applied probability calculations are Tacklind [373] (ruin probabilities) and Smith [350] (queueing theory). 7a Computing O(u) via differential equations The representation we use is essentially the same as the ones used in Sections 3 and 4. 3. A key tool is identifying poles and zeroes of transforms via WienerHopf factorization. (for some remarkable explicit formulas due to Paulsen & Gjessing [286]. See Fig.1) is the same as the r.T)1T2t . the ruin probability(u) was found in explicit form for the case of B being exponential.3a (1 0 T 2 + 1 T 102 (9I + 02 1 T)1) t P + 7.1t = /3a (9I .b*).3 is taken). 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. T).la. MATRIXANALYTIC METHODS . Much of the flavor of this classical approach and many examples are in Cohen [88].T)1)t = 8 (1 .h.5 is similar to arguments used in [29] for formulas in renewal theory.8. see Asmussen & Bladt [29]. In Corollary VII. cf. . 7.
t)).t2) be the matrix with ijth element P (mt2 =j I mtl = i). Note that in general >iEE Vi (U) < 1.tl < t2 < u.I] I 80 { tq f Q(v) dvl t1 1 . Ai(t) = P(mt = i). Figure 7.u)e = A(u)e (7.e. the definition of {m8} depends on the initial reserve u = Ro.1 A(0) = v(u) and A'(t) = A(t)(T + tv(u . Define further vi(u) as the probability that the risk process starting from RD = u downcrosses level u for the first time in phase i. Let P(tl. is no longer timehomogeneous. In fact. >iEE Vi (U) is the ruin probability for a risk process with initial reserve 0 and premium function p(u + •).7. P(tl. though still Markov. i. the A(t) and hence Vi(u) is available by solving differential equations: Proposition 7. t) is the vector of state probabilities for mt. Given the v(t) have been computed.1z I. t + s) . in contrast to Section 3. Proof The first statement is clear by definition. RESERVEDEPENDENT PREMIUMS 245 Rt l0 u . t2) = exp where Q(t) = ds [P(t. Also. O<.1 The difference from the case p(r) = p is that {m2}. Since v(u) = (vi(u))iEE is the (defective) initial probability vector for {m8}. we obtain V)(u) = P(m„ E E) = v(u)P(0. 0 < t < u.1) where A(t) = v(u)P(0. By general results on timeinhomogeneous Markov processes.
246 CHAPTER VIII.(tai + vi(u) E vj(u)tjp (u)  Q + vj (u)tjip ( u). Thus. Given A'. In the first case. or it stops between level u + p(u)dt and u. {mx} has jumps of two types. Given this occurs. the probability that level u is downcrossed for the first time in phase i is ai. Hence Q(t) _ T + tv(u . (7. two things can happen: either the current jump continues from u + p(u)dt to u. Proposition 7. The intensity of a jump from i to j is tij for jumps of the first type and tivj(u . A'(t) = A(t)Q(t) = A(t)(T + tv(u . dt]. 0 Thus. the probability that level u + p(u)dt is downcrossed for the first time in phase j is vj (u + p(u)dt).2 For i E E.Qdt) vi(u) + vi'(u)p(u)dt + p(u) dt E{tji+tjvi(u)}. vi. those corresponding to state changes in the underlying phase process and those corresponding to the present jump of {Rt} being terminated at level u . from a computational point of view the remaining problem is to evaluate the v(t). we get vi(u) = aidt + (1 . the probability of which is 1 . the interpretation of Q(t) as the intensity matrix of {my} at time t shows that Q(t) is made up of two terms: obviously.4) jEE jEE Proof Consider the event A that there are no arrivals in the interval [0. MATRIXANALYTIC METHODS However. the probability of downcrossing level u in phase i for the first time is E vj (u + p(u)dt) (Sji + p( u)dt • tji + p(u)dt • tjvi(u)) jEE vi(u) + vi' (u)p(u)dt + p(u) dt E {tji + tjvi(u)} jEE Collecting terms. given A.3dt. Given A.Sj i)p(u)dt • tji = Sji + p(u)tji dt.t and being followed by a downcrossing. 0 < t < u.t). whereas in the second case the probability is p(u)dt • tjvi(u). jEE .t) for the second.t)).(u) p ( u) = . the probability of downcrossing level u in phase i is 8ji(1 + p(u)dt • tii) + (1 .
say. RESERVEDEPENDENT PREMIUMS 247 Subtracting v. Now since both P(A n Bv) 3 0 and P"(A n Bv) . vi (U) = lim v= (u). . supRt>v l t<7 I where o. Thus.7.5).3 For any fixed u > 0. of (7. Then pv(r) p(r) r < v p r>v ' and (no matter how p is chosen) we have: Lemma 7.) + 0 as v + oo. we can first for a given v solve (7. P u which implies that v.0 as v + 00 we have P(A) P"(A) = P(AnBv)+P(AnBv) P"(AnB. starting from v"(v) = .) is the tail of a (defective) random variable so that P(Bv) + 0 as v 4 oo. refer to the modified process. (v) is given by the r. <oo.i7rT1/p.) P"(AnBv) = P(AnB. then P(A n Bv) _ P"(A n BV'). say. F" etc. When solving the differential equation in Proposition 7. we have p(r) = p = vi (u) 0aTe. Then P(B.. Since the processes Rt and Rt coincide under level B.h. To deal with this. (u) for any values of u and v such that u < v.2. denotes the time of downcrossing level u .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. From Section 3. we face the difficulty that no boundary conditions is immediately available. This yields v. (u) on both side and dividing by dt yields the asserted differential u equation.)P"(AnB. consider a modification of the original process {Rt} by linearizing the process with some rate p..4) backwards for {va (t)}v>t>o. Let p" (t). Rt . V . and similarly P"(Bv) . after a certain level v..s.^ 0..
1. To evaluate p1(u). However.e. v = u.z51(v)). Therefore u pl(vvueTa t 1. Then v(u) is the initial distribution of the undershoot when downcrossing level v given that the process starts at u..7) equals 01 (v . cf. The f iin in (7. the evaluation of Vi(u) requires q(u) = 1 . 7b Twostep premium rules We now assume the premium function to be constant in two levels as in VII. such that Rt coincide with RI under level v and with Rt above level v. MATRIXANALYTIC METHODS Next consider a sequence of solutions obtained from a sequence of initial values {v. as well p1(u). assuming u > v for the moment.1. 0 < u < v.1a. for u > v the distribution of v . p2. 1/n. < 0}).V" M 0 . i. Corollary 3.) are so. where v = inf It > 0 : Rt < v}. 3u etc.. We recall from Propositon VII. 2u. numerically implemented in Schock Petersen [288]) and the present one based upon differential equations require both discretization along a discrete grid 0.x) dx f v(u)eT xt dx .7) f o (the integral is the contribution from {R.1.248 CHAPTER VIII.. Let ii'( u) = a+'ie(T+ta +^)"e denote the ruin probability for R't where a+ = a+i) = laT1/pi. (7. The trapezoidal rule used in [288] gives a precision of 0(n 3). p(r) P..10 that in addition to the O'(•). (u)}. while the fourthorder RungeKutta method implemented in [30] gives 0(n5). the probability of ruin between a and the next upcrossing of v.q(v dx +( ) ) = ( ) ( q( )) vueTva (7. let v(u) = a+2ieiT +ta+>)(uv). Recall that q(w) is the probability of upcrossing level v before ruin given the process starts at w < v. T). which is available since the z/i'(. say.9.RQ (defined for or < oo only) is defective phasetype with representation (v(u). where.6) We may think of process Rt as pieced together of two standard risk processes RI and Rte with constant premiums p1. > 0} and the last term the contribu tion from {R. typically the complexity in n is 0(n2) for integral equations but 0(n) for integral equations.zp1(u)/(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. 2/n. r<v r > v. The algorithm based upon numerical solution of a Volterra integral equation (Remark VII. The precision depends on the particular quadrature rule being employed.v v(u)eTat 1 1 .
x) dx 1 ^(v) ( 1 .and A2 = 3 .v(u)eTVe . Then one gets X20 20 21 f 1ea1(u v) + 1 3 3 ^ A 2(u e .8) equals v v(u)eTxta+2) e(T+ta +))( vx)edx which using Kronecker calculus (see A.v) + (2^ + 3v2 ea'(u " .8) The integral in (7.2.e6v Let Al = 3 + 2V'2. Example 7. all quantities involved in the computation of b(u) have been found in matrix form.v(u ) eTV e J v(u)eTxtz/)l (v . I.e.2V"2.4) can be written as (Y(u) ®a+)e(T+t°+>)°1 (T ® (T .. so we consider the nontrivial case example p2 = 4 and p1 = 1.2.24ev . Since µB = 5/21. B is hyperexponential corresponding to 3 0 3 a(2 2)' T= ( 0 7 t.to+))11 {e{T®(Ttoy+ ))}„ . p2 < 3.1 from which we see that pl (u) = 1 + 1 249  1 .24e.21 = ? yields 0(u) = 1.u .e6u 35 .x) dx} V 1 1(v) f V v(u) eTxt.be the eigenvalues of T + to( 2 ).jl (t ®e) Thus. 01(u) _ 24 u + 35 e6u 1 35 e 4(u) _ 35 .01 (v .(7 The arrival rate is (i = 3. RESERVEDEPENDENT PREMIUMS 1 .4 Let {Rt } be as in Example 3. From Example 3.^1(v) 1 .7.v) 1eai(u v) + 7 7 1 e\2(u v) 1 3 ^') eA2 (u. (7.v(u)eTve).
) e sv + ( 2v/2.24e5v .1 V2 = 4e5"+6 35e6v . 21 3 In particular.24e5v .24es" .2 35e6v ./2) ea 1(u .21(35e6v .v)esv + 7 4_ 2. pi (u) = p12(u)/p1 l(u) where p1i(u) p12(u) 35e6v . and one gets 12e5" .1)' ?.24es" .. The analysis and the example are from Asmussen & Bladt .1. 192esv + 8 P1 .b(v) = 192esv +8 35e6v + 168esv + 7* Thus all terms involved in the formulae for the ruin probability have been exu plicitly derived.1 Thus. MATRIXANALYTIC METHODS From (7.7) we see that we can write pi (u) = v(u)V2 where V2 depends only on v. Notes and references [30]./2 ea1(u") .250 CHAPTER VIII.+ it (3 4'I 1 ea2(uv e1\2(u") 7 + ( 32 +4.
III. x 4 oo. the exponential change of measure techniques discussed in II. 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.4. For further examples. and instead we shall work within the class S of subexponential distributions .46 and at numerous later occasions require a light tail.N(µ. Some main cases where this lighttail criterion are violated are (a) distributions with a regularly varying tail. A rough distinction between light and heavy tails is that the m.B(x). L(tx)/L(x) 4 1.g. 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. B(x) = ex0 with 0<0<1. a2)) with density 1 e(logyFh) 2/2az . (b) the lognormal distribution (the distribution of eu where U . B(x) = L(x)/x" where a > 0 and L(x) is slowly varying.Chapter IX Ruin probabilities in the presence of heavy tails 1 Subexponential distributions We are concerned with distributions B with a heavy right tail B(x) = 1. For the definition . for all t > 0.f. see I.2b. x 2iror2 (c) the Weibull distribution with decreasing failure rate . For example. we require that B is concentrated on (0. oo ) and say then that B is subexponential (B E S) if 251 .
F(X1 > x. oo). x 3 00. That is. Thus . one can check that x x where U is uniform on (0.v. HEAVY TAILS B*2\ 2. In terms of r.1(b) is oo. X2) > x} C {X1 + X2 > x}. To capture the intuition behind this definition. 1). That is. (1. then P(X1>xI X1+X2>x)* 2. X2 > x) = 2B(x) .2. X2) > x}. X1 is w. X2) > x) is P(X1 > x) + P(X2 > x) . then (with high probau bility) so are both of X1. Then: (a) P(max(Xi. We later show: Proposition 1. the behaviour in the lighttailed case is illustrated in the following example: Example 1. (b) liminf BB(() ) > 2. note first the following fact: Proposition 1. The proof shows that the condition for B E S is that the probability of the set {X1 + X2 > x} is asymptotically the same as the probability of its subset {max(Xi.3 Consider the standard exponential distribution. 1/2 'typical' (with distribution B) and w. X2) > u x)/B(x) = 2. In contrast.2B(x).p. that is. Proof By the inclusionexclusion formula.'s X1. P(Xi <yI Xi+X2>x) 1B(y). X2 with distribution B. X2) > x) ^' 2B(x). proving (a). oo).1) then means P(X1 +X2 > x) 2P(Xi > x).2 If B E S. B(x) Here B*2 is the convolution square.v. the r.p. given X1 + X2 > x. B(x) ax.B(x)2 .'s. in the subexponential case the only way X1 + X2 can get large is by one of the Xi becoming large. the distribution of independent r. if X1 + X2 is large . As contrast to Proposition 1. Then X1 +X2 has an Erlang(2) distribution with density yeY so that B*2(x) xex. P(max(Xi.252 CHAPTER IX. . Thus the liminf in Proposition 1. we have {max(Xi. Since B is concentrated on (0. and thus the lim inf in (b) is at least lim inf P(max(Xi.v.1 Let B be any distribution on (0. X2 but none of them exceeds x. 1/2 it has the distribution of X1I X1 > x.
S)x + B(Sx)2 < lim sup B(x) xaoo B(x) lim sup 2L((1 x^oo .B(y) = 2.B E S. Let 0 < 5 < 1/2.1. then the overshoot X . SUBEXPONENTIAL DISTRIBUTIONS Here is the simplest example of subexponentiality: Proposition 1. Finally lim inf B(x .B(y)) . x]. Proposition 1. B( 0 .B*n(x . 253 Proof Assume B(x) = L(x)/xa with L slowly varying and a > 0.'s: if X .4 Any B with a regularly varying tail is subexponential. [In terms of r. or they both exceed Sx. The uniformity now follows from what has been shown for y = yo and the obvious inequality y E [0.yo]. If lim sup B(x .6)x)/((1 . Hence lim sup a+oo B*2(x) 2B((1 . and combining with Proposition u 1.] Proof Consider first a fixed y.S)x.5 If B E S. y] and (y. Using the identity B*(n+1)(x) = 1+ + 1)(x) 1+ 2 1 . If X1 + X2 > x.v.B*(n ) B(dz) (1. we therefore get lim sup B*2(x)/B(x) > 1+B(y)+ 1 . then either one of the Xi exceeds (1 .y)/B(x) > 1. we get limsupB*2(x)/B(x) < 2. yo] as X + 00. 1 < B(x ) B( x) Y) < B( 0). This follows since the probability of the overshoot to exceed y is B (x + y)/B(x ) which has limit 1.z B(x) . a contradiction.5)x)' + 0 _ 2 L(x)l xa (16) Letting S 10.2) B(x) B(x ) B(x) Jo with n = 1 and splitting the integral into two corresponding to the intervals [0.y)/B(x) > 1 since y > 0. We now turn to the mathematical theory of subexponential distributions. then B(B(x)y) * 1 uniformly in y E [0. we get BZ(x)) > 1 + B(y) + B(B()y) (B(x) .xIX > x converges in distribution tooo.1(b) we get B*2(x)/B(x) * 2.
This implies B(x) > c2e5x for all x.B*2 (x) B(x) (x .z ) (x ) = 1 + (^ B(x .y) sup v>o B(v) B(x) which converges to 0 by Proposition 1. The case n = 2 is just the definition.z) B(dz) 2B(x) o rv 2 0 2 using Proposition 1. Proof We use induction. so assume the proposition has been shown for n.z) B(dz) (n + O(e)) ^x JO B(x) (n + 0(0) I B (x) . Proposition 1. 0 Proof of Proposition 1. choose y such that IB*n(x)/B(x) .z) B(x) Here the second integral can be bounded by B*n(y) B(x) .B(x .2. Then by (1. we have by Proposition 1. Proof For 0 < 5 < e.z) B(dz) _y B(x) 111 Lx B . B*(n+1) (x I xy + Jxx y) W.7 If B E S. then for any n B*n(x)/B(x) * n. The first integral is y B(x .2).1) for all large n so that B(n) > cle6n for all n.z) B(dz).5 and the induction hypothesis.nI < e for x > y.254 CHAPTER IX. and this immediately yields the desired conclusions.5 and dominated convergence.5 that B(n) > e6B(n . O The following result is extremely important and is often taken as definition of the class S.6 If B E 8. x oo. 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 . b[c] = oo for all e > 0. B(x) \Jo _ B(x . then e"R(x) * oo. Given e > 0. HEAVY TAILS Corollary 1.. its intuitive content is the same as discussed in the case n = 2 above.
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) .9 Let A1. X2 be independent r.y)Ai(dy) v) f o .y)Ai(dy) = (x)o(1) (1. it follows that it is necessary and sufficient for the assertion to be true that JX_VA (x . Proposition 1. an+1 fX B*n( *n(x . an = supx>o B*n(x)/B(x). choose T such that (B(x)B*2(x))/B(x) < 1 + b for x > T and let A = 1/B(T).z) B(x) < 1 + A + an sup f x B(x .z) B(dz) x . Xi <= v Ai (x . A2 be distributions on (0.i). Proof Define 5 > 0 by (1+5)2 = 1+e. For any fixed v.4) .X1 > xv. Then by (1.ajB(x)Ai(v) = ajB( x)(1+o„(1)) (j = 3 . 0 Lemma 1. oo) such that Ai (x) _ aiB(x) for some B E S and some constants al.3) Using the necessity part in the case Al = A2 = B yields f xv B(x .y)B(dy) = B(x)ov (1)• v (1.X2 > xv) < A1(xv)A2(x v) . 0 Proposition 1. Proof Let X1. a2 with a1 + a2 > 0.y))/B(x).0 completes the proof. then there exists a constant K = KE such that B*n(x) < K(1 + e)nB(x) for all n and x.z) B(dz ) + sup < 1 + sup f x<T B ( x) x>T 0 B(x .2).8 If B E S.z) B(x . Then Al * A2(x) = P(X1 + X2 > x).ala2B(x)2 which can be neglected. Then Al * A2 (x) . e > 0. Since P(X1+X2 > x.(al + a2)B(x).5 easily yields P(X1 + X2 > x.1.v.'s such that Xi has distribution Ai.B(x . Combining these estimates and letting a 4.z) B(dz) < 1 + A + an(1 + d) . 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.
2A(x).Bl (x) + B2 (x) follows precisely as in the proof of Proposition 1.B(x) Proof Take Al = A. then A E S. with a > 0 and L1. That is. However.13 Let B have density b and failure rate A(x) such that . Proof Taking Al = A2 = A. then so is L = L1 + L2. In the regularly varying case.3) follows if CHAPTER LX. HEAVY TAILS 'VV B(x .y)Ai(dy) = B(x)o„(1).2aB(x) .2. V (1. i = 1.9).y)B(dy). of (1.5) Here approximately the last term is B(x)o„(1) by ( 1. u Corollary 1. u Corollary 1. We next give a classical sufficient (and close to necessary) condition for subexponentiality due to Pitman [290].v)B(v) + _'U Aq(x . . u It is tempting to conjecture that S is closed under convolution. Then L = L1 + L2 is slowly varying and B1 * B2(x) sim L(x)/x«. the l. f " By a change of variables. Recall that the failure rate A(x) of a distribution B with density b is A(x) = b(x)/B(x) Proposition 1.11 Let B E S and let A be any distribution with a ligther tail. L2 slowly varying. A2 = B so that a1 = 0.(x) is decreasing for x > x0 with limit 0 at oo.Bl (x) + B2 (x) when B1. a2 = 1.256 Now (1.4). B2 E S. 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.5) becomes x B(x .10 The class S is closed under tailequivalence. Then B E S provided fo "O exA(x) b(x) dx < oo. B1 * B2 (x) . it should hold that B1 * B2 E S and B1 * B2 (x) .12 Assume that Bi(x) = Li(x)lxa. A(x) = o(B(x)). L2 are slowly varying.Ai(x . if q(x) aB(x) for some B E S and some constant a > 0. That is. Hence Corollary 1.h.s.aiB(v)) = B(x)o„(1). whereas the two first yield B(x)(Ai(v) . it is easy to see that if L1. a1 = a2 = a yields A*2(x) .v)Ai(v) .
12. To illustrate how Proposition 1.y) y\(y)• The rightmost bound shows that the integrand in the first integral is bounded by ey"(v). Since ) (x .1)xcl1 .y) < yA(x . a(x) = ax01.1. the first integral has limit 1 . we can use the same domination for the second integral but now the integrand has limit 0 . subexponentiality has alrady been proved in Corollary 1. The middle bound shows that it converges to b(y) for any fixed y since \ (x .A(xy)A ( y). Jo For y < x/2. Then B(x) = eA(x). In the regularly varying case.(x . an integrable function by assumption. proving B E S. Define A(x) = fo .y) dy.2).3 < 1. Example 1. By (1. Thus B*2(x )/ B(x) .y) < A (y) for y < x/2.14 Consider the DFR Weibull case B(x) = ex0 with 0 <.U) /or) v 2x This yields easily that ex. x .`(x)b(x) is integrable.. the DFR Weibull distriu bution is subexponential.16 For L(x) slowly varying and a > 1. Further. Goldie & Teugels [66]): Proposition 1. Example 1. Thus.15 In the lognormal distribution.(y) dy. L(x) y° (a . and exa(x)b(x) = (3x01e(10)x9 is integrable.e009xv)2/2a2/(x 2irv2) logx ( ) 't ((logx . elementary but tedious calculations (which we omit) show that A(x) is ultimately decreasing. replace B by a tail equivalent distribution with a failure rate which is everywhere decreasing). Thus A(x) is everywhere decreasing. 0 A(x) . B*2(x) . the u lognormal distribution is subexponential. we first quote Karamata's theorem (Bingham.y) * 0. Thus by dominated convergence .A(y)a(y ) = ev'(y) b(y). Then b(x) = Ox0lexp.1 has limit 1 + 0.1 B(x) eA( x)A(xv )A(y)A(y) dy f B(x . SUBEXPONENTIAL DISTRIBUTIONS 257 Proof We may assume that A(x) is everywhere decreasing (otherwise.y ) b(y)dy = B (x) o ox _ J = ox/2 eA( x)A(xy )..A(x .13 works in this setting. Thus.A(y)\(y) dy + fox/ 2 eA(x ). f ' L(y) dy .
we get (1. We conclude with a property of subexponential distributions which is often extremely important: under some mild smoothness assumptions.1/A(x) and P(X ixil'Y (x) > y) * e'.1)])a 1 1 .1) and P(X (.1))^ ' (b) Assume that for any yo )t(x + y/A(x)) 1 A(x) uniformly for y E (0. Then: Proposition 1. However. Then 7(x) .ea b(x) is integrable. then B(x) .18 (a) If B has a density of the form b(x) = aL(x)/xa with L(x) slowly varying and a > 1. the overshoot properly normalized has a limit which is Pareto if B is regularly varying and exponential for distributions like the lognormal or Weibull. Thus exa(x)b(x) . (1 + y/(a . HEAVY TAILS Proposition 1.258 From this we get CHAPTER IX. f O B(y) dy . let X W = X . .6) EX(x) . 'y(x) = EXix>.x)+ _ 1 °° P(X > x) P(X>x )J L PX >y)dy 1 x L(y)/ydy L(x)/((a1)x'1) x )l ° J °° ( ()l a x a1 Further P ((a .1)X(x)/x > y) = P(X > x[1 + y/(a .E(X .a/x.y(x)B(x).13 may present a problem in some cases so that the direct proof in Proposition 1. then 7(x) x/(a .1)] I X > x) L(x[1 + y/(a . More precisely.L(x)/x" and )t(x) .1))a .17 If B has a density of the form b(x) = aL(x)/x°+1 with L(x) slowly varying and a > 1.)/Y(x) > y) (1 + y/(a . Proof ( a): Using Karamata's theorem.4 is necessary in full generality. (c) Under the assumptions of either ( a) or (b).1)]) xa L(x) (x[1 + y/(a . the monotonicity condition in Proposition 1. yo] .xjX > x.
f yl 0 0 = exp {y (1 + 0(1))} 0 fY A( x + u /A( x)) a(x) du } The property (1... Let St = Ei ` Ui .15.. Bo(x) = f0 B(y) dy / µB. P The proof is based upon the following lemma (stated slightly more generally than needed at present). Kliippelberg & Mikosch [134].1 If Bo E S. We assume p = /3µB < 1 and are interested in the ruin probability V)(u) = P(M > u) = P(r(u) < oo).2.2 Let Y1. Examples 1. Notes and references A good general reference for subexponential distribution is Embrechts. It is trivially verified to hold for the Weibull. and that for each Proof Recall from Section 1 that G*n (u) z > 1 there is a D < oo such that G*n(u) < G(u)Dzn for all u. with EzK < oo for some z > 1. Theorem 2 . Then P(Y1 + • • • + YK > u) . Lemma 2.EK G(u).and lognormal distributions .t be the claim surplus at time t and M = sups>0 St.nn. THE COMPOUND POISSON MODEL 259 We omit the proof of (c) and that EX (x) . then Vi(u) P Bo(u).n0 1•P(K= n)•n = EK.v. St > u}. be i.A(x) I X > x) = exp {A(x) . Y2. 0 G(u) L G(u) ..7) is referred to as 1/A(x) being selfneglecting.14. u a oo. d. We get p(yl+.1/. . i. 2 The compound Poisson model Consider the compound Poisson model with arrival intensity /3 and claim size distribution B. The remaining statement (1. r(u) = inf it > 0. . nG(u).(x).A(x + y/A(x))} =a(x) a(x + x) dx = ex p ex P .8) in (b) then follows from P (A(x)X (x) > y) = F(X > x + y/. 1. with common distribution G E S and let K be an independent integervalued r. cf.+YK> u) = ^•P(K = n)G* n(u ) . Recall that B0 denotes the stationary excess distribution.
x400 PBB(x) PB Leta+oo. lognormal .p) and EzK < oo whenever pz < 1.2) M = Yl + • • • +YK where the Yt have distribution Bo and K is geometric with parameter p. in our three main examples (regular variation . we have fx B(y)dy = a B0 (x) > lim inf lim inf x+oo B(x) . mathematically one must note that there exist (quite intricate) examples where B E S.1) In particular . Since EK = p/(1.µB(01 . Weibull) one has Bo(x ( B(x) . .3 If B E S.?(xµ 8 (x). However. x 4 00. The tail of Bo is easily expressed in terms of the tail of B and the function y(x) in Proposition 1.µ J ) .p)p'. and for the lognormal and Weibull cases it can be verified using Pitman 's criterion (Proposition 1. Bo E S. (2. u The condition Bo E S is for all practical purposes equivalent to B E S.1 is notoriously not very accurate. as well as examples where B ¢ S. In general: Proposition 2. HEAVY TAILS u using dominated convergence with >2 P(K = n) Dz" as majorant.x^ ) B(x) _ f or ( lox . The approximation in Theorem 2. _ B(x^sx Bo(x) µ8 I aoB(y )dy = (^) .18. Bo E S is immediate in the regularly varying case. the result follows immediately from Lemma 2. r(1/Q) xlQexp B(x) = ex' From this . Bo ¢ S. Proof of Theorem 2. The problem is a very slow rate of convergence as u ^ oo.1 is essentially due to von Bahr [56].1)xa1' vxe(109x11)2/202 2 +° /2 µB = eµ Bo(x) eµ+O2/2(log x)2 27r' = µB = F(1/0 ) Bo(x 1 ) . a]. u x+a Notes and references Theorem 2. then Bo(x)/B(x) + 00. The PollaczeckKhinchine formula states that (in the setup of Lemma 2.2. see Abate. Bo is more heavytailed than B . P(K = k) = (1. Borovkov [73] and Pakes [280].13).. Proof Since B(x + y)/B(x) * 1 uniformly in y E [0.260 CHAPTER IX.1. See also Embrechts & Veraverbeeke [136]. Note that in these examples .. For some numerical studies.
. G+ (A) = P(Sq+ E A. M = sup s$ . Thus G+ is the ascending ladder height distribution (which is defective because of PB < PA).9+ < oo) = P(S. i=1 .1 gives 1010.... To Bo(u) u + 00. T1 the ith interarrival time and Xi = U. Asmussen & Binswanger [27] suggested an approximation which is substantially better than Theorem 2. Define further 0 = IIG+II = P(r9+ < oo). let t9+ = i9(0) be the first ascending ladder epoch of {Snd> }. Snd) = Xl +. 1 Assume that (a) the stationary excess distribution Bo of B is subexponential and that (b) B itself satisfies B(x . Let U= be the ith claim . We assume positive safety loading. in [219] p. Then K M=EY.] The proof is based upon the observation that also in the renewal setting.. one may have to go out to values of 1/'(u) which are unrealistically small before the fit is reasonable. .1. [279].g. In [1].e. i. THE RENEWAL MODEL 261 Choudhury & Whitt [1].. Somewhat related work is in Omey & Willekens [278]. there is a representation of M similar to the PollaczeckKhinchine formula.3. 195 there are numerical examples where tp(u) is of order 105 but Theorem 2. p = iB /µA < 1. This shows that even the approximation is asymptotically correct in the tail. T+ < oo) where r+ = T1 + • • • + T.1 when u is small or moderately large. also a second order term is introduced but unfortunately it does not present a great improvement. Based upon ideas of Hogan [200].y + as usual denotes the first ascending ladder epoch of the continuous time claim surplus process {St}.. 3 The renewal model We consider the renewal model with claim size distribution B and interarrival distribution A as in Chapter V.y)/B (x) > 1 uniformly on compact y internals. Kalashnikov [219] and Asmussen & Binswanger [27]. E. (3. Then l/i(u) 1 P P [Note that (b) in particular holds if B E S. {n= 0. t9(u) = inf {n : Snd> > u} .Ti.1) this end .} Then ik(u) = F ( M > u) = P(i9 (u) < oo). The main result is: Theorem 3 .+ E A. + Xn.
