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 . . . . . . . . . . . . . . . . . . ... . . 281 2 Simulation via the PollaczeckKhinchine formula . . . . . . . . . . . . . 279 X Simulation methodology 281 1 Generalities . . . . . . .. . . . . . . . . . 350 Bibliography Index 363 383 . . 304 3 Large deviations . . . . . 261 4 Models with dependent input . . . . . . . 290 5 Regenerative simulation . . . . . . . . . . . . . .. . . . . . . . . . . . . . . 292 6 Sensitivity analysis . . . . . . . . . . . . . . . . . . . . . . . 326 Appendix 331 Al Renewal theory . . . . . . . . . . 331 A2 WienerHopf factorization . . . . . . .. . . . . . . 340 A4 Some linear algebra . . . . . . . . . . . . . .. . . . . . . . . .. . . . . . . . . 264 5 Finitehorizon ruin probabilities . . . . . .CONTENTS vii IX Ruin probabilities in the presence of heavy tails 251 1 Subexponential distributions . . . . .. . . 285 3 Importance sampling via Lundberg conjugation . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . 344 AS Complements on phasetype distributions . . . . . 306 4 The distribution of the aggregate claims . . . . . . . . . . .. . . . 259 3 The renewal model . . . . . . . . . . . 251 2 The compound Poisson model . . . . . . ... . . . . .. . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . 297 2 Further applications of martingales . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . .. . . . . . .. . 323 6 Reinsurance . . . . 294 XI Miscellaneous topics 297 1 The ruin problem for Bernoulli random walk and Brownian motion. . 316 5 Principles for premium calculation .. . . . . . . . . . . . . . The twobarrier ruin problem . 287 4 Importance sampling for the finite horizon case . . . . .. . . . . .. . . . . . .. . . . . . . . .. . .. . . . . . . . 271 6 Reservedependent premiums . . . . . . .
This page is intentionally left blank .
But the pace was much slower than expected. if the formulations occasionally give a different impression. and the series editor Ole BarndorffNielsen for their patience. As an excuse: many of these projects were related to the book. it would not be fair not to say that the practical relevance of the area has been questioned repeatedly. 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!). However. In particular. One reason for writing this book is a feeling that the area has in the recent years achieved a considerable mathematical maturity. and my belief was that this could be done rather quickly. and other projects absorbed my interest. Risk theory in general and ruin probablities in particular is traditionally considered as part of insurance mathematics. I have deliberately stayed away from discussing the practical relevance of the theory. Let me take this opportunity to thank above all my publisher World Scientific Publishing Co. 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. Apart from these remarks. Since I was to produce some handouts for the students anyway. the idea was close to expand these to a short book on the subject. It has obviously not been possible to cover all subareas. that it can only say something about very simple models and questions. this applies to longrange dependence which is intensely studied in the neighboring ix . The course was never realized. University of Copenhagen. and has been an active area of research from the days of Lundberg all the way up to today. Thus. which has in particular removed one of the standard criticisms of the area. I was invited to give a course on ruin probabilities at the Laboratory of Insurance Mathematics. the book is basically mathematical in its flavour. and the result is now that the book is much more related to my own research than the initial outline.Preface The most important to say about the history of this book is: it took too long time to write it! In 1991.
3. Asmussen. http:// www. IV. Chapters IXX then go in more depth with some of the special approaches for analyzing specific models and add a number of results on the models in Chapters IIIVII (also Chapter II is essentially methodological in its flavor). Hojgaard & Taksar [206]. the standard stochastic control setting of diffusion models has been considered. see in particular Michna [259].g. see also Schmidli [325] and the references in Asmussen & Taksar [52]. another by method.2. VI. VII.g. an area which is becoming increasingly important.1.4a. I intend to keep a list of misprints and remarks posted on my web page. for the effects on tail probabilities. 111. In the classical setting of CramerLundberg models. More recently.13.se Lund February 2000 Soren Asmussen . Concerning ruin probabilities. Willinger et al. The main motivation comes from statistical data for network traffic (e. [381]).13. Hojgaard & Taksar [35] and Paulsen & Gjessing [284].2 more properly). The book does not go into the broader aspects of the interface between insurance mathematics and mathematical finance.g. VII. Another interesting area which is not covered is dynamic control. incorporate 11.6 (to understand the PollaczeckKhinchine formula in 111.maths . IV. The present book is in between these two possibilities.2.se/matstat / staff/asmus and I am therefore grateful to get relevant material sent by email to asmusfmaths . Here is a suggestion on how to get started with the book. One is by model.13 and IX.5.lth. e.45. The rest is up to your specific interests. see e.15.13 and XI. Chapters IIIVII introduce some of the main models and give a first derivation of some of their properties. Finally. Resnick & Samorodnitsky [303] and references therein.89.lth.x PREFACE field of queueing theory. For a brief orientation. A book like this can be organized in many ways. IV. it has not been possible to incorporate more numerical examples than the few there are. some basic discussion can be found in the books by Biihlmann [82] and Gerber [157]. Good luck! I have tried to be fairly exhaustive in citing references close to the text. the first part of 11. I regret that due to time constraints. VIII. In addition. It is obvious that such a system involves a number of inconsistencies and omissions. X. for which I apologize to the reader and the authors of the many papers who ought to have been on the list. read Chapter I. IX. 111.14. some papers not cited in the text but judged to be of interest are included in the Bibliography. For a second reading.
5 from Asmussen [21] with permission from CRC Press. 5. Fig. Aarhus. Parts of II. More substantial remarks.6.2 by Rafal Kulik . many of which were pointed out by Hanspeter Schmidli .6 is reprinted from Asmussen & Schmidt [49] and parts of IX.1 and X. Fig.1 by Bjarne Hojgaard and the table in Example 111.8 . 1 is almost identical to Section 2 of Asmussen [26] and reprinted with permission of Blackwell Publishers.6 by my 1999 simulation class in Lund. 5 from Asmussen & Kliippelberg [36] with the permission from Elsevier Science . 111 . Section VII .3 are reprinted from Asmussen & Rubinstein [46] and parts of VIII. Schmidli & Schmidt [47] with the permission from Applied Probability Trust . 3 is reprinted from Asmussen & Nielsen [39] and parts of IX.4 from Asmussen. Parts of X. of which there are not many at this stage . Section VIII. Lund September 2001 Soren Asmussen Acknowledgements Many of the figures . not least the more complicated ones. IV. A number of other figures were supplied by Christian Geisler Asmussen . as well as some additional references continue to be at the web page.PREFACE xi The second printing differs from the first only by minor corrections. . were produced by Lone Juul Hansen . supported by Center for Mathematical Physics and Stochastics (MaPhySto).
This page is intentionally left blank .
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 . results and topics to be studied in the rest of the book.2) (O<t<T Ro=ul. For mathematical purposes.Rt.i(u.1) We also refer to t/) ( u) and 0(u. They are the main topics of study of the present book. and give a very brief summary of some of the models. we introduce some general notation and terminology. The probability O(u) of ultimate ruin is the probability that the reserve ever drops below zero. as defined in broad terms . is a model for the time evolution of the reserves of an insurance company. Letting T(u) = inf {t > 0 : Rt < 0} = inf It > 0 : St > u}. MT = sup St. (1.T) = P inf Rt < 0 I . We denote throughout the initial reserve by u = Ro. M = (1. it is frequently more convenient to work with the claim surplus process {St}t>0 defined by St = u . respectively.Chapter I Introduction 1 The risk process In this chapter . (1. A risk reserve process { Rt}t>o. (1. T) as ruin probabilities with infinite horizon and finite horizon .3) sup St.
(1. Thus.T) = F (MT > u) = P(r(u) < T). t] is finite. . However. We denote the interarrival times of claims by T2. That is.1 .pt.1. and Nt = min {n > 0 : 0rn+1 > t} = max {n > 0: Un < t}• The size of the nth claim is denoted by Un. and T1 is the time of the first claim. (1.2 CHAPTER I. say. Putting things together. INTRODUCTION be the time to ruin and the maxima with infinite and finite horizon. the time of arrival of the nth claim is an = T1 + • • • + Tn. we see that Nt Nt Rt = u + pt . the ruin probabilities can then alternatively be written as .b(u) = P (r(u) < oo) = P(M > u). T3..7) k=1 k=1 The sample paths of {Rt} and {St} and the connection between the two processes are illustrated in Fig.6) Sofar we have not imposed any assumptions on the risk reserve process. the number Nt of arrivals in [0.i(u. per unit time. the following setup will cover the vast majority of the book: • There are only finitely many claims in finite time intervals. (1. respectively. Figure 1.. St = E Uk .E Uk.5) i. 1. • Premiums flow in at rate p.
If 77 > 0.(.1 Assume that (1. (1. not discuss whether this actually corresponds to practice. t * oo.e.b(u) = 1 for all u. but as an approximation to the risk process rather than as a model of intrinsic merit. one may well argue that Brownian motion in itself could be a reasonable model. We shall not deal with this case either. allowing a countable infinity of jumps on Fig. Thus. however. It would appear obvious. VII. though many results are straightforward to generalize from the compound Poisson model. rl= pP P It is sometimes stated in the theoretical literature that the typical values of the safety loading 77 are relatively small. • Brownian motion or more general diffusions. We study this case in Ch. For the purpose of studying ruin probabilities this distinction is. then M < oo a..20%. that the insurance company should try to ensure 77 > 0. of course. we shall.1 the slope of {Rt} should depend also on the level).1. for example. say 10% .s. and hence O(u) < 1 for all sufficiently large u. and the basic ruin probabilities are derived in XI. however.1.8) holds. immaterial.s. on Fig. If 77 < 0. However. 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. one could well replace Rt by Rtnr(u) or RtA. a basic references is Gerber [127]. We shall discuss Brownian motion somewhat in Chapter IV. 1. and in fact: Proposition 1. 1.1. and hence . then M = oo a.. Some main examples of models not incorporated in the above setup are: • Models with a premium depending on the reserve (i. since any modeling involves some approximative assumptions. . 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. The models we consider will typically have the property that there exists a constant p such that Nt a E Uk k=1 p.8) The interpretation of p is as the average amount of claim per unit time. • General Levy processes (defined as continuous time processes with stationary independent increments) where the jump component has infinite Levy measure.) V 0.
and here (1. St In concrete models. (1. k=1 (1.10) Again.11) . However. 0(u.4 CHAPTER I.10) is a property which we will typically encounter. The simplest concrete example (to be studied in Chapter III) is the compound Poisson model. it is not too difficult to show that p as defined by (1. .Tp). If u oo. U2. where {Nt} is a Poisson process with rate . and independent of {Nt}. 0 We shall only encounter a few instances of a Cox process.. However. and independent of {(0(t). if {(3(t)} is nonergodic. If U1. then this limit is > 0 which implies St a$ oo and hence M = oo a.b(u) < 1 for all u when rl > 0. are i..2 (Cox PROCESSES) Here {Nt} is a Poisson process with random rate /3(t) (say) at time t. U2.8).v. in connection with risk processes in a Markovian or periodic environment (Chapter VI). .8) is given by ^t p = EU • lim it (3(s) ds t. If 77 < 0. and that .d. tb(u) = 1 for all u holds also when rl = 0. not all models considered in the literature have this feature: Example 1. Then the connection between the ruin probabilities for the given risk process {Rt} and those ^(u).Q (say) and U1. rl > 0.6EU (on the average.d.T) = i. then similarly limSt/t < 0. (1. this needs to be verified in each separate case. _ St __ k =1 Uk pt a4.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. we obtain typically a somewhat stronger conclusion. The simplest example is 3(t) = V where V is a r . (1.8) that F N. INTRODUCTION Proof It follows from (1. Nt)}.. .10) hold with p constant. zP(u . namely..s.i(u.T) for {Rt} is given by V)(u) = t/i (u).i.oo t 0 J (provided the limit exists). are i. M < oo a. with the most notable special case being V having a Gamma distribution.i. This case is referred to as the mixed Poisson process. Here it is easy to see that p = . t t p  p' t ^ oo.s. Thus p may well be random for such processes..3 Assume p 54 1 and define Rt = Rt1p.s. corresponding to the Pdlya process. namely that M = oo a. Proposition 1.
CLAIM SIZE DISTRIBUTIONS 5 The proof is trivial. Heilmann [191]. Hipp & Michel [198]. Embrechts et al.g. and we do not get near to the topic anywhere in this book. see also Chapter XI. Mitteilungen der Verein der Schweizerischen Versicherungsmathematiker and the Scandinavian Actuarial Journal. For mixed Poisson processes and Polya processes..g. Since { Rt } has premium rate 1. in a number of models. [330]. the research literature is often published in journals like Astin Bulletin . many results and methods in random walk theory originate from there and the area was ahead of related ones like queueing theory. De Vylder [110]. In the even more general area of nonlife insurance mathematics. [76]. another important early Swedish work is Tacklind [373]. the claim arrivals are Poisson or renewal at the same time). We roughly classify these into two groups . Notes and references The study of ruin probabilities. Sundt [354]. often referred to as collective risk theory or just risk theory. which is feasible since in most cases the process { Rt } has a similar structure as {Rt} (for example. Grandell [171]. Buhlmann [82]. Some main later textbooks are (in alphabetical order) Buhlmann [82]. Taylor [364]. Besides in standard journals in probability and applied probability. Insurance: Mathematics and Economics. The Swedish school was pioneering not only in risk theory. The term risk theory is often interpreted in a broader sense than as just to comprise the study of ruin probabilities. U2. Schmidli. Gerber [159]) has a rather different flavour. see e . lighttailed distributions (sometimes the term . the assumption > 0 is equivalent to p < 1. was largely initiated in Sweden in the first half of the century. Note that when p = 1. An idea of the additional topics and problems one may incorporate under risk theory can be obtained from the survey paper [273] by Norberg. Daykin et al. Note that life insurance (e. some main texts (typically incorporating some ruin theory but emphasizing the topic to a varying degree) are Bowers et al. and in fact p < 1 is the fundamental assumption of queueing theory ensuring steadystate behaviour (existence of a limiting stationary distribution).. the role of the result is to justify to take p = 1. Straub [353].. Schmidt & Teugels [307] and Seal [326]. we shall be able to identify p with the traffic intensity of an associated queue. Rolski. Some of the main general ideas were laid down by Lundberg [250].. Some early surveys are given in Cramer [91]. but in probability and applied probability as a whole. Daykin. in particular. Segerdahl [334] and Philipson [289]. while the first mathematically substantial results were given in Lundberg [251] and Cramer [91].2. [101]. Gerber [157]. the recent survey by Grandell [173] and references therein. Cox processes are treated extensively in Grandell [171]. 2 Claim size distributions This section contains a brief survey of some of the most popular classes of distributions B which have been used to model the claims U1. Pentikainen & Pesonen [101]. [134].
a simple stopping time argument shows that this implies that the conditional distribution of the overshoot ST(u) . and heavytailed distributions. 2a Lighttailed distributions Example 2. The crucial feature is the lack of memory: if X is exponential with rate 6.x given X > x is again exponential with rate b (this is essentially equivalent to the failure rate being constant). In contrast. and can also be interpreted as the (constant) failure rate b(x)/B(x). a fact which turns out to contain considerable information. 6 has density r(p)xPleax b(x) P and m. then the conditional distribution of X . for the compound Poisson model with exponential claim sizes the ruin probability . s<8.6 CHAPTER I. (2. Equivalently. 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'.8.f. B[s] is finite for some s > 0. the exponential distribution is by far the simplest to deal with in risk theory as well. i.3) . P B[s]= (8Is ) . if 1 °O AB Jbos x B(dx) > 0.g.2) = 0. B is heavytailed if b[s] = oo for all s > 0.u at the time of ruin given r(u) is again exponential u with rate 8. INTRODUCTION 'Cramertype conditions' is used). On the more heuristical side. but different more restrictive definitions are often used: subexponential.1 (THE EXPONENTIAL DISTRIBUTION) Here the density is b(x) = beax (2. Here lighttailed means that the tail B(x) = 1 .g.f.1) The parameter 6 is referred to as the rate or the intensity. In particular. For example in the compound Poisson model.2 (THE GAMMA DISTRIBUTION) The gamma distribution with parameters p. where B(bo. Example 2 .2 and /LB is the mean of B.B(x) satisfies B(x) = O(e8x) for some s > 0.e. regularly varying (see below) or even regularly varying with infinite variance. As in a number of other applied probability areas.O(u) can be found in closed form. the m.
one has r(bx. then X v Xl + • • • + X. where X1. u . among others. CLAIM SIZE DISTRIBUTIONS 7 The mean EX is p/b and the variance Var X is p/b2. p). .).) VarX1 (EX )2 p is < 1 for p > 1. is > 1.ate (b2 ): L• i=o In the present text. 0..d..v. if p is integer and X has the gamma distribution p. i = 1. or just the Erlang(p) distribution. This special case is referred to as the Erlang distribution with p stages.c. An important property of the hyperexponential distribution is that its s.y i=1 where >i ai = 1. p) = J tPletdt. B(x) = r(p) Asymptotically. X2. In particular.2..1) (or the 1/pth root if p < 1). 0 < ai < 1. the Gamma density (2. 2. we develop computationally tractable results mainly for the Erlang case (p = 1.v. P b(x) = r` aibiea..1 Poisson events in [0. p. x] so that B(x) = r` e. An appealing feature is its simple connection to the Poisson process: B(x) = P(Xi + • • • + XP > x) is the probability of at most p .. by Grandell & Segerdahl [175] and Thorin [369].i.. JP 1 B(x) r(p ) XP ie ax In the sense of the theory of infinitely divisible distributions. In particular. Ruin probabilities for the general case has been studied. The exact form of the tail B(x) is given by the incomplete Gamma function r(x. and exponential with rate d. p) °° where r (x. > 1 for p < 1 and = 1 for p = 1 (the exponential case).c. the squared coefficient of variation (s. u Example 2 . ..2) can be considered as the pth power of the exponential density (2.3 (THE HYPEREXPONENTIAL DISTRIBUTION) This is defined as a finite mixture of exponential distributions. . are i.
xo)+ is covered by the reinsurer). T) is called the representation.7) are possibly complexvalued but the parameters in (2.(2. which is slightly smaller but more amenable to probabilistic reasoning. 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 . We give a more comprehensive treatment in VIII.7) q1 b(x) = cjxieWWx + djxi cos(ajx)ea'x + > ejxi sin(bjx)e`ix . T) or sometimes the triple (E.8) are realvalued.f. but the current trend in applied probability is to restrict attention to the class of phasetype distributions. q2 q3 (2.6. We give some theory for matrixu exponential distribution in VIII. of which one is absorbing and the rest transient. the restriction T of the intensity matrix of the Markov process to E and the row vector a = (ai)iEE of initial probabilities.8 CHAPTER I. See XI. 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. However.. B(x) > 0 for x < xo) is of course a trivial instance of a lighttailed distribution.. a. INTRODUCTION Example 2 .6. the Erlang and the hyperexponential distributions. This class of distributions is popular in older literature on both risk theory and queues.1 and defer further details to u Chapter VIII.e. The density and c. The parameters of a phasetype distribution is the set E of transient states.6 (DISTRIBUTIONS WITH BOUNDED SUPPORT) This example (i. there exists a xo < oo such that B(x) = 0 for x > xo. B(x) = aeTxe where t = Te and e = (1 . 1)' is the column vector with 1 at all entries. (or. Important special cases are the exponential.5 (DISTRIBUTIONS WITH RATIONAL TRANSFORMS) A distribution B has a rational m. Example 2 . a rational Laplace transform) if B[s] _ p(s)/q(s) with p(s) and q(s) being polynomials of finite degree. Example 2 .d. . Equivalent characterizations are that the density b(x) has one of the forms q b(x) j=1 = cjxienbx. resp.f. it is notable from a practical point of view because of reinsurance: if excessofloss reinsurance has been arranged with retention level xo. are b(x) = aeTxt. The couple (a.4 (PHASETYPE DISTRIBUTIONS) A phasetype distribution is the distribution of the absorption time in a Markov process with finitely many states. equivalently.8) j=1 j=1 j=1 where the parameters in (2.g.
CLAIM SIZE DISTRIBUTIONS 9 2b Heavytailed distributions Example 2. a2). All moments are finite. (2.9 (THE PARETO DISTRIBUTION) Here the essence is that the tail B(x) decreases like a power of x. Example 2 .10) The loinormal distribution has moments of all orders.12) Sometimes also a location parameter a > 0 and a scale parameter A > 0 is allowed. However. we obtain the Weibull distribution B(x) = eCx'. It follows that the density is 't (1ogX . the tail is B (x ) 2 x.2. the mean u is eµ+a /2 and the second moment is e2µ+2o2.1.N(0. p is defined as the distribution of ev where V .9) which is heavytailed when 0 < r < I.8 (THE LOGNORMAL DISTRIBUTION) The lognormal distribution with parameters a2.u l b(x) = d dx or J ax lor 1 exp Asymptotically. the exponential distribution representing the simplest example since here b(x) is constant.pl = 1 W (logx . one being B(x) (1 + X)b(x) (1 + x)a+1' x > 0.13) u . Writing c = d/r.p a 1 (2.11) ex log logx 2r p 1 1 2 ( a ) f 1 (lox_P)2} (2. a)/A)a+1' x > a. x < a.N(p. and then b(x) = 0.1). b(x) = crx''le`xr. u Example 2 . in practice one may observe that b(x) is either decreasing or increasing and may try to model smooth (incerasing or decreasing) deviations from constancy by 6(x) = dx''1 (0 < r < oo). b(x) _ A(1 + (x a The pth moment is finite if and only if p < a . In particular.7 (THE WEIBULL DISTRIBUTION) This distribution originates from reliability theory. (2. (2. or equivalently as the distribution of a°U+µ where U . Here failure rates b(x) = b(x)/B(x) play an important role. There are various variants of the definition around.
in particular.17) where L (x) is slowly varying. in particular. the loggamma distribution (with exponent 5) and a Pareto mixture of exponentials.e. (2.'s of the form YX. another standard example is (log x)').16) 11 Example 2. x 4 oo (any L having a limit in (0.2).v. The simplest examples correspond to p small and integervalued. the density is { 3 (1 . i. Choudhury & Whitt [1] as the class of distributions of r.1)/p. 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). x + 00.12) (here L (x) * 1) and ( 2.15) x2 + 16x3 ) a3x/2) 3 (1 . where Y is Pareto distributed with a = (p . { s () 1s+3s29s3log(1+2s I p=3. Thus.(1 + 2x + 2x2)e2x) p = 2 (2. oo) is slowly varying . the loggamma distribution is a Pareto distribution. A = 1 and X is standard exponential. (2.14) The pth moment is finite if p < 5 and infinite if p > 5.10 (THE LOGGAMMA DISTRIBUTION) The loggamma distribution with parameters p. examples of distributions with regularly varying tails are the Pareto distribution (2. u . In general.11 (PARETO MIXTURES OF EXPONENTIALS) This class was introduced by Abate.13).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) .x6+lr(p) (2.10 CHAPTER I. The density is 8p(log x)pi b(x) .(1 + Zx + $ p = 3. satisfies L(xt)/L(x) 4 1. u Example 2 . 6 is defined as the distribution of et' where V has the gamma density (2. For p = 1.L( x ). B(x) = O(xP). INTRODUCTION Example 2.
At least as important is the specification of the structure of the point process {Nt } of claim arrivals and its possible dependence with the claims... Similar discussion applies to the distribution of the accumulated claims (XI.1. The reason is in part mathematical since this model is the easiest to analyze. U2. and based upon such information it seems questionable to extrapolate to tail behaviour. it will be seen that we obtain completely different results depending on whether the claim size distribution is exponentially bounded or heavytailed. though the proof of this is nontrivial.3. Thus. the knowledge of the claim size distribution will typically be based upon statistical data. We return to a closer study in IX.1) that any distribution with a regularly varying tail is subexponential. Namely. this phenomenon represents one of the true controversies of the area..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. one may argue that this difficulty is not resticted to ruin probability theory alone. 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.13 (THE SUBEXPONENTIAL CLASS OF DISTRIBUTIONS) We say that a distribution B is subexponential if xroo lim B `2^ = 2. and so is the Weibull distribution with 0 < r < 1.18) B(x) It can be proved (see IX. THE ARRIVAL PROCESS 11 Example 2. the subexponential class of distributions provide a convenient framework for studying large classes of heavyu tailed distributions. However. 3 The arrival process For the purpose of modeling a risk process . When studying ruin probabilities.. which each have a ( timehomogeneous) small rate of experiencing a . the claim size distribution represents of course only one aspect (though a major one). but can never be sure whether this is also so for atypical levels for which far less detailed statistical information is available. By far the most prominent case is the compound Poisson (CramerLundberg) model where {Nt} is Poisson and independent of the claim sizes U1. We give some discussion on standard methods to distinguish between light and heavy tails in Section 4f. for example the lognormal distribution is subexponential (but not regularly varying). but the model also admits a natural interpretation : a large portfolio of insurance holders . (2. From a practical point of view. Also.
not many detailed studies of the goodnessoffit of the Poisson model in insurance are available . The one we focus on (Chapter VI) is a Markovian environment : the environmental conditions are described by a finite Markov process {Jt }too.. To the author 's knowledge . so that . gives rise to an arrival process which is very close to a Poisson process. The compound Poisson model is studied in detail in Chapters III. Mathematically. where {/3 (t)}too is an arbitrary stochastic process . epidemics in life insurance etc.6. INTRODUCTION claim . the periodic and the Markov modulated models also have attractive features . we study this case in VI . its basic feature is to allow more variation (bursty arrivals ) than inherent in the simple Poisson process. see 11.3(t) fluctuating over time.d.8 (t) is a periodic function of t.e. when Jt = i.i. An obvious example is 3(t) depending on the time of the year (the season). 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. getting away from the simple Poisson process seems a crucial step in making the model more realistic. Some of them have concentrated on the marginal distribution of NT (say T = one year ). T2. In order to prove reasonably substantial and interesting results . which facilitate the analysis. IV (and. e. This model can be intuitively understood in some simple cases like { Jt} describing weather conditions in car insurance . are i. Nevertheless . 5. in Chapter VII). however. has some mathematically appealing random walk features . too general and one neeed to specialize to more concrete assumptions . A more appealing way to allow for inhomogeneity is by means of an intensity .. This applies also to the case where the claim size distribution depends on the time of the year or . Another one is Cox processes. I. the first extension to be studied in detail was {Nt } to be renewal (the interarrival times T1 . The difficulty in such an approach lies in that it may be difficult or even impossible to imbed such a distribution into the continuous setup of {Nt } evolving over time . to be studied in Chapter V. This model . Cox processes are. In others. Historically. found the Poisson distribution to be inadequate and suggested various other univariate distributions as alternatives .. such that 8(t) = . the negative binomial distribution.(3.12 CHAPTER I. radioactive decay (a huge number of atoms each splitting with a tiny rate ) and many other applications. and also that the ruin problem may be hard to analyze . with a common term {Nt} is a Markovmodulated Poisson process . in just the same way as the Poisson process arises in telephone traffic (a large number of subscribers each calling with a small rate).. The point of view we take here is Markov dependent random walks in continuous time (Markov additive processes ). However .g. but with a general not necessarily exponential distribution ). it is more questionable whether it provides a model with a similar intuitive content as the Poisson model. with the extension to premiums depending on the reserve. in particular to allow for certain inhomogeneities.
(4. More generally. it is a recurrent theme of this book to stress this connection which is often neglected in the specialized literature on risk theory. reliability. this amounts to Vo having the stationary distribution of {Vt}). interacting particle systems. In the setting of (4.1).1) permitting to translate freely between risk theory and the queueing/storage setting. 0(u) = P(V > u).'s like V is available. stochastic differential equations. A general release rule p(x) means that {Vt} decreases according to the differential equation V = p(V) in between jumps. A SUMMARY OF MAIN RESULTS AND METHODS 13 the environment (VI. The study of the steady state is by far the most dominant topic of queueing and storage theory. that quite often the emphasis is on computing expected values like EV. and here (4.6) .1) where V is the limit in distribution of Vt as t + oo. and the limit t 4 oo is the steadystate limit. with Poisson arrivals and constant release rule p(x) = 1. and which seems well motivated from a practical point of view as well. It should be noted. extreme value theory. it is desirable to have a set of formulas like (4. Thus. genetics models.4.v. Similarly. methods or modeling ideas developed in one area often has relevance for the other one as well. The M/G/1 workload process { Vt } may also be seen as one of the simplest storage models. however. dam/storage processes. the classical result is that the ruin probabilities for the compound Poisson model are related to the workload (virtual waiting time) process {Vt}too of an initially empty M/G/1 queue by means of . point processes and so on. time series and Gaussian processes. Some of these have a certain resemblance in flavour and methodology. 4 A summary of main results and methods 4a Duality with other applied probability models Risk theory may be viewed as one of many applied probability areas. and a lot of information on steadystate r. ruin probabilities for risk processes with an input process which is renewal. Mathematically. this gives only f0 O°i (u)du which is of limited . others being branching processes. The ones which appear most related to risk theory are queueing theory and dam/storage processes. A stochastic process {Vt } is said to be in the steady state if it is strictly stationary (in the Markov case. stochastic geometry.0 (u. R = p(R) in between jumps. queueing theory. others are quite different.1) holds as well provided the risk process has a premium rule depending on the reserve. In fact. Markovmodulated or periodic can be related to queues with similar characteristics.T) = P(VT > u).