FI(x) /IPG_I.B(x).y+ given r+ < oo).1) is equivalent to P(M > u) " . P(K = k) = (1 .. 0 The lemma implies that (3..g+ > x.y_ E A) the descending ladder height distribution (IIG II = 1 because of PB < P A) and let PG_ be the mean of G_. Let F denote the distribution of the Xi and F1 the integrated tail.2 F(x) . are independent of K and i.PBBo(x).d)) E A) denote the pre19+ occupation measure and let and U_ = Eo G'_" be the renewal measure corresponding to G_.oo. 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. u a 00. B(x) _ J O° B(B(x)y) A(dy) f 1 . 0] to the integral is O(F(x)) = o(FI(x)). 0] normalized by IPG_ I so that we should have to G+(x) . Let further 19_ _ inf {n > 0: S^d^ < 0} be the first descending ladder epoch. and hence FI(x) .. we will use the fact that the proof of (3.d. with distribution G+/9 (the distribution of S.3 G+ (x) . As for the compound Poisson model.i. cf. Then 0 0 F( x .1) holds for a general random walk satisfying the analogues of (a).Y2.3) and we will prove it in this form (in the next Section. x * oo.FI(u).y) dy = 1 Pi (X) oo IPG_ I . Lemma 3 .N.9)9'' and Y1.1 IPG_ I / F(x . x + oo. whereas for large y . the contribution from the interval (. Proof By dominated convergence and (b). A. FI (x) _ fz ° F(y) dy.Ti). x > 0.(. The heuristics is now that because of (b). Write G+( x) = G+ ( x. (b) and does not rely on the structure Xi = Ui . G_(A) = P(S. (3. Proof Let R+(A) = E E'+ ' I(S.262 CHAPTER IX. U_ (dy) is close to Lebesgue measure on (. oo) = F(S. d+ < oo). HEAVY TAILS where K is geometric with parameter 9.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_. Lemma 3 . A(dy) = 1.2).
u Proof of Theorem 3. choose N such that F(n .3. 0] x+00 FI(x) 00 + lim up 1 x) E F(x + n) U_ (n . F(Y= > x) FI(x)/(OIp _ 1).1.1.1. In the lattice case.I n=N (1 E)2 r00 F(x + y) dy + e) lim sup .F[s] = (1 .=1 BIp G_ I (1.G+[s]) .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. THE RENEWAL MODEL 263 We now make this precise. Similarly. > (1 . and that U_(n . (3.3.1)/F(n) < 1 + e for n > N (this is possible by (b) and Lemma 3. n] + 1/I µG_ I.O[s])(1 .1.9)IpG_ I Differentiating the WienerHopf factorization identity (A. n] is just the probability of a renewal at n. then by Blackwell 's renewal theorem U_ (n . Hence using dominated convergence precisely as for the compound Poisson model. and in the last that FI is asymptotically proportional to Bo E S.1 I .y) U_ (dy) 00 FI (x) < lim sup F(x) U(N. we can assume that the span is 1 and then the same conclusion holds since then U(n .UG_ I x. the proof is complete.(dy) fN FI ( x) + lim sup ZY00 N F(x . We then get lim sup G+(x) xro0 Fj(x) < lim sup X)00 o F(x . If G_ is nonlattice.0)0k k I(u) A.2) yields 00 F F I (u) P(M > u) _ E(1 .9) 1 .1. n] F1 ( n=N _1 1+e E F(x+n) 0 + limsup xr00 FI(x) FAG. Given e.2).e) z lim inf G+(x)  FI (x) Ip G_ I Letting a 10. n] < (1 + e)/1µc_ I for n > N. By Lemma 3.y) U.
a. allowing also for possible dependence between the arrival process and the claim sizes.So(u)) are available. on the set {M > u.l. S+9(u) .2. Notes and references Theorem 3. Mn < u}. Sty(u) . must attain a maximum > 0 so that P(M > u. u)) > P(w(u) < oo)(i lp (0))• On the other hand.IIG+II)µc_ = (1 .(1 .a.So( u)_1 < a) < P (w(u) < oo)j/i(0) < 0(0) P(M E (u . FJ(u) UBBO(U) PBo(u) N = (10)Ipc_I JUA .4 For any a < oo.4 on the joint distribution of (S. 10(0) But since P(M > u . Note that substantially sharper statements than Lemma 3. we have P(M E (u .a) N P(M > u).u)) = o(P (M > u)) = o(FI(u)). Proof Let w(u) = inf {n : Sid) E (u .AB iP We conclude by a lemma needed in the next section: Lemma 3 . 4 Models with dependent input We now generalize one step further and consider risk processes with dependent interclaim times. HEAVY TAILS µF = (1 . u)).Sty(u)_I < a} we have w(u) < oo. u).(u)+n .SS(u)}n=o... see Asmussen & Kliippelberg [36]. S+q(u) ..yiui_1.1 is due to Embrechts & Veraverbeke [136]. 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. with roots in von Bahr [56] and Pakes [280].264 and letting s = 0 yields CHAPTER IX. P(M > u.1)6+[0] .a.Se(u)_1 < a) = o(Fj(u)). . Therefore by Lemma 3. Then P(M E (u .a. and {Su.0)ua_ .
. We give here one of them.1 Note that no specific sample path structure of {St} (like in Fig. and the distribution of {Sxk+t . < 0 and EoX < oo where X = X2 .1. E0. For further approaches. Thus the assumption .1 = max k=0.. ..1.1 based upon a regenerative assumption. = Sx. See Fig.F*(X) = P0(Si > x) .X1 is the generic cycle... Define S. G(x) (4.4 below.Sxi}0<t<x2Xl .. (viewed as random elements of the space of Dfunctions with finite lifelengths) are i... see [47].n n=0. Theorem 4.1 except for the first one) is a random walk. {Sn}n=o..1) is assumed. MODELS WITH DEPENDENT INPUT 265 Various criteria for this to be true were recently given by Asmussen.. and apply it to the Markovmodulated model of Chapter VI.Sxk}o<t<xk+1xk is the same for all k = 1. (corresponding to the filled circles on Fig.. M = sup St.1) . Assume that the claim surplus process {St}t>o has a regenerative structure in the sense that there exists a renewal process Xo = 0 < Xl <. We let F* denote the Podistribution of Si. 0o(u) etc.i. We return to this point in Example 4. 4.. 4. t>0 S. The idea is now to observe that in the zerodelayed case. {SX1+t ..... The zerodelayed case corresponds to Xo = Xl = 0 and we write then F0.1 where the filled circles symbolize a regeneration in the path. M* = max S.. +1. Schmidli & Schmidt [47]. assume pp.4.d.. M.X2 < . 4. Figure 4.. such that {SXo+t  SXo}0<t< X 1Xo . examples and counterexamples.. 2..
Since clearly M(x) > Sl .. it suffices by (4. (4.2. Fo(Si > X).1 Assume that (4.2) to show F(M* > u) > 1. The one we focus on is Fo (Mix) > x) .2) Imposing suitable conditions on the behaviour of {St} within a cycle will then ensure that M and M* are sufficiently close to be tail equivalent. Proof Since M > M*.3) where Mnx) = sup o<t<xn +1 X. See Fig. (4..* i o<t<xn+1x.1) and (4.266 CHAPTER IX. 4. u p 00.2 Theorem 4.4) liminf u>oo F(M > u) .S.Sxn = sup Sxn+t .. jF11 F* (U).3) hold. Then '00 (u) = Fo(M > u) . HEAVY TAILS for some G such that both G E S and Go E S makes (3. (4. the assumption means that Mix) and Sl are not too far away.3) applicable so that F(M* > u) 141 F*(u). N N Xi=0 N Figure 4. Sxn +t ..
Mn +1 >aV(u n=1 00 S.S.4. Po(M* > u) .e)Po (M > u. Given e > 0. )) > (1 . We shall use the estimate Po(M > u) Miu^+ 1 < a) = o(Po (M > u)) (4. Theorem 4. M^xu)+l > a) .(u) .1 = limti00 St/t...: Sn > u} . 0 yields (4.a.. Letting first u + oo and next e . S. u))/P(M* = 0) = o(Po(M* > u)).4.. assume the path structure Nt St = EUit+Zt i=1 . Under suitable conditions . MODELS WITH DEPENDENT INPUT Define 79* (u) = inf {n = 1 .e)Po (M > U).4). x > a.Po (M* > u.E) Po ( n max St u... Mn+l > a V (u .1 can be rewritten as 00 (U) (4.e)Po (MMX> > x). .( u)1 > a) 00 1: Po(Mn<u. Let a > 0 be fixed.+Mn+1>u} 267 (note that {M> u} = {3(u) < oo}).6) 1 p pBo(u) u where B is the Palm distribution of claims and p . 2.. MW O(u)+1 < a) IN ( U n=1 A1.a. To this end.(1 . choose a such that Po(Si > x ) > (1 . u)} < P(M* E (u .Sn 0<t<x„+j ( 1 .5) which follows since Po (M > u. E (u .: S. Then by Lemma 3. /3(u) = inf{n=1.2.Sn+1Sn>aV(uSn*)) n=1 00 > (1E)EPo(Mn<u.
X both have tails of sup Zt. and the rest is just rewriting of constants: since p = 1+tlim St = 1+ .p) Ju P Bo(u) 1p 0 . are Fmeasurable and NX Po J:U=>x i=1 (iv) Po sup Zt > x / (0:5t<x o(B(x)) Then (4. we get 00 (u) 1 IPF. cf. Corollary 4.2 Assume that {St} is regenerative and satisfies (4.3PB.'s order EoNx • B(x).6 below. (ii) EozNX < oo for some z > 1. Assume further that (i) both B and Bo are subexponential. and also for Mix) since Nx FNX U.Q = EoNx/EoX. The same is true for Sl. Mix) < > UE + i=1 o<t<x Thus Theorem 4. Then the Palm distribution of claims is B(x) = E N Eo 0 I( U1 < x) .1 is in force. the proof of Lemma 4.6) holds with p = .v. i=1 (4. independent of {> CHAPTER IX. (iii) For some o field Y.I u J Po(Sl > x) dx 1 EoNxB(x) dx EoX(1 .7). oX (see Proposition A1.8) x Write . a4' 0. X and N. and ENX Ui .4).} and satisfying Zt/t N. since the tail of Zx is lighter than B(x) by (iv). Proof It is easily seen that the r. HEAVY TAILS N` U.268 with {Zt} continuous.
consider the periodic model of VI. 1) is Poisson with rate /3 = fo /3(s) ds so that (ii) holds. (iii) is obvious. X2 = 1. i=1 B = >2 7riaiBi i=1 and we assume p = 014 B = Ep ri/3ipB.4 Assume that St = Zt .0 (thus (iv) is trivial).. (i) holds. of claims arriving in [0.3 that (4. 3 The average arrival rate / and the Palm distribution B of the claim sizes are given by P P Q = ir i/i.6) u holds. Zt .t + EN'I Ui where {>N`1 Ui .9).5 Consider the Markovmodulated risk model with claim size distributions satisfying (4. Bo E S.e. 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. Assume that B E S. The key step of the proof is the following lemma. Bo E S. We now return to the Markovmodulated risk model of Chapter VI with background Markov process {Jt} with p < oo states and stationary distribution 7r.. > 0. in particular lighttailed.6) holds. . . X3 = 2. then (iv) holds since the distribution of supo<t<i Z(t) is the same as that of I Zl 1. . . we assume that B E S.3 As a first quick application. X3 = 2.t} is standard compound Poisson and {Zt} an independent Brownian motion with mean zero and variance constant a2. In particular. we conclude just as in Example 4. Thus we conclude that (4. Then (4. we will assume that lim B2(x) = ci x+oo G(x) for some distribution G such that both G and the integrated tail fx°O G(y) dy are subexponential . The number N. X2 = 1. and taking F = o. The arrival rate is /3i and the claim size distribution Bi when Jt = i. i.4.(NX). and for some constants ci < oo such that cl + • • • + c.. < 1. Theorem 4.6) holds.. The regenerative assumption is satisfied if we take Xo = Xi = 0. Taking again Xo = Xi = 0. We consider the case where one or more of the claim size distributions Bi are heavytailed. Again . MODELS WITH DEPENDENT INPUT 269 Example 4 . Example 4 . More precisely.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)..
6 Let (N1.270 CHAPTER IX. oo) and define p Ni Yx = EEX'i . 2 .. Markovmodulation typically decreases the adjustment coefficient y and thereby changes the order of magnitude of the ruin . 1. Then P P(Yx > x) . If Jo = i.. .. NP ) and X are . u Proof of Theorem 4. Assume EzN1+"'+Np < oo for some z > 1 and all i. .. . X > 0 a r.G( x ) > ciNi .. It follows by a slight extension of results from Section 1 that P P(Yo > x I Y) G( x) ci Ni. are independent with distribution Fi for Xij. cp with cl + > 0 it holds that Fi(x) .X i=1 j=1 where conditionally upon F the Xi. For lighttailed distributions. HEAVY TAILS Lemma 4 .}P. we can define the regenerations points as the times of returns to i.c'(x) where c = ciENi .ciG(x). . and the rest of the argument is then just as the proof of Corollary 4. P P P(YX and > x I."+Np .. .5. i =1 P(Yo > x I ^ ) < CG(x)zN1+ +Np for some C = C(z) < oo.Fmeasurable. An easy conditioning argument then yields the result when Jo is u random.F) < CG(x)zn'1+.v. and F a aalgebra such that (N1. i=1 P(Yx > x ^) < P(Y0 > x I. The same dominated convergence argument completes the proof. i1 = E\ G(x) In the general case.2.F) = P(Yo > X+x I •^) G (x +x)>2ciNi i=1 . Let {Fi}t=1 P be a family of distributions on [0. . oo) such that G E S and some c1. Thus dominated convergence yields ( P(Yo>x P(Yo>x . as x a oo. and that for some + cp distribution G on [0. i=1 Proof Consider first the case X = 0.. NP ) be a random vector in {0.. .^•) G(x) P ^ E ciNi = C.
As usual. Theorem 5.. FINITEHORIZON RUIN PROBABILITIES 271 probabilities for large u..i. states that under mild additional conditions.4.2 and Example 4. The present approach via Theorem 4. we let PN"N = P(.6) to hold in a situation where the interclaim times (T1.e. i.T2. For further studies of perturbations like in Corollary 4. i. That paper also contains further criteria for regenerative input (in particular also a treatment of the delayed case which we have omitted here).1.7). Then O(u) . The main result of this section.3. cf. An improvement was given in Asmussen & Hojgaard [33]. I T(u) < oo). In contrast.p)Bo(u). the discussion provides an alternative point of view to some results in Chapter IV. 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). Notes and references Theorem 4. Combined with the approximation for O(u). Schmidli & Schmidt [47].5 shows that basically only the tail dominant claim size distributions (those with c. 5 was first proved by Asmussen. as well as a condition for (4. and the final reduction by Jelenkovic & Lazar [213]. m is a (orfinite) . and independent of (T1. for lighttailed distributions the value of the adjustment coefficient y is given by a delicate interaction between all B. this is applied for example to risk processes with Poisson cluster arrivals. this should be compared with the normal limit for the lighttailed case. IV. in particular Proposition 2.4.5.7.. Floe Henriksen & Kliippelberg [31] by a lengthy argument which did not provide the constant in front of Bo(u) in final form. VI.. Essentially. cf. see Schlegel [316].d.4. ) form a general stationary sequence and the U.. Theorem 4.5 that the effect of Markovmodulation is in some sense less dramatical for heavytailed distributions: the order of magnitude of the ruin probabilities remains ft°° B(x) dx. We start by reviewing some general facts which are fundamental for the analysis.T2.4. Within the class of risk processes in a Markovian environment. ). 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. > 0) matter for determining the order of magnitude of the ruin probabilities in the heavytailed case. 5 Finitehorizon ruin probabilities We consider the compound Poisson model with p = /3pB < 1 and the stationary excess distribution Bo subexponential. cf.pl(1 . r(u) is the time of ruin and as in IV. this then easily yields approximations for the finite horizon ruin probabilities (Corollary 5.. Theorem 2.1 is from Asmussen. It follows from Theorem 4.
2) means ffh(a.s. follows by the substitution y = x .1 A compound Poisson risk process {Rt} and its associated claim surplus process {St} are in classical duality w . Sw(F. Proof Starting from Ro = x. k on E. (note that we allow x. Let G denote the distribution of ENt U. For the present purposes it suffices . HEAVY TAILS measure on E such that L for all measurable A C E and all t > 0. r. j. 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. for states i. St is distributed as y .h. {Rt}. resp. k as indicator functions.h.z. u For F C E. . w(Fc) < oo ) 'In general Markov process theory. however .z) dx G(dz) = ffh(y + z) k(y)dy G(dz).).t.s. oo). a familiar case is time reversion (here m is the stationary distribution). say. and starting from So = y. Rt is distributed as x + t .r. t.)k(x .s=j are the transition probabilities for {St}. .>N` Ui.2) for all bounded measurable functions h.00). to consider only the case Px(w(F`) = 0) 0. where we can take h. to the r. The simplest example is a discrete time discrete state space chain. The equality of the l.272 CHAPTER IX.y = Qx (. the whole of R and not as usual impose the restrictions x > 0.2) with t = 1 means m. We let QS be the corresponding distribution and Qx. Then (5. Say {St} is reflected Brownian motion on [0.= y. {St} and {Rt} are in classical duality w. y to vary in. Thus. in the terminology of general Markov process theory. x = 0+ and F = (0. Lebesgue measure. and (5. y = 0).rij = mjsji where r13. but the example of relevance for us is the following: Proposition 5. m. a main difficulty is to make sense to such excursions also when Px(w(F°) = 0) = 1. an excursion in F starting from x E F is the (typically finite) piece of sample path' {St}o<t<w(F°) I So = x where w(Fc) = inf It > 0: St 0 F} .t.t + EI U.
= y) Theorem 5 .3 The distribution of r(0) given r(0) < oo.y(2p21 .s. 5. . FINITEHORIZON RUIN PROBABILITIES 273 y E F (in discrete time. w(0.13AB < 1] Proof of Theorem 5.(0)_ = y < 0 is the same as the distribution of w(y) where w(z) = inf It > 0 : Rt = z}. i1.2. Sn = in = y. the one in (b) is the time reversed path. [note that w(z) < oo a. and we let Qy y refer to the time reversed excursion . when p = . z > 0. x = 0. Thus.. in E F.y() = P ({SW(F`)t} 0<t<w(F °) E So = x. 0].1 for the case F = (oo... The theorem states that the path in (b) has the same distribution as an excursion of {Rt} conditioned to start in y < 0 and to end in x = 0. We can then view Qy. In particular: Corollary 5...y as a measure on all strings of the form i0i1 . The theorem is illustrated in Fig ..SS(F.y = Qy Q.5. io = x. in with i0. But in the risk theory example (corresponding to which the sample paths are drawn). this simply means the distribution of the path of {Rt} starting from y and stopped when 0 is hit. That is.1 The sample path in (a) is the excursion of {St} conditioned to start in x = 0 and to end in y > 0. Qx y is the distribution of an excursion of {St} conditioned to start in x E F and terminate in y E F. . S.2 Qy.. in = y. Sw(Fo)_ should be interpreted as Sw(F^)_1).).itt) = P Px(w(Fc) < 00. .. 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). QR and QRy are defined similarly. Sw(F)1 = y) . /^s x (S1 = Z1. oo) = r(0) x= St y (a) Figure 5. Sn+1 E Fc) nx. Qx. .
R . in) = Qx... i0 = y...gilt' k=1 ii .rin_1in E Txj jEFC m21 s2120 m2252221 m in Ssn n1 mjSjx Mx m2p mil min1 jEF` 1 Sinin _ 1 .in E F. t' y and Qy x are measures on all strings of the form ipi l .... .. ..(F<)1 = Y) S S and Qx y( ipil . HEAVY TAILS E E Px (Si = 21i . S.. Si11 S 1 ..... in = x. note first that Pt' (R l = il... 2n) = Qx. Rn = in = x.. in)  Pt' (R1 = ii.....ik_1EF Sxin_1 .. 21 .in1 ..i„_iEF Similarly.. Rn+1 E F`) F (w(Fc) < 00...ii .ik1EF Similarly but easier Sxin_1 . Silt' E Sxik_1 . (Fc)1 = y) 00 CHAPTER IX.... Sn+1 E Fc) n=1 i1... Rn = in = x... . Rn+1 E FC) TioilTili2. . ....274 note that Fx(w(Fc) < 00. .y(inin _ E SYj jEF` 00 Sxik _1 .J (i. ..... MY Thus Qx(ioii ... To show Q y x (i0 i 1 .ik_1EF . S.. 2p).. R Qy x(2p21 ... i0) Q x.TI( 2n2n _1 .. . Silt' E SO k=1 i1... Si1y 00 jEF° E E 5xik_ 1 .. in = x.. in) = oo jEF^ Sxin1 .. in with 20.. in E F. 20 = y. 2p) when 20..... Si1y k=1 i1 . ... = in = y. Si l io E mjSjx.
')distribution of Z since P(Z>aIY>u) = 1 °° B(y) B(y + a) dy FLBBo(u) B (y) J°° (z) dy .2. see Fig.')density of Y is B(y)/[.p. that is.t.t. Y > y} . The formulation relevant for the present purposes states that Y has distribution Bo and that conditionally upon Y = y.r.v.r. Let Y = Yl = Sr+( 1) be the value of the claim surplus process just after the first ladder epoch . ST(o) > y.UBBo(u)]. Now the P(u.5. y > u. FINITEHORIZON RUIN PROBABILITIES 275 5b The time to ruin Our approach to the study of the asymptotic distribution of the ruin time is to decompose the path of { St} in ladder segments .')distribution of Yu is Bo"). To clarify the ideas we first consider the case where ruin occurs already in the first ladder segment . That is.2. P(o) ). 5. Z = Zl = ST+( 1)_ the value just before the first ladder epoch (these r.2. Z follows the excess distribution B(Y) given by B(Y) (x) _ B(y + x)/B(y). that is. Y > u). 7(0) < oo. P(") = P(.'s are defined w. the P(u.(o) > y} = {T(0) < oo.2 The distribution of (Y. Bo") is also the P(u. S. U T(O) = T (u) Y Figure 5. the distribution w. Z) is described in Theorem 111.B(a) +a PBBo(u) . 1 w . We are interested in the conditional distribution of T(u) = T(0) given {T(0) < oo. the case r (O) < oo.
1.i...1) of the last ladder segment can be estimated by the same approach as we used above when n = 1. r(u)/Z 4 1/(1 .p).. we get the same asymptotics as when n = 1. then by the subexponential property Yn must be large. . it therefore follows that T(u)/Z converges in Pi"'')probability to 1/(1 . Z1). 5. more precisely. Recall the definition of the auxiliary function y(x) in Section 1..T+(2). must be large and Z1.. a slight rewriting may be more appealing. z ^ oo. i. in particular of Z. .. . Zn_1 'typical' which implies that the first n1 ladder segment must be short and the last long.e. this in principle determines the asymptotic behaviour of r(u).276 CHAPTER IX. . . Zk be defined similarly as Y = Y1.. let r+(1) = T(0).3. Bo") ). . conditionally upon r+ (n) < oo. Y1 + • • • + Yn > u} denote the number of ladder steps leading to ruin and P("'n) = P(• I r(u) < oo.o be defined by w(z) = inf It > 0: Rt = z} where {Rt} is is independent of {St}.p) in Pi"'')distribution.d. 4 Assume that Bo E S and that (5. Hence Z. and YI. We now turn to the general case and will see that this conclusion also is true in P(")distribution: Theorem 5 .3 implies that the P("'1)distribution of T(u) = r(0) is that of w(Z). denote the ladder epochs and let Yk. Since w(z)/z a$. However.e.. Yn_1 'typical'. > u with high probability.. We let K(u) = inf In = 1.18(c) Bo")(yY (u)) + P(W > y) ( 5.. and since its dominates the first n .. Since the conditional distribution of Z is known (viz. cf. Z). That is . and distributed as (Y. i. 1/(1 . Then 7(u)/y(u) ^ W/(1 ..3) where the distribution of W is Pareto with mean one in case ( a) and exponential with mean one in case (b).3) holds.p) in F(u) distribution. (Y. Then Corollary 5. the random vectors (YI. It is straightforward that under the conditions of Proposition 1. are i..: r+ (n) < oo. 2. Z/'y(u) * W in Pi "' ')distribution . Now Bo E S implies that the Bo ")(a) + 0 for any fixed a. Zn). . The idea is now to observe that if K(u) = n. Z = ZI but relative to the kth ladder segment. In the proof. HEAVY TAILS Let {w(z)}Z^. Fig.p)..p) then yields the final result T(u)/y(u) + W/(1 . the duration T+ (n) ..r+ (n . K(u) = n). P(Z < a I Y > u) 3 0. Then.
I A'(u)) = P(u... +Yn1<u.3 In the following... FINITEHORIZON RUIN PROBABILITIES 277 16 Z3 Z1 r+(1) T+(1) T+(1) Figure 5. .n) (y1.u) E • I A'(u)) = Bo (n1) ®Bou) . I A"(u ))II + 0. > u}..Bo (ri1) ®B( . P (Yj. .. .Yl+ +Yf1>u}. A"(u) _ {K(u)=n} = {Y1+ P(.n). II ' II denotes the total variation norm between probability measures and ® product measure. Yn . suitably adapted). the condition on A'(u) A A"(u) follows from Bo being subexponential (Proposition 1. Proof We shall use the easily proved fact that if A'(u).2.Yn1iYn . Further. Y„1. P(. then IIP( I A'(u)) Taking A'(u) = {Y. .u) E •) .5 Ilp(u.u) II 0.5. Lemma 5. A"(u) are events such that P(A'(u) AA"(u)) = o(F (A'(u)) (A = symmetrical difference of events)..
see Fitzsimmons [144]).i.u has distribution Bout That is. Now use that if the conditional distribution of Z' given Y' is the same as the conditional distribution of Z given Y and JIF(Y E •) . + Y" > u) Flul (K (u ) = n) _ Cu) P"F(1'i +. Then according to Section 5a.1). the discussion just before the statement of Theorem 5.....r.p) < y).. .278 Lemma 5 .+y 1 p"F(Yn > u) P)Pn1 P/(1 ... Let {wl(z)}.. Similarly (replace u by 0). Thus F(u'n)(T(u) /7(u) > y) = F(u'n)((wl (Z1) + . +wn(Z n))l7( u ) > 1y) ^' P(u'n)(wn (Zn)/7(u) > y) 4 NW/(1 . Zn).d.. y > u. 2. n...2.. It therefore suffices to show that the P(u'")distribution of T(u) has the asserted limit..1 P PBo(u) • P(W/(1 .. k = 1....4). For Theorem 5. wk(Zk) has a proper limit distribution as u + oo for k < n. By Lemma 5.1. be independent random vectors such that the conditional distribution of Zk given Y. Y1 +. whereas wn(Zn) has the same limit behaviour as when n = 1 (cf..). Proof of Theorem 5. . the marginal distribution of Zk is Bo for k < n.4. . Proof Let (Y11. Z11). .. The same calculation as given above when n = 1 shows then that the marginal distribution of Zn is Bou). .7 O (u.6. . n_1 < u.. P(u) since by Theorem 2. {wn(z)} be i. Zn) E •) . and clearly Zi. HEAVY TAILS ((Z1'. Z' are arbitrary random vectors. Z. . (Y.' = y is BM. in particular his Proposition (2. .1 and Y„ . Y'. . . then 11P(Z E •) .t..P(Z' E •)II > 0 (here Y..6 IIPIu'n ) CHAPTER IX.P) Bo(u) for n = 1. .... The first step is to observe that K(u) has a proper limit distribution w.P(Y' E •)II * 0. n ...Bo (n1) ®Bo' 0.. in our example Y = (Y1. Notes and references Excursion theory for general Markov processes is a fairly abstract and advanced topic. copies of {w(z)}. Y") u etc. and that Yk has marginal distribution B0 for k = 1.y(u)T) . the F'distribution of r(u) is the same as the P'distribution of w1(Zl) + • • • + wn(Zn). the density of Yn is B(y)/[IBBO(u)]. Zn are independent.. .P) > y) Corollary 5.
Extensions to the Markovmodulated model of Chapter VI are in Asmussen & Hojgaard [33]. More precisely. Theorem 6 . nontrivial and we refer to Asmussen [22]. that fo p(x)1 dx < oo. We will show that the stationary density f (x) of {Vt} satisfies f (x) /B(x) r(x) We then get V. T) when T + oo with u fixed. Then 0 (u) Qf "O ^) dy. that MQ becomes large as consequence of one big jump. max VB>0I Vo=0^ o<s<t J11JJJ Lemma 6 .2 Define M. however. cf. claim size distribution B.B(u). 3. i. and premium rate p(x) at level x of the reserve. the results only cover the regularly varying case.(u) = P(V > u) = f f (y) dy .. Then P(MT > u) ./3Ea B(u). Asmussen & Teugels [53] studied approximations of i (u. u (6. The form of the result then follows by noting that the process has mean time Ea to make this big jump and that it then occurs with intensity /3B(u).(3 u u J B(y) dy .2. V. = supo<t<0. RESERVEDEPENDENT PREMIUMS 279 The results of Section 5b are from Asmussen & Kluppelberg [36] who also treated the renewal model and gave a sharp total variation limit result . The heuristic motivation is the usual in the heavytailed area.6. 6 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate /3. The rigorous proof is.e. x > oo. Proof of Theorem 6. Corollary II. the probability that is exceeds u is then B(u .1) The key step in the proof is the following lemma on the cycle maximum of the associated storage process {Vt}.1 Assume that B is subexponential and that p(x) > 00.y) . and define the cycle as a = inf{t>0: Vt=0.1. Assume for simplicity that {Vt} regenerates in state 0 . . one expects the level y form which the big jump occurs to be 0(1). p(Y) and the result follows.
q ( u)) 1 . u Notes and references The results are from Asmussen [22]. 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. .q(u) Now just use that p(x) * oo implies q (x) + 0. Then D(u) = f(u)p(u) and. D(u) = DQ(u)/µ. Hence f (u)r(u) = D(u) = Do(u) . by regenerative process theory.280 CHAPTER IX.P(MT > u) $B(u) Ft µ(1 . there exist constants c(u) 4 0 such that the limiting distribution of r(u)/c(u) given r(u) < oo is exponential. 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). where also the (easier) case of p(x) having a finite limit is treated . It is also shown in that paper that typically.
and this is the form in which the result of the simulation experiment is commonly reported. 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. Fox & Schrage [77]. 281 .2) is an asymptotic 95% confidence interval . Hence 1. The crude Monte Carlo ( CMC) method then amounts to simulating i.. . estimating z by the empirical mean (Z1 + • • + ZN)/N and the variance of Z by the empirical variance N s2 = E(Z{  N 2. Ripley [304].i. ZN.. where a2 = Var(Z )..Chapter X Simulation methodology 1 Generalities This section gives a summary of some basic issues in simulation and Monte Carlo methods .96s z f (1. 4Z).d. topics of direct relevance for the study of ruin probabilities are treated in more depth. z) 2 = Zit NE ii ii According to standard central limit theory . vrN(z .z) 4 N(0. replicates Zl. Rubinstein [310] or Rubinstein & Melamed [311] for more detail . We shall be brief concerning general aspects and refer to standard textbooks like Bratley.