Here O(u) = peryu where 3 is the arrival intensity.2). • The compound Poisson model with a claim size distribution degenerate at one point. • The compound Poisson model with some rather special heavytailed claim size distributions. the functions w x f d 1 exdx () . the ideal is to be able to come up with closed form solutions for the ruin probabilities 0(u). 4b Exact solutions Of course .v. A prototype of the duality results in this book is Theorem 11. Thus . • The compound Poisson model with premium rate p(x) depending on the reserve and exponential claim size distribution B. Here ?P(u) is explicit provided that .3. 3.T).1).1 . as is typically the case. The fact that Theorem H. though overlapping. B(x) = ebx. Similarly. • The compound Poisson model with constant premium rate p = 1 and B being phasetype with a just few phases .. 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.14 CHAPTER I. The infinite horizon (steady state ) case is covered by letting T oo.8.3. see Corollary III.p(y) y^ Jo p(x) can be written in closed form. p = 0/8 and y = 8 . have to some extent a different flavour. the two areas.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 ). . Vi(u.1. INTRODUCTION intrinsic interest .3. 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. see Boxma & Cohen [74] and Abate & Whitt [3]. which gives a sample path version of (4. Here Vi(u) is given in terms of a matrixexponential function ( Corollary VIII. 3. The qualifier 'with just a few phases ' refers to the fact that the diagonalization has to be carried out numerically in higher dimensions.g. e . see Corollary VII.1 is a sample path relation should be stressed : in this way the approach also applies to models having supplementary r.'s like the environmental process {Jt} in a Markovmodulated setting. which can be expanded into a sum of exponential terms by diagonalization (see.6.3. Example VIII.
f {eXp U LX 2. T) themselves. Here are some of the main approaches: Laplace transform inversion Often. it is easier to find the Laplace transforms = f e8 .7.u(y)/a2(y) dy} 4c Numerical methods Next to a closedform solution. Ab(u). relevant references are Grubel [179]. . a2 (x): Ip (u) = where S(u) = f °O exp {. Given this can be done. T) du dT 0 TO 00 in closed form than the ruin probabilities z/'(u).Lef$er function.b(u)du . We don't discuss Laplace transform inversion much. esuTb( u.S(u) 1S(oo) f °D exp {. 1). the second best alternative is a numerical procedure which allows to calculate the exact values of the ruin probabilities. 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. 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 . Embrechts. O(u.ff 2µ(y)/a2(y) dy} dx .2) is the natural scale. see VIII. [s.1) is the explicit form of the ruin probability when {Rt} is a diffusion with infinitesimal drift and variance µ(x). say the fast Fourier transform (FFT) as implemented in Grubel [179] for infinite horizon ruin probabilities for the renewal model.4. However. For the finite horizon ruin probability 0(u. the formulas ( IV. where Furrer [150] recently computed ii(u) as an infinite series involving the Mittag. T).1) are so complicated that they should rather be viewed as basis for numerical methods than as closedform solutions. 191). but are somewhat out of the mainstream of the area . T) can then be calculated numerically by some method for transform inversion.f f 2µ(y)/a2(y) dy} dx  (4. A notable fact ( see again XI. • An astable Levy process with drift . (u. Grubel & Pitts [132] and Grubel & Hermesmeier [180] (see also the Bibliographical Notes in [307] p. Also Brownian models or certain skip free random walks lead to explicit solutions (see XI . Abate & Whitt [2].
f. For the compound Poisson model with p = 1 and claim size distribution B with moment generating function (m. and carry out the solution by some standard numerical method. 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. Differential. most often it is more difficult to come up with reasonably simple equations than one may believe at a first sight. it states that i/i(u) .7. (4. 0(u) is then given in terms of a matrixexponential function euu (here U is some suitable matrix) which can be computed by diagonalization. dt] most often leads to equations involving both differential and integral terms. 4d Approximations The CramdrLundberg approximation This is one of the most celebrated result of risk theory (and probability theory as a whole). T) as the solution to a differential. U is explicit in terms of the model parameters. whereas for the renewal arrival model and the Markovian environment model U has to be calculated numerically. as the solution of linear differential equations or by some series expansion (not necessarily the straightforward Eo U'u/n! one!).p)/(13B'[ry] . and in quite a few cases (Chapter VIII).or integral equation.Ce"u. see VIII. which can equivalently be written as f3 [7] = 1 +13 . u * oo.and integral equations The idea is here to express 'O(u) or '(u. .3) where C = (1 .4) 00['Y]1)'Y = 0. One example where this is feasible is the renewal equation for tl'(u) (Corollary III. and in particular the naive idea of conditioning upon process behaviour in [0.g. In the compound Poisson model with p = 1.) B[s]. However. An example where this idea can be carried through by means of a suitable choice of supplementary variables is the case of statedependent premium p(x) and phasetype claims.3) in the compound Poisson model which is an integral equation of Volterra type.3.1) and y > 0 is the solution of the Lundberg equation (4.16 CHAPTER L INTRODUCTION Matrixanalytic methods This approach is relevant when the claim size distribution is of phasetype (or matrixexponential).
Large claims approximations In order for the CramerLundberg approximation to be valid.6) 4e Bounds and inequalities The outstanding result in the area is Lundberg's inequality (u) < e"lu.7 and IV. In the case of heavytailed distributions. However.4. when the claim size distribution is of phasetype. (4. often for all u > 0 and not just for large u. It has generalizations to the models with renewal arrivals. the exact solution is as easy to compute as the CramerLundberg approximation at least in the first two of these three models. some further possibilities are surveyed in 111 . . See Chapter IX. other approaches are thus required. but typically not very precise in their first naive implementation.6) are by far the best one can do in terms of finite horizon ruin probabilities '(u. a Markovian environment or periodically varying parameters. Diffusion approximations are easy to calculate. 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. and use the fact that first passage probabilities are more readily calculated for diffusions than for the risk process itself. the claim size distribution should have an exponentially decreasing tail B(x). J B dx.2. Approximations for O(u) as well as for 1(u. corrected diffusion approximations (see IV. T). In particular. T) for large u are available in most of the models we discuss. u > oo. Diffusion approximations Here the idea is simply to approximate the risk process by a Brownian motion (or a more general diffusion) by fitting the first and second moment. The CramerLundberg approximation is renowned not only for its mathematical beauty but also for being very precise. In fact. in such cases the evaluation of C is more cumbersome. However. This list of approximations does by no means exhaust the topic. in some cases the results are even more complete than for light tails. for the compound Poisson model ^(u) p pu In fact . For example. incorporating correction terms may change the picture dramatically.
In practice. UNT are i. in the compound Poisson model. However. which is a standard statistical problem since the claim sizes Ui. INTRODUCTION Compared to the CramerLundberg approximation (4. though not too many precise mathematical results have been obtained. UNT may be viewed as an interpolation in or smoothing of the histogram). as a general rule. . For example.. it has the advantage of not involving approximations and also. fitting a parametric model to U1.18 CHAPTER I.d. they have however to be estimated from data. .. obtained say by observing the risk process in [0. 4f Statistical methods Any of the approaches and results above assume that the parameters of the model are completely known. 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. it is a general principle that adding random variation to a model increases the risk. This procedure in itself is fairly straightforward. We return to various extensions and sharpenings of Lundberg's inequality (finite horizon versions. T]. e.. The standard suggestion is to observe that the mean residual life E[U . more importantly.g.U(k)) i =k+ i . However. This is proved for the compound Poisson model in 111. one may question whether it is possible to distinguish between claim size distributions which are heavytailed or have an exponentially decaying tail. to have smaller ruin probabilities than when B is nondegenerate with the same mean m. How do we produce a confidence interval? And.8. .k (U(`) .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. . When comparing different risk models. given NT. . . of being somewhat easier to generalize beyond the compound Poisson setting. 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. For example. can we trust the confidence intervals for the large values of u which are of interest? In the present author's opinion.) at various places and in various settings. lower bounds etc. the difficulty comes in when drawing inference about the ruin probabilities. say degenerate at m.3). one expects a model with a deterministic claim size distribution B. and to plot the empirical mean residual life 1 N . empirical evidence shows that the general principle holds in a broad variety of settings.i.
(5. The problem is entirely analogous to estimating steadystate characteristics by simulation in queueing/storage theory. The infinite horizon case presents a difficulty. However. good methods exist in a number of models and are based upon representing the ruin probability zb(u) as expected value of a r. . A main problem is that ruin is typically a rare event (i. having small probability) and that therefore naive simulation is expensive or even infeasible in terms of computer time. where U(1) < . and of course the method is relevant in risk theory as well. CONVENTIONS 19 as function of U(k).4. reference [14]. See further Embrechts.2. UN. For example. respectively. but is not very satisfying. Simulation may be used just to get some vague insight in the process under study: simulate one or several sample paths.. 4g Simulation The development of modern computers have made simulation a popular experimental tool in all branches of applied probability and statistics. to observe whether one or the other limiting behaviour is apparent in the tail. and in fact methods from that area can often be used in risk theory as well .v's) which can be generated by simulation. this is a straightforward way to estimate finite horizon ruin probabilities. and look at them to see whether they exhibit the expected behaviour or some surprises come up.. Truncation to a finite horizon has been used. < U(N) are the order statistics based upon N i. because it appears to require an infinitely long simulation.v. We look at a variety of such methods in Chapter X. Klnppelberg & Mikosch [134].. (or a functional of the expectation of a set of r.5.3 (or just VI. Thus Proposition 4.2. and also discuss how to develop methods which are efficient in terms of producing a small variance for a fixed simulation budget.d.. Still. 5 Conventions Numbering and reference system The basic principles are just as in the author's earlier book Applied Probability and Queues (Wiley 1987.e.i.3) or Section 3 of Chapter VI are referred to as Proposition VI. claims U1.. The chapter number is specified only when it is not the current one.3). in all other chapters than VI where we just write . in this book referred to as [APQ]).3) and Section VI. formula VI. . the more typical situation is to perform a Monte Carlo experiment to estimate probabilities (or expectations or distributions) which are not analytically available. formula (5.
20 CHAPTER L INTRODUCTION Proposition 4.v. IIGII the total mass (variation ) of a (signed ) measure G . B(x) = P(X < x) = fx. cumulative distribution function P(X < x) c. E expectation.f.o. EX2/(EX)2.29) refer to the Appendix. say a heuristic approxi1 + h + h2/2.B(x) = P(X > x) of B. b[s] is defined always if Rs < 0 and sometimes in a larger strip (for example. or a more precise one like eh .s.s. for a probability distribution IIGII = 1. .f. right hand side (of equation) r. oo).g.h. log E[s] where b[s] is the m.v. B(x) the tail 1 .p. random variable s. r. E. with probability Mathematical notation P probability. with respect to w. (moment generating function) fm e82B(dx) of the distribution B. formula (5.4.r. B(dy).i. then for 1s < 5). B[s] the m. If.ceax. i.Used in asymptotic relations to indicate that the ratio between two expressions is 1 in the limit.f. see under b[s] below. Abbreviations c. left hand side (of equation) m.2. if B(x) . w. i. and for a defective probability distribution IIGII < 1. h + 0.d. n i oo.g.g. A different type of asymptotics: less precise.d. B is concentrated on [0.c. cumulant generating function. .3) or Section 3.f.g. (A. References like Proposition A. independent identically distributed i.h.t.d. as for typical claim size distributions. moment generating function. The Laplace transform is b[s]. infinitely often l. mation.e.The same symbol B is used for a probability measure B(dx) = P(X E dx) and its c. squared coefficient of variation.g. n! 27r nn+1/2en.f.f. In particular.
CONVENTIONS {6B the mean EX = f xB(dx) of B ABA' the nth moment EXn = f x"B(dx) of B. . a2) the normal distribution with mean p and variance oa2. (the dimension is usually clear from the context and left unspecified in the notation). matrices have uppercase Roman or Greek letters like T. 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. of numbers. Xt_ the left limit limstt X8f i. R(s) the real part of a complex number s. In the Frenchinspired literature. often the term 'cadlag' (continues a droite avec limites a gauche) is used for the Dproperty. a. intensity interpretation. the value just before t. the ith entry is 1 and all other 0. Usually. .5. i.. and column vectors have lowercase Roman letters like t. row vectors have lowercase Greek letters like a. A. though slightly more complicated. . the processes we consider are piecewise continuous. oo) the space of Rvalued functions which are rightcontionuous and have left limits. i. only have finitely many jumps in each finite interval. 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. the ith unit row vector is e'i. Unless otherwise stated. N(it. 21 D [0.e.e. Thus. an example or a remark.oi denotes the column vector with the xi as components Special notation for risk processes /3 the arrival intensity (when the arrival process is Poisson).A] means E[XI(A)].. 7r. all stochastic processes considered in this book are assumed to have sample paths in this space.e. Notation like f3i and 3(t) in Chapter VI has a similar . 0 marks the end of a proof. E[X. Usually. I(A) the indicator function of the event A. xa. (xi)diag denotes the diagonal matrix with the xi on the diagonal (xi)row denotes the row vector with the xi as components (xi). Matrices and vectors are denoted by bold letters. F o r a given set x1.
ry The adjustment coefficient.5. Notation like BE and B(t) in Chapter VI has a similar. 111.1. 'q the safety loading .g.1. I. FL. . cf. cf. or quantities with a similar time average interpretation. EL the probability measure and its corresponding expectation corresponding to the exponential change of measure given by Lundberg conjugation. cf. p the net amount /3pB of claims per unit time. VI. e.22 CHAPTER L INTRODUCTION B the claim size distribution. though slightly more complicated. J the rate parameter of B for the exponential case B(x) = eby. I.5. interpretation.
the level of the exposition is. The reader should therefore observe that it is possible to skip most of the chapter. 5 on random walks and Markov additive processes can be skipped until reading Chapter VI on the Markovian environment model. strictly speaking. in particular at a first reading of the book. however. The topic is. however. 5) are. When encountered for the first time in connection with the compound Poisson model in Chapter III. the relevance for the mainstream of exposition is the following: The martingale approach in Section 1 is essentially only used here. in most cases via likelihood ratio arguments. The general theory is. a parallel selfcontained treatment is given of the facts needed there. fundamental ( at least in the author' s opinion) and the probability involved is rather simple and intuitive. Due to the generality of the theory. More precisely. in part. The duality results in Section 3 (and. All results are proved elsewhere . used in Chapter VI on risk processes in a Markovian (or periodic) environment. The likelihood ratio approach in Section 2 is basic for most of the models under study. not crucial for the rest of the book. 23 . Sections 4.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. however. Sections 4. somewhat more advanced than in the rest of the book.
Our first result is a representation formula for O(u) obtained by using the martingale optional stopping theorem . {e'YS° }t>0 is a martingale. T(u) > T] .5 can be skipped. T) = P(T(u) < T). (b) St a$ oo on {T(u) = oo}.)AT = E [e7ST(°).(u). T(u) < T] + E [eryST . however.1 Assume that (a) for some ry > 0. Proposition 1. claim size distribution B and p = . using r(u) A T invokes no problems because r(u) A T is bounded by T).u denote the overshoot.. V) (u. and in the limit (1. and the ruin probabilities are ip(u) = P (T(u) < oo).2) takes the form 1 = E [e'ys().QµB < 1.0. (1. T(u) < oo] + 0 = eryuE [e7Vu). 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. As usual.2) As T > oo. Example 1 . f1 . StUit. Let e(u) = ST(u) . Thus N. We get 1 = Ee7So = E e'Y S(.T(u) < cc] = e7uE {e7f(u) I T(u) < cc] z/. The more general Theorem 6.24 CHAPTER II.2 Consider the compound Poisson model with Poisson arrival rate .1 is basic for the study of the compound Poisson model in Chapter III. SOME GENERAL TOOLS AND RESULTS The ladder height formula in Theorem 6. the second term converges to 0 by (b) and dominated convergence (e7ST < eryu on {r(u) > T}). 1 Martingales We consider the claim surplus process {St} of a general risk process. the time to ruin r(u) is inf It > 0 : St > u}.
with common distribution B (and independent of {Nt}). of the normal distribution.Q and the U.Ft = a(S" : v < t).u + x is again just exponential with rate S.5 For the compound Poisson model with B exponential.1) shows that Eels. the ruin probability is O(u) = pe. and p =. the martingale property now follows just as in Example 1.= e"(') where K(a) = .2(c)). and (b) follows from p < 1 and the law of large numbers (see Proposition III.1.1) . Proof Since c(a) = /3 (B[a] .d.1.(„)_ = x is that of a claim size U given U > u .Q(B[a] . are i. Thus.1 is satisfied. Thus.ap.3/6 < 1. Then {St } is Brownian motion with variance constant o2 and drift p < 0. Eeas° = e"(°) where n(a) = a2a2/2 .f.g. B(x) _ edx. By standard formulas for the m.a it is immediately seen that y = S . From this it is readily seen (see III. Since {St} has stationary independent increments.x. MARTINGALES 25 where {Nt} is a Poisson process with rate . and thus Ee'rs° = 1. and thus Ee7s° = 1. O(u ) < e7". it follows that E [e7st+v I J] = e"rstE [e7(st+vSt) I Ft] = e7StEe"rs° = elst where .r" where y = S . From this it is immediately seen that the solution to the Lundberg equation ic(y) = 0 is y = 2p/a2. 1.4 (LUNDBERG ' S INEQUALITY ) tion 1 .2.Q.1 are satisfied. and thus by the memoryless property of the exponential distribution . 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. The available information on this jump is that the distribution given r(u) = t and S.1. u Corollary 1. Since {St} has stationary independent increments. condition (a) of Proposition 1.a.. the conditions of Proposition 1.a = a .3 Assume that {Rt} is Brownian motion with variance constant o. Example 1 .2 and drift p > 0.1) . Now at the time r(u) of ruin {St} upcrosses level u by making a jump ./3. the conditional distribution of the overshoot e(u) = U .i. u Corollary 1. A simple calculation (see Proposition III.6a for details) that typically a solution to the Lundberg equation K(y) = 0 exists.
the parameters of the two processes can be reconstructed from a single infinite path. P are then singular (concentrated on two disjoint measurable sets). F(S) = P(S) = 1. u Notes and references The first use of martingales in risk theory is due to Gerber [156].1) 'though not always: it is not difficult to construct a counterexample say in terms of transient Markov processes. (2. A]. Two such processes may be represented by probability measures F. However.F). on (DE. as shown by the following example this setup is too restrictive: typically'.6 N S = { lim Nt I t +00 t gJ are both in F. A E Ft. Proof Just note that ^(u) = 0 by continuity of Brownian motion. we look for a process {Lt} (the likelihood ratio process) such that P(A) = E[Lt. P on (DE. [172]. which we equip with the natural filtration {. Theorem 2. Thus the sets S = I tlim +oot t =. F). Embrechts.v. oo).6 If {Rt} is Brownian motion with variance constant a2 and drift p > 0. I.Ft}too and the Borel afield F. cf. and in analogy with the theory of measures on finite dimensional spaces one could study conditions for the RadonNikodym derivative dP/dP to exist. hence so is Nt = limfyo N2`i.. 2 Likelihood ratios and change of measure We consider stochastic processes {Xt} with a Polish state space E and paths in the Skorohod space DE = DE[0. then S and S are disjoint . More recent references are Dassios & Embrechts [98]. .e. Delbaen & Haezendonck [103] and Schmidli [320]. Grandell & Schmidli [131]. Grandell [171]. and is further exploited in his book [157]. B. and F.3 below). A somewhat similar u argument gives singularity when B $ B. 0 and claim size distributions B. P correspond to the claim surplus process of two compound Poisson risk processes with Poisson rates /3.1 Let F. The interesting concept is therefore to look for absolute continuity only on finite time intervals (possibly random. and by the law of large numbers for the Poisson process . The number Nt F) of jumps > e before time t is a (measurable) r. then z/'(u) = e7" where 'y = 21A/a2. Example 2 . But if a $ ^ .26 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Corollary 1.
P) such that LLt = 1. Lets < t. Lt < 0] can only be nonnegative if P(A) = 0.r.t. A] = E[Lt. 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 . then there exists a unique probability measure Pon . the restriction of P to (DE.A] = EE[LtI(A)IF8] = EI(A)E[LtIFB] = EI(A)L8 = PS(A).F8] = L8 and the martingale property.2) Proof Assume first G C {T < T} for some fixed deterministic T < oo. Then Ft (A) = E[Lt. Then Lt > 0 and ELt = 1 ensure that Pt is a probability measure on (DE.. A]. then { 1 P(G) = EG . F) such that ELt = 1.Pt)) The following result gives the connection to martingales. F the Borel o•field and P a given probability measure on (DE. (i) If {Lt}t> o is a nonnegative martingale w. 1 J (2. then {Lt} is a nonnegative martingale w. ({Ft} .Ft}. A E F8. under the assumptions of (ii) we have for A E rg and s < t that A E Ft as well and hence E[L8. implies that E[LtI. By the martingale property. G ] = E [LT . If r is a stopping time and G E PT. Proof Under the assumptions of (i). This proves (i).1) holds.F such that (2.1) and nonnegativity by letting A = {Lt < 0}. we have E [ LTIFT]1 = LT on {T < T}.2(i). G C {T < oo}. The truth of this for all A E Y.2. Then P(A) = E[Lt. _.t. Ft).3 Let {Lt}. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 27 (i.Pt}adapted process {Lt}t>o (2. that the restriction of P to (DE. ({. Proposition 2.r. Hence the family {Pt} is t>o consistent and hence extendable to a probability measure F on (DE. if for some probability measure P and some {. Finally.r. G l ] = E [_I(G)E[LTIFT] ] = E { _I(G)Lr ] = P(G).t. P be as in Proposition 2.Y) such that P(A) = Pt(A).Tt) is absolutely continuous w. define P by Pt (A) = E[Lt. (ii) Conversely. .1) holds. ELt = 1 follows by taking A = DE in (2. Hence E [_ .e. using the martingale property in the fourth step. A]. .F). Conversely.2 Let {Ft}t>o be the natural filtration on DE. A E Ft . A E F.
Xt = St.1) is that it seems in general easier to deal with the (unconditional) expectation E[eryVu). Consider a (timehomogeneous) Markov process {Xt} with state space E.1: Corollary 2.3 we obtain a likelihood ratio representation of the ruin probability V) (u) parallel to the martingale representation (1. t. . T(u) < oo]. and this problem will now be studied. we have F(G) = V )(u). applying (2. In the context of ruin probabilities.4) Proof Letting G = {r(u) < oo}.4) compared to (1. Rt) or Xt = (Jt.2) by noting that 1 = ersr(„) = e1'ue7Ou).3) to G of{r < T} we get 1111 F(Gn {r <T}) = E[ 1 .28 CHAPTER II.t. the natural filtration {. 5) for processes with some randomwalklike structure.t. r(u) < oo] occuring there than with the (conditional) expectation E[e'r{(u ) Jr(u) < oo] in (1. we need the concept of a multiplicative functional..Rt} the claim surplus process and {Jt} a process of supplementary variables possibly needed to make the process Markovian. say. Xt = (Jt. of (2.s. Now just rewrite the r.4 Under condition (a) of Proposition 1. is nonnegative and has Ey Lt = 1 for all x.1) in Proposition 1.r.1). First we ask when the Markov property is preserved. each F.Ft} . St). (2.r. To this end. first in the Markov case and next (Sections 4. both sides are increasing in T.2) follows by monotone convergence. {St} = {u . we assume for simplicity that {Xt} has Dpaths. in continuous time (the discrete time case is parallel but slightly simpler).h. Lr(u) 11 The advantage of (2. where {Rt} is the risk reserve process.Gn {r <T} . The crucial step is to obtain information on the process evolving according to F. 1 Since everything is non negative.O(u) = eryuE[e'YC(u). SOME GENERAL TOOLS AND RESULTS In the general case .. (2. u From Theorem 2. one would typically have Xt = Rt. is Markov w.1. and letting T * oo. A change of measure is performed by finding a process {Lt} which is a martingale w. The problem is thus to investigate which characteristics of {Xt} and {Lt} ensure a given set of properties of the changed probability measure. For the definition.
5) is equivalent to Ex[Lt+8Vt+8] = E8[Lt • (L8 o 91)Vt+8] for any .7).v. Zt. t] * [0. o 9tI.(Xtitl) with all t(i) < t + s. since Ext [L8Y8] = E[(Y8 o et)(L8 o 8t)I.v. By definition of Px. t and let Px be the probability measure given by t.7).6) implies (2. 0 . o 9t = V.t.[Y. Similarly.Pt+8measurable r.r.F8measurable r.8) which is the same as (2.Y8f t < s.v.'s of the form Zt • (Y8 o 0t) comprises all r.6) for any . the Markov property can be written E. this in turn means Ex[Lt+8Zt(V8 oet)] = Ex[LtZtExt[L8Y8]].Ft] = Ex.5 Let {Xt} be Markov w.'s of the form fl' f.5) Pxa. nonnegative and Lt+8 = Lt•(Lso9t) (2. Indeed. Vt+e.. where Ot is the shift operator.T9measurable Y8. let {Lt} be a nonnegative martingale with Ex Lt = 1 for all x.2.7) for any Ftmeasurable Zt and any . LIKELIHOOD RATIOS AND CHANGE OF MEASURE 29 on DE and define {Lt} to be a multiplicative functional if {Lt} is adapted to {. A].(A) = Ex [Lt. Then the family {Px}xEE defines a timehomogeneous Markov process if and only if {Lt} is a multiplicative functional.v. The precise meaning of this is the following: being .Ft+8measurable. Ex[Lt+8Zt(Y8 o et)] = Ex[LtZt(Y8 o 0t)(L8 o Bt)]. or. (2. t. (2. Y8.Ft }. ({Xt+u} 0<u<8) Theorem 2.YB] for any Ftmeasurable r. s. for all x. The converse follows since the class of r.. which in turn is the same as Ex[Lt+8Zt • (V8 o Bt)] = Ex[Lt • (L8 o 91)Z1 • (Y8 o et)] (2. since Zt • (Y8 o Ot ) is .5) are Tt+e measurable.Ft]. the natural filtration {Ft} on DE.s. oo). and then L.Ftmeasurable. which is the same as Ex[Zt(Y8 o Bt)] = E8[ZtEx. Proof Since both sides of (2. (2.v. Lt has the form Lt = 'Pt ({x }0<u<t) for some mapping cot : DE[O. (2.
. t. The result is a sample path relation. t] = LtExt L8 = Lt. We work on a finite time interval [0. {Vt} .. The storage process {Vt }o<t<T is essentially defined by time reversion. More precisely . In between jumps.. R = p(R)). CN. see Dynkin [128] and Kunita [239]. and just after time or* {Vt} makes an upwards jump of size UU = UN _k+l. T] in the following setup: The risk process {Rt}o<t<T has arrivals at epochs or. the premium rate is p(r) when the reserve is r (i. Thus R = Ro + f p(R8) ds .Ft] = LtE[L8 o 9t I. . u Notes and references The results of the present section are essentially known in a very general Markov process formulation.5 can be found in Kuchler & Sorensen [240]..e. .At where At = k: vk <t U. (3. reflection at zero and initiar condition Vo = 0.. and the time to ruin is 7(u) = inf {t > 0: Rt < 0}. 0 < vl < . aN where or* = T UN_k+l. In between jumps. The corresponding claim sizes are Ul. then Remark 2. SOME GENERAL TOOLS AND RESULTS to define a timehomogeneous Markov process. .. Indeed. < aN < T. . we shall establish a general connection between ruin probabilities and certain stochastic processes which occurs for example as models for queueing and storage. Ro = u (say). with a proof somewhat different from the present one. } E[Lt+B I. UN. A more elementary version along the lines of Theorem 2. 3 Duality with other applied probability models In this section..30 CHAPTER H..6 For {u .. the arrival epochs are Qi. and thus for the moment no parametric assumptions (on say the structure of the arrival process) are needed.. The formulation has applications to virtually all the risk models studied in this book.. (using the Markov property in the second step) so that the martingale property is automatic. it xEE suffices to assume that {Lt} is a multiplicative functional with Ex Lt = 1 for all x..1) The initial condition is arbitrary.
(3. V = p(V)).____•_.2) k: ok <t and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0... Note that these definitions make {Rt} rightcontinuous (as standard) and {Vt} leftcontinuous.__.e.AT_t... In particular. That is.11 4. :. Theorem 3.3.. .3) (u) Proof Let rt' denote the solution of R = p(R) subject to r0 = u.1) we have Vt = At  f P(Vs)ds where A= U= AT ..x.___ . (3...T) = P(VT > u).. DUALITY WITH OTHER APPLIED PROBABILITY MODELS 31 decreases at rate p(r) when Vt = r (i.1. The sample path relation between these two processes is illustrated in Fig. V)(u. {Vt} remains at 0 until the next arrival).T) = inf Rt < 0 P (O<t<T P(r(u) < T) be the ruin probability._: 1} 0 011 =T01N ^N3 To 0 011 014 01N Figure 3.__. Then rt°) > rt°) for all t when u > v. 3._. instead of (3.1 The events {T(u) < T} and {VT > u} coincide.1 Define r(u) = inf It > 0: Rt < 0} (r(u) = oo if Rt > 0 for all t < T) and let ii(u.
Proof Let T ^ oo in (3. Then the storage process {Vt} has a proper limit in distribution. Resnick & Tweedie [79]. we have r(u) > T.3 and being i. 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.T l . and then '0 (u) = P(V > u). this represents a model for storage. Some further relevant more general references are Asmussen [21] and Asmussen & Sigman [51].2 from Harrison & Resnick [188].32 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Suppose first VT > u (this situation corresponds to the solid path of {Rt} in Fig. we can repeat the argument and get VoN_1 > Ra2 and so on. Then similarly VVN = r0. = r(VT) . Thus we may think of {Vt} as having compound Poisson input and being defined for all t < oo. The arrival epochs correspond to rainfalls. Historically. one may feel that the interaction between the different areas has been surprisingly limited even up to today. see Siegmund [344].and left continuity is immaterial because the probability of a Poisson arrival at any fixed time t is zero). 3. .U1 = Rol. if nothing else n = N). Theorem 3. and in between rainfalls water is released at rate p(r) when Vt (the content) is r. Historically. We get: Corollary 3.3). with distribution B.1 and its proof is from Asmussen & Schock Petersen [50].i.d.Ul < roil  Ul = RQ„ Va1V_1 < RQ2. 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. u Notes and references Some main reference on storage processes are Harrison & Resnick [187] and Brockwell.2 Consider the compound Poisson risk model with a general premium rule p(r). Hence RQ„ > 0 for all n < N. we have RQ„ < 0 so that indeed r(u) < T. Suppose next VT < u (this situation corresponds to the broken path of {Rt} in Fig.U1 > roil . Corollary 3. Hence if n satisfies VVN_n+1 = 0 (such a n exists. Nevertheless. the distinction between right. and since ruin can only occur at the times of claims. say V.1 with Ro = u = ul). Then Vo.1 with Ro = u = u2). Then the time reversibility of the Poisson process ensures that {At } and {At } have the same distribution (for finitedimensional distributions. u A basic example is when {Rt} is the risk reserve process corresponding to claims arriving at Poisson rate . and so on. if and only if O(u) < 1 for all u. 3. and a general premium rule p(r) when the reserve is r.
. N min (Y1 + + Yn). Z2. where the Yi are i . the Lindley process Wo.. is defined by assigning Wo some arbitrary value > 0 and letting Wn+1 = (Wn + Zn+1)+• (4...1 in terms of Lindley processes . .1. has a proper limit W in distribution as n + oo... Z2 . just verify that the r .. Here F is a general probability distribution on R (the special case of F being concentrated on {1.. Let further N be fixed and let Wo.w. i..i. For discrete time random walks . = Xo + Y1 + • • • + Y. . ZN = . WN be the Lindley process generated by Z1 = YN.. ... .. Z2 = Y2.. W2.2 The following assertions are equivalent: (a) 0(u) = P(r(u) < oo) < 1 for all u > 0. as long as the random walk only takes nonnegative values.....1.4.1. generated by Z1.Yl min (YN .2) (for a rigorous proof. hits (oo. Most often... 0).. with common distribution F (say). I. Xo = 0..1)). and is reset to 0 once the r. if Wo = 0 then (Z1+•••+Zn) WN = Zl+•••+ZN.h.... N min (Y1 + • • • + YNn) n=0. n=0..2) satisfies the same recursion as in (4. Then the events {r(u) < N} and {WN > u} coincide. RANDOM WALKS IN DISCRETE OR CONTINUOUS TIME 33 4 Random walks in discrete or continuous time A random walk in discrete time is defined as X.2).s......1 Let r(u) = inf In: u + Y1 + • • • + Yn < 0}.1.N From this the result immediately follows. 0 Corollary 4.YNn+1) n=0. In particular.e.d... . W1. For a given i. Proof By (4.. (b) 1/i(u) = P(•r(u) < oo) > 0 as u * oo. . {Wn}n=0..1) Thus {Wn}n=o... .. of (4.1.Y1 according to Wo = 0...N (4...1. Z2 = YN_1 i . R valued sequence Z1.... (c) The Lindley process {WN} generated by Zl = Y1. W1.. Theorem 4. . there is an analogue of Theorem 3. Z2. can be viewed as the reflected version of the random walk with increments Z1.min n=0. .d.. WN = YN .. . evolves as a random walk with increments Z1i Z2. 1} is often referred to as simple random walk or Bernoulli random walk)...