Say that Var(Z') = Var(Z)/2. Letting Z' = E[Z I Y]. generated at the same time as Z. it is straightforward to use the CMC method to simulate the finite horizon ruin probability z = i. 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).282 CHAPTER X. there are others which are widely used in other areas and potentially useful also for ruin probabilities. Therefore. We survey two methods which are used below to study ruin probabilities. one can argue that unless Var(Z') is considerable smaller than Var(Z). so that Z' is a candidate for a Monte Carlo estimator of z.b(u. variance reduction is hardly worthwhile. This is a classical area of the simulation literature. 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. conditional Monte Carlo and importance sampling. We mention in particular ( regression adjusted) control variates and common random numbers. writing Var(Z) = Var(E [Z I Y]) + E(Var[Z I Y]) . T): just simulate the risk process {Rt} up to time T (or T n 7(u)) and let Z be the indicator that ruin has occurred. we then have EZ = EZ = z. typically by modifying Z to an alternative estimator Z' with EZ' = EZ = z and (hopefully) Var(Z') < Var(Z). an added programming effort. Sections 24 deal with alternative representations of Vi(u) allowing to overcome this difficulty. and in most cases this modest increase of N is totally unproblematic. Z = I inf Rt < 0 (0<t<T = I('r(u) < T). Typically variance reduction involves both some theoretical idea (in some cases also a mathematical calculation). 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'. and many sophisticated ideas have been developed. and a longer CPU time to produce one replication. However. SIMULATION METHODOLOGY In the setting of ruin probabilities. v. lb Variance reduction techniques The purpose of the techniques we study is to reduce the variance on a CMC estimator Z of z. Further.
e. . (1. i. Nevertheless. one would try to choose P to make large values of Z more likely. Variance reduction may or may not be obtained: it depends on the choice of the alternative measure P. L such that z = EZ = E[LZ].3). However. Then z Var(LZ) = E(LZ)2 . it gives a guidance: choose P such that dP/dP is as proportional to Z as possible. To this end. using the CMC method one generates (Z1. Thus we cannot compute L = Z/z (further. .E [Z Z]2 = z2 .z2 = 0.. .3) Thus. Thus. 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.1. it appears that we have produced an estimator with variance zero. the argument cheats because we are simulating since z is not avaliable analytically.zrs.zrs) = 2 1 N 2 2 2 i=1 i=1 N > Lt Zi . This may also be difficult to assess . L = z/Z (the event {Z = 0} is not a concern because P(Z = 0) = 0). but tentatively. it may often be impossible to describe P in such a way that it is straightforward to simulate from P). L1). GENERALITIES 283 and ignoring the last term shows that Var(Z') < Var(Z) so that conditional Monte Carlo always leads to variance reduction. and the problem is to make an efficient choice.. (ZN.v. In order to achieve (1. a crucial observation is that there is an optimal choice of P: define P by dP/dP = Z/EZ = Z/z. even if the optimal change of measure is not practical.[E(LZ)] = E Z2 Zz .96 sis v^ N 2 1 where srs = N j(LiZi . the obvious possibility is to take F and P mutually equivalent and L = dP/dP as the likelihood ratio. LN) from P and uses the estimator N zrs = N > L:Zj i=1 and the confidence interval zrs f 1.
However. a confidence interval of width 10 4 may look small. The CMC method leads to a variance of oZ = z(1 .284 CHAPTER X. the optimal P is the conditional distribution given A. We then . Thus. and further it is usually not practicable to simulate from P(•IA). Again. 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 . let z(u) = P(A(u)). SIMULATION METHODOLOGY 1c Rare events simulation The problem is to estimate z = P(A) when z is small .5 or even much smaller . say of the order 103 or less. z I.z) 1 > 00.e. Z z V5 In other words .z) which tends to zero as z ^ 0. we may try to make P look as much like P(•IA) as possible. i.96oz /(zV) = 0.96 2 z2 z increases like z1 as z .e.1. as is the case of typical interest.B = iP(AB) = P(BIA).z) 1001. To introduce these. in terms of the halfwidth of the confidence interval. but if the point estimate z is of the order 105.1. We shall focuse on importance sampling as a potential (though not the only) way to overcome this problem.96 2Z ( 1 . This leads to the equation 1. Two established efficiency criteria in rare events simulation are bounded relative error and logarithmic efficiency. An example where this works out nicely is given in Section 3. it does not help telling whether z is of the magnitude 104. if z is small. Z = I(A) and A is a rare event. assume that the rare event A = A(u) depends on a parameter u (say A = {r(u) < oo}). N . The optimal change of measure ( as discussed above) is given by P(B) = E [ Z] i. For each u. and let Z(u) be a Monte Carlo estimator of z(u). In ruin probability theory. A = {T(u) < T} or A = {r(u) < oo} and the rare events assumption amount to u being large. the issue is not so much that the precision is good as that relative precision is bad: oZ z(1 .e. However.100 . 10 . u + oo. 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. say 10%.. assume that the A(u) are rare in the sense that z(u) * 0..0. I.
Var(Z(u)) hm sup U+00 z (u) 2E < oo (1. this means that the sample size N = NE(u) required to obtain a given fixed relative precision (say a =10%) remains bounded. . but as a CMC method . The algorithm gives a solution to the infinite horizon problem . where Z = I(M > u) may be generated as follows: 1. Generate X1. F(K = k) = (1 .1. If M > u. the mathematical definition puts certain restrictions on this growth rate. P(K = k) = (1 . let Z +. which gives a logarithmically efficient estimator .log z(u) of (1.p)pk. with common density bo(x) = B(x)/µB and K is geometric with parameter p..4).1) may be written as V) (u) = P(M > u). i.p)pk. O (u) = z = EZ. SIMULATION VIA THE POLLACZECKKHINCHINE FORMULA 285 say that {Z(u)} has bounded relative error if Var(Z(u))/z(u)2 remains bounded as u 3 oo. Let M . it is appealing to combine with some variance reduction method . where X1. are i. The term logarithmic comes from the equivalent form . X2. XK from the density bo(x).log Var(Z(u)) lim inf > 2 u+oo . We shall here present an algorithm developed by Asmussen & Binswanger [ 271.X1 + + XK. Thus.4) for any e > 0. 2 Simulation via the PollaczeckKhinchine formula For the compound Poisson model.d.0. where M = X1 + • • • + XK. . . so that NE (u) may go to infinity..(2.2. Generate K as geometric. and in practice. Therefore . Notes and references For surveys on rare events simulation. let Z +. it is not efficient for large u .e. see Asmussen & Rubinstein [45] and Heidelberger [190]. This allows Var(Z(u)) to decrease slightly slower than z(u)2. However. logarithmic efficiency is almost as good as bounded relative error. 3. 2. the PollaczeckKhinchine formula III.. According to the above discussion. 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. Otherwise..i.
286 CHAPTER X.XK1] = EBo(uX1 ..X(n_1)) Bo(X(„_l) V X) Bo(X(n1)) .1) V)(u) .. So. To see this. and let Z(2)(u) = _ P (SK B0((u > u I X(l). and considering only the remaining ones.XK_1).+XK > uIXl. compute Y = u .2.. K > 2] = P2p(Xl > x) = P2Bo(x) (here we used that by positivity of the X. As a conditional Monte Carlo estimator .X(2). y < 0).L(x)/x`' with a > 0 and L(x) slowly varying. . A first obvious idea is to use conditional Monte Carlo: write i. Thus.....X(2).Xl .. ... .. However. For the simulation.p)Bo(x). we generate only X1.. Xl > x. < X(K) throw away the largest one X(K). This calculation shows that the reason that this algorithm does not work well is that the probability of one single Xi to become large is too big.SK1)2.. and that Bo(y) = 1. SIMULATION METHODOLOGY when the claim size distribution B (and hence Bo) has a regularly varying tail.S( K_1)) V X(K1)) / Bo(X(K 1)) where S(K_l) = X(1) + X(2) + • • • + X(K_1).p/(l . .. asymptotically it presents no improvement : the variance is of the same order of magnitude F(x). conditional probability.. Then (cf. just note that EZ(1)(u ) 2 > E[Bo (x . we thus generate K and X1i . 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.b(u) = P (Xl +•••+XK>u) = EF[Xl + .. XK. assume in the following that Bo(x) .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.. form the order statistics X(1) < X(2) < .. Theorem IX.....X1 . XK1. note first that To check the formula for the P(X(n) > x I X(1).. Z(1) (u) has a smaller variance than Zl (x).X(K1)) ... X1 + + XK_ 1 > x when X1 > x. Z(1)(u) is defined as 0).
in the renewal or Markov. the continuoustime process {St} is simulated by considering it at the discrete epochs {Qk} corresponding to claim arrivals. Then the algorithm given by { Z (2) (u) } is logarithmically efficient. Notes and references The proof of Theorem 2.. Compute y > 0 as solution of the Lundberg equation 0 = K(y) = )3(B[y] . 3 Importance sampling via Lundberg conjugation We consider again the compound Poisson model and assume the conditions of Ce7". B.3. . that is.u is the representation 0(u) = e7sr(u) overshoot (cf.Khinchine formula and importance sampling . X . l)) BO(X(n1)) Theorem 2 . X (. X(n1)) Bo((x . BL instead of 0.y. The algorithm is sofar the only efficient one which has been developed for the heavytailed case. .. Asmussen . X(2).S(n_1) I X(1). and simulate from FL._1) > P(X(n) > _ X X(1). l)) ..5). for the purpose of recording Z(u) = erysr(u).. X(n1)) P(X(TZ) + S(.. 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) . . Binswanger and HOjgaard of [28] give a general survey of rare events simulation for heavy tailed distributions . IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 287 We then get P(S" > x I X( 1). using 13L. X(2). and that paper contains one more logarithmically efficient algorithm for the compound Poisson model using the Pollaczeck. Thus.. BL by I3L = /3B[y].S (n1)) V X (. For practical purposes. use the the CramerLundberg approximation so that z(u) = '(u) = e7"ELe7E(") where ^(u) = ST(") . and we refer to [27]. However .1 is elementary but lengty. . . 1 Assume that Bo (x) = L(x)/x° with L(x) slowly varying. BL(dx) = e7sB(dx)/B[y]. Also in other respects the findings of [28] are quite negative: the large deviations ideas which are the main approach to rare events simulation in the lighttailed case do not seem to work for heavy tails. the algorithm for generating Z = Z(u) is: 1. X(2). and define )3L. 111. .modulated model P(r+ < oo) and G+ are not explicit ).
r(u) < oo) = 1. b different from . It resolves the infinite horizon problem since FL(. Let FL (dx) = 'For the renewal model.i. one must restrict attention to the case 4µB > 1. Let S .288 2. The answer is no. the discussion at the end of Section 1b. In fact: Theorem 3. M(u) = inf {n : S„ > u}. b) # (/3L.7 tell that P(. SIMULATION METHODOLOGY 3. If S > u. namely ELery£("). r(u) < oo) and FL (both measures restricted to. The estimator is then M(u) /3eQT' dB Z(u) (Ui) j=1 )3 e $Ti dB where M(u) is the number of claims leading to ruin. Let X1.. so that changing the measure to FL is close to the optimal scheme for importance sampling .F. let S.T. with distribution F. the results of IV.QL. and avoid simulating the known part e7". 4. Generate T as being exponential with parameter . u It is tempting to ask whether choosing importance sampling parameters .. B) is not logarithmically efficient when (/3. to deal with the infinite horizon problem . cf. return to 3.1) (simulated with parameters ^3. We may expect a small variance since we have used our knowledge of the form of 0(u) to isolate what is really unknown. + X.. The proof is given below as a corollary to Theorem 3. Proof Just note that EZ(u)2 < e . and the change of measure F r FL corresponds to B > BL.Q. A > AL as in Chapter V. and assume that µF < 0 and that F[y] = 1. Let Sf0 CHAPTER X. = X1 + .. P'[y] < oo for some ry > 0.1 The estimator Z(u) = e'rs* "u) (simulated from FL) has bounded relative error.. . Xi = U. BL could improve the variance of the estimator . X2. Ti. There are various intuitive reasons that this should be a good algorithm.S+U . be i.l3 and U from B. We formulate this in a slightly more general random walk setting '. Otherwise.2 The estimator (3. . BL). More precisely.2ryu _ z (u)2/C2.. The algorithm generalizes easily to the renewal model .3.(u)) are asymptotically coincide on {r(u)} < oo. and we have: Theorem 3. In detail .. let Z F e_'s.d.
+ KM(u))} = exp {ELM(u)(E . Jensen's inequality and Wald's identity yield EpZ(u)2 > exp {EL(K1 + . (3.P = FL. Proof The first statement is proved exactly as Theorem 3 ..2 > 0. IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 289 e7yF(dx). .2) dF Theorem 3.2) (simulated with distribution F of the X3 has bounded relative error when . is not logarithmically efficient. where e' = EL Iog dFL (Xi) > 0 by the information inequality. let F be an importance sampling distribution equivalent to F and M(u) dF Z(u) _ I (Xi) . EFZ(u)2 EFZ(u)2 lim sup z(u)2eeU = lim cop C2e2.2'X1 .3. it thus follows that for 0 < e < e'/ELXi. For the second. When F # FL.... By the chain rule for RadonNikodym derivatives. Here ELK.3 The estimator (3. The importance sampling estimator is then Z( u) = e'rSM( ). F(XM(u)). 1. where Kl og (X) (j) 2 ) = log dFL (Xi) ..+KM(u)}. = c'.. Since ELM(u)/u + 1/ELXi.i. .d. are i. More generally..yu = G.2ryELXi.2ryELXi)} .. EFZ(u)2 = EeW2(FIF) = Ep [W2(FIFL)W2(FLIF)] = EL [W2 ( FIFL)w(FLIF)] = ELexp {Kl+.. write W(F IF) _ F(XI). Since K1.yu+elu u +oo etry' 1 > lim up C2e2. K2.
The optimality result Theorem 3. 4 Importance sampling for the finite horizon case The problem is to produce efficient simulation estimators for '0 (u. The queueing literature on related algorithms is extensive . Consider compound Poisson risk process with intensities /3'.T". generic interarrival times T' . T) with T < oo.2. The easy case is y > 1/k'(y) where O(u. then the estimator Z(u) = e7Sr(°)I(r(u) < yu) (simulated with parameters /3L.3'eO'x f f P (U" . The results of IV. see e. from 3' P(U'T'>x) ^ = ^ eQ'zB (z) dz. In fact: Proposition 4.'(y) or y > 1/r. SIMULATION METHODOLOGY u Proof of Theorem 3.4. U' . Then according to Theorem 3.290 which completes the proof. we write T = yu.i. The extension to the Markovian environment model is straightforward and was suggested in Asmussen [ 16]. Next. all that needs to be shown is that if U' . with the present (shorter and more elementary) proof taken from Asmussen & Rubinstein [45].T".T" > x) J /3"e0 yB (x + y) dy = . U' . B" and generic claim sizes U'. In [13]. we conclude by differentiation that Bo(x)=B' (x)forallx > 0.T' = U" .g. 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 . First by the memoryless distribution of the exponential distribution . then /3' B' = B". As in IV.1 If y > 1/ic'('y). the references in Asmussen & Rubinstein [45] and Heidelberger [190].4 indicate that we can expect a major difference according to whether y < 1/r.T" has a left exponential tail with rate /3'. claim size distributions B'. Further discussion is in Lehtonen & Nyrhinen [245]. /3'eQ'YR'( x + y) dy = . T".3. BL) has bounded relative error. so that one would expect the change of measure F 4 FL to produce close to optimal results.3"eQ x 0 J eQ zB (z) dz x (x > 0) and /3' = /3". /3". U".T' has a left exponential tail with rate /3' and U" . yu) is close to zk(u). . This immediately yields / = 3".T' D U" .B'=B".e.'(y). optimality is discussed in a heavy traffic limit y 10 rather than when u + oo.1 is from Lehtonen & Nyrhinen [244]. CHAPTER X.
2) Since the definition of ay is equivalent to Eay r(u) . the result follows as in the proof of Theorem 3.8 has a stronger conclusion than (4.1). Further .to g x ( u ) u u so that (1. IMPORTANCE SAMPLING FOR THE FINITE HORIZON CASE 291 Proof The assumption y > 1/n'(y) ensures that 1fi(u.8). one would expect that the change of measure F Pay is in some sense optimal. Let Qy2 = .yu. We next consider the case y < 1/r. Bounding u ELZ(u)2 above by a7u.yk(ay) determines the order of magnitude of z'(u. Remark 4 .1. The corresponding estimator is Z(u) = eavS' ( u)+T(u)K (ay)I(T( u) < yu).(u) * 1 (Theorem IV. that ryy = ay .log Var(Z(u)) l im of . 7y (4. and that ryy > ry.(ay). T( u) < yu] e2ryyuEay le 2ay^(u).4. yu) is of order of magnitude a71. (4. Proof Since ryy > y. yu)/z. yu) = eayu Eay Leay^(u)+r(u)K(ay). we have ic(ay ) > 0 and get Eay Z(u)2 = Eay [e  2aySr( u)+2r(u )r.yy> 2 .log 4')u) 4 u (Theorem IV. T(u) < yu] e Hence by (4.log Var(Z(u)) _ .5) follows. yu) in the sense that .4. T(u) < yu] . Bay) is logarithmically efficient.'(7).3) (simulated with parameters /gay. (4.3) and we have: Theorem 4.O(u. lim inf uoo 27yu .1) so that z(u) = zP(u. .4.2 The estimator (4. 3 Theorem IV. and in fact.4. We recall that ay is defined as the solution of a'(a) = 1/y.1) which is all that is needed here can be showed much easier .1).
1) where the identity for Vi(u) requires that Vt has a limit in distribution V.a yu +l/ur' (av)Ei`av reav^(u)+(T(++)(U) yu .4). However.o . SIMULATION METHODOLOGY Vara„ (r(u))/u so that (T(u) .1) may be useful. there exists a dual process { V t} such that i.7y x(u) > hm inf u+Oo U .o. (5.a vt(u).Qyu1/2 < T(u) C yu e.1) is used to study Voo by simulating {Rt} (for example.u1/2 < r(u) < yu Le ] l = e. and (5. yu .uaoo U That lim sup < follows similarly but easier as when estimating En.1): then by Proposition A1. the algorithm in Section 3 produces simulation estimates for the tail P(W > u) of the GI/G/1 waiting time W).4.ryyu +oy u1/2K'(av)Eo l v 1/2) where the last step follows by Stam's lemma (Proposition IV. Hence lira inf log ryyu + vyu 1/2 tc(ay) . Z (u)2 above.b(u. In Asmussen [13]. yu .3.2).(av)Eav l e.2) . 0 Notes and references The results of the present section are new.1) (see Proposition IV. In most of the simulation literature (say in queueing applications). we believe that there are examples also in risk theory where (5.. One main example is {Vt} being regenerative (see A.T) = P O<t<T inf Rt < 0 = P(VT > u). 5 Regenerative simulation Our starting point is the duality representations in 11. related discussion is given in a heavy traffic limit q J. zi(u) = INV. '%(u) = P I info Rt < 0) = P(VV > u). > u) = E f I(VV > u) dt 0 (5.yu)/(uyu1/2) . 0 rather than when u 3 oo.4.292 CHAPTER X. N(0.3: for many risk processes {Rt }. Then z(u) = Eay Z(u) > Eay avS'(u)+T( u)k(av 1 ).yu1/2 <1 T(u) < yu l r > e7vu +avul/ 2r. the object of interest is {Vt} rather than {Rt}.
The method of regenerative simulation. To derive confidence intervals . Therefore . Z2 .3) . is the cycle length. EZ2'i = z2 = E Thus.h (zl. > u) (and more general expectations Eg(V.5. i (^(u) . Z(N) are i . Z2'> the time during the cycle where { Vt} exceeds u and zj = EZJ').d.. and Z2'>) where Zi'i = w.t(u)) 4 N(0. Vh = (8h/8z1 8h/ 8z2). i. + Z2N)) . z2) z2/z1 yields Vh = (z2/z2 1/zl). j = 1. provides estimates for F ( V. which we survey below . letting J0 'o I (Vt > u) dt . . Then (Z1z1i Z2z2 ) 4 N2(0. Thus. )). the regenerative estimator z%(u) is consistent.. consider first the case of independent cycles . Z2 = N (X21' + . oh) for h : R2 ^ R and Ch = VhEVh. (u) ?2 = E fo I(Vt > u) dt = 0( u ) zl Ew as N > oo. .. record Zi'i = (Z1'). REGENERATIVE SIMULATION 293 where w is the generic cycle for {Vt}.E).. For details .. Then Z(1). EZ1'i = z1 = Ew. + Z1N>) .. Simulate a zerodelayed version of {V t } until a large number N of cycles have been completed. Thus the method provides one answer on to how to avoid to simulate { Rt} for an infinitely long time period. Taking h(zl.... Zl the LLN yields Z1 a$' Z(1) +.+Z(N) z 1. a standard transformation technique (sometimes called the delta method) yields 1 V 2 (h (Zi. let E denote the 2 x 2 covariance matrix of Z('). For the ith cycle.. 2. Z2 a4* z2.. Z1 = (Zl1i +. z2)) > N(O. 02) (5.
z^ i=1 so a2 can be estimated by 2 2 = 72 S11+ 12 S22 .C)dx = f w(x) d( f ( x. v. asymptotic estimates were derived using the renewal equation for z /i(u). However . () dx = E[SZ] f(X. the expectation z = EZ of a single r. Notes and references The literature on regenerative simulation is extensive. Z of the form Z = ^p(X) where X is a r . say risk processes with a complicated structure of the point process of claim arrivals and heavy tailed claims . Here are the ideas of the two main appfoaches in today 's simulation literature: The score function ( SF) method . with distribution depending on a parameter (.5) Z1 Z1 Z1 and the 95% confidence interval is z1 (u) ± 1. see e. There is potential also for combining with some variance reduction method.96s/v"N. SIMULATION METHODOLOGY 2 Eli = Z2 z1 + 2 E22 .2. in some situations it may be the only one resolving the infinite horizon problem .g. to evaluate the sensitivity z/i( (u ) = (d/d() 0(u) where ( is some parameter governing the risk process . () depending on C. Then z(() = f cp(x) f (x.v.2 E1 2 z1 z1 Z2 The natural estimator for E is the empirical covariance matrix N S = N 1 12 (ZW . 6 Sensitivity analysis We return to the problem of 111 .0 . Rubinstein [310] and Rubinstein & Melamed [311].Z) ^Z(=) . We here consider simulation algorithms which have the potential of applying to substantially more complex situations.g S12 (5. The regenerative method is not likely to be efficient for large u but rather a brute force one. In 111.9. Let X have a density f (x. () dx so that differentiation yields zS d( fco(x)f(x. consider an extremely simple example . () dx f Ax) (dl d()f (x' () f ( z.294 where 01 2 CHAPTER X. 9. Before going into the complications of ruin probabilities .
1). this phenomenon is particularly unpleasant since indicators occur widely in the CMC estimators . ()) d( hc(U. /3o is M(u) Oe (3T: < oo) . C).v. ()) h((U.t. /3 is 0. however .r. SZ is an unbiased Monte Carlo estimator of z(.t. () = log U/(2. the Poisson rate /3 in the compound Poisson model. ()) is 0 w .6. IPA will estimate zS by 0 which is obviously not correct. ( where h( (u. () _ (eSx. p. for some Co = (o(U). () Thus. say W(x) = I(x > xo) and assume that h(U. To see this. cp' (h(U. For IPA there are. () = (8/8()h (u. For the SF method. nonpathological examples where sample path derivatives fail to produce estimators with the correct expectation. () is increasing in C. Then . giving h( (U. ()) is 0 for C < Co and 1 for C > Co so that the sample path derivative cp'(h(U.() d( is the score function familiar from statistics . Thus. ()). So assume that a r. if f (x. C) f(X. cp(h(U. For example . A related difficulty occurs in situations involving the Poisson number Nt of claims: also here the sample path derivative w. with density 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. 11 /3oeOoT. one can take h (U. C)). () is an unbiased Monte Carlo estimator of zS. just take cp as an indicator function . Example 6 . this is usually unproblematic and involves some application of dominated convergence . Thus . Let M(u) be the number of claims up to the time r(u) of ruin (thus. Infinitesimal perturbation analysis (IPA) uses sample path derivatives. Then z(() = Ecp(h(U. () can be generated as h(U.1 Consider the sensitivity tka(u) w. () = d log f (X.r. The following example demonstrates how the SF method handles this situation. In the setting of ruin probabilities . r(u) = Tl + • • • +TM(u)). = E [`d (h(U. () = . zc = E [d co(h(U. () where U is uniform(0. SENSITIVITY ANALYSIS where 295 S = (d/d()f (X. one. I(r(u) .log U/(. The likelihood ratio up to r(u) for two Poisson processes with rates /3.
r.T(u)) I(T(u) < co) ] .r.t. To resolve the infinite horizon problem . SIMULATION METHODOLOGY Taking expectation. In the setting of ruin probabilities. We recall (Proposition 111. 4) that V5. ) we have VarL(ZQ(u)) ZO(u)2 O(u2)e2 u2e2ryu yu . since ELZp(u)2 < (M(U) _T(u) \ 1 2 a2ryu = O(u2)e27u. in part for different measures of risk than ruin probabilities. There have been much work on resolving the difficulties associated with IPA pointed out above.t. . the risk process should be simulated with parameters . We then arrive at the estimator ZZ(u) = (M(u) .9 . Thus. j3 and letting flo = 0. change the measure to FL as when simulating tp(u). Example 6.1 is from Asmussen & Rubinstein [46] who also work out a number of similar sensitivity estimators. different parameters.296 CHAPTER X.3 (u) (to generate Zp (u). for different models and for the sensitivities w. BL). However.3L.0(1) so that in fact the estimator Zf(u) has bounded relative error.T(u)) e7uerVu) for ?P.3 (u) is of the order of magnitude ue7u. a relevant reference is VazquezAbad [374]. differentiating w. the estimation of z(ip(u) is subject to the same problem concerning relative precision as in rare events simulation . we get 1 M(u) 00(u) = E (_Ti)I(T(U)<) E [(M(u) . whereas for the SF method we refer to Rubinstein & Shapiro [312]. 0 Notes and references A survey of IPA and references is given by Glasserman [161] (see also Suri [358] for a tutorial).
+• • •+X. Oa(U ) can also be a useful vehicle for computing t/i(u) by letting a * oo.. are i. T(u. and {1. 'Note that in the definition of r(u ) differs from the rest of the book where we use r(u) = inf {t > 0 : Rt < 0} ( two sharp inequalities ). X2. a) = r(u) A T+(a). defined as Ro = u (with u E {0.. in the Bernoulli random walk example below. wherel T(u) = inf {t > 0 : Rt < 0} .Chapter XI Miscellaneous topics 1 The ruin problem for Bernoulli random walk and Brownian motion.. }). Consider first a Bernoulli random walk..i. as e.. R„ = u+X.g. in most cases . with P(Xk = 1) = 9. either this makes no difference (P(R. Y'a(U) = P(T (u) = r+(a)) = 1 . 297 .(u) is defined as the probability of being ruined (starting from u) before the reserve reaches level a > u. where X1. a) = r(u)). That is. Besides its intrinsic interest .P(•r(u... The twobarrier ruin problem The twobarrier ruin probability 0. . T+(a) = inf It > 0 : Rt > al.1..(u) = 0 ) = 0) or it is trivial to translate from one setup to the other...1}valued .d.
Conditioning upon X1 yields immediately the recursion 'a(1) = 19+00a(2).a) Y.e.y] = 1. 7/la(a .1) = (19)4/'0(a3)+9ba(a1). = (1 . = z°Va(u) + za(1  . and in view of the discrete nature of a Bernoulli random walk we write z = e7.1. By optional stopping.4) by ea(u+Xl+.+Xn) F[ a]n n=0....1) is solution.(u) I\ e = 1 oa ' ()i a = u.o» = z°P (RT ( u.298 CHAPTER XI. u + 1.a(u)). one elementary but difficult to generalize to other models. i.1) o If 0 = 1/ 2. (1. tba(2) _ (1 .1 For a Bernoulli random walk with 0 0 1/2.. MISCELLANEOUS TOPICS Proposition 1.(1B)u oJ 0..(4.. In a general random walk setting . zu = EzRO = EzRT(u. We choose a = ry where ry is the Lundberg exponent. C1_0\a.r(u. and insertion shows that ( 1.. then 'Oa(u) _ au a We give two proofs . u Proof 2. where a is any number such that Ee°X = F[a] <oo... Wald's exponential martingale is defined as in 11.o)T/la (1) + 8z/'u(3). The martingale is then {zuzXl+•••+X„ } = {zR° }. The Lundberg equation becomes 1=F[ry]=(19)+9z.2) Oa(a . z and the solution is z = (1 . Proof 1.o)'t/1a(a .2). and the other more advanced but applicable also in some other settings. the solution of F[.a) = 0) + zap ( R.0)/0.