. a Markovian change of measure as in Theorem 2. For a random walk. . + Z. Remark 4 .l. Thus the assertion of Theorem 4.g. Y1) has the same distribution as (Y1.the result is a sample path relation as is Theorem 3. Similarly. 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. assumption on the Z1. w. YN in Theorem 4.. (d) #..s. (Yi + • • • + Yn) > oo a. 176) but appears to be rather intractable. By Kolmogorov's 01 law. W v m and P(W > u) = P (m > u) = 0(u). . SOME GENERAL TOOLS AND RESULTS (d) m = inf.1 have the same distribution for n = 0..ooa. the condition 00 F(YI+•••+ Yn<0)<00 n=1 is known to be necessary and sufficient ((APQ] p.o. (e)Yi+•••+Yn 74 .N so that WN _P4 M = supra=0. the Y1. YN).. ZN or. Combining these facts gives easily the equivalence of (a)(d).2 and Theorem 4. equivalently... (Z1 + • • • + Zn) = m and P(W > u) = P(M > u) = i (u ). . . Px to Fn are equivalent (have the same null sets) so that the likelihood ratio Ln exists. .s.1. or M < oo a. . 0 By the law of large numbers.. Proof Since (YN. then the restrictions of Fx. the Lindley processes in Corollary 4.. doubly u infinite (n'= 0.l.34 CHAPTER II.. . N. The following result gives the necessary and sufficient condition for {Ln} to define a new random walk: . .1 is actually not necessary .... Clearly....5 does not necessarily lead to a random walk: if....d.=o. In that case . .. In general..g. . e..1. Next consider change of measure via likelihood ratios. .s.1..s . (e). One then assumes Yn to be a stationary sequence.. either M = oo a.i. a sufficient condition for (e) is that EY is welldefined and > 0.. ±1.) sup n=0.1 is equivalent to WN D MN = (Z1 + .) and defines Zn = Yn. ±2..2. The converse follows from general random walk theory since it is standard that lim sup (Y1 + • • + Yn) = oo when Y1 + • • • + Yn 74 oo. there is a more general version of Corollary 4.1.3 The i.
. 100 ).5 corresponds to a new random walk with changed increment distribution F(x) = e'(a) Jr e"'F(dy) . We get: Corollary 4. RANDOM WALKS INDISCRETE OR CONTINUOUS TIME 35 Proposition 4. u A particular important example is exponential change of measure (h(y) = e°y'(") where r.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.4 Let {Ln} be a multiplicative functional of a random walk with E_L.4) ({Ln} is the familiar Wald martingale ..3) holds for n = 1. for some function h with Eh(Y) = 1.s.4. Since L1 has the form g (Xo.s..f. cf.5 Consider a random walk and an a such that c(a) = log F[a] = log Ee° ' is finite. In particular.. we get L2 = L1 (L1 o91 ) = h(Y1)g(X1. the random walk property implies Ex f (Y1) = Eo f (Y1 ). For n = 2. implying g (x.Y2) = h(1'i)h(I'a). of F). Then the change of measure in Theorem 2. and define Ln by (4. Y1). Conversely. Y ) f (Y)] = E[g(O. (4.g. Y) f (Y)] for all f and x. y ). and so onforn =3.nrc(a )} (4. Y) = h(Y ) a.g. Then the change of measure in Theorem 2.3) holds. Breiman [78] p. the changed increment distribution is F(x ) = E[h(Y). .3) 1Pxa. The corresponding likelihood ratio is Ln = exp {a (Y1 + • • • + Yn) .. where h (y) = g(0. (a) = log F(a] is the c. e. this means E(g(x. In that case. h(Yn) (4. Proof If (4. Y < x].. = 1 for all n and x.5 corresponds to a new random walk if and only if Ln = h(Y1) .4).
The appropriate generalization of random walks to continuous time is processes with stationary independent increments (Levy processes). however.7) for all e > 0. given by a the increment distribution F(x) = P(Xn+l . with premium rate p. or imbedded into continuous time processes . say by recording the reserve or claim surplus just before or just after claims (see Chapter V for some fundamental examples). a Brownian component {Bt} (scaled by a variance constant) and a pure jump process {Mt}. (4. we are .Xn < x). The simplest case is 3 = JhvMM < oo.1). {Xt} can be written as the independent sum of a pure drift {pt}. v2 = 0 and v = 3B). Xt (n)t(n1) being independent whenever t(O) < t(1 ) < . However.36 CHAPTER II. i. An equivalent characterisation is {Xt} being Markov with state space R and E [f (Xt+e . SOME GENERAL TOOLS AND RESULTS Discrete time random walks have classical applications in queueing theory via the Lindley process representation of the waiting time . say the beginning of each month or year .6) More precisely. this description needs some amendments.Ft] = Eof (X. < t(n) and with Xt( i)_t(i_l) having distribution depending only on t(i) . see Chapter V. In continuous time.).Xt)I.e. . In risk theory... the claim surplus process for the compound Poisson risk model . the interpretation is that the rate of a jump of size x is v(dx) (if f of Ixlv (dx) = oo. corresponds to a process with stationary independent increments and u = p.t(i . {Xt} is a random walk. they arise as models for the reserve or claim surplus at a discrete sequence of instants. The traditional formal definition is that {Xt} is Rvalued with the increments Xt(1)_t(o). (4. Roughly. 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. a positive measure on R with the properties e J x2v(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 . . the tradition in the area is to use continuous time models. e f x:IxJ>e} v(dx) < oo (4..5) Note that the structure of such a process admits a complete description. but we omit the details ). In discrete time. Xt =Xo+pt+oBt+Mt. the pure jump process is given by its Levy measure v(dx). Xt(2)_t(l).
3 and decreases linearly at rate 1 in between jumps. Proposition 4.min Xt (4.10) .7 If {Xt} has stationary independent increments as in (4. Corollary 4. virtual waiting time refers to Vt being the amount of time a customer would have to wait before starting service if he arrived at time t (this interpretation requires FIFO = First In First Out queueing discipline: the customers are served in the order of arrival). O(u.1)v(dx) (4. Here workload refers to the fact that we can interpret Vt as the amount of time the server will have to work until the system is empty provided no new customers arrive. V E [0.v.s. Furthermore. RANDOM WALKS IN DISCRETE OR CONTINUOUS TIME 37 almost solely concerned with the compound Poisson case and shall therefore not treat the intricacies of unbounded Levy measures in detail.Q and distribution B of the service times of the arriving customers. has upwards jumps governed by B at the epochs of a Poisson process with rate . where c(a) = ap + a2a2/2 + f 00 provided the Levy measure of the jump part {Mt} satisfies f".6 In the compound Poisson risk model with constant premium rate p(r) .4.1. Chapter III. A different interpretation is as the workload or virtual waiting time process in an M/G/1 queue.8) O<t<T (assuming Wo = Xo = 0 for simplicity). VT + V for some r.6). cf. i. defined as a system with a single server working at a unit rate. and b(u) = P(V > u). WT = XT . is easily seen to be f3pB < 1.] Processes with a more complicated path structure like Brownian motion or jump processes with unbounded Levy measure are not covered by Section 3. then Ee'(xtxo) = Eoeaxt = etx(a). where VT is the virtual waiting time at time T in an initially empty M/G/1 queue with the same arrival rate /3 and the service times having the same distribution B as the claims in the risk process. (ex . Then the storage process {Vt} has constant release rate 1. First assume in the setting of Section 3 that {Rt} is the risk reserve process for the compound Poisson risk model with constant premium rate p(r) = 1. T) = P(VT > u). Now consider reflected versions of processes with stationary independent increments.e. having Poisson arrivals with rate . jxJ v(dx) < oo.2). and the reflected version is then defined by means of the abstract reflection operator as in (4. oo]. [The condition for V < oo a.
Proof For the first statement .Xo is necessarily infinitely divisible when {Xt} has stationary independent increments.g. Then l e" (a) = Eo [ Li ea "] = eK (9)Eo {e ( a+9)x1 J I = er(a+o )K(B) R(a) = K(a + 0) . let e" (a ) = Eoeaxl. In particular.6) are µ = µ + Oo2 . of an infinitely divisible distribution (see. we use the characterization (4. . Eea(µt + QBt) = et{aµ +a2OZ/2}.11) (eax . use the representation as limit of compound Poisson processes. 5 has stationary independent increments as well. then the changed parameters in the representation (4..10) is the LevyKhinchine representation of the c. Then the Markov process given by Theorem 2. u Note that (4.g. t. v(dx) = e9xv (dx). 8 Assume that {Xt} has stationary independent increments and that {Lt} is a nonnegative multiplicative functional of the form Lt = g(t.1)eexv(dx). Xt Xo) with E2Lt = 1 for all x.Xt)g(s. if Lt = e9(xt . e.1. Chung [86]).5) and get E [f(Xt+B .Xt)I Ftl = Eof (X8)g(s.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 . Xt +B . This is of course no coincidence since the distribution of Xl .f.Xt)I'Ftl = E [f(Xt+B . we show in the compound Poisson case ( IlvIl < oo) in Proposition III.38 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Proof By standard formulas for the normal distribution. (4.Xt)L8 o 0tIFt] = E [f (Xt+s . X8) = Eof (X8)L8 = Eof (X8)• For the second.xo)tk ( e). Q2 = v2. Theorem 4 .1 that E eaMt = exp fmoo In the general case .1 ) v(dx) . By explicit calculation .
..10 Let Xt be the claim surplus process of a compound Poisson risk process with Poisson rate .. St)} where {Jt} is a Markov process with state space E (say) and the increments of {St} are governed by {Jt} in the sense that E [f (St+8 .(3B(dx). B have densities b.o[f (S8)g(J8)].Ft] = Ejt.1) For shorthand .St)g(Jt+s)I. it is then easily seen that Lt = H dB(Ui) i:o. b = a = 0) the changed process is the claim surplus process of another compound Poisson risk process with Poisson rate . MAP stands for the Markovian arrival process discussed below.g. we write Pi.3 =. v(dx) _ . As for processes with stationary independent increments . MARKOV ADDITIVE PROCESSES 39 Remark 4.(3 = . Example 4 . then the martingale {eex(t)tk(e)} is the continuous u time analogue of the Wald martingale (4. the corresponding claim sizes . Recalling that U1. 0. abbreviated as MAP in this section2. 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..5. dB/dB = b/b when B.0. . corresponding to p = 1.3 and claim size distribution B # B.l3 and claim u size distribution B. Thus (since µ = p = 1. where . and let the Px refer to the claim surplus process of another compound Poisson risk process with Poisson rate.8.3B[B].2. Example 4 . the structure of MAP's is completely understood when E is finite: 2and only there .0 in the following..3 and claim size distribution B.4). b with b(x) > 0 for all x such that b(x) > 0). B(dx) = B[9] B(dx). Then we can write v(dx) _ /3eOxB(dx) = / (dx).11 For an example of a likelihood ratio not covered by Theorem 4. U2. <t whenever the RadonNikodym derivative dB/dB exists (e. Ei. Ei instead of P2. u 5 Markov additive processes A Markov additive processes. (5. are the arrival times and U1. one reason is that in parts of the applied probability literature.9 If X0 = 0. a = 0. is defined as a bivariate Markov process {Xt} = {(Jt. .
jEE• On an interval [t. 1 J1 ='^])iJEE = (Fij[a])i . Y1 E dx) where Y„ = S„ . As a generalization of the m.jEE (here pij = Pi(J1 = j)) and the probability measures Hij(dx)=P(Y1 EdxlJo=i. let {Jt} be standard Brownian motion on the line.it = A. As an example. with the Y„ being interpreted as interarrival times. a jump of {Jt} from i to j # i has probability qij of giving rise to a jump of {St} at the same time. v.o(Ji = j. In addition. consider the matrix Ft [a] with ijth element least Ei . oo). t+s) where Jt .i. the distribution of which has some distribution Bij. this means that the MAP can be simulated by first simulating the Markov chain {J„} and next the Y1.f. a MAP is the same as a semiMarkov or Markov renewal process. Jn = j. {Jt} is specified by its intensity matrix A = (Aij)i. Y2. SOME GENERAL TOOLS AND RESULTS In discrete time. Fn[a] = F[a]n where P[a] = P . If all Fij are concentrated on (0.6) depending on i.J1=j)= Fij (dx) Pij In simulation language. 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.) If E is infinite a MAP may be much more complicated.40 CHAPTER H.[a) = (Ei[easl. {St} evolves like a process with stationary independent increments and the parameters pi.9 EE = (iii&ij[a])i j EE .1 For a MAP in discrete time and with E finite. (That a process with this description is a MAP is obvious.. In continuous time (assuming Dpaths).. Proposition 5. a MAP is specified by the measurevalued matrix (kernel) F(dx) whose ijth element is the defective probability distribution Fij(dx) = Pi.g.Sr_1. .. An alternative description is in terms of the transition matrix P = (piA.. which we omit and refer to Neveu [272] or cinlar [87]. the converse requires a proof. by generating Yn according to Hij when J„_1 = i. vi(dx) in (4..
Bij (i. Jt = j] Ejesh'^ + E Ak j hEi [ease . j E E) and So = 0. Jt = k] { xk kEE j la] . we infer that in the discrete time case the . Jt = j] is given by etK[a].5. vi(dx) (i E E). qij.1) } (recall that qjj = 0). a= . which in conjunction with Fo[a] = I implies Ft[a] = etK[a) according to the standard solution formula for systems of linear differential equations. this means that F't+h [a] = Ft[a] II+h(rc(i)(a)) +hA+h(Aijgij(Bij[a]1)) I. pi. up to o( h) terms.2 Let E be finite and consider a continuous time Markov additive process with parameters A.4c). u In the following.1)) . Sn ) yields Ei[easn+ '. where K[a] = A+ (r. MARKOV ADDITIVE PROCESSES Proof Conditioning upon (Jn. Then the matrix Pt[a] with ijth element Ei [east. kEE Jn = k]Ek[e"Y" . u Proposition 5. \ diag Ft[a] = Ft[a]K. J1 = A which in matrix formulation is the same as Fn+1 [a] = Fn[a]F[a]. In matrix formulation . aSt h = (1 + Ajjh) Ei [east . 00 r(i) (a) = api + a2ot /2 + f (e° . Then. Jt = k] { 1 . Proof Let {Stt) } be a process with stationary independent increments and pa rameters pi .1 )v(dx).(')(a)) diag + (). assume that the Markov chain/process {Jt} is ergodic.ijgij(Bij[a] .qkj + k?^j qkj Bkj [a] } = Ei [east. Jn+1 = A] = 41 Ei[ e 5„. By PerronFrobenius theory (see A. Jt = j] (1 + htc (j) (a)) j + Ak j qk j (Bk +h E Ei [east . vi(dx). 013 .
its derivatives are 'asymptotic cumulants'. Corollary 5.3 Ei [east.r. cf. Yrh(a ) = 1. Proposition 5.7.etx It then follows that E feast+^(t+v)K(a)h(a) I ^tl l . and write k = k(°). just note that [a]h(a) = eietK (a)h(a) = etK(a)h(a). of a random walk. Jeast.c(a) (and h(")). 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.5 EiSt = tK'(0) + ki . Then h(°) = e. . Proof For the first assertion.Eikjt = ttc'(0) + ki . The function ic(a) plays in many respects the same role as the cumulant g.4c).4 Eie"sth(a) = h=a)et?("). (5.e=e°tk.h(a)vva)etw(a).tK(a)h(a) J jj it L o is a martingale.42 CHAPTER II. Corollary 5. as will be seen from the following results. we are free to impose two normalizations. The corresponding left and right eigenvectors v("). and we shall take V(a)h(a) = 1. We also get an analogue of the Wald martingale for random walks: Proposition 5. 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. Since v("). h(") are only given up to a constants. a.4.2) where 7r = v(°) is the stationary distribution.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). cf. h(") may be chosen with strictly positive components. Proof By PerronFrobenius theory (see A. Furthermore. u Let k(a) denote the derivative of h() w. In particular. Jt = j] .f. Corollary 5.t.