TWO BARRIERS 299 and solving for 4/la(u) yields t/ia(u) = (za . pa( u) _ u Corollary 1. then Vi(u) = 1.5) .1) for p # 0. } yields e7u = Ee7R° = e°Wa(u) + e7a(1 . BROWNIAN MOTION. (1.u) = et(a2 /2 +aµ) the Lundberg equation is rye/2'yp = 0 with solution y = 2p..4 For a Brownian motion with drift u > 0. Applying optional stopping to the exponential martingale {e7R.ba(u) = e2µa .1). i1(u) = e211 .} 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)). Corollary 1.1 yields 't/la(u) = (a .2) is trivial (z = 1). then Proof Since 'Oa (U)  au a Eea(R°. u Proposition 1. If 9 = 1/2. RANDOM WALK. and solving for 9/la(u) yields Z/)a(u) = (e 76 . (1. thenz1 (u)=1.2 For a Bernoulli random walk with 9 > 1/2.1 If p = 0.u)/u. If p<0.e7u)/(e7° . 1h (u) = a el u \1 If 9 < 1/ 2.1).1. Then for p 0 0.zu)/(za . . However.• a2µa e2µu .3 Let {Rt} be Brownian motion starting from u and with drift p and unit variance . If p = 0. {Rt} is itself a martingale and just the same calculation as in the u proof of Proposition 1.0a(u)). Proof Let a+ oo in (1. {R.
a) = a on {r (u.a) = r+ (a)} and similarly for the boundary 0.5a). For most standard risk processes . Ic 5ry 'pa(u) Using y = 6 .6 If the paths of {Rt} are upwards skipfree and 7//(a) < 1. (u) _ O(u) . Here is one more case where this is feasible: Example 1.0(a) 0 < u < a. and hence e7u = Ee7Ro E [e7R(.7) .a ) < 0) + e 7aF ( R (u.4).5 Consider the compound Poisson model with exponential claims (with rate. valid if p < 1 (otherwise .a) = a) = 5 y = P (R (u.616).e7a (u) = 6 /0 . However. MISCELLANEOUS TOPICS u The reason that the calculations work out so smoothly for Bernoulli random walks and Brownian motion is the skipfree nature of the paths.a) < 0) + e7°P (R(u. VIII. (1.3.7).0 (u) (where u p =. 5). however. passing to even more general cases the method quickly becomes unfeasible (see.e7a Again . CHAPTER XI. Here the undershoot under 0 is exponential with rate 5. say. implying R(u. this immediately yields (1. .. letting a * oo yields the standard expression pe7u for .a) = a ) + e ' ° ( 1 .7/la(u)).300 Proof Let a * oo in (1. 0. the paths are upwards skipfree but not downwards. a) I R(u a ) < 0] P (R(u .+^a(u))^(a) If 7k(a) < 1. 1 . and thus one encounters the problem of controlling the undershoot under level 0. we obtain 'Oa a7u . 7/'(u) = 1). 7O(u) = 7/la(u) + (1 .vi(a) Proof By the upwards skipfree property. It may then be easier to first compute the onebarrier ruin probability O(u): Proposition 1.
8 ). and hence Pµ('r(u) E dT) = Eo [e µsr(.. and (1 . + µ2T) } . P(MT > u) = P(ST > u) + P(ST < u. T ) = P(T(u) < T ) = 241.µ%T (1. 10) follows then by straightforward differentiation.9) = 2P(ST > u). (1. in particular symmetric so that from time r(u) (where the level is level u) it is equally likely to go to levels < u and levels > u in time T . Here {St } is Brownian motion with drift 0 (starting from 0).f.1a for computing ruin probabilities for a twostep premium function.2 . Hence P(MT>u. of r(u) are ( U2 Pµ (T(u ) E dT) = 2^T 3/2 exp µu .4) I = .8) Proof In terms of the claim surplus process { St} = {u . Then the density and c.r(u).Rt}.ST<u) = P(MT>u. We now return to Bernoulli random walk and Brownian motion to consider finite horizon ruin probabilities.10) Pµ (T(u) < T) !.11 ) is the same as (1.T) P(MT > u) where MT = maxo<t<T St.8 Let {Rt} be Brownian motion with drift . = 1 . ( 1..ST>U). For µ # 0.. Corollary 1. For the symmetric (drift 0) case these are easily computable by means of the reflection principle: Proposition 1. the density dPµ / dP0 of St is eµsttµ2/2..µ so that {St} is Brownian motion with drift µ . T(u) E dT. MT > U) = P(ST > u) + P(ST > u) (1. 0(u.. (i).11) VIT ) Proof For p = 0. (1.µ T I + e2µ"4) ( . MT > u) = P (ST > u) + P (ST > u. = eµuTµ2/2Po (T( u) E dT) 2 eµuTµ2/2 u T3/2 ex p u 27r p 12 T . BROWNIAN MOTION. we have ili(u. TWO BARRIERS 301 Note thas this argument has already been used in VII. RANDOM WALK.7 For Brownian motion with drift 0.1.)_ _( u)µ2 /2.d.
(1. We assume that u(x) and a2 (x) are continuous with a2 (x) > 0 for x > 0.s. is zero for all u > 0 but that nevertheless Rt ^4 0 (the problem leads into the complicated area of boundary classification of diffusions. MISCELLANEOUS TOPICS which is the same as (1.T) = P(ST = u) + 2P (ST > u).9 For Bernoulli random walk with 9 = 1/2. Thus. Proof The argument leading to ( 1. oo) with drift µ(x) and variance a2 (x) at x. Vi(u.g... Here {2T( (v}TT)/2) v=T. as defined in (1. is finite for all x > 0.13) with 0 as lower limit of integration. Breiman [78] or Karlin & Taylor [222] p. We finally consider a general diffusion {Rt} on [0. The same argument as used for Corollary 1. T are integervalued and nonnegative. and (1.10) and that the value at 0 is 0. that 0(u). such that the drift µ(x) and the variance a2(x) are continuous functions of x and that a2(x) > 0 . The expression for F ( ST = v) is just a standard formula for the u binomial distribution. and in a similar spirit as in VII. see e.h. e.10 Consider a diffusion process {Rt} on [0.11) then follows by checking that the derivative of the r. the behaviour at the boundary 0 is more complicated and it may happen.T+2. is (1. 226). S(oo) = f c s(y)dy.3 we can define the local adjustment coefficient y(x) as the one 2µ(x)/a2(x) for the locally approximating Brownian motion.12) is the same as ( 1.. whenever u. S(x) = f x s(y)dy..T2. 0 0 (1.g.302 CHAPTER XI. If this assumption fails.10).T)dx. as defined above as the probability of actually hitting 0.T (1.13) The following results gives a complete solution of the ruin problem for the diffusion subject to the assumption that S(x). Theorem 1. u Small modifications also apply to Bernoulli random walks: Proposition 1.12) P(ST = v) = 0 otherwise. close to x {Rt} behaves as Brownian motion with drift µ = u(x) and variance a2 = a2(x).9) goes through unchanged.8 also applies to the case 9 54 1/2.9). but we omit the details. oo). Let s(y) = ef0 ry(.
Letting a T oo and considering the cases S(oo) = oo. 15) i.ba. If b < u < a.S(u) (1.16) yields 4b (u) = 1 . then 0 < 2l. elementary calculus shows that we can rewrite L as Lq(u) d 1a2 (u)s(u)d [ s (u) ? ] .S(b) Proof Recall that under mild conditions on q.(u) < 1 for all u > 0 and ^ S^ Conversely. (1. b = 0.S(u)/S(a).b('u) = Eu .b(a) = 0 then yield the result.b(Rdt). we can ignore the possibility of ruin or hitting the upper barrier a before dt. A good introduction to diffusions is in Karlin & Taylor [222].b = 0 implies that VQ b/s is constant. the function S(x) is . 0 in (1.16) S(a) . A classical reference for further aspects of Bernoulli random walks is Feller [142]. RANDOM WALK. S(oo) < oo separately u completes the proof.13) is finite for all x > 0.b = a+/3S. Then YIa.14) fails.b(u) be the probability that {Rt} hits b before a starting from u. Letting b J. [117].6 to Markovmodulated models . (1 . TWO BARRIERS 303 for x > 0.10.16). Further references on twobarrier ruin problems include Dickson & Gray [116]. so that Y)n. E„ q(Rdt) = q(u)+Lq(u)dt.14) S(oo) < 00.b(u) + L.e LVa. 191195 for material related to Theorem 1. Assume further that S (x) as defined in (1. 1'. .b('u) = Eu &0. Using s'/ s = 2p/a2.b(b) = 1.10. BROWNIAN MOTION.1.17) Hence L. In view of (1.0(u) = 1 for all u > 0. For generalizations of Proposition 1. The obvious boundary conditions '0a. where Lq(u) = 0'22u) q "(u) + p(u)q(u) is the differential operator associated with the diffusion. i. Notes and references All material of the present section is standard.b(u) = S(a) .e. see in particular pp. and we get Wo.11 Let 0 < b < u < a and let t&0. Wa. if (1.ba. O. then. 0 Proof of Theorem 1.b (Rdt) = Oa. see Asmussen & Perry [42].b(u)dt. Lemma 1. If (1.
Lo I.4. Lo is a martingale (cf. 7y = ay . Markovmodulated Brownian models .. (2. y > . Remark 11.aRo .2) C_e7u < t(u) < C+e _7u.5. but by duality. (2. Another basic quantity is the speed measure M .(7) . 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. yu) where W (ay) = y.5) A martingale proof of (2. (2. They all use the fact that ( tx(a) l ( eaRt = eau + aSttx(a) < e7yu. (2. The emphasis is often on stationary distributions .t&(u. 1 y < k (y). and here are alternative martingale proofs of the rest . equivalently. 111 .(. where C_ = B(x) _ B(x) sup 2no fy° e7(Y )B(dy)' f2e7(Y2)B(dy)' C+ i/i(u.(a) (2.4. variance 0. defined by the density 1/va(u)s(u) showing up in (1.5): _ z/'(u) < e 7u.1.3) < e 7yu. See Asmussen [20] and Rogers [305] for some recent treatments and references to the vast literature.)AT . with the drift and the variance depending on an underlying Markov process .304 CHAPTER XI.17). much of the literature dels with the pure drift case.1 ) was given already in II.3.1) (2.ytc (ay).2.9 ) and optional stopping applied to the stopping time r(u) A T. one works instead with a lower limit 5 > 0 of integration in (1. correponding to piecewise linear paths or . yielding eau = Ee. information on ruin probabilities can be obtained .aR. IV.o•K(a) = Ee . which is motivated from the study of modern ATM (asynchronous transfer mode ) technology in telecommunications.(T(u)AT) r.13)). yu) '+/1(u) . MISCELLANEOUS TOPICS referred to as the natural scale in the general theory of diffusions (in case of integrability problems at 0.6) .4) I.6. is currently an extremely active area of research.
we have tc(ay) > 0 and we can bound (2. yu))• Letting T + oo yield e_ayu > eyur4ay)(0(u)  Notes and references See II.yu) Y Similarly for (2. we have ic(ay ) < 0 and use the lower bound E [e7Rr („). and the proof of the lower inequality is similar. Proof of ( 2. when Rt_ = r. (2.1.2.7R. dr) e 7( yr)B(dy) B(r) f oo o 0 r > H(dt. Rt has distribution B(r + dy)/B(r). Let H(dt.6) below by 1 E Le7Rr(.yu))• b(u.yuk (ay)(u&(u. dr) denote the conditional distribution of (T(u).3). u Proof of (2.1 .f. it follows easily from (2. Equivalently.E [e.. eyuk (ay) = e7yu e > eyu"(ay ) ij(u.6). FURTHER APPLICATIONS OF MARTINGALES 305 (we cannot use the stopping time r(u) directly because P(r(u) = oo) > 0 and also because the conditions of the optional stopping time theorem present a problem).B(r))/B(r).T)  V.(u. dr JO Zoo ) f e7'B(r + dy) B(r) Jo ^00 ^00 H(dt.4). A claim leading to ruin at time t has c.2): As noted in Proposition II.3). RT(u)_) given r(u) < oo.(ay)I T(u) < yu] P(r(u) < yu) (using RT(u) < 0).( u ) I T(U) < 00] .4): We take a = ay in (2. For (2.1.T(u)K(ay) I yu < r(u) < T] F(yu < r(u) < T) > e. dr) 1 = 1 I0 /o C+ C+ From this the upper inequality follows. y > r.6) with = 'y that eyu . Hence E [e7Rr (u) Jr(u) < ool ^00 H( dt.)r(u)r. .d. (B(y) . so that i/1(uL yu) < eayu .
nroo n n /// Note in particular that (3. For example.3na with a < 1. which in the setting of (3..1 We will go into some more detail concerning (3. in being capable of treating many models beyond simple random walks which are not easily treated by other models . we will write fn 1.306 CHAPTER XI.1) is an example of sharp asymptotics : . logarithmic asymptotics is usually much easier to derive than sharp asymptotics but also less informative .nn or C2e. ri./n E I) for intervals I C R. og For sequences fn. and gave sharp asymptotics for probabilities of the form P (S. gn with fn + 0 . .1) amounts to the weaker statement lim 1 log P I Sn > x I = 17. then P C S. gn 4 0. logarithmic asymptotics .1) does not capture the \ in (3.the correct sharp asymptotics might as well have +.^ e nn 1 > x n 0o 2xn (3. Example 3. the parameter will be u rather than n).1) where we return to the values of 0.1). Thus. not quite so much in insurance risk.. large deviations results been. e. v2 later.g. if x > EX1. 1) but only the dominant exponential term . however . The last decades have seen a boom in the area and a considerable body of applications in queueing theory. Cramer considered a random walk Sn = X1 + . such that the cumulant generating function r. (3. However .2). large deviations results have usually a weaker form. (3.gn if nioo lim 109 fn = 1 log gn (later in this section. and that a considerable body of theory has been developed. Thus .?n typically only give the dominant term in an asymptotic expression . its generality. The classical result in the area is Cramer's theorem.means (as at other places in the book) that the ratio is one in the limit (here n * oo). The limit result (3. Accordingly. cle .. + X. The advantage of the large deviations approach is. ..(B) = log EeOX 1 is defined for sufficiently many 0. 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.2) can be rewritten as F (Sn/n > x) 1g a'fin.
i. replacing Sn in the exponent and ignoring the indicator yields the Chernoff bound P Sn > x 1 < e°n (3. 2 where o2 = o2(x) = rc"(0).(0)) e 307 (other names are the entropy. Define . the sup in the definition of rc* can be evaluated by differentiation: rc*(x) = Ox . we have P(nx < Sn < nx + 1. of P(X1 E dx) = E[e9X1K. nx < Sn < nx + 1.r. if we replace Sn by nx + o / V where V is N(0. exponential change of measure is a key tool in large deviations methods. LARGE DEVIATIONS Define rc* as the convex conjugate of rc.q = rc* (x).4 enn +1.sseo f which in conjunction with (3.3.2). More precisely.t. since Sn is asymptotically normal w.rc(0) where 0 = 0(x) is the solution of x = rc'(0). (3. S rtn > x 1. the LegendreFenchel transform or just the Legendre transform or the large deviations rate function).9S„+n' ( 9). rc*(x) = sup(Ox . Since P nn > x) = E {e_8 ' ( 9).425.960/) * 0.96o /] > 0.(e)i XI E dx].r. we get P(Sn/n > x) E [e9nx +nK(9)9" '. P with mean nx and variance no. and hence for large n P(Sn/n > x) > E [e.4) n Next. which is a saddlepoint equation .1). In fact.4) immediately yields (3. Most often. V > 0 e.3) is put equal to x.the mean rc'(0) of the distribution of X1 exponentially tilted with 0.tin f o') o e9o^y 1 1 ey2/2 dy 21r = etin 1 Bo 27rn .e.
. Sanov's theorem which give rare events asymptotics for empirical distributions. 260 for details. Xn given by Fn(dxl.v. asymptotics for probabilities of the form P ({S[nti/n}o<t<l E r) for a suitable set r of functions on [0. which is of similar spirit as the dicussion in VII..e < 8 < y + e. We shall need: Lemma 3 .p > 7 < zn. (ii) lim supn.3 For each i > 0.. however. In the application of large deviations to ruin probabilities.308 CHAPTER XI. Ee9X n < oo for e < 0 < e. . commonly denoted as is the saddlepoint approximation.s.. Assume that there exists 'y... Mogulskii's theorem which gives path asymptotics. see Jensen u [215] or [APQ] p.h. The substitution by V needs.e < 8 < y + e.o log Ee9Sn /n. We further write µ = tc'(ry). to be made rigorous.dxn) = 05nKn(7)Fn(dx1.. and write Sn = X1 + • • • + Xn.. and the WentzellFreidlin theory of slow Markov walks.3. . 1) and no such that Sn .2 (GLYNN & WHITT [163]) Let X1. For the proof. Pn Sn1 . that is.dxn) where Fn is the distribution of (X1i . there exists z E (0. be a sequence of r.'(u) )Ng a"u. we introduce a change of measure for X1.. (iii) #c (8) = limn./^ >7 < zn n for n n0. (iv) tc(ry) = 0 and r.. integrates to 1 by the definition of Icn).'s. Xn) and sn = x1 + • • • + xn (note that the r. n Icn(0) exists and is finite for ry . which is a version of Cramer's theorem where independence is weakened to the existence of c(O) = limn. Further main results in large deviations theory are the GartnerEllis theorem.. X2.... . Then i/. r(u) = inf {n : Sn > u} and o(u) = P('r(u) < oo). is differentiable at ry with 0 < K'(y) < 00.... 1].1). e > 0 such that (i) Kn (0) = log Ee°Sn is welldefined and finite for 'y . we shall concentrate on a result which give asymptotics under conditions similar to the GartnerEllis theorem: Theorem 3 . .. MISCELLANEOUS TOPICS which is the same as (3.
can be chosen strictly negative by taking 9 small enough.ne(p+ 17).91) + o(O ) as 0 J. LARGE DEVIATIONS Proof Let 0 < 9 < e where a is as in Theorem 3..Bµ . This establishes the first claim of the lemma . ( U) P(S..Kn(7)e'n (p(O +7))/p I Ee geXn]1/q where we used Holder's inequality with 1/p+ 1/q = 1 and p chosen so close to 1 and 0 so close to 0 that j p(0 +.+r7) < zn for n > no.077 n^oo n and by Taylor expansion and (iv ).2..m(7). it is easy to see that the r. log zl'(u)/u > 'y. > 1 +17] m(7). We first show that lim inf„_. h. is of order .2. can be chosen strictly negative by taking p close enough to 1 and 0 close enough to 0. Then V.n e(µ +o)w"(7) [Eep(B +7)Sn]1 /p [EegoX.YS. Let r7 > 0 be given and let m = m(77) = [u(1 + 77)/µ] + 1. for Sn.. This proves the existence of z < 1 and no such that Pn (Sn/n > µ. P n(Sn/n > {c+77) < e no(µ 309 +n)Enees n +n)elcn(B +7).ne(µ limsup 1 log Pn (Sn/n > µ + 17) < ic(9 + ry) . mµ Sm > u] km e7Sm+n.s.h. 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. u Proof of Theorem 3.W. 0. Since I EeqOX „ ] 1/q is bounded for large n by (ii).. The rest of the argument is as before.]1/q = e. h.s.µ?7 . S.n m µ 1 + rl . The corresponding claim for Pn(Sn/n < µ .s. the r .71 < e and jq9j < e.3.y) .> .+r.n > u ) = [ Em [em Em 1e. 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.77) follows by symmetry (note that the argument did not use µ > 0). in particular the r. S. Clearly.r (7) n = e.
P(T(u) = n) < P(Sn > u) = En [e7S.I < µl1 1+77 I M 1_ 1+277 S.n Yµ 1 + m + r ('Y) } U n \ 77 m µ µ7 1 < 1+ 77 ) Here E. Obviously.7) so that n(b) I1 < e'Yu E en..(•) goes to 1 by Lemma 3.6) for some z < 1 and all n > n(E). we get lum inf z/i(u) 1 +12r7 >_ ry + 77 Letting r7 J...n(ry)/u 4 0andm/u* (1 + r7)/µ.(Y). Sn > u] < eYu+Kn(7)pn(Sn > u) (3. logO(u)/u > ry. 14 = = E Lu(16)/aJ+1 Lu(1+6)/µJ+l = n) and n(S) is chosen such that icn('y )/n < 6 A (. and since Ic.YS +^c CHAPTER XI.3.. 0 yields liminfu __. I > IL exp `S.310 ]Em I e. For lim supu.. this is possible by (iii).. 3. MISCELLANEOUS TOPICS (7). we write P(T(u) = n) = Il + I2 + I3 + I4 'i/I(u) _ E00 n=1 where n(b) Lu(10/µJ Ii = 1: F(T(u) = n). I2 = F(T(u) = n). (iv) and Lemma 3.+wn(7).0 log i'(u )/u < 'y.log z) /2 and Sn Fn\ n >lb+S) <Zn. n=1 .. n=1 n=n(b)+1 00 Lu(1 +6) /µJ 13 F( T (u) = P(T(u) n). Pn \ > la+ 8 I < zn (3.
10) 00 I4 < E F(Sn_1 < u.3. > u) Lu(1+6) /µJ +l 00 )^n 'YSn+kn (7) . u . LARGE DEVIATIONS Lu(16)/µJ 311 I2 < e"u n=n(6)+1 e'n(Y)P(Sn > u) < Lu(16)/µJ ^.' 1 + b) n e7u x 1 /2 1 n x n / 2x (3. Sn1 C U. eryu en logz/2p n nt n. C 26u `p / +1 I e6u(1+6)/µ (3. S. µ n=n(6)+1 \ 1u(16)/µ1 00 1 zn < e7u E Z n/2 < e(U xn/2 E n=n(6)+1 n=0 eYu = 1 .11) [u(1+6)/µJ+1 1  Thus an upper bound for z/'(u) is n(6) e'Yu n=1 eKn (7) + 2 + (28U + 1) e6u(1+6)/µ Fi 1 zl /2 and using (i).zl/z en6 [u(1 +6)/µJ 1u (1 +6) /µJ ekn(7) < e' 13 < C" E Yu l u(16)/lij+1 Lu(16)/µJ+l1 < e7U Finally. we get lim sup log u/00 O (U) < y + b(1 + b) U Letbl0. 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 .
Letting c11 = maxn<n.u(1+b)/rc'(7)). cf.7' a"ju. For 12. Then for n large. IV.(u) = I1+I2+I3+I4'^ ery( u).9) can then be sharpened to x LQuJ /2 I2 < e7u 1 .312 CHAPTER XI. u(1 + b)/i(7)) Proof Since V. we need to redefine n(b) as L. > u) < e"' E eIsn = ectueKn (a+'Y)Kn(7) where 0 < a < e and a is so small that r. we replace the bound P(Sn > u ) < 1 used in (3.3ui where . ryue«iu . this is straightforward since the last inequality in (3. MISCELLANEOUS TOPICS The following corollary shows that given that ruin occurs. 4 there is an aj > 0 and a cj < oo such that Ij < c3e. it suffices to show that for j = 1. For 14. u . the last steps of (3.4. we have rcn (a + 7) < 2n^c(7 + a) < 4narc' (7).xl/2 to give the desired conclusion.4/3rc'(y) > 0.Q is so small that w = 1 ..('+'Y). 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.z 1/z For I1.. say n n1. e'. I2.b)/i(7). For I. it holds for each b > 0 that 0(u) 1' g F(T(u) E (u(1 .8) by P(S. the typical time is u/rc'(7) just as for the compound Poisson model.2. Corollary 3.4 Under the assumptions of Theorem 3.11 ) can be sharpened to x 4 [u(1+6)/µJ /2 1 . (7 + a) < 2arc'(7). 2. 13 = P(T (u) E (u(1 b)l^ (7).
1.1.5 Assume the Xn form a stationary Gaussian sequence with mean p < 0.. Xk+l) k=1 00 naoo n provided the sum converges absolutely.(O) = 9µ+02 for all 9 E R..14) is needed in both examples ...f.3(s) at time s. r. 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.e. whether P ( sup St > u ltg a ^" 0<t<oo // (3.. Obviously many of the most interesting examples have a continuous time scale. Thus the total reward in the interval [0. Hence z z\ 2 z nrn(9) _ n Cn0p+BZn/ * .LARGE DEVIATIONS 313 Example 3 . 11 Inspection of the proof of Theorem 3. i. To verify these in concrete examples may well present considerable difficulties. for the ruin probability z/'h(u) of any discrete skeleton {Skh}k=0.12) k=0. t] is Rt = E V (Un) n: o„ <t . 2 is in force with y = 2p/wz. Assuming that the further regularity conditions can be verified.v. Let {Nt}t>0 be a possibly inhomogeneous Poisson process with arrival rate . Theorem 3. the key condition similar to (iii).2 shows that the discrete time structure is used in an essential way. An event occuring at time s is rewarded by a r. The problem is whether this is also the correct logarithmic asymptotics for the (larger) ruin probability O(u) of the whole process.. and in fact. criteria are given in Duffield & O'Connell [124]. we shall give two continuous time examples and tacitly assume that this can be done. If {St}t> 0 is the claims surplus process.g. V(s) with m. The reader not satisfied by this gap in the argument can easily construct a discrete time version of the models! The following formula (3. and we conclude that Theorem 3 . (iv) becomes existence of a limit tc(9) of tct(9) _ log Ee8S° It and a y > 0 with a(y) = 0.2 then immediately yields the estimate log F( sup Skh > u) a7u (3.3.'(y) > 0..13) One would expect this to hold in considerable generality. but nevertheless.. 09(9).
Q„) . leading to St = At(1+77) Joo t S8 ds. MISCELLANEOUS TOPICS are the event times.s). 0 Example 3 . Thus. n: o.6 We assume that claims arrive according to a homogeneous Poisson process with intensity 0 . nondecreasing and with finite limits Un as s T oo ( thus. More precisely. Kt (0) t (Ee9U"it8i J0 . one would take p(t) = (1 + rt)At/ t.g. Un represents the total payment for the nth claim). <t which is a shotnoise process.1) . Thus. = U„ ( t . 0 and since EeOUn(8) + Ee°U^ as s * oo.0 and assume there are y. (3. It is interesting and intuitively reasonable to note that the adjustment coefficient ry for the shot . . e. We further assume that the processes {U1(s)}8>0 are i.Lundberg model has the larger ruin probability. derive .t. we have S. where Ft = a(A8 : 0 < s < t).. if the nth claim arrives at time a. the best estimator of /3µB based upon Ft. Of course .15) . Most obviously. O'n +S] is a r . a differential equation in t). it contributes to St by the amount Un(t .1) ds rt (3..1) ds . i. the CramerLundberg model implicitly assumes that the Poisson intensity /3 and the claim size distribution B (or at least its mean µB) are known.v. At = .1) ds . Since the remaining conditions of Theorem 3.14). the Cramer. we have rct (9)/t 4 ic (9)..314 where the an CHAPTER XI. An apparent solution to this problem is to calculate the premium rate p = p(t) at time t based upon claims statistics .14) (to see this . e > 0 such that ic('y) = 0 and that r. We let ic (9) = 3(EeWU° .noise model is the same as the one for the Cramer Lundberg model where a claim is immediately settled by the amount Un.'`1 U. then the payments from the company in [on.d. Example 3.. Then logEeOR° = J0 /3(s)(^8(9) .9t = /3 J t (Ee8U° i8l . Thus by (3. If the nth claim arrives at time Qn = s. 7 Given the safety loading 77. Un(s).2 are trivial to verify. (9) < oo for 9 < 'y + C. assuming a continuous premium inflow at unit rate.9t. the above discussion of discrete skeletons).It. Of course. we conclude that Cu) log e7 u (cf. but that a claim is not settled immediately. this is not realistic . is At .
typically the adaptive premium rule leads to a ruin probability which is asymptotically smaller than for the CramerLundberg model .log Oi are i. To see this .17) K(a) f o 1 O (a[I + (1 + 77) log u]) du )3. equivalently. the Vi = . It then follows from (3.20) (3.(1 + 17)0µB = 0.i. (3. which yields eau f 1 t(1+n )audtl = E r Ee°Y = E [O(1+n)aueaul = E [eau J L Jo J L1+(l+r))aUJ .(1 +i) f > i= 1 s ds = E Ui 1 . we conclude that t. one has y > y' (3. 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.d. Thus.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 . Ui Nt / t 01i 315 St = Ui . uniform (0.b(u) IN a'Yu (cf. and since the remaining conditions are trivial to verify.e.d.19) with equality if and only if U is degenerate. rewrite first rc as te(a) _ /3E 1 1 +(1+77)aUJ eau 1 . standard exponential . (3.i.3./3. we have Nt t N.1) or .(1 + r7) log t (3.2 hold. Indeed. i.14) that rt _ 13 Jo _ (a [1_( i+77)log]) ds_flt = t (a) (3.16) i=1 o i=1 Let ict (a) = log Eeast .1) . LARGE DEVIATIONS With the Qi the arrival times.18) Thus (iii) of Theorem 3. the solution of /3(Eelu .
In addition to Glynn & Whitt [163]. For notational simplicity.19).(1 + ri)y*x is convex with k(oo) = 00. using that Ek(U) = 0 because of (3. For Example 3.d. 11 Notes and references Some standard textbooks on large deviations are Bucklew [81].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. a* (s) are convex with tc'(0) < 0 . x > x0. see Nyrhinen [275] and Asmussen [25].2.1 .316 CHAPTER XI. MartinL6f [256]. with common distribution B and independent of Nt. [245]. 0 < x < x0. MISCELLANEOUS TOPICS Next.2 expressing the finite horizon ruin probabilities in terms of the distribution of A.20) is due to Tatyana Turova. Dembo & Zeitouni [105] and Shwartz & Weiss [339]. k(0) = 0. k'(0) < 0. [257] and Nyrhinen [275]. and k(x) < 0. 4 The distribution of the aggregate claims We study the distribution of the aggregate claims A = ^N' U.xo. assuming that the U. we are interested in estimating P(A > x) for large x. are i. . so there exists a unique zero xo = xo(r7) > 0 such that k(x) > 0. This is a topic of practical importance in the insurance business for assessing the probability of a great loss in a period of length t. rc*' (0 ) < 0. at time t.7. Further. the study is motivated from the formulas in IV. say one year. we then take t = 1 so that p. y = y* can only occur if U . the proof of (3. Further. The main example is Nt being Poisson with rate fit. this in turn yields y > y*. see also Nyrhinen [275] for Theorem 3. In particular. = P(N = n) = e(3an However. the function k(x) = e7*x . though we do not always spell this out. Lehtonen & Nyrhinen [244]. much of the analysis carries over to more general cases. This implies n(y*) < 0. and since tc(s). Therefore e7'U _ k(U) E [1+(1+77)y*U] . Further applications of large deviations idea in risk theory occur in Djehiche [122]..i.