8 Also for E being infinite (possibly uncountable ). subtraction yields Vary St = tic"(0) + O(1). More precisely.5. For the second . (5.6 For any stopping time T with finite mean.5. Ee"st typically grows asymptotically exponential with a rate ic(a) independent of the initial condition (i.Eikjr . the existence of exponential moments is assumed ) but can be made rigorous by passing to characteristic functions. In the same way. there is typically a function h = h(") on E and a ^c(a) such that Ey a"st t" (") * h(x). the distribution of Jo). we differentiate (5. t a oo.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.4) .4. Corollary 5.g. Squaring in Corollary 5. [E. 43 Ei [Steast h(a) + east k^a)1 = et"(a) (kia) + tic (a)hia)) . ] = t2tc (0)2 + 2tK'(0)vk + ttc"(0) + O(1) . summing and letting a = 0 yields E„ [St + 2Stkj. u The argument is slightly heuristic (e. 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.3) Let a = 0 and recall that h(°) = e so that 0=°) = h(o) = 1.e. tam E tSt a (0).St]2 = t2/c'(0 ) 2 + 2ttc (0)vk . for a random walk: Corollary 5. MARKOV ADDITIVE PROCESSES Proof By differentiation in Proposition 5. t im v^"St = '(0) Proof The first assertion is immediate by dividing by tin Corollary 5. one obtains a generalization of Wald's identity EST = Er • ES. Remark 5 .+ k.2ttc (0)Evkjt + 0(i).a) + ttc (a)2hia ) Multiplying by v=.. E=ST = tc'(0)E7.") }) .7 No matter the initial distribution v of Jo.. (5. .
44 CHAPTER IL SOME GENERAL TOOLS AND RESULTS for all x E E. We then want to determine h and x(a) such that Ejeasth (Jt) = etK(a)h(i).s..6. we take the martingale property as our basic condition below (though this is automatic in the finite case). Jt = (s+t) mod 1 P8a. Remark 5. From (5.5.9 The condition that (5.6. 0) = h(i )( 1 + ttc(a)). An example beyond the finite case occurs for periodic risk processes in VI. however. xEE .e.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 5.5) is a martingale can be expressed via the infinitesimal generator G of {Xt } = { (Jt. where {Jt} is deterministic period motion on E = [0. For t small .3b and Remark VI. In view of this discussion . gha(i. G is defined as Gf (x) = lim Exf (Xt) . let ha(i. h(Jo) Lo is a Px martingale for each x E E.(9) {Lt}t>o = . St) } as follows. this leads to h(i) + tcha( i.f (x) tyo t provided the limit exists. and the family {f LEE given by Theorem 2. Given a function h on E. First. Then {Lt } is a multiplicative functional.for the present purposes.6) We shall not exploit this approach systematically. in particular that f is bounded.5) is a martingale . u forsEE). 1) (i. s) = ea8h(i). Usually. some extra conditions are imposed. V. see.5 defines a new MAP..4 that { h(Jt) easttK(a) L o (5. however. St)} be a MAP and let 0 be such that h(Jt) OStt.10 Let {(Jt. (5. 0 Proposition 5. this is. inconvenient due to the unboundedness of ea8 so we shall not aim for complete rigour but interpret C in a broader sense. 0) = n(a) h(i).
(0)j. In particular. Ai = µi + 0Q.c(0)e = tc(0)e .10 is given by P = eK(e) Oh e) F[e]Oh('). .11 Consider the irreducible case with E finite. That 0 < qij < 1 follows from the inequality qb <1.1) eft ea' f ij (dx) = Hij (dx) Hij [0] .. That the rows sum to 1 follows from Ae = Oh(e) K[O]h(B) .11 below in the finite case. 1 + q(b . Bi [0] Remark 5.7(dx) Bij [0] Bij(dx) in the continuous time case .7) In particular. this gives a direct verification that A is an intensity matrix: the offdiagonal elements are nonnegative because Aij > 0.tc(0)e = 0 .St)sl(e) h(Jt) 45 The proof that we have a MAP is contained in the proof of Theorem 5. vi (dx) = f3 Bi(dx) with .1) holds for the P. and by A = Oh(°) K [0]Oh(e ) vi(dx) = e"xvi (dx).ic(0)e = ic(0)Oh e) h(e) . MARKOV ADDITIVE PROCESSES Proof That { Lt} is a multiplicative functional follows from L8 ogt = h(Jt+s) es(St+ . 0 < qij < 1 and Bij [0] > 0.12 The expression for A means h(e) Aij = hie) Aij [1 + gij(Bij[0] i 0 j. (5. one can directly verify that (5. u Theorem 5.1) . if vi(dx) is compound Poisson.Qi < oo and Bi a probability measure. 0<b<oo.5. Here Oh(e) is the diagonal matrix with the h=e) on the diagonal. qij = r. in the discrete time case. Bi. then also vi (dx) is compound Poisson with e Ox ^i = /3iBi[0].. We omit the details. Bi(dx) = Bi(dx). ^i = of qij Bij [0] 1 + qij ( Bij [0] . 0<q<1. In the infinite case . Then the MAP in Proposition 5.
a = 0 in (5. this means that Ft[a] = etw ( e)Ohc) Ft[a + 9]oh (e) (5. v= . it follows that indeed the normalizing constant is H1 [0]. This shows that F. (dx) of a process with stationary independent increments follows from Theorem 4. this implies k[a] = A 1 ) (K[a + 0] . Jt = j] = hie) . Yi E dx..tc(0)' )Ah() = Oh(o) K[a + 0]Oh() . .8) h(. are probability measures .8.11.8) yields et'[a] = Ohie )et (K[a +e]K(e)I)Oh(°) By a general formula (A. First note that the ijth element of Ft[a] is etK(e)Ej [e(a+B)st E:[east Jt = j] = Ej[Lteas' .t. is absolutely continuous w. In matrix notation . Hence the same is true for H=j and H.e) Consider first the discrete time case .8).. (dx).tc(0)I. . Jl = j) = Ei[Lt. H1. Further Fib (dx ) = P=(YI E dx.46 CHAPTER II. Letting a = 0 yields the stated expression for A. Here the stated formula for P follows immediately by letting t = 1. F:j with a density proportional to eei .tc (') (0) corresponds to the stated parameters µ.r. since Hij. SOME GENERAL TOOLS AND RESULTS Proof of Theorem 5. v.13) for matrixexponentials . Jl = j] :(Yi E dx. Jt = A. Ji = j) h(e) eeyK(B)p h(8) h(e) eexK ( h=e) e)Fi.tc(') (8)/ d)ag h 7 Aiiii (Bii[a + 0] . Now we can write K[a] =A+A ) ( K[a + 0] .Bay [0]) That k(') (a + 0) . in continuous time ( 5.. Similarly.K [O])Oh(e) (0) l + ( A + (tc(') (a + 0) .
and is typically defective.1). 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].. [225]. 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).1) = Aij4ij(Bij[a] . Much of the pioneering was done in the sixties in papers like Keilson & Wishart [224]. has no mass on (oo.+ < x.3 for an infinite E are given by Ney & Nummelin [266]. an extensive bibliography on aspects of the theory can be found in Asmussen [16]. Write r+ = T(0) and define the associated ladder height ST+ and ladder height distribution by G+(x) = 11 (S.. oo). 6 The ladder height distribution We consider the claim surplus process {St } of a general risk process and the time 7. however. however. .6. [262] in discrete time. [226] and Miller [260]. see also Fuh & Lai [149] and Moustakides [264].e. 0]. i. Note that G+ is concentrated on (0. h. is slightly less general than the present setting. Though the literature on MAP's is extensive. Notes and references The earliest paper on treatment of MAP's in the present spirit we know of is Nagaev [265].6. which. [261]. 7+ < oo).7). For the Wald identity in Corollary 5. Conditions for analogues of Corollary 5.)Ajjgij(Bij[a+0] . THE LADDER HEIGHT DISTRIBUTION 47 Finally note that by (5.Bij[0]) = hjel)ijgijBij[0](Bij[a] . < x) = 11 (S. the literature on the continuous time case tends more to deal with special cases. hardly a single comprehensive treatment.(u) = inf {t > 0 : St > u} to ruin in the particular case u = 0 . there is.
at present we concentrate on the first ladder height. 0 f T+ (6. Also. For the proof of Theorem 6. = ST+(1) Figure 6. see Fig. Recall that B(x) = 1 .(3B(x ) = pbo(x) on (0. a fact which turns out to be extremely useful.5 below. the second ladder point is ST+(2) where r+(2) is the time of the next relative maximum after r+(1) = r+.1. In any case. there are only finitely many).i. On Fig. Here bo(x) _ B(x)/µB.ST+(1) and so on.1) The interpretation of R+(A ) is as the expected time {St} spends in the set A before T+. 6. g(y)R+(dy) = E f g(St)dt.2. To illustrate the ideas. 0]. 6..1 The term ladder height is motivated from the shape of the process {Mt} of relative maxima. The first ladder step is precisely ST+. o 00 (6. Thus. The main result of this section is Theorem 6.d.1. it follows that for g > 0 measurable.2) . the second ladder height (step) is ST+(2) . where basically only stationarity is assumed.e. by approximation with step functions . In other cases like the Markovian environment model. they have a semiMarkov structure (but in complete generality.48 CHAPTER K. Theorem 6 .T+ > t)dt = E f 0T+I(St E A) dt. G+ is given by the defective density g + (x) =. In simple cases like the compound Poisson model.e. the ladder heights are i. the sum of all the ladder steps (if rl > 0. the dependence structure seems too complicated to be useful).B(x) denotes the tail of B. SOME GENERAL TOOLS AND RESULTS M ST+(2) Sr. and the maximum M is the total height of the ladder. R+ is concentrated on (oo.1. which gives an explicit expression for G+ in a very general setting. i.. oo). i. has no mass on ( 0. we shall first consider the compound Poisson model in the notation of Example 1.00 ). 1 For the compound Poisson model with p = 01LB < 1. define the prer+occupation measure R+ by R+(A) = E f o "o I(St E A.
St S* t a Figure 6.O<t<T). 6.St<0. THE LADDER HEIGHT DISTRIBUTION Lemma 6 . 0 < t < T.ST<St. since the distribution of the Poisson process is invariant under time reversion.ST<St. 49 Proof Let T be fixed and define St = ST . has the same distribution as {St}o<t<T.2.0<t<T) = F(ST E A. see Fig.2 R+ is the restriction of Lebesgue measure to (00.2(b): r+ < t Thus. ST < ST_t.T+>T) = P(STEA.O<t<T) = P(STEA. That is. 0]. {St }o<t<T is constructed from {St}o<t<T by timereversion and hence. . P(STEA.6.ST_t.2(a): T+ > t Figure 6. 0 < t < T) P(STEA.
St).50 CHAPTER II.3 Lemma 6 .. But since St 4 oo a.y) (here we used the fact that the probability of a jump at u t is zero in the second step. That is. G+(A) = Q f 0 B(A . for A C (0.. oo).2) in the last).St _).t dT.St _)I(r+ > t). and (6. .T+ > t] dt 0 T+ _ /3E f g( St) dt = 0 f g(y) R+(dy) 0 00 where g(y) = B(A . we get G+ (A) = f 00 /3 dt E[B(A . U + St_ E A. this is just the Lebesgue measure of A.3 where the bold lines correspond to minimal values.y)R+(dy) 00 Proof A jump of {St} at time t and of size U contributes to the event IS.3 G+ is the restriction of /3R+*B to (0. oo). T+ > t] 0 _ /3 f E[B( A . The probability of this given { Su}u<t is B(A .r. Fig. Figure 6. E A} precisely when r+ > t. and since the jump rate is /3. 6. it follows that R+ (A) is the expected time when ST is in A and at a minimum at the same time . SOME GENERAL TOOLS AND RESULTS Integrating w. s. cf.
4 (the points in the plane are (ak . assuming basically stationarity in time and space. The marked point process . 6 . h]} /h (by stationarity. In the stationary case.* ) and the second the mark (the claim size Uk ). The traditional representation of the input sequence {(TT. 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.e. this does not depend on h). THE LADDER HEIGHT DISTRIBUTION 51 Proof of Theorem 6.. as a point process on [0. . {St+8 . .:T1 +•••+Tk <t}. Fig. i. Uk) (k = 1. oo) x (0.3 yields g+ (x) = .S8 )t> o = {St }t>o for all s > 0.. 0 Generalizing the setup. Nt St =>Uk k=1 t where Nt = max{k = O. we consider the claim surplus process {St }t>o of a risk reserve process in a very general setup..1 With r+(y) = I(y < 0) the density of R+.. i.. We call M * stationary if M* o B8 has the same distribution as M* for all s > 0. the first component representing time (the arrival time o.Q f r+(x . The sample path structure is assumed to be as for the compound Poisson case: {St*} is generated from interclaim times Tk and claim sizes Uk according to premium 1 per unit time. this is equivalent to the risk process {St*} being stationary in the sense of (6. cf. U k)} k=1 a is as a marked point process M *. obviously.z)B(dz) _ f I(x < z)B(dz) _ f (x).s > 0).4) are (ak. we define the arrival rate as E# { k : ak E [0 .T+ < oo).s. 6 . 2.1. The points in the plane (marked by x on Fig.) where ak = Ti + • • • + Tk .6.M o 08 shifted by s is defined the obvious way.4).e. Uk) for those k for which ak . Lemma 6. oo).
s. See. most often one takes h = 1). . SOME GENERAL TOOLS AND RESULTS M* U. We represent M by the sequence (Tk. where T is the first arrival time > 0 of M and h > 0 an arbitrary constant (in the literature.4 Given a stationary marked point process M*.. Assume {Jt} irreducible so that a stationary distribution 7r = (1i)iGE exists. As above .2. letting h J. h] and the sum approximately ^o(M*)I(ul < h). h] Eco(M*) = 1 E f co(M o Bt)dt. where TI = 0. 0. and let T = T2 denote the first proper interarrival time . V(M* o eak ). we define its Palm version M as a marked point process having the conditional distribution of M* given an arrival at time 0 . k: vk E [0. This more or less gives a proof that indeed (6. of (6. The two fundamental formulas connecting M* and M are Eco(M) = aE E. h.5) represents the conditional distribution of M* given vi = 0. Sigman [348] for these and further aspects of Palm theory.4 Consider a finite Markov additive process (cf. vi(dx) = .5) does not depend on h.e. o.52 CHAPTER II... Example 6 . e. i 1 U2 Us 1_ 0 or Q2 $ U3 *1 L 0 7 X I 11 1 Figure 6. = 0 . Oh becomes the approximate probability F(ri < h) of an arrival in [0.. i. Uk) k=1. the r. Section 5) which has pure jump structure corresponding to pi = a = 0.g.QiBi(dx). Note also that (again by stationarity) the Palm distribution also represents the conditional distribution of M* o Ot given an arrival at time t.
p. aij for (i. qij when {Jt} jumps from i to j and have mark distribution Bij.O fo "o F(x)dx = .OEU0.s.*(0) with initial reserve u = 0 is p = /3EU0.6 Under the assumptions of Theorem 6.p. the distribution of Ul) is the mixture B = E aii Bi + aij Bij J = j#i !i J..oo a. let the arrivals and their marks be generated by {Jt} starting from Jo = j. After that. the ruin probability . This follows by noting that iP*(0) = IIG+JI = J0 "o g+(x)dx = .6. THE LADDER HEIGHT DISTRIBUTION 53 Interpreting jump times as arrival times and jump sizes as marks. Then the ladder height distribution G+ is given by the (defective) density g+(x) = . Note in particular that the Palm distribution of the mark size (i. .= i.p.5. A stationary marked point process M* is obtained by assigning Jo distribution Tr. First choose (Jo_.OF(x). the probability aij of Jt . It follows that we can describe the Palm version M as follows .O for i # j. having the Palm distribution of the claim size and F (x) = F(Uo < x) its distribution . 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 . we get a marked point process generated by Poisson arrivals at rate /3i and mark distribution Bi when Jt = i. v. we note: Corollary 6.6iBi + Aijgij Bij j#i iEE iEE 0 Theorem 6 ./. and that p = 0EU0 < 1. and by some additional arrivals which occur w. 5 Consider a general stationary claim surplus process {St }t>o.O for i = j and iriAijgij/.e. Jo) w. Assume that St * . Before giving the proof. Jt = j is iri(3i /. If Jt_ = i. j) and let the initial mark Ul have distribution Bi when i = j and Bij otherwise. let U0 be a r. an arrival for M* occurs before time t + dt w.
oo)). The sample path relation between { Su } and { Su } amounts to S„ = St .(left limit) when 0 < it < t and is illustrated on Fig . in (oo.5). in (0. the arrival times 0 < 0'1 < Q2 < . A standard argument for stationary processes ([78] p. 2.0<u<t) = P(St EA. SOME GENERAL TOOLS AND RESULTS V` (0) = E E Uk k: ak E [0.0<u<t) = P(StEA..5. 0).St <. .. oo) x (0 . 0<u<t) = P(St EA. T+ = t given the event At that an arrival at t occurs .o<u<t where a claim arrives at time t and has size Uo. moves down linearly at a unit rate in between jumps and starts from S0 = U..e.o. { Su}0<u<t is distributed as a process {Su} . CHAPTER H.Su_ <0. It follows that for A C (0. The last property is referred to as insensitivity in the applied probability literature. 105) shows that one can assume w. Then clearly * G+ (A) = P(ST+ E A) = Consider a process { f p(t)f3dt.. are point processes on (oo . Let p(t) be the conditional probability that ST+ E A..54 By (6.Mt)..5.Q_k and has size U_ k. .o.). that M* and M have doubly infinite time (i.l.. (k = St}t>o 1. has a very simple interpretation as the average amount of claims received per unit time .Su< 0. h.St*_ u. and the kth preceding claim arrives at time t . Now conditionally upon At .St<Su.$St_ u.0<u<tIAt) = P(St EA. oo ) and the arrival times 0 > 0_1 > a_2 > . We then represent M by the mark (claim size ) Uo of the arrival at time 0.1] here the r . which makes an upwards jump at time . oo) p(t) = P(St EA..0<u<tIAt) = P(St EA. 6. the mark at time Qk is denoted by Uk.Su<0. .s. Proof of Theorem 6.g.A. 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.