1 Assume that lim8T8. no(a) = logE9e'A = rc(a + 9) . e9x+K(°) P(A > x) B 2ir /3 B" [9] Proof Since EBA = x. A > x) eex+K(e ) ee AB°[ely 1 ev2/2 dy 0 2^ 00 9x+p(e) e ezez2/(2BZpB „[9)) dz 9 27r/3B" [9] fo eex+w ( e) oo z x)] ] 0 27r /3B" [9] o e 9 2 /3B" [9] J eex+w(B) dz .ic(9) = . THE DISTRIBUTION OF THE AGGREGATE CLAIMS 317 4a The saddlepoint approximation We impose the Poisson assumption (4. The exponential family generated by A is given by Pe(A E dx) = E [eeA K(9). A > x)] = eex+K( e)E9 [e .3e(bo[a] . The analysis largely follows Example 3.1). only with 0 replaced by a9 and B by B9. A E dx] . Then as x * oo.[s])3/2 = 0. For a given x. Then Ee"A = e'(") where x(a) _ 0(B[a] . 818' where s' = sup{s : B[s] < oo}."(0) = .x)//3B"[9] is standard normal.3B"[9]. Hence P(A > x) = E e [e9A+ ic(9). In particular. This shows that the Pedistribution of A has a similar compound Poisson form as the Fdistribution. we define the saddlepoint 9 = 9(x) by EBA = x. B"[s] = oo.9(Ax). Proposition 4.1.1). B"' [s] lim (B".3B[9] and Be is the distribution given by eox B9(dx) = B [9] B(dx).1) where )30 = .e. i. (4. Vare(A) = s.2) implies that the limiting Pedistribution of (A . K'(0) _ ic'(9) = x.4.
4) . In particular. b is gammalike. some regularity of the density b(x) of B is required.2 If B is subexponential and EzN < oo for some z > 1. 4b The NP approximation In many cases . For details.2i and that (A . where q(x) is bounded away from 0 and oo and h (x) is convex on an interval of the form [xo. 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 . just the same dominated convergence argument as in the proof of Theorem 2.(3µB)/(0µB^))1/2 has a limiting standard normal distribution as Q ^ oo. 2).Q{AB (4.l3pB.318 CHAPTER XI. For example. or.ycix °ie6x B.3) The result to be surveyed below improve upon this and related approximations by taking into account second order terms from the Edgeworth expansion.2) is often referred to as the Esscher approximation. Furthermore 00 b(x)Sdx < oo for some ( E (1. either of the following is sufficient: A. The present proof is somewhat heuristical in the CLT steps. Var(A) _ ^3p. B covers distributions with finite support or with a density not too far from ax° with a > 1.1). 1 . Y satisfies 9(u) ti eu2/2(1 + ibu3) (4. The (first order) Edgeworth expansion states that if the characteristic function g(u) = Ee"`}' of a r.x') where x' = sup {x : b(x) > 0}. then P(A > x) . b(x) = q(x)eh(z). For example. Thus . large x. Jensen [215] and references therein. In fact. and (4. bounded with b(x) . under the Poisson assumption (4.(D X . [138]. MISCELLANEOUS TOPICS It should be noted that the heavytailed asymptotics is much more straightforward. Remark 4 . see Embrechts et al.v.1 goes all the way back to Esscher [141]. it holds that EA = . it is quite questionable to use (4. A covers the exponential distribution and phasetype distributions. leading to P(A > x) :.e. the distribution of A is approximately normal .3) and related results u for the case of main interest .EN B(x). Notes and references Proposition 4. b is logconcave.1 yields: Proposition 4. more generally. For a rigorous proof. i.
Thus if EY = 0.EY)3. so that 1(u) 3 exp { . are small.equantile in the distribution of Y.s. of (4. (4.. K4 . . Let Y = (A . Heuristically. the NP (normal power) approximation deals with the quantile al_E. the CLT for Y = Y6 is usually derived via expanding the ch. THE DISTRIBUTION OF THE AGGREGATE CLAIMS where b is a small parameter. s.. then P(Y < y) 4(y) .. .6) . .h. In concrete examples .2 ^ \1 . If the distribution of Y is close to N(0. the density of Y is 1 °° _ eiuy f(u) du 2x _.EA)/ Var(A) and let yl_E. Rather than with the tail probabilities F(A > x). u5. one needs to show that 163. (4.e.2K3 + 4i 64 + . A particular case is a.. as u2 u3 u4 9(u) = Ee'uY = exp {iuci . zl_e be the 1 . however. If this holds .3!). K3 = E(Y .. K2 = Var (Y). are the cumulants ...i 3 K3 } Pt^ exp .f.2X2 . and from this (4. in particular. which is often denoted VaR (the Value at Risk). where Kl .: EA + zl_E Var(A) .5) follows by integration. resp.99.i 6 r 1 3 so that we should take b = ic3/6 in (4.1).. Remark 4. Var(Y) = 1 as above .2 2 .6(1 .5) may be negative and is not necessarily an increasing function of y for jyj large.5)..5) is obtained by noting that by Fourier inversion. f °o 9(y) = 1 e'uye u2/2(1 + iSu3) du 27r _ cc(y) . defined as the the solution of P(A < yle) = 1 .4.y2)^P(y)• 319 Note as a further warning that the r. and so as a first approximation we obtain a1_E = EA + yle Var(A) .3& (y). ylE should be close to zl_E (cf.c2i.l = EY. the standard normal distribution.5 (y3 . one expects the u3 term to dominate the terms of order u4.
3ni /3 . K5 . 4c Panjer 's recursion Consider A = constants a. n = 1.1)EY3.zlE )w(zl _E) = which combined with S = EY3/6 leads to q^ 1 Y1 ...zi. the kth cumulant of A is /3PBk' and so s.E(/3PB^1 )1^2 + s(z1E .E . b = /3 for the Poisson distribution with rate /3 since Pn = Pn1 n! n (n . k3 is small for large /3 but dominates 1c4..zl E)^o(zl E) .E)A1 l E) 1 E 4)(yl E) ^' .5(1 .S(1 ..5) by noting that the 4.k = /3µB^1 / (. For example..zl E)V(zl_E) .EA)3 a1_E = EA + z1_E(Var (A))1/2 + 1 Var(A) Under the Poisson assumption (4. as required . this yields the NP approximation 6(Z1 _E . [101].y2)cp( y) term.S(1 . b such that EN 1 U%.zlE)W(zlE) 1 . Another main reference is Daykin et at. and assume that there exist n ) Pn_i . let pn Pn = (a+ = P(N = n).EA ) / Var(A).1)^ 2) µ'E Notes and references We have followed largely Sundt [354].7) as 1 (3) a1E = Qµa +z1 .. that [101] distinguishes between the NP and Edgeworth approximations.E + (yl. however.. .1)! n ^eQ . 21 .6 (1 . This leads to t( yl E) . MISCELLANEOUS TOPICS A correction term may be computed from (4. In particular .6pBki) d/2.E = z1E + S(zi_E .320 CHAPTER XI. Note. Using Y = (A .1) E (A . this holds with a = 0.1).E )Azl E) 4(z1E) + ( ylE .yi. We can rewrite (4.(y) terms dominate the S(1 .
12) we get for j > 0 that fj n a b + n p nlgj *n 00 U I n 1 *n = E a+bUi=j pn19j n=1 j i=1 CC) n Ui EE n=1 Ia +b Ul i=1 =j pn_1 . .10) f o = po. j = 1. . fj = P(A = j). which would consist in noting that (in the case go = 0) fj = pn9jn n=1 (4. n.. .4 Assume that B is concentrated on {0..13) but only O(j2) for Proposition 4.. . By symmetry.. j = 1. (4.. E[a +bU=I >Ui =j l i=1 J (4. .. The expression for fo is obvious. j = 0. Hence by (4.4 is that the algorithm is much faster than the naive method. THE DISTRIBUTION OF THE AGGREGATE CLAIMS 321 Proposition 4. (4. the complexity (number of arithmetic operations required) is O(j3) for (4. .12). then j (a + b!) 1ag k_1 3 gkfj. if go = 0.. 2. u Proof of Proposition 4. Since the sum over i is na + b.. Then fo = >20 9onpn and fi = 1 E In particular. (4. fj = E (a+ b k =1 )9kfi_k . .4. n = k=n1 9k(n1 )9j k • (4.. 1.4.14) is independent of i = 1. and calculating the gj*n recursively by 9*1 = 9j.13) Namely.. (4.11) Remark 4. . the value of (4. 2.. j1 g..9).5 The crux of Proposition 4.} and write gj = 2 . 2.12) where g*n is the nth convolution power of g.4.14) is therefore a + b/n.k . .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).
this implies 7* > 7.4). Eexp 7 [E [Ui .h(Ui)] . (6.3). X2 = U .P } < 1 = Eexp E[Ui.P I = EC [7]N.h(u) and any P satisfying (6.b)+ is referred to as excessofloss reinsurance and plays a particular role: Proposition 6. we get N 1 = Eexp ry E[Ui ii .P]}. This follows by taking Xl = U A b. Then for any local retention function u .P > EexP{7[X . . (6.P. u But since ry > 0. that 01[ry] < 0[y] where 0[y] = Ee'r(U^') . Assuming for simplicity that the Ui are i. Remark 6. expectations like those in (6. however. REINSURANCE Proof Define N 329 h* (x) = E > h(Ui) X = x .b)+ with b determined by E(U .4) and u g(x) = e7x in Ohlin's lemma.4.6).h(Ui)] . Applying the inequality Ecp(Y ) > EW(E (YIX )) (with W convex ) to W(y ) = eryy.h( UU) = EN • E[U . The arrangement used in practice is.h(U)].d. we get EX = EN • EU. as often local as global.h(U) (as in the proof of Proposition 6.i.4).6.6 Assume the Ui are i.h(Ui)] .h(Ui)P JJJ l:='l {ry ] or.4). the excess ofloss rule hl (u) = (u .5) holds trivially.6) u where C[ry] = Ee'r(u4(u)). then (6.5 Because of the independence assumptions . y = Ei [Ui .5) reduce quite a lot.d.b)+ = Eh(U) (and the same P) satisfies 71 > ry..h * (X) . and so on. ry* > 0 because of (6. Proof As in the proof of Proposition 6. N E X . Local reinsurance with h(u) = (u . ' ii (6. it suffices to show that Eexp {ry ii 'UiAb. i. appealing to (6.
The present proof is from van Dawen [99].g. The original reference for Ohlin's lemma is Ohlin [277].330 CHAPTER XI. see also Sundt [354]. . MISCELLANEOUS TOPICS Notes and references The theory exposed is standard and can be found in. See further Hesselager [194] and Dickson & Waters [120].many texts on insurance mathematics. [76]. Bowers et at. e. Heilman [191] and Sundt [354].
1) (here U(t) = U([0. 2h. The associated renewal measure U is defined by U = u F*" where F*" is the nth convolution power of F.. not concentrated on {h. The point process is called a renewal process if Yo. stating that U(t+a)U (t) ^ a. some condition is needed: that F is nonlattice.. of interarrival times and the time Yo = To of the first arrival (that is. note in particular that U({0}) = 1.. Lebesgue measure for some n > 1).r.. when t is large. all have the same distribution. t] is denoted by Nt.. . the distribution of Yo is called the delay distribution. The mathematical representation is either the ordered set 0 < To < T1 < . + U2 where U1 is a finite measure and U2(dt) = u(t)dt where 331 .. are independent and Y1.} for any h > 0.e. i. the renewal process is called zerodelayed. denoted by F in the following and referred to as the interarrival distribution. Technically.. t +a]).. t]) so that U(t + a) . .. If F satisfies the stronger condition of being spreadout (F*' is nonsingular w . Lebesgue measure dt normalized by the mean to of F.U(t) is the expected number of renewals in (t. The number max k : Tk_j < t of renewals in [0.. That is. Then Blackwell 's renewal theorem holds.. then Stone 's decomposition holds : U = U. If Yo = 0. t 00 (A. Y1. Y2. Y2. . The renewal theorem asserts that U(dt) is close to dt/µ. Y. U(A) is the expected number of renewals in A C R in a zerodelayed renewal process. . = T„ .T„_1). of epochs or the set Y1.Appendix Al Renewal theory la Renewal processes and the renewal theorem By a simple point process on the line we understand a random collection of time epochs without accumulation points and without multiple points.t.
z(x) = 0. µF (A.3) Further.2) Z(u) = J0 u z(x)U(dx).i". wee shall need the following less standard parallel to the key renewal theorem: Proposition A1. Then Z(u) 4 z(oo).2). see [APQ] Ch. and that F has a bounded density2.out.5) 2This condition can be weakened considerably .a. A weaker (and much easier to prove) statement than Blackwell's renewal theorem is the elementary renewal theorem. IV). the asymptotic behavior of Z(u) is given by the key renewal theorem: Proposition A1.4) If F is spread. then Z(u) i f0 z(x)dx .1 if F is nonlattice and z (u) is directly Riemann integrable (d. Note in particular that F is spreadout if F has a density f.9. oo). (A. u u PF 4 00.e. but suffices for the present purposes . z(u) a known function.4) that z is Lebesgue integrable with limZ. Equivalently. (A. that z(u) has a limit z(oo) (say) as u 4 oo.i.332 APPENDIX u(t) has limit 1/µ as t 4 oo. Both result are valid for delayed renewal processes. then it suffices for (A.R. IV). 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. (A. in convolution notation Z = z + F * Z. resp.2 Assume that Z solves the renewal equation (A. the statements being EN(t + a) . U Z(u . Under weak regularity conditions (see [APQJ Ch. stating that U(t)/t > 1/p. and F(dx) a known probability measure . i. In 111.EN(t) . (A.x)F(dx).
k+t }t>o is independent of To. The kth cycle is defined as {XTk+t}o<t<Yk . this covers discrete Markov chains where we can take the Tn as the instants with Xt = i for some arbitrary but fixed state i. 0 PF µF 11 In risk theory. • . . equivalently. To this end. .. The simplest case is when {Xt} has i. Tk and {Xt }o<t<Tk • For example. refer to the zerodelayed case. Here the relevant F does not have mass one (F is defective). Note. Y1 . Assuming that y can be chosen such that f °° Ox F(dx) = 1. however.x)u(x) dx = z(u( 1 . the postTk process {XT. the present more general definition is needed to deal with say Harris recurrent Markov chains. .. this expression is to be interpreted as a random element of the space of all Evalued sequences with finite lifelengths. 1c Regenerative processes Let {T. The distribution F of Y1.APPENDIX 333 Proof The condition on F implies that U(dx) has a bounded density u(x) with limit 1/µF as x * oo.2) by e7x to obtain Z = z +P * Z where Z(x) = e'Y'Z(x). Eo etc.. Y2... {Tn} if for any k. a basic reason that renewal theory is relevant is the renewal equation II. that F is a probability measure. . However.. and its distribution does not depend on k. where the Tn are the instants where a customer enters an empty system (then cycles = busy cycles).t))u(ut) dt 0 0 J f z(oo) • 1 dt = z(OO). .5a. .d. Hence by dominated convergence.e. or many queueing processes. multiply (A.t. The property of independent cycles is equivalent to the postTk process {XTk+t}t>0 being independent of To. This program has been carried out in III. i.(3. A regenerative process converges in distribution under very mild conditions: . T1. z(x) = e7xz(x). Tk (or. Z(u) U = 1 u 1 u f z(u . is called the cycle length distribution and as before. of Yo. T1.. cycles. that the existence of y may fail for heavytailed F. F(dx) = e7xF(dx). Yk ).. asymptotic properties can easily be obtained from the key renewal equation by an exponential transformation also when F(dx) does not integrate to one. We let FO.. A stochastic process {Xt}t>0 with a general state space E is called regenerative w.3) satisfied by the ruin probability for the compound Poisson model. we let µ denote its mean. results from the case fo F(dx) = 1 can then be used to study Z and thereby Z.} be a renewal process. However.i.r.
r. (b) If in addition Var(Ul ) < oo.+ X.6) id Cumulative processes Let {Tn} be a renewal process with i. An example is Zt = fo f (X8) ds where {Xt} is regenerative w. Otherwise .ZTOI < 00. 0<t<Yi then Zt /t a$• EU1/µ. but in fact.t. just the same proof as there carries over to show: Proposition A1.0 be cumulative w.r. We denote the limiting r.. and we have: holds more generally that (rl(t). then e (t) .ZT Then: (a) If E sup I ZTo+t .e. oo). Then {Zt}t^. {i7(t)} are Markov with state spaces (0..r. resp .3 Consider a regenerative process such that the cycle length distribution is nonlattice with p < oo. oo). {Tn}.r.0 is called cumulative w. are i.v.. i.3.4 Let {Zt}t^.t.Tk : t < Tk} as the age.t : t < Tk}.. If p = oo. for n = 1.i.. cycles (we allow a different distribution of the first cycle).e. e(t )) . Y1) le Residual and past lifetime Consider a renewal process and define e ( t) as the residual lifetime of the renewal interval straddling t.. and q(t) = sup It .oo (i.tEU1/µ)/f has a limiting normal distribution with mean 0 and variance Var(Ui) + (!)2Var (Yi)_ 2EU1 Cov(U1. µ 0 If F is spreadout.. C).d.. then (Zt .i.334 APPENDIX Proposition A1. 2. C(t) and ij (t) both have a limiting stationary distribution F0 given by the density F (x)/p.. [0. This is the case considered in [APQ] V..'s by e. fi (t) = inf {Tk . Then it (ii.ZT }0<t<Y„+. where the distribution of X. is given by Eg(Xoo) = 1 E0 f Ylg (Xt)dt.d. under the condition of Blackwell's renewal theorem. assume that p < 00 and define Un = ZT}1 .. Then {e(t)}. Then Xt Di X.. P(C ( t) < a) 4 0 for any a < oo) and ij (t) * oo. {Tn}.t. then Xt . (A. in total variation. {Tn} if the processes {ZT +t .
(1 V)W) where V. Proof The number Nt of renewal before t satisfies Nt/t a4' p.APPENDIX 335 Theorem A1. Y1 > t] 4 0. Since the maximum Mn of n i.y) = f U(t . In the general case. (b) the joint distribution of (ri. assume first the renewal process is zerodelayed.t. Yo > 0] + f Eo^ (t .t. we can bound e(t) by M(t) = max {Yk : k < 2t/p}. W are independent. Y1i Y2.dy )z(y) < c ^ l z(k) Eoe(t 0 0 k=o where c = sup. For the second.6 Consider a renewal process with µ < oo.'s with finite mean satisfies Mn/n a$• 0 (BorelCantelli).^(t))} as a regenerative process. . Hence t t lt ) = f U(dy)z(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).. use t E^(t)/t = E[Yo .i.U(x) < U( 1)). we used: Proposition A1. the sum is o(t) so that Eo£(t)/t + 0 . are not i.d. and the equivalence of (a) with (b)(d) is an easy exercise..5 Under the condition of Blackwell's renewal theorem. the joint distribution of (rl. U(x + 1) . ^) is given by the following four equivalent statements: (a) P (77 > x.4. Yl > t]. and the conditional distribution of ri given l. EC(t)/t + 0.y)P(Yo E dy) . l:) is the same as the distribution of (VW. if in addition EYo < oo.. (c) the marginal distribution of q is FO. 1) and W has distribution Fw given by dFw/dF(x) = x/pF. r. the first statement follows. Then Eo^(t) satisfies a renewal equation with z(t) _ E[Y1 . = z is Foz) The proof of (a) is straightforward by viewing {(r. ^ > y) = 1 f +Y (z)dz.(t). Then fi(t)/t a4' 0 and. In IV. (d) the marginal distribution of ^ is FO.d.i.v. Since z ( k) < E[Yi . but governed by a Markov chain {Jn} (we .U(x) (c < oo because it is easily seen that U(x + 1) . Hence for t large enough. V is uniform on (0. 0 If Markov renewal theory By a Markov renewal process we understand a point process where the interarrival times Yo .
7 Consider a nonlattice semiregenerative process.. < yIJ) = Fij( y) on {Jn= i. . These facts allow many definitions and results to be reduced to ordinary renewal. Jn = i is the same as the P. Yn.. .. = io for some arbitrary but fixed reference state io E E. A2 WienerHopf factorization Let F be a distribution which is not concentrated on (oo.. be i. distribution ofjXt}t>o itself where Pi refers to the case Jo = i. oo). oo)... IT. is given by Eg(X00) = 1 YO vjEj f g(Xt) dt µ jEE o where p = ujEEViAj. Jo. . Jn_1.. We call r+ (T_) the strict ascending (weak descending) ladder epoch and G+ (G_) the corresponding ladder height distributions.336 APPENDIX assume here that /the state space E is// finite) in the sense that P(Y.. Notes and references Renewal theory and regenerative processes are treated. . Y1. 0] or (0 . . .and regenerative processes.i .T_ < oo).. .+ < x. Jn +1=j} where J = a(JO. e..jEE is a family of distributions on (0. the Markov renewal process if for any n.r. T_=inf{n>0: Sn<0}. the conditional distribution of {XT„+t}t>o given Yo. Let X1. .. r+ < oo). Assume that uj = EjYo < oo for all j and that {J„} is irreducible with stationary distribution (v3)jEE. Sn = X1 + • • • + Xn the associated random walk.t. in [APQ].. G+(x) = P(S. X2.d.t. namely {Twk } where {Wk } is the sequence of instants w where Jo. G_(x) = P(ST_ < x. A Markov renewal process {Tn} contains an imbedded renewal process. A stochastic process {Xt}t>o is called semiregenerative w. where the distribution of X.g. J1 i .. Further: Proposition A1.}.} is nonlattice (it is easily seen that this definition does not depend on i). and define r+=inf{n>0: Sn>0}.. Alsmeyer [5] and Thorisson [372]. The semiregenerative process is then regenerative w. For example.) and (Fij )i. with common distribution F. .. Then Xt 4 Xo. .r. the semiregenerative process is called nonlattice if {T.
n=0 The basic identities are the following: Theorem A2. n=0 n=0 00 00 and the T+. F(A . .r. F(A) is the contribution from the event {T_ = 1} = {X1 < 0}. A C (0.. S. we may rewrite (a) as G_ (A) = G+(A) = F(A) + (G+ * G_)(A).S. oo).>0. In (A. m<j<n}. .=n w=m i Figure A.=EGn. A C (0.S. Proof Considering the restrictions of measures to (oc. Sr_ _1 is at its minimum .g. we consider the last such time (to make w unique) so that {w=m. define w as the time where the preT_ path S1. n 0 R_(A) = E I(Sn E A). G+. A C (oo. (A. >0.T_=n} = {S.. On {T_ > 2}.7).G+ * G_: (b) G_ (A) = f °° F(A . F(A) + (G+ * G_)(A).7) follows since G+(A) = 0 when A C (oo.x)R_ (dx).1 (a) F = G+ + G_ .and r_ preoccupation measures T+1 r_1 R+(A) = E E I(Sn E A). oo) (A. oo). A C (oo. More rigorously. 0<j<m. 0] and (0.. (d) R+ = U_. G_. (e) R_ = U+.x)R+(dx).APPENDIX 337 Probabilistic WienerHopf theory deals with the relation between F. 0). U.7) (A. (c) G+(A) = f °. 0]. the renewal measures U+=>G+.8) (e.1 . 0]). u .
XnEAx) 00 f 0 f 0 00 00 1: F(A .+ E du)P(S..1)._ E A . Sr_ E Adu) (s ee again Fig . Sn1 E dx) n=1  F(A . ST_ E A . (b) follows from 00 G+ (A) _ E F(Sn E A. 0 < k < n. . m it follows (see Fig. E du) = P(T_=nm.+ E du) E P(S.du) (G+ * G)(A)• C llecting terms. SnEAIS.Sn_1Edx. S.x)R+(dx)._ = n . S.3 8 APPENDIX Reversing the time points 0.u) f0m m=1 n=m+1 00 J0 OO P(S. A. . (A... A. and reversing the order of summation yields P(T_ > 2.>0. ST+Edu). . r+ = n) n=1 n=1 0  C0 E fF(Sk< 0..x)P(Sk < 0.0<k<ri . and the proof of (A.= n.1) that P(Sj Sn.F(r_n_mSrEA_u). ST_ E A) P(T+ = m. 0<j<m..3. clearly (Sj Sm>0.._ E A) n1 f P(r_=nw=m Sm EduSrEA) m=1 n1 F(r+=mSr+Edu).1. m < j <n...8) is similar. SmEdu) = P(T+=m. Aso. It follows that for n > 2 F (7.m.7) follows. m=1 f S mming over n = 2.
consider a fixed n and let Xk = Xn_k+l. . Again. such developments motivate the approach in Chapter VI on the Markovian environment model. P(SnEA . then T+ = inf It > 0 : St = 0} is 0 a. a number of related identities can be derived. In this generality of..0<k<n. G_ are trivial.SnEA) = P(Sn<Sk.SnEA) is the probability that n is a weak descending ladder point with Sn E A. Since G+ is concentrated on (0. H+ (s) = 1G+[s] is defined and bounded in the halfplane Is : ERs < 0} and nonzero in Is: Rs < 01 (because IIG+lI _< 1). which is basic for the PollaczeckKhinchine formula.'s.1. see for example Bingham [65].G_[s]) (A. being concentrated at 0. The classical analytical form of the WienerHopf problem is to write 1 . the analogue of a random walk is a process with stationary independent increments (a Levy process. For example.F[s] = (1 . and using timereversion as in (d) to obtain the explicit form of R+ (Lebesgue measure)..SnEA) = P(Sn<Sk.9) whenever F[s]. The present proof of Theorem A2.O<k<n. see e. 6+ [s]. u Notes and references In its above discrete time version. and the proof of (e) is similar.1). 0].APPENDIX 339 and the proof of (c) is similar. the survey [15] by the author and the extensive list of references there.g. Summing over n yields R+ (A) = U_ (A). oo). is based upon representing G+ as in (b).6.s. Then for A C (oo. this holds always on the line its = 0. we can rewrite (a) as 1 . if {St} is Brownian motion. Nevertheless. there are direct analogues of Theorem A2. In continuous time.g. WienerHopf theory is only used at a few places in this book. E.1(a) is from Kennedy [228].Sn_k.1. and similarly H_ (s) = 1 .0<k<n. For (d). G_ [s] are defined at the same time.g. Sk = X1 + • • • + Xk = Sn .T+> n) = P(Sk < O.f. Another main extension of the theory deals with Markov dependence.4). 11.0+[s])(1 .G_ [s] is defined and bounded in the halfplane is : ERs > 01 and nonzero in Is : ERs > 0}. In discrete time. there is no direct analogue of Theorem A2. and G+. cf. However. u Remark A2.O<k<n.P as a product H+H_ of functions with such properties.2 In terms of m. and sometimes in a larger strip. it serves as model and motivation for a number of results and arguments in continuous time. the derivation of the form of G+ for the compound Poisson model (Theorem 11.SnEA) = P(SnSn_ k.
and eQ can then be computed as the mth power (by squaring if = 2). write eQ = (eK)m where = Q/m for some suitable integer m (this is the scaling step).13) henever A is a diagonal matrix with all diagonal elements nonzero. Here are. _I 0 (A. one needs to compute matrix inverses Q1 and matrix exponentials eQt ( r just eQ ). ere A is the eigenvalue of largest absolute value. It is seen from Theorem VIII. JAI = max {Jjt : µ E sp(A)} and sp(A) is the set of all eigenvalues of A (the spectrum).10) d dteAt = AeAt = eAtA (A. Here it is standard to compute matrixinverses by GaussJordan el imination with full pivoting . however . To circumvent this. 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).11) A f eAtdt = eA.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. whereas there is no similar single established a proach in the case of matrix exponentials. three of the c rrently most widely used ones: xample A3. if m is s fficiently large. 1.12) eA'AO = Ale AA (A.5 that when handling phase type distributi ons.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. Thus. Some fundamental properties are the following: sp(eA) = {e' : A E sp(A)} (A. 0 . Eo Kn/n! converges rapidly and can be evaluated without p oblems.
APPENDIX 341 Example A3.. One then can reduce to p linear differential equations by noting that k = ZQ. condition upon the number n of Poisson events in [Olt])  Example A3. Zo = h). construction of {Xt} by realizing the jump times as a thinning of a Poisson process {Nt } with constant intensity 77.14) holds is therefore that the tstep transition matrix for {fft} is eQt = E ent (. Let vi.e. i. The probabilistic reason that (A.7t) n=0 n! u °O n Pn (to see this.. In practice. 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.. Zo = a (Z = QZ. .2 (UNIFORMIZATION) Formally.15) Then it is easily checked that P is a transition matrix . However . vp be the corresponding left .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}. To this end. i.]t)n (A.4 (DIAGONALIZATION) Assume that Q has diagonal form..3 i (A. the procedure consists in choosing some suitable i > 0. what is needed is quite often only Zt = TreQt (or eQth) with it (h) a given row (column) vector. some jumps are dummy in the sense that no state transition occurs ). . letting P = I + Q/i and truncating the series in the identity = e17t 00 Pn(. p different eigenvalues Aj i .. The approach is in particular convenient if one wants eQt for many different u values of t.3 (DIFFERENTIAL EQUATIONS) Letting Kt = eQt.e. assume that Q is the intensity matrix for {Xt} and choose q with rt > max J%J = max qii• 1. Ap. Here is a further method which appears quite appealing at a first sight: Example A3 . 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 . 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.
Nevertheless.16) (A. Example A3. and vihi ¢ 0. vi.... Complex calculus : Typically. Qhi = vihi. we can take H as the matrix with columns hl. There are.. Then P P Q = > Aihivi = E Aihi (9 vi.. In view of this phenomenon alone care should be taken when using diagonalization as a general tool for computing matrixexponentials. hi have been computed. the eigenvalue... (A. two serious drawbacks of this approach: u Numerical instability : If the A5 are too close.18) Namely. v5Q = Aivi. and we may adapt some normalization convention ensuring vihi = 1.18) contains terms which almost cancel and the loss of digits may be disasterous. Then vihj = 0. under the conditions of the PerronFrobenius theorem). (A.342 APPENDIX (row) eigenvectors and hl. and writing eQt as eQt = He°tH1 = H (e\it)di. hp.. however. and we need to have access to software permitting calculations with complex numbers or to perform the cumbersome translation into real and imaginary parts. i= 1 i=1 P P (A. and hence A2 is so because of A2 = tr(Q). D = ) 2 2 . say A = (Ai)diag. some cases remain where diagonalization may still be appealing.17) eQt = E e\`thivi = E ea:thi ® vi. not all ai are real. this last step is equivalent to finding a matrix H such that H1QH is a diagonal matrix. i # j. of largest real part is often real (say. hp the corresponding right (column) eigenvectors. The phenomenon occurs not least when the dimension p is large.g H1. i=1 i=1 Thus. we have an explicit formula for eQt once the A j. Everything is nice and explicit here: 411+q2+D' )12_g11+q2^^ where (411422z + 4412421. say Al.5 If Q= ( 411 ( q21 q12 q22 is 2 x 2.
6 A particular important case arises when Q = q1 qi ) q2 q2 J is an intensity matrix.20) ir = q2 ql qi +q 2 9l +q2 (A. 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. where (A.e.19) Example A3 . v2 and h2 can be computed in just the same way. However. Of course.7 Let 3 9 2 14 7 11 2 2 . k  C k2 ) =b ( A1 q 1 Q11 / where a . i.k1). it is easier to note that 7rh2 = 0 and v2k = 1 implies v2 = (k2 . replacing ai by A2. The other eigenvalue is A = A2 = q1 . l ab (g12g21 + (A1  411) 2) = 1.q.APPENDIX 343 Write 7r (= v1) for the left eigenvector corresponding to a1 and k (= hl) for the right eigenvector. Then Al = 0 and the corresponding left and right eigenvectors are the stationary probability distribution 7r and e. b are any constants ensuring//Irk = 1. u Example A3. 1) . h2 = Thus.21) Here the first term is the stationary limit and the second term thus describes the rate of convergence to stationarity. Then 7r = (ir1 7r2 ) = a (q21 Al .
A2 = 3/2 .344 Then D= 2+ 11)' 7 T4 2 =52.. They are most often constructed by imposing some additional properties .22) Note that in this generality it is not assumed that A is necessarily square.satisfying AAA = 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_.23) .6. (A+A)' = A+A. and a generalized inverse may not unique.5 . A+AA+ = A+.11/2 + 5 1. but only that dimensions match . (AA+)' = AA+.. 2 2 1=ab(142+(1+2)2 ) = tab. for example AA+A = A. Generalized inverses play an important role in statistics. (A. (A. APPENDIX x1 3/2 .11/2 . e_6u A4 Some linear algebra 4a Generalized inverses A generalized inverse of a matrix A is defined as any matrix A.
and define D = (A .23) is called the MoorePenrose inverse of A. 0 01 In applied probability.. ( Q . = 0 where m < p is the rank of A.25) . Am+1 = .1Q = Q(Q .I) (A. and can define /ail 0 0 0 0 0 0 A+ = C A' 0 0 0 C' .= (I .. Assume that a unique stationary distribution w exists . Rather than with generalized inverses .P + e7r)1 (here ( I .e.1 Let A be an irreducible intensity matrix with stationary row vector it.. one then works with Q = (Q .D + O(ebt).1 goes under the name fundamental matrix of the Markov chain). .eir ).g.P + e7r ). one is also faced with singular matrices . are ordered such that Al > 0.24) = te7r .g.APPENDIX 345 A matrix A+ satisfying (A. Am > 0.. if A is a possibly singular covariance matrix (nonnegative definite). E.. and exists and is unique (see for example Rao [300]). Here is a typical result on the role of such matrices in applied probability: Proposition A4. These matrices are not generalized inverses but act roughly as inverses except that 7r and e play a particular role . . (A.eir)1 = I .e ® 7r)1.. (I .P). _ A. then there exists an orthogonal matrix C such that A = CDC' where 0 0 D = AP Here we can assume that the A . Then for some b > 0. lt o eAx dx = te7r + D(eAt .eir )1. most often either an intensity matrix Q or a matrix of the form IP where P is a transition matrix.ew.
I) . Then A(O) _ B(O) = 0. 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. Note that h ® it has rank 1.26) follows by integration by parts: t f t /' xeAx dx = [x {xe7r + D(eAx .2 Let it be a row vector with m components and h a column vector with k components. . (A. the formulas involving O(e6t) follow by PerronFrobenius theory.I) (A.s.D + D2 + O(ebt). in block notation i2h A®B= ( a11B a21 B a12B a22 B Example A4. I.h. it follows that h ® it is the k x m matrix with ijth element hi7rj ..J {xe^r + D(e .27) Proof Let A(t).DZ(ent .2e7r .I)} dx. .h.91a(2) .s. o Finally. Interpreting 7r. 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.24). the rows are proportional to it. resp. of (A. see below. h as 1 x m and k x 1 matrices. Equivalently. B(t) denote the l.3 Let 2 A= 4 3 Vf' N7 5 )' B= ( 8 ). B'(t) = e7r + DAeAt = eir + (I . and the columns to h. (A. the r.346 t APPENDIX 2 xe Ax dx = eir + t(D + e7r) + D(eAt .I)}. respectively. ()®(6 f 6/ 7f 8^ 7 8 )=! ^)( 6 7 8 )=(6^ 7^ 8^) \ u Example A4. For example. h ® it reduces to hit in standard matrix notation. and in fact any rank 1 matrix can be written on this form.e.26) 2 = 2 e7r + tD .eir)eAt = eAt = A'(t).
31). it follows that e® ® e B An _ 0o oo oo Bn 7 I F n! = ` k! (I .3f 4v/.29) If A and B are both square (k1 = ml and k2 = m2). 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.5v/.50 6 7 6 4f 4.A9.APPENDIX 347 Then A®B = 2 f 20.3V8.3v'6.(A. then v1B1h1 and v2B2h2 are real numbers. (A. (A B)' = eA®B e! L 1=0 0 . and v1B1h1 • v2B2h2 = v1B1h1 ® v2B2h2 = ( v1(&v2 )( B1(&B2 )( h1(&h2 ) . and the number of such factors is precisely given by the relevant binomial coefficient. (AED B)1 = (A®I+I(9 B)l is the sum of all products of t factors.30) eA+B = eAeB function generalizes to Kronecker notation (note that in contrast typically only holds when A and B commute): Proposition A4.3vV/72f 20.5v'8 5vf9 11 A fundamental formula is (A1B1C1) ®(A2B2C2) = (A1 (9 A2)(B1 (9 B2)(C1®C2). if Al = vi. (A.4vf. each of which is A ® I or I ® B. such a factor is Ak (&B 1k according to (A.29). A2 = v2 are row vectors and C1 = h1.4 eA® B = eA ®eB. Using (A. then the Kronecker sum is defined by A(1) ®A(2) = A(1) ®Ik2 + k ®A(2).31) Indeed. if A ® I occurs k times. C2 = h2 are column vectors.28) In particular.k)! ( n0 n=0 t=0 k=0 J _ ® Ak ®Blk r ^.
s. represents ces Q( 1). n2 n1 ) {X(2) } are independent Markov chains with transition matrices P(1). in the definition (A. Q(2). { On the other hand. Yt(2 ) }.32).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). Let further it.32) is the intensity matrix of the bivariate continuous Markov process {Yt(1). first term on the r .348 APPENDIX Remark A4. resp . {Yt(1). P8 = exp {sQ} = exp {s (Q(1) ®Q(2)) } . where transition matrix of the bivariate Markov chain {X n1). (A. p = P(1) ® {X }.6 Suppose that A and of B. Thus . v whenever a is an eigenvalue of A and 0 is an eigenvalue be any row vectors and h. and Q = Q(1) ® Q (2) = Q(1) ® I + I ® Q(2) (A.I)(h ® k). k any column vectors. P(t) Yt(2) }. A special case of Proposition A4. 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 +. h. Let P8f P(Sl). From what has been said about matrices of {Yt( 1). Ps 1) = exp {sQ ( 1) } > p(2 ) = exp {sQ(2) } can therefore be rewritten as Taking s = 1 for simplicity . independent Markov chains. X ) }.4 can easily be obtained by probabilistic be the sstep transition reasoning along the same lines . the same time. the {Yt(2) } transitions in the {Yt(1) } component and the second transitions in the component . 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).3 < 0 Lemma A4 . Yt(2) where independent Markov processes with intensity matri{y(2) } are {Y(1) }. Then 2 0 ire At h • ve Bt kdt = (^®v)(A®B)1(e A®Ba .33) . { 1't(1) }.
. E (0. il. .7 Let A be a p x pmatrix with nonnegative elements. then IN < Ao for all A E sp(A). n.g.. which can be found in a great number of books. 4c The PerronFrobenius theorem Let A be a p x pmatrix with nonnegative elements. . the integrand can be written as ( 7r (9 v)( eAt ® eBt )(h ®k ) = ( 7r ®v)(eA (DBt)(h (& k). . p there should exist io.. Here is the PerronFrobenius theorem. h can be chosen with 3By this. (b) if in addition A is aperiodic.The PerronFrobenius theorem has an analogue for matrices B with properties similar to intensity matrices: Corollary A4. i. [APQ] X. f o r each i. and the corresponding left and right eigenvectors v. ao). and if we normalize v. . in such that io = i.1 and references there (to which we add Berman & Plemmons [63]): Theorem A4. h can be chosen with strictly positive elements. Then: (a) The spectral radius Ao = max{JAI : A E sp(A)} is itself a strictly positive and simple eigenvalue of A. Now note that the eigenvalues of A ® B are of the form a +. j = 1.3 whenever a is an eigenvalue of A and 3 is an eigenvalue of B. . > 0 for k = 1. .29). and appeal to (A. ...APPENDIX 349 Proof According to (A.34) Note that for a transition matrix. Then the eigenvalue Ao with largest real part is simple and real. and the corresponding left and right eigenvectors v.8 Let B be an irreducible3 p x pmatrix with nonnegative offdiagonal elements. = j and atk_li. then An = Aohv+O(µ") = Aoh®v+O(µ") for some u. so that by asssumption A ® B is u invertible. We call A irreducible if the pattern of zero and nonzero elements is the same as for an irreducible transition matrix. we mean that the pattern of nonzero offdiagonal elements is the same as for an irreducible intensity matrix..12). Similarly. . h = e and v = 7r (the stationary row vector). h such that vh = 1. see e. . (A. we have AO = 1.. A is called aperiodic if the pattern of zero and nonzero elements is the same as for an aperiodic transition matrix. That is.
.(ti)ding. To this end.2). Example A3. Then for any (3.8 is most often not stated explicitly in textbooks.35) for some p E (oo. one can consider A = 77I + B where rl > 0 is so large that all diagonal elements of A are strictly positive (then A is irreducible and aperiodic).1.350 APPENDIX strictly positive elements. we have A0 = 0. let t = (ti)iEE # 0 have nonnegative entries and define T(°) = aQ . Furthermore. Proposition A5. the condition is that t is small compared to Q. but is an easy consequence of the PerronFrobenius theorem. let {Yti°i } be a Markov process with initial distribution a and intensity . not only in the tail but in the whole distribution. if we normalize v. h such that vh = 1.(ti)diag where Q = T + (ti)diag is a proper intensity matrix (Qe = 0). 10) and use the formula me at e Bt = e 00 Antn = e . relate the eigenvalues of B to those of B via (A. Bi° (x) + at*x Proof Let { 4 } be the phase process associated with B(a) and (°) its lifelength. A5 Complements on phasetype distributions 5a Asymptotic exponentiality In Proposition VIII. h = e and v = 7r (the stationary row vector). I. T(°)) is asymptotically exponential with parameter t* _ r EiEE aiti as a 4 oo.n t AL n=0 n! (cf. Corollary A4. The content is that B is approximately exponential if the exit rates ti are small compared to the feedback intensities tij (i # j). Ao).(3. then eBt = ea0thv + O(eµt) = eA0th ® v + O(et t) (A. it was shown that under mild conditions the tail of a phasetype distribution B is asymptotical exponential. For example. the analogy of this procedure with unformization.1 Let Q be a proper irreducible intensity matrix with stationary distribution a. Note that for an intensity matrix. note that we can write the phase generator T as Q . The next result gives a condition for asymptotical exponentiality.e. the phasetype distribution B(a) with representation (.8.
2 Pi (c(a) > x. a' = a .(a) > x . We can think of ( ( a) as the first event in an inhomogeneous Poisson process ( Cox process ) with intensity process matrix aQ . it states that the state.g. Then a(a'V)/a (aV) a' 1. Proof Assume first ti > 0 for all i and let I. in fact . and this easily yields a(x)/x a' 1/t*. we get dx F (Idx = j) = (1 + qij t )Sij + qij dt.APPENDIX 351 ((1) etc.aE where 0 < e < 1). Since JJ(. = YQ(x).a' + oo (e. By the law of large numbers for Markov processes .jEE. has a limit distribution: Proposition A5.)_ = Y(a) = 1'aS(a) = Ya(av)^ it follows that Pi ((. We can assume that Jta) = Yt(°). a'/a + 1.YQ(av) = j) Pi ( ci(a'V) > x. prove a somewhat more general result which was used in the proof of Proposition VI. v/ t.Yj(av) = j f . J(()) _ = i) + at•x t tt' . {t Y( a) } v>0 . a . We shall . and that Yt(a) = Yat for all t.x (1 . dx/ti] or not.bij) Hence the intensity matrix of { Ix} is (qij/ti)i.1. Then {Ix} is a Markov process with to = Yo. In addition to the asymptotic exponentiality.9. and write Yt = Yt(1). from which the phase process is terminated . 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}. t < (a). fo tY dv/t a$' t*. Conditioning upon whether { Yt} changes state in [0. Let further V be exponential with intensity V and independent of everything else. Hence O ((a) aa. from which it is easily checked that the limiting stationary distribution is (aiti/t*)iEE• Now let a' 4 oo with a in such a way that a' < a. J^O)_ = j) Pi (v(aaV) > x.
a = b = (bk)k=1. 1 k=1 1 0 otherwise.. See also Korolyuk.. zkbk is za(I . u Notes and references Propositions A5.. P. is discrete phasetype..+ at*x • a't' L ` at t* t* J Reducing the state space of {Ix } to {i E E : t. A distribution B on {1. = 0 for one or more i. (b) the generating function b[z] _ E' . k>1. (c) the nth moment k 1 k"bkis 1)"n!aP"p. so we shall be brief. Indeed. 2.1 and A5. However. the simplest discrete phasetype distribution: here E has only one element.p)k1 p. an easy modification of the argument yields finally the result for the case where t. ' pk 0 k>1 11 Theorem A5. .zP)'p. Gnedenko & Kovalenko [164] and Glasserman & Kou [162]).Pe. let E and Pkj j=k1. .2 do not appear to be in the literature. say bk = 0. 2. 5b Discrete phasetype distributions The theory of discrete phasetype distributions is a close parallel of the continuous case.352 rr Ia(a'V) Ei I ( > x) P APPENDIX L at (Yo (aV) ..x k > K. a).} is said to be discrete phasetype with representation (E. > 0}. Keilson [223].5 Let B be discrete phasetype with representation (P.3 As the exponential distribution is the simplest continuous phasetype distribution. K}. a) if B is the lifelength of a terminating Markov chain (in discrete time) on E which has transition matrix P = (p..4 Any discrete distribution B with finite support. Example A5.. k = 1. these results are in the spirit of rare events theory for regenerative processes (e.. and thus the parameter p of the geometric distribution u can be identified with the exit probability vector p.j) and initial distribution a. Then P is substochastic and the vector of exit probabilities is p = e . with point probabilities bk = (1 . Penev & Turbin [238]. Then: (a) The point probabilities are bk = aPklp. so is the geometric distribution. .. Et II I a(a^V) > x) at' . Example A5...g.
as is seen by minor modifications of Example A5. Then {Jt} has lifetime U1 + U2 .6.{ 0. The discrete counterpart is the negative binomial distribution with point probabilities bk k1) (1 k = r. A reduced phase diagram (omitting transitions within the two blocks) is am E(1) t(1) a(2) (2) t(2) Figure A.r + 1.1 This corresponds to a convolution of r geometric distributions with the same parameter p. Then the convolution B = B1 * B2 is phasetype with representation (E. and piece the processes together by it = 41) 0<t<U1 U1 < t < U1 + U2 2U.7 (THE NEGATIVE BINOMIAL DISTRIBUTION) The most trivial special case of Example A5. resp.T(1)). Jt t > U1 + U2.. resp. U2... and hence the negative binomial distribution is discrete phaseu type..a(1). and a=1). B2 be phasetype with representations (E(1).T(2)). (E(2). T= ( 0 T(2) ) (A. A. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2). .36) in blockpartitioned notation (where we could also write a as (a (1) 0)). r . { Jt 2) } with lifetimes U1 .6 is the Erlang distribution Er which is the convolution of r exponential distributions. _ i E E(1) T(1) t(1)a(2) i E E(2) .a(2). 11 Example A5.APPENDIX 353 5c Closure properties Example A5. a.2 The form of these results is easily recognized if one considers two independent phase processes { Jt 1) }. a' . initial distribution a and phase generator T.6 (CONVOLUTIONS) Let B1.
T. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2).10 (GEOMETRIC COMPOUNDS) Let B be phasetype with representation (E.354 APPENDIX Example A5. In risk theory. i E E(1) T 0 I (A.p)pn1. Then the mixture B = 9B1 + (1 .a(1). Example A5. Equivalently.0)a(2))).0)a(2) E(2) Figure A. if U1. A reduced phase diagram is 0a(1) E(1) A . Example A5.O)B2 (0 < 0 < 1) is phasetype with representation (E. and consider B(") = fA B(a) v(da) where v is a probability measure on A.p)pn1B*n. Thus. U2.T(2)). with common distribution and N is independent of the Uk and geometrically distributed with parameter p.. resp. a. P(N = n) = (1 .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. then C is the distribution of Ul + • • • + UN. a. we need to restart the phase process for B w. (E(2). this means that a = (Oa(1) (1 . p at each termination. Let B(") be the corresponding phasetype distribution. i E E(2) 0 T(2) =IT (in blockpartitioned notation.T(1)).p. T) and C = EO°_1(1 ..i.3 In exactly the same way.'). a mixture of more than two phasetype distributions is seen to be phasetype. a reduced phase diagram is f a E t Figure A. B2 be phasetype with representations (E(1).4 . are i.0)ai2).d.E) where a(°) = fAa(a)v(da). and o'i Oa.37) (1) (1 . To obtain a phase process for C.a(2).. Then it is trivial to see that B(") is u phasetype with representation (a(").8 (FINITE MIXTURES) Let B1. one obvious interpretation of the claim u size distribution B to be a mixture is several types of claims.
T).APPENDIX 355 and C is phasetype with representation (E..g. 12 (PHASETYPE COMPOUNDS ) Let fl. it follows by mixing (Example A5.11 (OVERSHOOTS) The overshoot of U over x is defined as the distribution of (U . a. U2 be random variables with distributions B1. Thus the representation is (E(1) x E(2).T + pta). U2. then C is the distribution of U1 + • • • + UN. Note that this was exactly the structure of the lifetime of a terminating renewal u process. then U1 +• is phasetype with representation (E. are i. Equivalently. let the initial vector be a ® v and u let the phase generator be I ® T + P ® (ta). Example A5 . with common distribution B and N is independent of the Uk with P(N = n) = f. It is zeromodified phasetype with representation (E.x)+.X)+ is zeromodified phasetype with representation (E. (E(2). let {Jtl)}. For U1 A U2. . let the phase space be E x F = {i j : i E E. X independent of U. resp.2.T) where F[T] = J0 "o eTx F(dx) u is the matrix m. 13 (MINIMA AND MAXIMA ) Let U1.. To obtain a phase representation for C .. v.°. T(1) ® T(2)). we then let the governing phase process be {Jt} _ {(411 Jt2))} 2) interpreting exit of either of {4 M }.°_1 f„ B*?l. { 4 } as exit of {Jt}. Example A5. of F.d. Indeed. If we replace x by a r.. say v. T + pta). To see this. then U1 + • • + UN is zeromodified phasetype with representation (a. cf.7. P). f2.v..1.f. Example A5 . T(2) ). .2. v. be the point probabilities of a discrete phasetype distribution with representation (E.. if U1. T + ta. if {Jt} is a phase process for U. i. { Jt2) } be independent with lifetimes U1.. then Jy has distribution aeTx. Corollary VIII.TWWW). E).a(1). resp.9) that (U . T) and C = F. cf. Proposition VIII.aF[T]. U2. If U1 has a different initial vector.aeTx.. let B be a continuous phasetype distribution with representation (F. a(2). but the same T.T) if U is phasetype with representation (E. Minor modifications of the argument show that 1. a. Then the minimum U1 A U2 and the maximum U1 V U2 are again phasetype. if B is defective and N + 1 is the first n with U„ = oo. say with distribution F. a. j E F}. B2 of phasetype with representations (E('). a(1) ® a(2 ). +UN 2.
. and let Bn be the Erlang distribution E. That is. relies more on matrix algebra than the probabilistic interpretation exploited here). and vice versa.B(bk) I < 1/n for n > k. with weight pi(n) for xi(n). The general case now follows easily from this. we can assume that ID. cf. the fact that any distribution B can be approximated arbitrarily close by a distribution with finite support. Here are the details at two somewhat different levels of abstraction: (diagonal argument ..2) } to go on (on E(2)) when { i 1) } exits. Thus the state space is E(1 ) x E(2) U E(1) U E( 2). there is a sequence {B.xq(n)(n)}.. Proof Assume first that B is a onepoint distribution. Hence it is immediate that Bn 4 B. elementary) Let {bk} be any dense sequence of continuity points for B(x). Example A5. r # oo.8. and the closedness of the class of phasetype distributions under the formation of finite mixtures. oo).. i= 1 C.} of phasetype distributions such that Bn 3 B as n + oo.(bk) + B(bk) for all k. 5d Phasetype approximation A fundamental property of phasetype distributions is denseness .(Sn) with Sn = n/b. By the diagonal argument (subsequent thinnings). Let the support of Dn be {xl(n). Then we must find phasetype distributions Bn with B. The mean of B„ is n/Sn = b and the variance is n/Sn = b2/n. q(n) q(n) pi(n)a . Now we can find first a sequence {Dm} of distributions with finite support such that D.. we need to allow { Jt.(bk)'. any distribution B on (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 . the initial vector is (a(1) (& a (2) 0 0). oo) can be approximated 'arbitrarily close' by a phasetype distribution B: Theorem A5.n = I:pi(n)Er v ( __ ) n) ) a= 1 . Then from above. say degenerate at b.356 APPENDIX For U1 V U2.. see Neuts [269] (where the proof.14 To a given distribution B on (0..(bk) + B(bk) for all k as n * oo..(n) = D. however.
k < n. and that cp is known to be continuous. Corollary A5. oo).n( b k ) . But To is the class G of all distributions on [0.r.. replications). oo) approximation Assume that we can compute a functional W(B) when B is phasetype. x 4 oo. there is a sequence {Bn} of phase type distributions such that Bn Di B as n 4 oo and f ' f. and we can take Bn = Cr(n).. oo) and any fl. For a general Bo. that this procedure should be used with care if ^p(B) is the ruin probability O(u) and u is large. oo) * [0.. compute W(B) and use this quantity as an approximation to cp(B0). k < n. . oo) such that f (x) = O(e«x). Hence G C PET and L = PIT. 2. however..14 is fundamental and can motivate phasetype assumptions.d.t.n (bk) . one would use the B given by some statistical fitting procedure (see below)..B(bk )I < . In particular. Then ICr( n ). the class CO of all discrete distributions. if information on Bo is given in terms of observations (i.i. 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..e.(x)Bf. for some a < oo.15 To a given distribution B on (0 . we can then approximate Bo by a phasetype B.APPENDIX 357 Hence we can choose r(n) in such a way that ICr( n).( dx) * f r f{(x)B(dx). i. u 2 (abstract topological ) The essence of the argument above is that the closure (w. It should be noted. If Cpl (B) and ^02(B) are weakly continuous. u Theorem A5.D(bk)I < n. E E. then it is immediate that WI(B) = p2(B) for all distributions B on [0... PIT contains all finite mixtures of onepoint distributions. i = 1. Let E be the class of functions f : [0. say on the claim size distribution B in risk theory. the topology for weak convergence) PET of the class PET of phasetype distributions contains all onepoint distributions.n. Since PET is closed under the continuous operation of formation of finite mixtures. f2.
there is a sequence {Bn} of phase type distributions such that Bn Di B as n + oo and all moments converge. i = 1.2 .. n.39). 2.. n B=az. f° xtBn(dx ) * f °° x`B( dx).. Now returning to the proof of (A. f00 fi(x)Cr.38 ).(dx) > J fi(x)B(dx). ... if f (x ) = e°x. oo).... and the case of a general f then follows from the definition of the class E and a uniform integrability argument.n(dx) < 1+. and hence we may choose r(n) such that L 9l) f (x)Cr(n). . TO (A..14 Dn has been chosen such that 00 1 °° f fi(x)D n(dx ) < 1++ '  o \ n o f fi(x)B(dx). By (A.n(dx) + f 0 fi(x)Dn(dx). then cc f (x)Bn ( dx) = (?!c ) e'= .. 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.  APPENDIX B implies that 00 o o 00 n. we may assume that in the proof of Theorem A5.. Bn=En z f f (x)Bn(dx) fof (x)B(dx) = ° (A.358 Proof By Fatou' s lemma. n.oo J fi(x)B.38) We first show that for each f E E.f (x)B(dx).16 To a given distribution B on (0 .f ' f (x)B(dx). liminf B. . i = 1. i=1.f (z) = f = 1 1 1 1n/ o . \\ 0 Corollary A5.. .39) Indeed. i = 1. ..
./3). (N or a given distribution Bo. the problem thus arises of how to fit a phasetype distribution B to a given set of data (1.16.} of phasetype distributions such that Bfz + B as n * oo and Yn 4 ry where ryn = y(Bn.. then Bn['Y + ei] * B[y + ei] > 1 + 7 Q implies that 'yn < ry + ei for all sufficiently large n . there is substantial advantage in assuming the claim sizes to be phasetype when one wants to compute ruin probabilities. and therefore the following result is highly relevant as support for phasetype assumptions in risk theory: Corollary A5.. 0 as i * oo. . The adjustment coefficient is a fundamental quantity. the adjustment coefficient 'y = 7(B.l3µb < 1. oo) with B[y +e] < oo for some e > y = 7(B. If ei > 0. . e ) and ei J. Proof Let fi(x) = el'r+E. from a more conceptual . lim sup ryn < 7. lim inf > is proved similarly. (N.18 In the setting of Corollary A5. one can obtain 7(Bn. For practical purposes.17 To a given /3 > 0 and a given distribution B on (0. . there is a sequence {B. Notes and references Theorem A5./3) is defined as the unique solution > 0 of B[y] = l+y/j3. We shall formulate the problem in the slightly broader setting of fitting a phasetype distribution B to a given set of data (1i .14 is classical. 5e Phasetype fitting As has been mentioned a number of times already.3). I. /3) = ry for all n.. but are certainly not unexpected. The present section is a survey of some of the available approaches and software for inplementing this. the loggamma or the Weibull have been argued to provide adequate descriptions of claim size distributions. the remaining results may be slightly stronger than those given in the literature. .> y for some sequence {ei} with ei E (0. However.e. This is motivated in part from the fact that a number of nonphasetype distributions like the lognormal. and in part from the fact that many of the algorithms that we describe below have been formulated within the setup of fitting distributions. . O We state without proof the following result: Corollary A5. .APPENDIX 359 In compound Poisson risk processes with arrival intensity /3 and claim size distribution B satisfying .
defined by the absence of loops in the phase diagram . A number of approaches restrict the phase type distribution to a suitable class of mixtures of Erlang distributions . g. [317] ) has considered an extension of this setup.. [70]) restrict attention to acyclic phase type distributions . we do not not want to perform matrix calculus in hundreds or thousands dimensions). (N is the empirical distribution Be. the number of phases required for a good fit will typically be much larger.360 APPENDIX point of view the two sets of problems are hardly different : an equivalent representation of a set of data (1 . and in practice this sets a limitation to the usefulness (the curse of dimensionality . Of course.'s).g. the L1 distance between the c . A method developed by Bobbio and coworkers (see e. and used a nonlinear programming approach . 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. d.. B„ The problem is that the constructions of {B„} are not economical : the number of phases grows rapidly.f. Johnson & Taaffe considered a mixture of two Erlangs (with different rates ) and matched (when possible ) the first three moments .g. It seems therefore a key issue to develop methods allowing for a more general phase diagram.d. Asmussen & Nerman [38] implemented maximum likelihood in the full class of phasetype distributions via the EM algorithm . Schmickler (the MEDA package. giving mass 1 /N to each S=. three for a mixture of two Erlangs ).g . risk theory. and as fitted distribution we may take B. at a a number of selected points . cf. The likelihood function is maximized by a local linearization method allowing to use linear programming techniques. one could argue that the results of the preceding section concerning phasetype approximation contains a solution to our problem : given Bo (or Be). for some suitable large n. [216] ). and this is what matters when using phasetype distributions as computational vehicle in say renewal theory. e . we have constructed a sequence { B.} of phasetype distribution such that Bo. In a series of papers (e. The observation is that the statistical problem would be straightforward if the whole ( EAvalued) phase process { Jtk)} o<t<( k associated with each observa . The earliest such reference is Bux & Herzog [85] who assumed that the Erlang distributions have the same rate parameter.. a program package written in C for the SUN workstation or the PC is available as shareware. and we next describe two such approaches which also have the feature of being based upon the traditional statistical tool of like maximum likelihood. . .g. 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. [202].. reliability or queueing theory. The characteristics of all of these methods is that even the number of parameters may be low (e.
the methods of [70] and [38] appear to produce almost identical results. it is easy to see that N (k Ea(n). eieT(n)((k. .... EN where ai = N 1 I ((k) = i) tii=i iEE.T (n)(TiI(1. In fact. Thus. .x)t(n) 1 and this and similar expressions are then computed by numerical solution of a set of differential equations. it seems open whether the restriction to the acyclic case is a severe loss of generality. Nii = = .T(n) (Ti ^^ 1. e. jEEA.(N) = E Ea(n). N Ti = I(J= i) dt.g. (N) tJk Ea ( n). (N ) (^ 54 k )+ and similarly for the cn+1) The crux is the computation of the conditional expectations. = j) f k=1 k =1 tE[0. .T(n) (Nik IC1.(k] (Ti is the total time spent in state i and Nii is the total number of jumps from i to j). .. since this is parameterdependent.. The general idea of the EM algorithm ([106]) is to replace such unobserved quantities by the conditional expectation given the observations. In practice. . E. one is lead to an iterative scheme.. then the estimators would be of simple occurenceexposure type..APPENDIX 361 tion Sk was available. (n+1) _ Ea (n)..g..T(n) k=1 I (Jti) dt o \f a(n)eT(n )(kt(n) N f:i a(n)eT(n)xei .