NIt)dt .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. and since by assumption St * oo a. Fig.5..6.5 where it = { St < Su. time instants corresponding to such minimal values have been marked with bold lines in the path of { St}. cf. A sample path inspection just as in the proof of Lemma 6 . Thus. the support of L has right endpoint U0. 6. THE LADDER HEIGHT DISTRIBUTION 55 { A Su}0<u<t U0 U0 \t tt u>0 N U_1 Figure 6. 6. and we let L(dy) be the random measure L(A) = fo°° I(St E A. In Fig. 0 < u < t } is the event that { Su } has a relative minimum at t . the left endpoint of the support is oo.s. Uo]. t a oo. G' (A) = 3 f P(St E A. Since So = U0. Mt)dt = i3EL(A) o"o . 2 therefore immediately shows that L(dy) is Lebesgue measure on (oo.
56 CHAPTER II. [263] (a special case of the result appear in Proposition VI.1).2.6 is Bjork & Grandell [67]. [147].5 is due to Schmidt & coworkers [48]. 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. . A further relevant reference related to Corollary 6.
4. St = uRt = EUi t. and independent of {Nt}. • the claim sizes U1.e. U2. and assume that • { Nt}t>o is a Poisson process with rate j3. Panjer's recursion ( Corollary XI. {Rt} and the associated claims surplus process {St} are given by Nt Nt Rt = u+t EUi. i.. It is worth mentioning that much of the analysis of this chapter can be carried over in a straightforward way to more general Levy processes . • the premium rate is p = 1. i.6) and simulation methods ( Chapter X).4 below ..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. exact matrixexponential solutions under the assumption that B is phasetype (see further VIII. with common distribution B. For finite horizon ruin probabilities . Some possibilities are numerical Laplace transform inversion via Corollary 3. 3). say. i=1 i=1 An important omission of the discussion in this chapter is the numerical evaluation of the ruin probability. .d. A common view of the literature is to consider such processes as perturbed compound Poisson risk processes . being of the form Rt = Rt+Bt + Jt where {Rt } is a compound 57 . see Chapter IV. are i. Thus .
Rt. Dufresne & Gerber [126].1 is the expected claim surplus per unit time. and that B(k)[0] = Pak).g. [324].1). for (d) just note that the kth cumulant of St is tic(k) (0).1 (a) ESt = t(13µ$ . Write pB^) = EUn' YB = Pali = EU.t = t(p . and Schlegel [316].. Proof It was noted in Chapter I that p .58 CHAPTER III.s. Furrer [150].f. we shall start by giving the basic formulas for moments. we get Ee8st = 00 e8t c` Ee8 (U1+.t = E E [ U k k=1 k=1 Nt .1) . We do not spell out in detail such generalizations. say stable Levy motion.1). (c) Ee8St = et" (8) where c(s) = f3(B[s] .t = fltpB .'s etc. 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).t = E[Ntµs] . (b) Var St = t. m. cumulants . e . 0 . Schmidli [319]. of the claim surplus St . where K(k) (0) is the kth derivative of is at 0. and this immediately yields (a). THE COMPOUND POISSON MODEL Poisson risk process.6pBa).1) = t(p .. (d) The kth cumulant of St is tf3p(k) for k > 2.+Uk)P(Nt = k) k=O e8t k=O B[s]k . For (c). A more formal proof goes as follows: Nt r Nt ESt = E > U k . 1 Introduction For later reference.u .)3t (fit' k t} = etk(8) exp {st '3t + B[s]f Finally. See e.g. P = PAB = 1/(1 + rl) Proposition 1. {Bt} a Brownian motion and {Rt} a pure jump process.
1. (b) If 77 < 0.. The connections to random walks are in fact fundamental. (a) No matter the value of 77. (c) If 77 > 0.. Obviously. The right hand inequality in (1.3EU01 = 1µs where rt is the safety loading. For the proof. where Tk is the time between the kth and the (k . then lien inft. however. which is often used in the literature for obtaining information about {St} and the ruin probabilities. obviously 0(u) = F(maxk Sok > u). Proof We first note that for u. we have Sok .3) is proved similarly.. meaning that the increments are stationary and independent.1 = . and there are at least two ways to exploit this: Recalling that ok is the time of the kth claim.3 If nh < t < (n + 1)h.2 (DRIFT AND OSCILLATION) St/ta3'p1 ast >oo.V.d. For example. INTRODUCTION 59 The linear way the index t enters in the formulas in Proposition 1.. then St 00.S„ attains its minimal value when there are no arrivals in (u. then St> SnhV>Snhh. S„+V > S„ . In particular. the Uk . We return to this approach in Chapter V.i. . lim supt.1)th claim.4. if t = nh + v with 0 < v < h. v > 0. we need the following lemma: Lemma 1. The point of view in the present chapter is. Here is one immediate application: Proposition 1.Tk are i. Sn+0 . St = oo. cf.Tk. In this way. rather to view {St} directly as a random walk in continuous time. St = oo.1 is the same as if {St} was a random walk indexed by t = 0. 2. (d) If 17 = 0. Indeed. then Snh .1. so that {Sok } is a random walk with mean EUET = EU. we get a discrete time random walk imbedded in the claim surplus process {St}. II. and the value is then precisely v.Sok_l = Uk . u + v]. then St 4 co.h < St < S(n+1)h + h.
3. Corollary 1. and (b). THE COMPOUND POISSON MODEL Proof of Proposition 1. hence by induction i... Snh u = 00 (the lemma is not needed for (d)). h A similar argument for lim sup proves (a).4 The ruin probability 0(u) is 1 for all u when 77 < 0. lim supn_...1. Proof The case of 17 < 0 is immediate since then M = oo by Proposition 1. . {Snh}n=o. we get lim inf St t>oo t nroo nh<t<(n+1)h t = lim inf inf St h l++m of Sn 7t h = ESh = p .p.6 Often it is of interest to consider size fluctuations.s. u 307). There is also a central limit version of Proposition 1.2.60 CHAPTER III.3.1).2. Thus using Lemma 1. this case can be reduced to the compound Poisson model by an easy operational time transformation u T1(t) where T(s) = )3 fo M(t)dt. and hence by the strong law of large numbers. h. Remark 1 . Notes and references All material of the present section is standard. is a discretetime random walk for any h > 0.. Snh/n a4' ESh = h(p ... it suffices to prove 4'(0) = F(M > 0) < 1. 0 Snh = 00. or by a general result on discrete skeletons ([APQ] p. if P(M > 0) = 1. and hence it folz lows from standard central limit theory and the expression Var(St) = tf3pB (Proposition 1.5 The limiting distribution of St .t . where the size of the portfolio at time t is M(t).o. However. Assuming that each risk generates claims at Poisson intensity /3 and pays premium 1 per unit time. This contradicts u St400. If rl > 0.1) as t 4 oo is normal vtwith mean zero and variance )3µsz) Proof Since {St}t>o is a Levy process (a random walk in continuous time). it is seen that upcrossing occurs at least twice. Considering the next downcrossing (which occurs w..1..1.. Part (d) follows by a (slightly more intricate) general random walk result ([APQ].2: Proposition 1. at least once. is a discrete time random walk. {Snh}n=o._. 169) stating that lim infra. (c) are immediate consequences of (a). p. and < 1 for all u when 77 > 0. then {St} upcrosses level 0 a. The general case now follows either by another easy application of Lemma 1. 2h. For any fixed h.1(b)) that the assertion holds as t 4 oo through values of the form t = 0. 1 since St 4 oo) and repeating the argument.
we may view the ladder heights as a terminating renewal process and M becomes then the lifetime. 0 Alternatively. (2.34 or A. As a vehicle for computing tIi(u). Note that this .2. Fig. and we further get information about the joint conditional distribution of the surplus and the deficit. and we shall here exploit the decomposition of the maximum M as sum of ladder heights. Combined with i/i(u) = P ( M > u).6.. that r(0) < oo) is Bo: taking y = 0 shows that the conditional distribution of (minus) the surplus ST(o). The decomposition of M as a sum of ladder heights now yields: 00 Theorem 2 .1. the ladder heights are i.. [APQ] Ch. p < 1. n=0 (2. 11. the formula for the distribution of M follows .1 provides a representation formula for 0(u).P) E PnBon(u) .e. It is crucial to note that for the compound Poisson model. Theorem 2.IIG+II) (the parenthesis gives the probability that there are no further ladder steps after the nth ). where G+ is given n=0 by the defective density g+ (x) = 3B (x) = pbo(x) on (0. This follows simply by noting that the process repeats itself after reaching a relative maximum. equivalently. Note that the distribution B0 with density bo is familiar from renewal theory as the limiting stationary distribution of the overshoot (forwards recurrence time ). The 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. We assume throughout rl > 0 or. 1e. The expression for g+ was proved in Theorem 11. 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}. B(x)/aB. we can rewrite the PollaczeckKhinchine formula as 00 (u) = P (M > u) = (1 . 1 The distribution of M is (1. Thus .1) is not entirely satisfying because of the infinite sum of convolution powers. Summing over n. oo ).1) representing the distribution of M as a geometric compound. d.IIG +II)EG+ .6. but we shall be able to extract substantial information from the formula. cf. which we henceforth refer to as the PollaczeckKhinchine formula. Here bo(x) _ Proof The probability that M is attained in precisely n ladder steps and does not exceed x is G+ (x)(1 . i.1. IV. nevertheless.
see Schmidli [323] and references there. ST(o) > y. Beekman [61]. the form of G+ is surprisingly insensitive to the form of {St} and holds in a certain general marked point process setup. cf. 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. ST(o )) given r (0) < oo is the same as the distribution of (VW. (c) the marginal distribution of ST(o)_ is Bo . Theorem 2 . 1) and W has distribution Fw given by dFyy/ dB(x) = x/µB.6.5. cf. 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). In the risk theory literature. 2 The joint distribution of (ST(o )_. (a) 11 (ST(o)_ > x.6. Feller [143] or Wolff [384].62 CHAPTER III. 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. The proof of Theorem 11. [62]. in this setting there is no decomposition of M as a sum of i. We assume rt > 0 throughout. and the conditional distribution of ST(o)_ given ST(o)_ = z is Bo z) The proof is given in IV. . see for example [APQ].just after ruin. As shown in Theorem 11.i. ladder heights so that the results do not appear not too useful for estimating 0(u) for u>0. 7r(0 ) < oo) = Q 3 Special cases of the PollaczeckKhinchine formula The model and notation is the same as in the preceding sections. Theorem 2.5. there is a general marked point process version. However. V is uniform on (0. f +b (b) the joint distribution of (ST( o). For the study of the joint distribution of the surplus ST(u)_ just before ruin and the deficit ST(„). ST(o)) is given by the following four equivalent statements: B(z) dz.1 is traditionally carried out for the imbedded discrete time random walk. the PollaczeckKhinchine formula is often referred to as Beekman 's convolution formula.V)W) where V. W are independent. (1 . where it requires slightly more calculation.d. Theorem A1.2(a) is from Dufresne & Gerber [125]. cf. Asmussen & Schmidt [49]. (d) the marginal distribution of ST(o)_ is B0.
1 e ax = n1 (n . But claims are exponential . then. however .e. Let r ( x) be the failure rate of M at x > 0. the result follows . 1 . also be seen probabilistically without summing infinite series . Integrating from u to oo.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. Alternatively. the formula for P(O) holds in a more general setting. a further relevant reference is Bjork & Grandell [67].6.1)1 00 ( 1 .. I.p) E pn S n x n. use Laplace transforms. B0 is exponential with rate S and the result can now be proved from the Pollaczeck Khinchine formula by elementary calculations . 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. Thus .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.2 If B is exponential with rate S.3.p. 0 .O)e(b0)x.p)pSe a ( l v)x = p( S . The result can. Bon is the Erlang distribution with n phases and thus the density of M at x > 0 is (1 . For a failure at x. and hence this overshoot has the same distribution as the claims themselves . Thus r(x) = S(1 .0(u) = pe(aA)" Proof The distribution Bo of the ascending ladder height ( given that it is defined ) is the distribution of the overshoot of {St} at time r+ over level 0.p. the current ladder step must terminate which occurs at rate S and there must be no further ones which occurs w. 3b Exponential claims Corollary 3. As shown in 11.p) = S .
and conditioning upon S.1) For a heavytailed B.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 .p + G+ * Z(u) = 1 .64 CHAPTER III. u + oo. the survival probability Z(u) = 1 .+ >u. A variety of proofs are available .y)f3 (y) dy. T+ <00) (3.4) zu P(M > u . THE COMPOUND POISSON MODEL In VIII. (3.4) is similar (equivalently. 0 Proof Write o (u) as P(M>u) = P(S. Then the first term on the r.y)G+(dy ) = f U V(u . E.3) below.3).s.T+ <oo)+P(M> u.3)).3) Equivalently. Corollary 3.3.1 p pBo(u).1.i(u) satisfies the defective renewal equation Z(u) = 1 .T+ <oo). (Example VIII.+ <u. u . (3. 2 is one of the main classical early results in the area.y)G+(dy) For the last identity in (3.S.g. 3c Some classical analytical results Recall the notation G+(u) = f^°° G+(dx).+ = y yields P(M>u. if 3 = 3 and B is a mixture of two exponential distributions with rates 3 and 7.S.h. The case of (3. (u) 35eu + 35e6u. just insert the explicit form of G+.4) can be derived by elementary algebra from (3.2). (3. and weights 1/2 for each. II.2) Notes and references Corollary 3.y)/3B (y) dy. (3. then 24 1 V. We mention in particular the following: (a) check that ip (u) = pe (60)u is solution of the renewal equation (3. is ?7+ ( u). cf. we use the PollaczeckKhinchine formula in Chapter IX to show that b(u) .3.p + f u Z(u . (b) use stopped martingales .+ <U. we show that expression for /'(u) which are explicit (up to matrix exponentials) come out in a similar way also when B is phasetype.