This page is intentionally left blank .
G. Sparre Andersen (1957) On the collective theory of risk in the case of contagion between the claims . Whitt ( 1992) The Fourierseries method for inverting transforms of probability distributions . Choudhury & W. Dover. Anantharam (1988 ) How large delays build up in a GI/GI11 queue. Abramowitz & I. [3] J. Stoch. Skand.L. [15] S. Proc. Optimizations Software. Probab. Queueing Systems 16. [7] E. In: Limit Theorems and Related Problems (A. [13] S. 311338. Aktuar Tidskr. Asmussen (1987) Applied Probability and Queues. Skand. 253267. [14] S. B. [9] G. Teubner. Whitt (1998) Explicit M/G/1 waitingtime distributions for a class of longtail service time distributions . 53100. G. ibid.). The case of equal risk sums. Transactions XVth International Congress of Actuaries. Abate & W. 191223. [8] G. 219229. [12] S. Appl. New York. Adv.Bibliography [1] J. pp. Stegun ( 1972 ) Handbook of Mathematical Functions (10th ed. Asmussen (1985) Conjugate processes and the simulation of ruin problems. Asmussen ( 1982 ) Conditioned limit theorems relating a random walk to its associate. Borokov. 363 . 38.). New York. John Wiley & Sons. 143170. Whitt (1994) Waitingtime tail probabilities in queues with longtail service time distributions . Abate & W. Appl. Alsmeyer (1991) Erneuerungstheorie. The Mathematical Scientist 14. 20. 57. II. [11) S. Arfwedson ( 1954) Research in collective risk theory. 213229. Queueing Systems 5 . 101116.A. Asmussen (1984) Approximations for the probability of ruin within finite time . Scand. [6] V. Queueing Systems 10. Abate. with applications to risk reserve processes and the GI /G/1 queue. Act. [5] G. 345368. ed. 587. 14. New York. 3157 . AT&T. 1984 . [10] K. 37. 1985. Chichester New York. J. Preprint. Aktuar Tidskr. [4] M. Arndt (1984) On the distribution of the supremum of a random walk on a Markov chain . Asmussen (1989a) Aspects of matrix WienerHopf factorisation in applied probability. Stuttgart. [2] J. Arfwedson ( 1955) Research in collective risk theory.
Scand. 193226. Boca Raton. Asmussen. Hojgaard (1996) Finite horizon ruin probabilities for Markovmodulated risk processes with heavy tails. Asmussen (1992c) Stationary distributions for fluid flow models and Markovmodulated reflected Brownian motion. 2943. Ann. Andronov. [24] S. 69100. 772789. eds. Chakravarty.K. Appl. Asmussen (1989b) Risk theory in a Markovian environment. SIAM Review 40. 27. Asmussen (1992b) Light traffic equivalence in single server queues. 303322. [28] S. 354374. Advances in Queueing: Models. Binswanger (1997) Simulation of ruin probabilities for subexponential claims. 8. [17] S. J. Hojgaard (1999) Approximations for finite horizon ruin prob abilities in the renewal model. K. Hojgaard (2000) Rare events simulation for heavytailed distributions. Scand. [21] S. Astin Bulletin 27. Appl. Asmussen (1998b) Extreme value theory for queues via cycle maxima. Alfa & S. Asmussen (1999) On the ruin problem for some adapted premium rules. 1996. Act. Proc. [31] S. [23] S. Asmussen (2000) Matrixanalytic models and their analysis. Marcel Dekker. Bernoulli 6. 313326. 2. [20] S. Ann. Bladt (1996) Renewal theory and queueing algorithms for matrixexponential distributions. Proc. Asmussen & K. 25. [29] S. [32] S. [27] S.M. [22] S. J. A. Appl. stationary distributions and first passage probabilities. Appl. 79102. Probab. Stoch. 297318.). J. [33] S. Act. Extremes 1 . CRC Press. Asmussen (1991) Ladder heights and the Markovmodulated M/G/1 queue. 2149. 189201. Dshalalow ed. 96107. Methods V Problems (J. Probab. 313341. Frey. T. [26] S. Probabilistic Analysis of Rare Events (V. Statist. . 4966. J. 555574. Th.). 137168. Asmussen. Asmussen.S.). Riga Aviation University. [18] S. Asmussen (1995) Stationary distributions via first passage times. Schmidt (1995) Does Markovmodulation increase the risk? ASTIN Bull. Scand. 315. [30] S. Kliippelberg (1994) Large claims approximations for risk processes in a Markovian environment. 1999 . Probab. Random Processes 2. Stoch. 1936. [19] S. 37. Asmussen (1998a) Subexponential asymptotics for stochastic processes: extremal behaviour. Scand. 20. Florida. Asmussen & M. Rolski & V.364 BIBLIOGRAPHY [16] S. New York. [34] S. Ann. 106119. Asmussen & M. MatrixAnalytic Methods in Stochastic Models (A. Asmussen (1992a) Phasetype representations in random walk and queueing problems. Floe Henriksen & C. Act. Binswanger & B. Bladt (1996) Phasetype distributions and risk processes with premiums dependent on the current reserve. eds. 1989 . L. Asmussen & B. Asmussen & B. 54. Kalashnikov & A. Stochastic Models 11. [25] S. Asmussen (1998c) A probabilistic look at the WienerHopf equation.
Appl. Th. Asmussen & H. Asmussen & C. Florida. Prob 32 . [49] S. Asmussen & S. [43] S. Asmussen & C. Insurance: Mathematics and Economics 10. Probab. Proc. Taksar (2000) Optimal risk control and dividend distribution policies . [52] S. 115. CRC Press. 31. Inf. Olsson (1996). Asmussen & R. [40] S. 422447.A. Nerman & M. Stoch. Johnson. Statist. Asmussen & T. Asmussen & G. Hojgaard & M. Asmussen & D. Schock Petersen (1989) Ruin probabilities expressed in terms of storage processes. Wiley. (to appear). Koole (1993). Proc.Y. Example of excessoffloss reinsurance for an insurance corporation. Res. 120. 30. Appl.A. 58. [47] S. Adv. Probab. Rubinstein (1999) Sensitivity analysis of insurance risk models. Appl. Asmussen & V. Schmidli & V. Asmussen. [51] S. Appl. [44] S. Rolski (1994) Risk theory in a periodic environment: Lundberg's inequality and the CramerLundberg approximation. 421458. Adv. Stochastic Models 8. Sigman (1996) Monotone stochastic recursions and their duals. 299324. Asmussen & K. [39] S. B. 103125. 64. [45] S. J. Eng. [37] S.BIBLIOGRAPHY 365 [35] S. O'Cinneide (2000/2001) Matrixexponential distributions [Distributions with a rational Laplace transform] Encyclopedia of Statistical Sciences.Y. [48] S. Rubinstein (1995) Steadystate rare events simulation in queueing models and its complexity properties. [36] S. [46] S. 736755. Appl. Rolski (1991) Computational methods in risk theory: a matrixalgorithmic approach. Statistica Neerlandica 47. Schmidt (1995) Ladder height distributions with marks. Math. Management Science 45.). Insurance: Mathematics and Economics 20. Appl. Asmussen & V. Schmidt (1999) Tail approximations for nonstandard risk and queueing processes with subexponential tails. . 23. Asmussen. 20. 429466. Asmussen & M. Fitting phasetype distributions via the EM algorithm. [38] S. Probab. Read eds.M. H. Boca Raton. Asmussen . 10. Supplementary Volume (Kotz. Marked point processes as limits of Markovian arrival streams. 11251141. Probab. Scand. O'Cinneide (2000/01) On the tail of the waiting time in a Markovmodulated M/G/1 queue. 365372. J. [41] S. Sc. Finance and Stochastics 4. J. 410433. Opns. Kliippelberg (1996) Large deviations results for subexponential tails. 913916. with applications to insurance risk Stoch. Perry (1992) On cycle maxima. [42] S.). Advances in Queueing: Models. Schmidt (1993) The ascending ladder height distribution for a class of dependent random walks. Dshalalow ed. 19. Oper. 259274. Asmussen & R. Asmussen & C. O. Methods 8 Problems (J. first passage problems and extreme value theory for queues. Res. 105119. 419441. Nielsen (1995) Ruin probabilities via local adjustment coefficients . [50] S. Taksar (1997) Controlled diffusion models for optimal dividend payout. Asmussen & T.
Teugels ( 1996) Convergence rates for M /G/1 queues and ruin problems with heavy tails . Actuaries 21. Berlin. Act. Probab. [62] J.J. 4148. Crijns (1987) Upper bounds on ruin probabilities in case of negative loadings and positive interest rates. 275279. [68] T. [63] A. 832840. [59] O. Baker (1977) The Numerical Solution of Integral Equations. 705766. [58] O. [54] K. 610. Schmidli (1995) Saddlepoint approximations for the probability of ruin in finite time. [66] N. 661667. BarndorffNielsen & H. Adv.B . Act. Probab. Appl. 1975. [69] P. 1995. Halsted Press. [64] P. Bjork & J. Beekman (1974) Two Stochastic Processes. SpringerVerlag.A. Stochastic Models 10. Beekman (1985) A series for infinite time ruin probabilities. 190204. Cambridge. [70] A. J. Athreya & P. Bingham. Bingham (1975) Fluctuation theory in continuous time.M. Springer . Cambridge University Press. Soc. Act. J. Trans. Borovkov (1976) Asymptotic Methods in Queueing Theory.H. de Waegenaere (1990) Simulation of ruin probabilities. 9. Scand. Scand. Billingsley (1968) Convergence of Probability Measures. 77111. Boogaert & A. [65] N. Oxford. 129134. Scand. 1985 . J. [57] C. 148156. New York. C. [61] J. Appl. Cox (1972) A low traffic approximation for queues.A. [72] P. J. Ney ( 1972) Branching Processes . Asmussen & J.H. Berman & R. Clarendon Press. Teugels (1987) Regular Variation. 1988 . Bobbio & M. [55] B. J.L. Grandell (1988) Exponential inequalities for ruin probabilities in the Cox case. 169186. Wiley. Plemmons (1994) Nonnegative Matrices in the Mathematical Sciences.366 BIBLIOGRAPHY [53] S.H. New York. 9599. [56] B. Boogaert & V. Bloomfield & D. Grandell (1985) An insensitivity property of the ruin probability. Act. Appl. Scand. Probab. Insurance: Mathematics and Economics 9. Wiley. [60] J. [73] A. J. Act. . J. [67] T.L. SIAM. Insurance: Mathematics and Economics 6. Scand. Beekman (1969) A ruin function approximation. Telek (1994) A bencmark for PH estimation algorithms: results for acyclic PH. Insurance: Mathematics and Economics 4. 33. 221232.T. 7. 1974. Goldie & J. von Bahr (1974) Ruin probabilities expressed in terms of ladder height distributions. [71] P. New York. 11811190. BarndorffNielsen (1978) Information and Exponential Families in Statistical Theory. von Bahr (1975) Asymptotic ruin probabilities when exponential moments do not exist.R. Bjork & J.
Resnick & R. 177204.W. Itasca.R. verve. De Vylder (1986) Ordering of risks and ruin probabilities. [84] D. Control AC28. Malgouyres (1983) Large deviations and rare events in the study of stochastic algorithms . Nesbitt (1986) Actuarial Mathematics .W. 62. [90] D. Wahrscheinlichkeitsth. S. Bucklew ( 1990) Large Deviation Techniques in Decision. 93121. Z. Syst. The Society of Actuaries . Herzog (1977) The phase concept : approximations of measured data and performance analysis. . Wiley. II. Stockholm. [85] W. Bux & U. M. Aut. 127130.M. Probab.L. Geb. Stockholm. Bratley. [93] K. Burman & D .J. Croux & N. Amsterdam.L.R. IEEE J. Tech. New York. Springer. SpringerVerlag. 313319. 2338.R. Jones & C. Fox & L. Veraverbeke (1990). Cohen (1998) The M/G/1 queue with heavytailed service time distribution. 24.U. [86] K. 749763.. [91] H. J. J.J. NorthHolland.I. Adv. Philos. Reading.Y . Insurance : Mathematics and Economics 5.A. 14.BIBLIOGRAPHY 367 [74] O. [75] O. Fort & G. Cox (1955) Use of complex probabilities in the theory of stochastic processes. B. Oper. 14331453. Smith ( 1983) Asymptotic analysis of a queueing system with bursty traffic. 105119.).L. Breiman ( 1968) Probability. Biihlmann ( 1970) Mathematical Methods in Risk Theory. 907920. Soc. Illinois.C. Bell. Proc. [80] F. [77] P.W. Cramer (1955) Collective risk theory. Boxma & J. [88] J. Insurance : Mathematics and Economics 9. Cottrell. Nonparametric estimators for the probability of ruin. The Jubilee volume of Forsakringsbolaget Skandia. Computer Performance (K. Broeck . [83] D. Bowers . Amsterdam. Simulation and Estimation. AddisonWesley. Queueing Systems 33. H. [89] M. Cohen (1982) The Single Server Queue (2nd ed. Cohen (1999) Heavytraffic analysis for the GI/G/1 queue with heavytailed service time distributions . Res. Chandy & M. Brockwell . [76] N. Berlin. Jr. D. Gerber. New York. 3539. Boxma & J. [92] H. 392433. Sel. Hickman. Academic Press. Burman & D. Schrage (1987) A Guide to Simulation. Smith ( 1986) An asymptotic analysis of a queueing system with Markovmodulated arrivals . Goovaerts & F.) NorthHolland.).C. 51. Qinlar (1972) Markov additive processes. Skandia Jubilee Volume. Appl. Chung (1974) A Course in Probability Theory (2nd ed.L. [82] H.Y. 16. Cambr. Cramer (1930) On the Mathematical Theory of Risk. 34. [87] E. Area Commun. Reiser eds. [79] P. New York San Francisco London.A. IEEE Trans. [78] L. Tweedie (1982) Storage processes with general release rule and additive inputs . [81] J. J.
[100] A. [101] C. 7083 (1969). Chapman & Hall. Appl. Suppl. Res. Insurance: Mathematics and Economics 10.P. Insurance: Mathematics and Economics 7. 88101. Csorgo & J. De Vylder (1996) Advanced Risk Theory.D. English version published in Skand. [110] F. T. Daley & T. Soc. J. Act. van Dawen (1986) Ein einfacher Beweis fur Ohlins Lemma. Devroye (1986) Nonuniform Random Variate Generation. Springer. Deheuvels & J. Steinebach ( 1991 ) On the estimation of the adjustment coefficient in risk theory via intermediate order statistics . Blotter der deutschen Gesellschaft fur Versicherungsmathematik XVII. Statist. 85116. New York. J. Insurance: Mathematics and Economics 3. 181217. Daley & T. Tidskr. [108] F. 121131. 227229. [112] F.. [97] D. Boston. [105] A. De Vylder (1977) A new proof of a known result in risk theory. Haezendonck (1987) Classical risk theory in an economic environment. [102] P. Rolski (1984) A light traffic approximation for a single server queue.. Stockholm. De Vylder (1978) A practical solution to the problem of ultimate ruin probability. Delbaen & J. Comp. J.B. Goovaerts (1988) Recursive calculation of finitetime ruin probabilities. Embrechts (1989). N. 583602. Goovaerts (1984) Bounds for classical ruin probabilities. Roy. Pentikainen & E.368 BIBLIOGRAPHY [94] M. Insurance: Mathematics and Economics 6. 201206. Oper. 5771. 624628. [99] R. [98] A. Stochastic Models 5. 1990 . Scand. [111] F. De Vylder & M. Scand. J. Laird & D. [107] L. 3. [106] A. 27. [103] F. Dassios & P. Insurance: Mathematics and Economics 4. Dempster. 435436. 135159. De Vylder & M. Math. Editions de 1'Universite de Bruxelles. Jones and Bartlett. Rolski (1991) Light traffic approximations in queues. . [95] M. Delbaen & J. Appl. Daykin. 9 . Csorgo & J. Steinebach (1990) On some alternative estimators of the adjustment coefficient in risk theory. Teugels ( 1990) Empirical Laplace transform and approximation of compound distributions . Math. Rubin (1977) Maximum likelihood from incomplete data via the EM algorithm. Forsiikringsmatematiska Studier Tillagnade Filip Lundberg.L. Actuarial J. Martingales and insurance risk.M. Zeitouni (1992) Large Deviations Techniques and Applications. Aktuar. 17. Res. Oper. [104] F. 16. 114119. [96] D. 3750. Dembo & O. Haezendonck (1985) Inversed martingales in risk theory. Pesonen (1994) Practical Risk Theory for Actuaries. Probab. Math. 22. Davidson (1946) On the ruin problem of collective risk theory under the assumption of a variable safety loading (in Swedish). [109] F.
C. 1993 . [115] D. Dickson & J. Dickson & C. [123] E.C. Waters (1996 ) Reinsurance and ruin . Scand. Grandell & H. Scand. 177192. Oper. Res. 191207. Embrechts ( 1988 ). Hipp (1998) Ruin probabilities for Erlang (2) risk processes. Scand. . O'Connell (1995) Large deviations and overflow probabilities for the general singleserver queue.M. Dufresne & H. J. Insurance : Mathematics and Economics 22.K.M. 131138.B.R.M.C. [118] D. Regterschot (1988) Conditional PASTA. J. 6180. 1984 . Act. [120] D. Springer . Math. 251262.C.U. Gerber & E. J. Shiu (1991) Risk theory with the Gamma process. [117] D. Proc. Dickson (1992) On the distribution of the surplus prior to ruin . British Actuarial J.C. Act.BIBLIOGRAPHY 369 [113] D. Duffield & N. Harrison & A.M.R. [114] D.M. Scand. 5159.C. Dickson & H.J. Insurance : Mathematics and Economics 10. 1993 . Dickson & H. Dickson & C. Scand. 363374.G. 1.C. Gray (1984) Exact solutions for ruin probability in the presence of an upper absorbing barrier. Scand. 1741. Hipp (1999) Ruin problems for phasetype(2) risk processes. Camb. Insurance : Mathematics and Economics 7. Dickson (1994) An upper bound for the probability of ultimate ruin. H. Letters 7. J. Scand.R. [128] E.C. with applications. Embrechts . [116] D. [124] N.M. 4259. [121] D. Gerber (1988) The surpluses immediately before and at ruin. and the amount of claim causing ruin. Insurance: Mathematics and Economics 19. J. [131] P. van Doorn & G. Act. 118. J.S. 193199.W. Berlin Gottingen Heidelberg. Gray (1984) Approximations to the ruin probability in the presence of an upper absorbing barrier. 4962. 105115. Soc. [127] F. Dickson (1995) A review of Panjer's recursion formula and its applications. 1975 . J.R.U. Insurance: Mathematics and Economics 7. Dufresne & H. [122] B. [130] P. 174186.M. 2000 . Act. Philos. Insurance: Mathematics and Economics 25. Ruin estimates for large claims . Schmidli (1993) Finitetime Lundberg inequalities in the Cox case .M. Taylor ( 1975) A diffusion approximation for the ruin probability with compounding assets. Astin Bulletin 21. [126] F. 147167. Dufresne. Gerber (1991) Risk theory for the compound Poisson process that is perturbed by a diffusion. [119] D. J. Act.J. [125] F.U. Djehiche (1993) A large deviation estimate for ruin probabilities. 229232. 269274. [129] D . J. 1994 . 3745. Act. 1984 .C. Insurance : Mathematics and Economics 11. Act. 107124.M.C. Dynkin (1965) Markov Processes I.M. Dickson & J. Emanuel . Waters (1999) Ruin probabilities wih compounding.
J. Heidelberg.). Adv.M. Insurance: Mathematics and Economics 10. 5974.). Math. Probab. [134] P. [146] E.K. 5975. [136] P. Adv. Th. Fuh (1997) Corrected ruin probabilities for ruin probabilities in Markov random walks. Reprinted 1948 as 'The theory of probabilities and telephone conversations' in The Life and work of A. 175195. Feller (1966) An Introduction to Probability Theory and its Applications I (3nd ed. 404422. Embrechts & J.370 BIBLIOGRAPHY [132] P. Sci. 26.L. 3339. Rel.A. Embrechts & T. [135] P. Astin Bulletin 16. Schmidt (1992) An insensitivity property of ladder height distributions.W. Pitts (1993) Some applications of the fast Fourier transform in insurance mathematics . Embrechts.W. Veraverbeke (1982) Estimates for the probability of ruin with special emphasis on the possibility of large claims . Act. Danish Academy Tech. Statistica Neerlandica 47. Proschan & D. Jensen. Furrer (1998) Risk processes perturbed by astable Levy motion. R. . [149] C. [141] F. Teugels (1985). Feller (1971) An Introduction to Probability Theory and its Applications II (2nd ed. C. 131137. New York. first passage times and moments of ladder variables in Markov random walks. with applications. Wiley. Esary. Embrechts. 269274. Franken. Esscher (1932) On the probability function in the collective theory of risk. Villasenor (1988) Ruin estimates for large claims. [140] J. Erlang 2. 29. Arndt & V.K. Walkup (1967) Association of random variables. M. Statist. Springer. New York.D. Appl. Lai (1998) Wald's equations. Ann. Appl. Probab. [142] W. Frenz & V. Fields 75. Embrechts & H. J. Akt. Probab. [143] W. Kliippelberg & T. Probab.L.D. Scand. Fitzsimmons (1987). U. Appl. Schmidli (1994) Ruin estimation for a general insurance risk model. [150] H. 14661474. 695712. Probab. 159178. Tidsskr. Embrechts. 5572. [144] P. Mikosch (1991). Fuh & T. 17. D. [137] P. 35. Adv. Insurance: Mathematics and Economics 7. 8190. Schmidt (1982) Queues and Point Processes. Griibel & S. [138] P. 181190. Frees (1986) Nonparametric estimation of the probability of ruin. Erlang (1909) Sandsynlighedsregning og telefonsamtaler . Wiley. Nyt Tidsskrift for Matematik B20. [148] C. J. On the excursions of Markov processes in classical duality. 623637. Approximations for compound Poisson and Polya processes. [147] M. Mikosch (1997) Modelling Extremal Events for Finance and Insurance. Konig.D. A bootstrap procedure for estimating the adjustment coefficient. Maejima & J. 616624. [133] P. 38. Probab. Embrechts & N. F. 566580. Appl. Appl. 29. 1998 . Skand. [145] P. Insurance : Mathematics and Economics 1. Trans.J. [139] A. Wiley. J.
BIBLIOGRAPHY 371 [151] H. Kluwer. [155] H. [154] H. [161] P. [163] P. Gani. 1523. [157] H. Furrer.. Gerber (1970) An extension of the renewal equation and its application in the collective theory of risk.U. ETH Zurich. Springer. 31A. Gerber (1986). [153] H. Z. van Heerwarden & T. Gerber (1971) Der Einfluss von Zins auf die Ruinwahrscheinlichkeit. J. Act. [160] H. Michna & A. Scand. [166] M. 7778. [169] J. Kou (1995) Limits of first passage times to rare sets in regenerative processes. [165] M.G. 28. F. Griibel (1996) Perpetuities with thin tails. Suppl. [156] H. Ann. 2336. Galambos & J. Gerber (1979) An Introduction to Mathematical Risk Theory. A.U. In Studies in Applied Probability (J. Schweiz. Scand. Probab. J. Appl. J. 205216. Furrer & H.W. . Grandell (1977) A class of approximations of ruin probabilities. Skand.). Glasserman & S. Scand. [152] H. Huebner Foundation Monographs. NorthHolland. Ver. Weron (1997) Stable Levy motion approximation in collective risk theory. Amsterdam.M.). Birkhauser . Goldie & R. Vers.S.U. Probab. Mitt. 97114. contained in the authors PhD thesis. Math. 1978 . NorthHolland. 424445. Kovalenko (1989) Introduction to Queueing Theory (2nd ed. Grandell (1978) A remark on 'A class of approximations of ruin probabilities'. Kaas. London. Appl.71. Vers. R. Boston. Bauwelinckx (1990) Effective Actuarial Methods. [159] H. [168] J. Life Insurance Mathematics.U. [162] P. Schweiz. [167] C.J. Goovaerts. Haezendonck (1984) Insurance Premiums. Insurance: Mathematics and Economics 20.V. [164] B. 1981 . Goovaerts. 3752. Whitt (1994) Logarithmic asymptotics for steadystate tail probabilities in a singleserver queue. Probab. 131156. Mitt. [158] H. Insurance: Mathematics and Economics 15. Gerber (1973) Martingales in risk theory. Act. Unpublished. Gerber (1981) On the probability of ruin in the presence of a linear dividend barrier. S. 6370. Ver. Math.. Insurance : Mathematics and Economics 7.U. eds. 205210. Glynn & W. Aktuarietidskrift 1970 . Act. de Vylder & J. Amsterdam. 5.J. 463480.N. J. Schmidli (1994) Exponential inequalities for ruin probabilities of risk processes perturbed by diffusion. Basel. Dordrecht. Furrer (1996) A note on the convergence of the infinitetime ruin probabilities when the weak approximation is astable Levy motion. 1977. Glasserman (1991) Gradient Estimation via Perturbation Analysis. Gnedenko & I.U. University of Pennsylvania. 105115. Adv. Appl.E. 73. Gerber (1988) Mathematical fun with ruin theory.U.
[187] J. Harrison & S.V. [181] D. Griibel & R. AkademieVerlag. 14. [188] J. M. Archiv v. [184] H . 33. Resnick (1977) The recurrence classification of risk and storage processes . [182] A. Insurance: Mathematics and Economics 26.372 BIBLIOGRAPHY [170] J. New York. Tinbergen Institute Research Series 20. 7590. M.131135. E. 143158. Preprint. Opns . Grandell ( 1992) Finite time ruin probabilities and martingales . .243255. Silvestrov (2000) Cram€rLundberg approximations for nonlinearly perturbed risk processes . M. [189] A.O. Proc. 355365. ed. [179] R. Stoch. [171] J. 197214. 6779. SpringerVerlag. Res. 1. Ibragimov. 3 . Proc.and Sozialforschung 6. Liet. Math. Appl. Informatica 2. Aktuar Tidskr. 347358. Rink. [186] J. Grandell (1997) Mixed Poisson Processes. [178] B. Opns . 14. 5 . Statistical Algorithm 265. Hadwiger (1940) Uber die Wahrscheinlichkeit des Ruins bei einer grossen Zahl von Geschiiften . Harrison & S. SpringerVerlag. Grandell (1979) Empirical bounds for ruin probabilities. Winter School on Stochastic Analysis and Appl . 566576. Proc .S. Appl.I. [175] J. Grigelionis (1996) Lundbergtype stochastic processes. 3041. Res. Skand. M.A. Probab. Resnick (1976 ) The stationary distribution and first exit probabilities of a storage process with general release rule . Stoch.I. Berlin. Matem. Th. [174] J. [172] J.. [176] B. Grandell & C. van Heerwarden (1991) Ordering of Risks: Theory and Actuarial Applications. Applied Statistics 40. Harrison (1977) Ruin problems with compounding assets . 332. Probability Theory and Mathematical Statistics (I. Segerdahl (1971) A comparison of some approximations of ruin probabilities. KTH. [177] B. Harrison & A. Appl. Lemoine (1977) Limit theorems for periodic queues. Probab. [183] M. Appl. mathematische Wirtschaft . Amsterdam. [173] J. Korolyuk (1969) On the joint distribution of a process with stationary increments and its maximum . Grandell (1999) Simple approximations of ruin functions. Grigelionis ( 1992) On Lundberg inequalities in a Markovian environment. Grigelionis ( 1993) Twosided Lundberg inequalities in a Markovian environment.J. Math. 5766. Gut (1988 ) Stopped Random Walks . Theory and Applications. 167176. Grandell (1990) Aspects of Risk Theory. 400409. Astin Bulletin 29. 8. [185] J. [180] R. Gusak & V. Chapman & Hall. 54. Grnbel (1991) G/G/1 via FFT. Gyllenberg & D.). Hermesmeier (1999) Computation of compound distributions I: aliasing errors and exponential tilting. J.
[191] W. Z. [202] 0. Michel (1990) Risikotheorie: Stochastische Modelle and Statistische Methoden. 18. New York.R. 8996. [207] D. [198] C. Insurance: Mathematics and Economics 5. Karlsruhe. Staalhagen (1999) Waiting time distributions in M/D/1 queueing systems. 6. Hermann (1965) Bin Approximationssatz fiir Verteilungen stationarer zufalliger Punktfolgen.).. Heidelberger (1995) Fast simulation of rare events in queueing and reliability models. Proc. No. on Asymptotic Statistics (P.V. 259268. 123151. [204] T. 305313. Hesselager (1990) Some results on optimal reinsurance in terms of the adjustment coefficient. Wiley. Prob. Electronic Letters 35. 3. verve. Probab. 29. Ann. Hipp (1989a) Efficient estimators for ruin probabilities. Taksar (2000) Optimal proportional reinsurance policies for diffusion models. Heilmann (1987) Grundbegriffe der Risikotheorie. Hojgaard & M. J. Hoglund (1990) An asymptotic expression for the probability of ruin within finite time. Asmussen & 0. Nerman (1992) EMPHT . 4385. Appl. Ann. 285292. Astin Bulletin 19. [206] B. . 11631174. Mathematical Statistics. 25. Hill (1975) A simple general approach to inference about the tail of a distribution. Insurance: Mathematics and Economics 5. Statist. [200] M. The ruin problem for finite Markov chains. [203] T. 285293.BIBLIOGRAPHY 373 [190] P. [205] T. [193] U. [201] L. Probab. Nachrichten 30. [197] C. Klugman (1984) Loss Distributions.a program for fitting phasetype distributions. Verlag Versicherungswirtschaft e. Scand. J. Hipp & R. Versicherungswirtschaft. Huskova. J. [195] B. Willekens (1986) Estimates for the probability of ruin starting with a large initial reserve. J. Herkenrath (1986) On the estimation of the adjustment coefficient in risk theory by means of stochastic approximation procedures. 1998 . Foruth Prague Symp. Hipp (1989b) Estimators and bootstrap confidence intervals for ruin probabilities. Studies in Statistical Quality Control and Reliability 1992 :4. [196] C. 5770. Hogg & S.L. 378389. ACM TOMACS 6. 377381. Appl. Hogan (1986) Comment on corrected diffusion approximations in certain random walk problems. 12981310. Iversen & L. Chalmers University of Technology and the University of Goteborg. [194] 0. Mandl & M. [199] R. Ann. Karlsruhe.V. 19. Act. Act. Hoglund (1991). Math.L. Wahrscheinlichkeitsth. Horvath & E.B. Probab. Hoglund (1974) Central limit theorems and statistical inference for Markov chains. S. Scand. Geb. Iglehart (1969) Diffusion approximations in collective risk theory. [208] V.M.A. 1990 . [192] U. Haggstrom. eds. 23. 166180. 8095.
Oxford. Soc. Insurance: Mathematics and Economics 5. Amer. [227] J. N.C. Kalashnikov (1999) Bounds for ruin probabilities in the presence of large claims and their comparison. Keilson & D. Cambridge Philos. [215] J. Kalashnikov (1997) Geometric Sums: Bounds for Rare Event with Applications. Soc. Wishart (1964). 165167. Keilson & D. [220] V. Probab. [211] J. [213] P. 711743. Reinhard (1985) Probabilities de ruine pour une classe de modeles de risque semiMarkoviens. 561563. 547567. Kemperman (1961) The Passage Problem for a Markov Chain. 173190. Stochastic Models 5. 187193. 325247. Keilson & D.M.G. Taaffe (1989/90) Matching moments to phase distributions. ibid. [210] D. Ann. [216] M. Munksgaard. Goovaerts (1986) General bound on ruin probabilities. N. Keilson (1966) A limit theorem for passage times in ergodic regenerative processes. [214] A.B. 60. Janssen (1980) Some transient results on the M/SM/1 special semiMarkov model in risk and queueing theories. [222] S.M. NEC Research Institute.J. Janssen & J. Lazar (1998) Subexponential asymptotics of a Markovmodulated random walk with a queueing application.G.J. 6. Taylor (1981) A Second Course in Stochastic Processes. [225] J. Proc. Stochastic Models 1.374 BIBLIOGRAPHY [209] D. Copenhagen. Scand. Cambridge Philos. 259281. Jensen (1953) A Distribution Model Applicable to Economics. Kao (1988) Computing the phasetype renewal and related functions. [223] J. Kaas & M. Princeton. 118. 31. Johnson & M. [226] J. Karlin & H. University of Chicago Press. Astin Bull. 116129. Cambridge Philos. Appl. Act. Kluwer.M. 283306. Technometrics 30. Academic Press.R. Proc. ibid.M. 8793. J. Math. 37. Wishart (1964) Boundary problems for additive processes defined on a finite Markov chain. 61.A. 239256. [219] V. 1996.L. [228] J. 35. Appl. [218] V. 11. J. [224] J. 4151. Jagerman (1991) Analytical and numerical solution of Volterra integral equations with applications to queues. Soc. 63. . J. [217] R.M. [221] E. Act. J. 3. 6. Manuscript. Proc. Kalashnikov (1996) Twosided bounds of ruin probabilities. Kennedy (1994) Understanding the WienerHopf factorization for the simple random walk. 123133. [212] J. Statist. Jagerman (1985) Certain Volterra integral equations arising in queueing. Addenda to for processes defined on a finite Markov chain. Wishart (1964) A central limit theorem for processes defined on a finite Markov chain. 866870. Jelenkovic & A.H. Clarendon Press. Probab. Astin Bull.P. 15.G. Jensen (1995) Saddle Point Approximations. New York. Chicago.
C. SpringerVerlag. Lindley (1952) The theory of queues with a single server. Act. [245] T. J. 1998 . Lipsky (1992) Queueing Theory .BIBLIOGRAPHY 375 [229] J. Cybernetica 4. Macmillan. Insurance: Mathematics and Economics 8. Soc. Kingman (1961) A convexity property of positive matrices . [242] A. I. Nyrhinen (1992a) Simulating levelcrossing probabilities by importance sampling. Sorensen (1997) Exponential Families of Stochastic Processes. J. 889900. Lecture Notes in Mathematics 511. Philos. 25. Lehtonen & H. Klnppelberg (1993) Asymptotic ordering of risks and ruin probabilities. J.F. 6075. Camb. 125147. Kliippelberg (1988) Subexponential distributions and integrated tails. Probab. Latouche & V. 154168. Kuchler & M. 60 . Mikosch (1995) Modelling delay in claim settlement. 48.. Klnppelberg (1989) Estimation of ruin probabilities by means of hazard rates. Stochastic Models 7. Scand. B24. Seminaire de Probabilties X . 4958. Stadtmiiller (1998) Ruin probabilities in the presence of heavytails and interest rates . [238] V. [248] D. 132141. Lehtonen & H.F. Kingman (1964) A martingale inequality in the theory of queues. Bernoulli 1 . Kunita (1976) Absolute continuity of Markov processes. New York. Scand. 24. Insurance : Mathematics and Economics 12. 133135 (in Russian). Lemoine (1981) On queues with periodic Poisson input. Scand. (235] C.C. 390397. Appl. Korolyuk. Klnppelberg & T. Appl. [246] D. Kingman (1962) On queues in heavy traffic. Probab. Statist. [237] C.F. [231] J. Adv. Quart. J. Philos. 279285. 283284. [244] T. Lucantoni (1991) New results on the single server queue with a batch Markovian arrival process. Actuarial J. Kliippelberg & U. 259264. 18. Penev & A. [247] L. [241] G. 359361. Mikosch (1995) Explosive Poisson shot noise with applications to risk retention. Lemoine (1989) Waiting time and workload in queues with periodic Poisson input. J. Probab. [239] H. Math.J. 4477. Nyrhinen (1992b) On asymptotically efficient simulation of ruin probabilities in a Markovian environment. Act.a Linear Algebraic Approach. 277289. SpringerVerlag. Ramaswami (1999) Introduction to MatrixAnalytic Methods in Stochastic Modelling. [243] A. Turbin (1973) An asymptotic expansion for the absorption time of a Markov chain distribution. Prob. 1995 . Proc. SIAM.J. Roy. Cambr. Appl. Proc.F.C. [232] C. [234] C. Klnppelberg & T. 858874. 146.P. J. Quart. [233] C. Appl.S. 383392. [230] J. Soc. [240] U. [236] C. . Soc. 26. Oxford 12. J.
S. 161174.S. Moustakides ( 1999) Extension of Wald 's first lemma to Markov processes. with an application . Michna ( 1998) Selfsimilar processes in collective risk theory. Miyazawa & V. Nagaev (1957) Some limit theorems for stationary Markov chains. Hoist eds. Stoch. 29 . J. Neuts ( 1990) A single server queue with server vacations and a class of nonrenewal arrival processes . a useful concept in risk theory. J. 378406. Probab. Statist. Appl. Nummelin (1987) Markov additive processes I. Th. Soc. Appl. 834841. Almqvist & Wiksell. Act. Opns.F.V. 32. Lucantoni . Anal. Adv. [265] S. Act. Eigenvalue properties and limit theorems . J. [252] A. Meier ( 1984). [266] P. 2. 2939. Insurance : Mathematics and Economics 5. 36.376 BIBLIOGRAPHY [249] D.D. [256] A. Ney & E. 58 .M. Res. J. [260] H.K. . 286298. 1261 (0?)1270. Probab. J. Appl. Miller (1962 ) A matrix factorization problem in the theory of random variables defined on a finite Markov chain . Proc. 1994. MartinL3f (1986) Entropy. K. Act. [257] A. [258] K . [259] Z . Europ. [254] V . 561592. Stockholm. .V. 676705. Makowski ( 1994) On an elementary characterization of the increasing convex order . Malinovskii ( 1994) Corrected normal approximation for the probability of ruin within finite time. MeierHellstern & M . Mammitzsch ( 1986 ) A note on the adjustment coefficient in ruin theory. Englunds Boktryckeri AB. Probability and Mathematical Statistics (A. 268285. Probab. Cambridge Philos. 31. Proc. J.D. Probab . Ann. 1986 . [262] H . 429448. 1996 . Lundberg (1926) Forsdkringsteknisk Riskutjdmning. II Aterforsdkring av Kollektivrisker . Miller ( 1962 ) Absorption probabilities for sums of random variables defined on a finite Markov chain . 124147. Math. MartinL6f (1983) Entropy estimates for ruin probabilities . Appl.). 15. [253] V. Malinovskii ( 1996) Approximation and upper bounds on probabilities of large deviations of ruin within finite time. Cambridge Philos. F. Miller ( 1961 ) A convexity property in the theory of random variables defined on a finite Markov chain. Probab. [263] M . 223235. Math. Uppsala. [250] F. Scand. 22. [251] F. Schmidt ( 1993) On ladder height distributions of general risk processes . [261] H. J. Scand. 147149. Ann. 4859. Appl. Probab. [255] V. Gut & J. Scand. 3. 58.D. Soc. Lundberg (1903) I Approximerad Framstdllning av Sannolikhetsfunktionen. 370377. 11 . [264] G. 763776.K. A fitting algorithm for Markovmodulated Poisson processes having two arrival rates . Ann. Appl.
Gjessing (1997a) Optimal choice of dividend barriers for a risk process with stochastic return on investments. Adv. Proc. [271] M. [274] H . Paulsen (1998) Ruin theory with compounding assets . J. 46. [281] J.BIBLIOGRAPHY 377 [267] M .K. J. [276] C. Insurance: Mathematics and Economics 20. Math. [275] H. [283] J. IEICE Trans. Statistics 21. Neuts ( 1981 ) MatrixGeometric Solutions in Stochastic Models. Stoch. Johns Hopkins University Press. Pakes (1975) On the tail of waiting time distributions. 21. Proc. Omey & E. Marcel Dekker. [278] E. Omey & E. [280] A. Gjessing (1997b) Present value distributions with applications to ruin theory and stochastic equations.a survey. [279] E. Paulsen & H. 733746. New York. Nyrhinen (1999) Large deviations for the time of ruin. [269] M . [285] J. 12551265. O'Cinneide (1990) Characterization of phasetype distributions. Probab.G. Neuts ( 1978) Renewal processes of phase type. 157. Paulsen & H. E75B . Abh. Stoch. 16. 339353. [270] M . Appl. Neuts (1992) Models based on the Markovian arrival process. [284] J. [277] J. Proc. Appl. Baltimore . 555564. Norberg ( 1990) Risk theory and its statistics environment . Naval Research Logistics Quarterly 25. Appl. Paulsen ( 1998) Sharp conditions for certain ruin in a risk process with stochastic return on investments.K. Paulsen (1993) Risk theory in a stochastic economic environment. Commun. [272] J . 10081026. Appl.A. Probab. 12. Neveu ( 1961 ) Une generalisation des processus a accroisances independantes. [268] M . Astin Bulletin 5 . Stochastic Models 3 . 445454. 71. Neuts ( 1989) Structured Stochastic Matrices of the M/G/1 Type and their Applications. 75. Insurance: Mathematics and Economics 22. 273299. [282] J . 764779. Stoch. Appl. 327361. 249266. [273] R.F.F. Appl. Probab. 30. Appl. Sem. Appl. Willekens (1987) Second order behaviour of distributions subordinate to a distribution with finite mean . 135148. Proc. . Neuts (1977) A versatile Markovian point process. Ohlin (1969) On a class of measures for dispersion with application to optimal insurance.F. Probab. Models 6. Hamburg.F. London. 123144. Willekens (1986) Second order behaviour of the tail of a subordinated probability distributions. J. Nyrhinen ( 1998 ) Rough descriptions of ruin for a general class of surplus processes. Stoch. 36. Stoch. 316.F. 215223. 311342.
Act. Math.G. Prabhu (1980) Stochastic Storage Processes. Pitts.U. Scand. [305] C.J. [295] N. J. 568576. Philipson (1968) A review of the collective theory of risk. Opns.G. Adv. 30. 465483. New York. [297] R.M. Rolski (1987) Approximation of periodic queues. Rogers (1994) Fluid models in queueing theory and WienerHopf factorisation of Markov chains. Res. 11. Aktuar. 390413. 215246. Austr. Resnick & G. [294] N. [289] C. Wiley. 12. 235243. Astin Bulletin XIV. 820827.K. 28. Res. de Smit (1986) The queue M/G/1 with Markovmodulated arrivals and services.L. Ramsay (1984) The asymptotic ruin problem when the healthy and sick periods form an alternating renewal process.J. 29A. Ann. Prabhu (1961).K. 32. Appl. Probab. [304] B. Ann. Pyke (1959) The supremum and infimum of the Poisson process. Gjessing (1997c) Ruin theory with stochastic return on investments. [303] S. Paulsen & H. [299] C. Oper. [291] S. Prabhu & Zhu (1989) Markovmodulated queueing systems. Probab . J. Insurance: Mathematics and Economics 16.. Skand.U. 147159. Math. 117133. 29. Insurance Risk. Griibel & P. 19. Adv. 965985. 46. [287] F. [296] N. [292] S.M. Schock Petersen (1989) Calculation of ruin probabilities when the premium depends on the current reserve. Ripley (1987) Stochastic Simulation .378 BIBLIOGRAPHY [286] J. [301] G. Heidelberg. [290] E. [300] C. Appl. Queues. Probab. Wiley. Adv.U. Regterschot & J. Math. [298] V. 139143. 4. R. . 537555. Probab. 2330. Samorodnitsky (1997) Performance degradation in a single server exponential queueing model with long range dependence.H. 691707. [288] S. Pitman (1980) Subexponential distribution functions. Wiley. Math. 1989 . and Dams. Prabhu (1965) Queues and Inventories. On the ruin problem of collective risk theory. [302] J. Ann. Pellerey (1995) On the preservation of some orderings of risks under convolution. Appl.M. Math. Appl. Insurance: Mathematics and Economics 3. Statist. Rao (1965 ) Linear Statistical Inference and Its Applications. Appl. 2343. J. Soc. Queueing Systems 5. 757764. New York. Berlin. [306] T. 222261. [293] N. Ramaswami (1980) The N/G/1 queue and its detailed analysis. Stat. Reinhard (1984) On a class of semiMarkov risk models obtained as classical risk models in a markovian environment. Tidskr.A. Inst. 4568. Statist. 61. Probab. Springer. Pitts (1994) Nonparametric estimation of compound distributions with applications in insurance. Ann. Embrechts (1996) Confidence bounds for the adjustment coefficient.R. 337347.U.
J. Appi. Schmidli. . Probab. [311] R. Statist. Th.Y. 145165. 5. Rubinstein & A. 7. 1997. Shapiro (1993) Discrete Event Systems: Sensitivity Analysis and Stochastic Optimization via the Score Function Method. [318] H. J. [314] T. [310] R. [320] H. 4068. Stochastic Models 6. [319] H. Rubinstein (1981) Simulation and the Monte Carlo Method. 795829. Schmidt & J. Insurance: Mathematics and Economics 16. Ross (1974) Bounds on the delay distribution in GI/G/1 queues. Appl. Wiley. Insurance: Mathematics and Economics 22. H. Appl. 21. V. 262286. 155188.L. Act. 417421. 687699. [313] M. [315] T. Schlegel ( 1998) Ruin probabilities in perturbed risk models. 135149. Ann. [325] H . Melamed (1998) Modern Simulation and Modelling. [324] H . [322] H. Adv. J. Scand. Seal ( 1969) The Stochastic Theory of a Risk Business. Statist. Proc. Schmickler ( 1992 ) MEDA : mixed Erlang distributions as phasetype representations of empirical distribution functions . 5. Stochastic Models 10. Schmidli ( 1999a) On the distribution of the surplus prior and at ruin. 365388. 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. 431447 [316] S. 4857. Schmidli ( 1997b) An extension to the renewal theorem and an application in risk theory. Science Culture Technology Publishing . 227244. Operationsforsch. Astin Bulletin 29. Rolski. Wiley. Ryden (1996) An EM algorithm for estimation in Markovmodulated Poisson processes. [326] H. [312] R. Stoch. Ryden (1994) Parameter estimation for Markov modulated Poisson processes. [317] L. [323] H. Schmidli ( 1996) Martingales and insurance risk.Y. 5. [309] H. Stochastic Models 10 . Schmidli ( 1999b) Perturbed risk processes: a review . Comp.Y. 93104. Rubinstein & B. Act. Math. Scand. 2001 . [321] H. 11. J. Teugels (1999) Stochastic Processes for Insurance and Finance.M. 121133. Schmidli ( 1995 ) CramerLundberg approximations for ruin probabilities of risk processes perturbed by a diffusion . Schmidli ( 2001 ) Optimal proportional reinsurance policies in a dynamic setting . Wiley. Schmidli ( 1994) Diffusion approximations for a risk process with the possibility of borrowing and interest. Wiley. [308] S. Schmidli ( 1997a) Estimation of the Lundberg coefficient for a Markov modulated risk model . Lecture Notes of the 8th Summer School on Probability and Mathematical Statistics ( Varna). Probab . Probab.BIBLIOGRAPHY 379 [307] T. Rudemo (1973) Point processes generated by transitions of a Markov chain. Data Anal. Singapore. 131156. Wiley.
Shwartz & A. 72. [336] B. Scand. 11. Aktuar Tidsskr. 75. . [334] C: O. [335] B . Ann. 1955. 383413. Chapman and Hall. Segerdahl (1959) A survey of results in the collective theory of risk. [340] E. [347] K. Stockholm. Chapman & Hall.380 BIBLIOGRAPHY [327] H. 171178. [328] H. 429442.A. Risk teory and the single server queue . Seber (1984) Multivariate Observations. [333] C: O. Shaked & J. Grenander). Appl. Siegmund (1979) Corrected diffusion approximations in certain random walk problems. Statist.L. Math. New York.L. Ann.S. [330] H. Verein Schweiz. Mitt. Segerdahl (1955) When does ruin occur in the collective theory of risk? Skand.S. 11. Queueing Systems 11. Th. t). 1974. Math. Aktuar Tidsskr. Skand. . Probab . Wiley. Probab. 61. [329] H. Verein Schweiz Versich. [338] E. 914924. Seal (1974) The numerical calculation of U(w. Sigman ( 1994 ) Stationary Marked Point Processes: An Intuitive Approach. 191197. [331] G. 4. 4. Act. Mitteil. Adv. 121139. 2236. Stochastic Models 6. Seal ( 1972). 4383.F. Sigman (1992) Light traffic for workload in queues. L. 21. 279299. [344] D. [346) D . 673684. Act. 77100. Appl. Siegmund (1975) The time until ruin in collective risk theory. [337] M.the Harald Cramer volume (ed. Scand. Sengupta ( 1989) Markov processes whose steadystate distribution is matrixgeometric with an application to the GI/PH/1 queue. the probability of nonruin in an interval (0. 1987. [332] C : O. Segerdahl ( 1942) Uber einige Risikotheoretische Fagestellungen .L. 157166. Probab. Versich. Almqvist & Wiksell. Siegmund (1976) Importance sampling in the Monte Carlo study of sequential tests. [343] D. In Probability and statistics . t). 243249. Shantikumar (1993) Stochastic Orders and Their Applications. Academic Press.S. 72. Versich. Verein Schweiz. J. J. Weiss (1995) Large Deviations for Performance Analysis. Insurance: Mathematics and Economics 8. Sengupta (1990) The semiMarkov queue: theory and applications. [348] K . pp. Seal (1978) Survival Probabilities. Math. Siegmund (1976) The equivalence of absorbing and reflecting barrier problems for stochastically monotone Markov processes .W.W. Siegmund ( 1985) Sequential Analysis . Seal (1972) Numerical calculcation of the probability of ruin in the Poisson/Exponential case. [339] A. SpringerVerlag. Shtatland ( 1966) On the distribution of the maximum of a process with independent increments.G. 159180. [342] D. [345] D. Appl. 483487. Probab. Adv. 701719. Shin (1987) Convolution of uniform distributions and ruin probability. Shin (1989) Ruin probability by operational calculus. [341] E. Wiley. Mitt.
5776. [353] E. Karlsruhe. Hoesman ( 1989) Moderate. [352] D. 1974. Astin Bulletin 12.). Soc 49. Adv. Thorin (1977). Jewell ( 1981 ) Further results on recursive evaluation of compound distributions . Scand. Insurance: Mathematics and Economics 16. 1976. Thorin (1982) Probabilities of ruin.J. 65102. Sundt & J. 722. 149162. Teugels (1997) The adjustment coefficient in ruin estimates under interest force . Insurance: Mathematics and Economics 1. [362] G. Suri ( 1989) Perturbation analysis : the state of the art and research issues explained via the GI/G/1 queue. Straub (1988) NonLife Insurance Mathematics. Taylor (1979) Probability of ruin under inflationary conditions or under experience ratings .V.C. J. [364] G. New York. Probab. Verlag Versicherungswirtschaft e. [360] G. Insurance: Mathematics and Economics 19. Taylor (1986) Claims reserving in NonLife Insurance. J.C. [351] D . Cambridge Philos.. 1978. Appl.). 114137. Stoyan (1983) Comparison Methods for Queues and Other Stochastic Models (D. 1980 . Slud & C. [350] W. Sundt ( 1993) An Introduction to NonLife Insurance Mathematics (3rd ed. 8594. [366] O. [358] R. Act. [365] J. J. Act. Scand. John Wiley & Sons. NorthHolland. [368] O. [363] G. 725741. [355] B . Daley ed. Astin Bulletin 16 . Sundt & J. 449461. J. Teugels (1982) Estimation of ruin probabilities . 197208.J.L. Proceeedings of the IEEE 77.L. Scand. Proc. Act. [356] B . Taylor (1979) Probability of ruin with variable premium rate. 8199. Teugels ( 1995 ) Ruin estimates under interest force .. Stanford & K.and largedeviations probabilities in actuarial risk theory. New York. Sundt & W. Scand. 1982 . [361] G.A. Actuarial J. 65102. Thorin (1986) Ruin probabilities when the claim amounts are gamma distributed. Taylor (1978) Representation and explicit calculation of finitetime ruin probabilities. [354] B.S. Springer. 235254. Taylor (1976 ) Use of differential and integral inequalities to bound ruin and queueing probabilities . Astin Bulletin 24. 163175. . [359] L.BIBLIOGRAPHY 381 [349] E. Unpublished manuscript. Scand. [357] B. 21. [369] O. Wiley. J. Stroinski ( 1994) Recursive method for computing finitetime ruin probabilities for phasedistributed claim sizes.L.C. Act.C. 2739. Smith ( 1953) Distribution of queueing times.L. Scand. Takhcs (1967) Combinatorial Methods in the Theory of Stochastic Processes. Ruin probabilities prepared for numerical calculations.C. Thorin (1974) On the asymptotic behaviour of the ruin probability when the epochs of claims form a renewal process. 118. Act. [367] O.
Haezendonck eds. Unpublished Ph. J.382 BIBLIOGRAPHY [370] 0. .). Oper. J. [375] N. [371] 0. Tidskr. Astin Bulletin IX. 31. Willmot (1994) Refinements and distributional generalizations of Lundberg's inequality. 37. Wolff (1990) Stochastic Modeling and the Theory of Queues. Taqqu. Wilson (1995) Selfsimilarity in highspeed traffic: analysis and modeling of ethernet traffic measurements. PrenticeHall. Submitted. Thorin & N. R. [378] H. [376] A. Skand. VazquezAbad (1999) RPA pathwise derivatives of ruin probabilities. Aktuar. Goovaerts (1983) The influence of reinsurance limits on infinite time ruin probabilities. Probab. Res. SpringerVerlag. M. 1983 . 137153. Wikstad (1977) Numerical evaluation of ruin probabilities for a finite period. Tacklind (1942) Sur le risque dans les jeux inequitables. Insurance: Mathematics and Economics 15. [377] V. Veraverbeke (1993) Asymptotic estimates for the probability of ruin in a Poisson model with diffusion. Wikstad (1977) Calculation of ruin probabilities when the claim distribution is lognormal. Wallace (1969) The solution of quasi birth and death processes arising from multiple access computer systems. De Vylder. F. In: Premium Calculation in Insurance (F. 1942. Waters (1983) Probability of ruin for a risk process with claims cost inflation. 231246. [380] W. Wiley. Stationarity and Regeneration. Astin Bulletin VII. [382] G. 148164. [372] H.R. [381] W. 4963. M. [383] G. University of Michigan. Lin (1994) Lundberg bounds on the tails of compound distributions. Reidel. Whitt (1989) An interpolation approximation for the mean workload in a GI/G/1 queue. Willinger. Willmot & X. Appl. D. [379] M.E. Wald (1947) Sequential Analysis. Act. Goovaerts.W. Sherman & D. Insurance : Mathematics and Economics 13. 6785. 5762. Van Wouve. [384] R. thesis. Statistical Science 10. J. [374] F. Dordrecht Boston Lancaster. Thorisson (2000) Coupling. De Vylder & M. [373] S. Thorin & N. 743756. Scand. 142.E. 936952.
37. 111117. 7879.6779.301 central limit theorem 60 . 170173. 278 gamma distribution 67. 9496. 135. 226. 3436. 196201 inverse Gaussian distribution 76. 86. 271274. 15.135. 205. 17.160167.285292. 91. 1112. 180182.228229.150.269. 301 Kronecker product.203.137141. 4851.217. 141144.307312 compound Poisson model 4. 119. 7079.318320 change of measure 2630. 207 heavytailed distribution 6. 248 WienerHopf 144 interest rate 190. 217.86. 97129. 201214. 8283 hyperexponential distribution 7.121129. 117128. 3032. 110113.249250 integral equation 16 Lindley 143 renewal 64. 89. 302303 diffusion approximation 17.182. 2526. 39. 162164.293294.and sum 221.178184. 138139. 5796.328330. 7179.287292. 17. 360 excursion 155156.299. 1415. 316323 Bessel function 102.251280 heavy traffic 76. 332333 Volterra 192194. 14. 9293. 7475. 318319 Erlang distribution 7. 341.242. 80 81.185187.226.100.308 CramerLundberg model: see compound Poisson model cumulative process 334 dams: see storage process differential equation 16.Index adjustment coefficient 17. 40. 117127 corrected 121127 duality 1314.200201. 5.4447. 189. 201 Brownian motion 3 . 323 Coxian distribution 147. 308.9899. 239. 218 Cox process 4.272.281.203. 79. 12 CramerLundberg approximation 1617.359 aggregate claims 103106.292293 Edgeworth expansion 113. 245248. 97. 227229. 2425. 1819.259261. 9396. 283. 3334.346349 383 .314316. 122. 361 diffusion 3.249. 3839.
287291 INDEX matrix equation . 25. 113114.336339 Laplace transform 15. 149. 229 M/M/1 101 Markovmodulated 185187 periodic 187 martingale 2426. 267269 Panjer's recursion 320323 Pareto distribution 910. 154. 15.108. 7179. 230.285287 queue 14 . 5758.269271.234. 295. 38. 71.227230. 257. 171. 175 light traffic 8183 Lindley integral equation 143 process 3334. 6970.161. 41. 144.298299. 42. 16. 176185. 14. 245 M/G/1 13. 179 NP approximation 318320 Palm distribution 5253.348 terminating 215216.238.315 inequality 1718.160161. 227228. 203204. 44. 100. see also sensitivity analysis phasetype distribution 8. 304305 random walk 3336. 261264. nonlinear 155. 269 PerronFrobenius theory 4142. 9899.178182. 25. 2730.161. 142 likelihood ratio : see change of measure lognormal distribution 9. 157. 96. 106108. 108109.287.180. 80. 137139. 133.201. 3947.339 large deviations 129. 271274.218221. 234 matrixexponential distribution 240244 matrixexponentials 14. 16.261264. 185187 GI/G/1 141144 M/D/1 6667 equation 16. 162. 37. 133.302. 52 53. 112113.350361 Poisson process Markovmodulated 12 periodic 12. 32. 99. 3947.297299.152160. 3639. 108 life insurance 5.240244. 65.234240. 138.340350 multiplicative functional 2830. 39. 251. 132133. 178 modulation 12. 35. 4446.336339 . 134.349 350 perturbation 172173. 203 Markov additive process 12. 134135.146148.174. 35. 260 Lundberg conjugation 6979 . 306316 Levy process 3.384 ladder heights 4756.123.139141. 6162. 38.134135.259261. 86 periodicity 12.148. 59. 145187. 44. 141144. 7576.128129.275278.161164. 176185 nonhomogeneous 60 PollaczeckKhinchine formula 6167. 213214.304 process 2830.288290.215250.
87. see also matrixexponential distribution regenerative process 264 268. 332333 model 12. 9693.262263. 261264 reservedependent premiums 14. 152. 186187 renewal process 131. 280. 160. 335336 sensitivity analysis 8693.279280 Rouche roots 158. 147.336339 workload 13.244250. 168172 storage process 13. 177 timereversion 14. 31. 294296 shotnoise process 314 simulation 19. 223226. 253.INDEX 385 waiting time 141. 244. 338 utility 324. 326330 Weibull distribution 9. 331336 equation 64. 307308. 123. 37. 229234. 251280 time change 4.314. 8386. 317318 semiMarkov 147. 120 statistics x. 222.154157. 213. 238 saddlepoint method 115117. 1819. 141144. 60. 5455. 186187 virtual: see workload rational Laplace transform 8. 333334 regular variation 10. 11. 146. 3032. 292294. 251. 233234. 279280 subexponential distribution 11. 89. 12. 281296 stable process 15. 256258. 240. 260 reinsurance 8. 7475. 260 WienerHopf theory 144. 189214. 172173. 257. 233. 131144. 327 . 4950. 174. 107.186. 191192. 162. 251.273274.359361 stochastic control x stochastic ordering 18.
P'i yfliother approximations (e. y finite horizon ruin probabilities.Vol. Some i (l I JL I J r of the topics are Lundberg's inequality." Short Book Reviews ISBN 9810222939 mi u inn i nun I I I I I I i in u www..Advanced Series on Statistical Science & Applied Probability .com 2779 he 9 "789810ll22293211 .. "This book is a must for anybody working in applied probability. worldscientific. It is a comprehensive treatment of the known results on ruin probabilities. Special features of the book are the emphasis on change of measure techniques. phasetype distributions as a computational vehicle and the connection to other applied probability areas like queueing theory. the ^W A l \ i l ' ''' CramerLundberg approximation. Markovmodulation or periodicity. 2 A I 11 JjVb l' i  i Yj .T [Ail i The book is a comprehensive treatment of  I i I \ classical and modern ruin probability theory. exact solutions.g. extensions of the classical compound Poisson model to allow f o r reservedependent premiums. I 1! Ruin Probabilities . for heavytailed claim size distributions).
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.