f.g.5 The first two moments of M are 2 EM . 0 Notes and references Corollary 3.8) Proof This can be shown. which yields the survival probability as 00 f u }t Z(u) = f f3eRtdt 0 from which (3. see e.P)PB 2(1 .s . eau B(u) du = f PB 3PB SPB 0 o (3.7) and Corollary 3. [APQ] pp. e. Embrechts.4) can be derived by elementary but tedious manipulations. In fact. . The approach there is to condition upon the first claim occuring at time t and having size x . either of these sets of formulas are what many authors call the PollaczeckKhinchine formula. Of course. We omit the details (see.p)s .3 .P)pB' (3.Ee8M) f ao e8' ( u)du = a8uP (M > u)du = 0 o 1 ( 1+ (1 . 206207).3 is standard . Some relevant references are Abate & Whitt [2].p = (1 . by analytical manipulations (L'Hospital's rule) from (3. g.(3 . numerical inversion of the Laplace transform is one of the classical approaches for computing ruin probabilities.3 .5) Proof We first find the m.3..(3B[s] 1 .4 The Laplace transform of the ruin probability is 65 fo Hence Ee8M 00 e8uiP(u)du . [APQ] pp.1 Bo[s] = f oc.5). Griibel [179] and Thorin & Wikstad [370] (see also the Bibliographical Notes in [307] p. (3. for example. Corollary 3. 111112 or Feller [143].p)2 3(1 ./3B[s] .g.7). Also (3.Ps s(.5). Bo of B0 as m e8u B(du) = B[s] ./3B[s] which is the same as (3.. In view of (3.5 can be found in virtually any queueing book. it is not surprising that such arguments are more cumbersome since the ladder height representation is not used.7) s +.p) E p"Bo[s]" = 1 .)3B[s]) (3. Griibel & Pitts [132].s .6) 00 = (I .PPB2) EM2 = PPB) + QZPBl 2(1 . SPECIAL CASES OF POLLACZECKKHINCHINE Corollary 3. 191).pBo[s] no (1 .p)s s /3 .
u)]k k! k0 The renewal equation (3. we may assume p = 1 so that the stated formula in terms of the survival probability Z(u) = 1 .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./3Z(u . Assume (3.h. Z^ =eR(k.9) shown for n .1).Q (k 1 k= n  [O(k .3+ 18+ J0 Z(uy).Q) k=0 k! E e0( = /32(u) .u) [p(k .)3(1 . For n < u < n + 1.u)]k d 1 u) _ a) n ( du ( k! (1  .u)]k1 ku+1) [/3( k .1)! k=1 u1 .u) [N(k .Q) 3e. differentiation yields Z'(u) _ /3Z(u) which together with the boundary condition Z(0) = 1/3 yields Z(u) _ (1/3) eAu so that (3.4) for Z( u) means f lhu Z(u) = 1. THE COMPOUND POISSON MODEL 3d Deterministic claims Corollary 3.6 If B is degenerate at p. then p) 1: ep(k u/.9) follows for 0 < u < 1.1). of (3. differentiation yields Z(u) _ /3Z(u) .3I( 0<y<1)dy Z(y)/3I(0<u.9)./32(u .u + 1 )]k = QZ(u) .z/'(u) takes the form Z(u) L^J L. . eO('u) [)3(k .u/p)]k ko k! Proof By replacing {St} by {Stu/p} if necessary.u) a)Qea" + (1 .1 < u < n and let Z(u ) denote the r.u)]k k! (1 L3) 1: e_O(ku) NIN (k (k .66 CHAPTER III.s.
g. co(a) = rc(a + 9) .2) (Here 9 is any such number such that r. F and c. say t = 1: recall from Proposition 1. See also Iversen & Staalhagen [208] for a discussion of computational aspects and further references.1 that c(a) = /3(B[a] . (4. We could first tentatively consider the claim surplus X = St for a single t.d.1) or equivalently.3) by t.) The adaptation of this construction to stochastic processes with stationary independent increment as {St} has been carried out in 11.2) shows that the solution is Ox [O]0].Qe(Bo[a] . we just have to multiply (4.2).f.6 is identical to the formula for the M/D/1 waiting time distribution derived by Erlang [139]. The question then naturally arises whether ie is the c.3B[9].f. 0 Notes and references Corollary 3.4) works as well.g.4. corresponding to a compound Poisson risk process in the sense that for a suitable arrival intensity 00 and a suitable claim size distribution BB we have no(a) = rc(a + 9) . of F9. in terms of the c. we set up .3B = .(9). CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 67 Since Z(n) = 2(n) by the induction hypothesis. and define rce by (4.4. but will now be repeated for the sake of selfcontainedness.f. B9(dx) = B[9] B(dx).(9) is welldefined.r.rc(9) = .4) .1) . or equivalently BB[a] = B[^+ Repeating for t 54 1.a.g. it follows that Z(u) = 2(u) for n<u<n+1.1) .f. and thus (4. (4. Formalizing this for the purpose of studying the whole process {St}.a. K(a) = logEe'X = 109f 00 eaxF(dx) = logF[a]. 00 the standard definition of the exponential family {F9} generated by F is FB(dx) = e°xK(e)F(dx). 4 Change of measure via exponential families If X is a random variable with c. (4. The answer is yes: inserting in (4.
the PBT) are mutually equivalent on.5) for the density of n i.68 CHAPTER III.0e and claim size distribution Be. (4.d. with T taking the role of n) is the analogue of the expression exp{8(x1 + • • • + xn) .Tic (0)} . Xn). then EBZ = E [Ze9ST _T"(9)I .6) F(G) = Po (G) = EB [exp {BST + Ttc(0)} . The following result (Proposition 4. .1 Let P be the probability measure on D[0. in particular the expression (4.7) now follows by taking Z = eBST+TK(e)I(G) u Theorem 4 .. EeeBSt + tk(B) = 1. v(Xi. and thus (4. the corresponding expectation operator is E9. and PBT) the restriction of PB to FT. Z is measurable w.7) Proof We must prove that if Z is FTmeasurable. oo) governing a given compound Poisson risk process with arrival intensity.8) By standard measure theory. Then P(G) = Fo(G) = EB [exp {BST + TK(O)} . ..t.nr. (4. Then FB denotes the probability measure governing the compound Poisson risk process with arrival intensity.FTn) = Q(SkTIn : k = 0.g.5) for the density. .4). G C {T < oo}. The identity (4. and dP(T) dP^T) That is.2.10) . .(9)} (4. Ti(a)/n.1) and multiply from 1 to n).2 For any fixed T. Proposition 4. G]. Let FT = o(St : t < T) denote the o•algebra spanned by the St.8) follows by discrete exponential family theory. replications from Fe (replace x by xi in (4. But let Xk = SkT/n .t.r. (4. n) for a given n.S(k_1)Tln. and define 09.9) Proof We first note that for any fixed t. t < T.r.3 and claim size distribution B. (4. .d. with common c.1.3 Let T be any stopping time and let G E FT.. THE COMPOUND POISSON MODEL Definition 4. Then the Xk are i.f. for G E FT. it suffices to consider the case where Z is measurable w. = exp {BST .FT. BB by (4.i. . (4. G].i.
.r is deterministic. Ee [exp { BST +Trc(9)} I(G) FT)] = 1. the typical shape of rc is as in Fig.7 1 Some discussion further supporting this statement is given in the next section.r)rc(9)}I .1 It is seen that typically) a ry > 0 satisfying 0 = r. so that PG = EeE0 [exp { 9ST+Trc(9)}I(G)I FT)] = Ee [exp { BST + rrrc(O)} I(G)EB [ exp {9 (ST .ST) + (T . Letting T t oo and using monotone convergence then shows u that (4. (0) + rc'(0)a = 0 + ES1 a = a (p .(Y) = 13(B['Y] . GT C_ Jr < T}.g.10).7) holds. . Then G E FT.5.1) _ 1 + a. Thus. Then GT = G n Jr < T} satisfies GT E FT.1(a). t = T .FT]] = EB [exp { BST + Trc(9)} I(G)] . The behaviour at zero is given by the first order Taylor expansion c(a) r. Thus by (4.9) holds with G replaced by GT. (4. Given FT.9) holds for G as well. 5 Lundberg conjugation Being a c.1) . 77 Thus. c(a) is a convex function of a. LUNDBERG CONJUGATION 69 Now assume first that G C Jr < T} for some deterministic T. and hence (4. Now consider a general G. subject to the basic assumption ij > 0 of a positive safety loading. (a) rc (a) (b) KL(a) 'Y 'Y Figure 5.f. 5. according to what has just been proved.
the Lundberg exponent.3.2 s As support for memory. Thus B[7] = 6/. . an equivalent version illustrated in Fig. (5. 5. u It is a crucial fact that when governed by FL.2) 7 Figure 5. Fig.1) .3) cf.1(b).3.s). and (4. Note that KL (a) = /L (BL [a] . Fig.70 CHAPTER III. Example 5 .1) (or (5.g. (5. e. b[s] = 5/(b .1(b).1) is precisely what is needed for one of the terms in the exponent . An established terminology is to call y the adjustment coefficient but there are various alternatives around.2 is B(7) = 1 + ^. G = {T(u) < oo} in Theorem 4.QL instead of /37 and so on in the following .4) ELS1 = #L(0) cf. It is then readily seen that the nonzero solution of (5.4) yields /3L = b and that BL is again exponential with rate bL =. Taking T = r(u). Equation (5. (5.2)) is 7 = 5/3. Thus. the claim surplus process has positive drift > 0. 5. we write FL instead of F7.1) is known as the Lundberg equation and plays a fundamental role in risk theory .3.1 Consider the case of exponential claims. Lundberg conjugation corresponds to interchanging the rates of the interarrival times and the claim sizes.a = i(a + 7). THE COMPOUND POISSON MODEL exists . we further note that ( 5. 5.
(oo) (in the sense of weak convergence w.+ E A} in Theorem 4. where C .Ce7u as u 4 oo.1 (5.G+ L) (x) G+L) (x) IL(+) µ+L) L) where G+L) is the FL. see A . Proof Just note that e(u) > 0 in (5.7) 0 and all that is needed to check is that ( 5. T(u) < oo] .G+L)(x)) dx ry^+L) J 00 f 0 (1 . G = {S. (5.ascending ladder height distribution and µ+ its mean. (5. take first 0 = ry. we therefore have ELe7t(u) + C where C ELe7 (00) = µ+) f e7(1 . Since a7' is continuous and bounded.u be the overshoot and noting that PL(T(u) < oo) = 1 by (5. 0 Theorem 5 .3 (THE CRAMERLUNDBERG APPROXIMATION) i'(u) .5. (5.8) .5). Then P(ST+ E A) = EL [exp { 7S?+} .1.e7x)G+(dx). e(u) has a limit i.3. LUNDBERG CONJUGATION 71 to vanish so that Theorem 4. To this end. V) (u) = P(T(u) < oo) = EL [exp {ryS.P Y j o' xeryxOB (x) dx /3k [Y] .1e. PL ) with density 1 .3 takes a particular simple form.7) is the same as (5.2 (LUNDBERG'S INEQUALITY) For all u > 0. ST+ E A] .t.4).1p .6) Proof By renewal theory. V)(u) < e7u.6 ). Letting e(u) = ST(u) . which shows that G(L) (dx) = e7xG +(dx) = e7x /3 (x) dx. T = T+.(u)} .5) Theorem 5 .r. we can rewrite this as 0(u) = e"ELe7^(u).
4 Consider first the exponential case b(x) = Seax.1 .1)) and 7µ+L) = 'y/3 [7] 7 1/0 = /3B ['y] . we get L where 00 (1 .11) so that I 7B ['Y](B[7]1) BI [7]Q VP (7) 72 7 (using (5. A direct proof of C = p is of course easy: B ['y] d S S (S7 )2 d7S y S 02' C 1p 1p _ 1p /3B' [7] 2 1 P1 p.12) Example 5 .1) (5.1 = ^7 The accuracy of Lundberg's inequality in the exponential case thus depends on how close p is to one. but some tedious (though elementary) calculations remain to bring the expressions on a final form. Then 0(u) = pe(a_Q)u where p = /3/S. u .10) VW = JI c* e° (x) dx = a (B[a] .e.3 (this was found already in Example 5.8) yields +L) J0 xel'B ( x) dx (5. Noting that SIG(L)II = 1 because of (5. this solves the problem of evaluating (5.1 above) and that C = p. of course. (5. or equivalently of how close the safety loading 77 is to zero.")G + (dx ) = 1  J0 00 3B(x) dx = 1p.4). From this it follows. that 7 = S . Using (5.72 CHAPTER III.7). . THE COMPOUND POISSON MODEL In principle.
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)
.
numerical evidence (e. or equivalently that /3 is only slightly smaller than /3max = 1/µ8.1 As /3 f Nmax.b(u) = p*e.80 CHAPTER III. cf./3)] 1 .1.8µBo  Ss' u where 6 = µB/µBo = 2µa/µB 2) . (/3max . we have according to the PollaczeckKhinchine formula in the form (3.s(/3max .1.3 and Corollary 3.P . Proposition 1.)3 )PBo µB  . Mathematically. That is./3)PBo PB .3 )1 } _ 1p 1 . Notes and references The approximation (7. THE COMPOUND POISSON MODEL These three equations have solutions 9/3µB2)3 30µa2)2 3µa2) (3) P+ (3) ' 0 . Though of course it is based upon purely empirical grounds.p . Grandell [171] pp. the premiums exceed only slightly the expected claims. but has an obvious interpretation also in risk theory: on the average. the approximating risk process has ruin probability z. we shall represent this situation with a limit where /3 T fl but B is fixed. 7c The heavy traffic approximation The term heavy traffic comes from queueing theory.p = (/3max 0)µB.2) was suggested by De Vylder [109]. 1924./3)M converges in distribution to the 2a exponential distribution with rate S = B' Proof Note first that 1 .g. and hence the ruin probability approximation is b(u) e(bAln)u.7) that Ee$(Amex /j)M _ 1p _ 1p Eo [s (0max 1 .Ps(/3max . Letting Bo be the stationary excess life distribution. p* _ . [174]) shows that it may produce surprisingly good results.PBo [s (/3max . heavy traffic conditions mean that the safety loading q is positive but small.3* /S. Proposition 7.p + p { 1 1p ti 1 .(3)2 P PB 2µB 2µB Letting /3* = /3/P.(bA*)".
It is worth noting that this is essentially the same as the approximation (2) z/i(u) Ce. In the setting of risk theory. but has an obvious interpretation also in risk theory: on the average . or equivalently that 0 is small compared to µB . then P(u) 4 e6„ Proof Write z'(u) as P((/3max .4) suggested by the Cramer Lundberg approximation and Proposition 6. [APQ] Ch. .1 1 .ze a2unµB laB (7. VARIOUS APPROXIMATIONS FOR THE RUIN PROBABILITY 81 Corollary 7. Mathematically.2.p _ 2rl11B PB p. 2 provides the better mathematical foundation. light traffic conditions mean that the safety loading rl is positive and large .Q T /3max.ryu .p. obviously Corollary 7. The present situation of Poisson arrivals is somewhat more elementary to deal with than the renewal case (see e . That is . light traffic is of some interest as a complement to heavy traffic . as well as it is needed for the interpolation approximation to be studied in the next subsection./3)u * v.7. Notes and references Heavy traffic limit theory for queues goes back to Kingman [230]. Numerical evidence shows that the fit of (7.B AB ) 6()3max _'3) = However . in risk theory heavy traffic is most often argued to be the typical case rather than light traffic .g. the first results of heavy traffic type seem to be due to Hadwiger [184].l3)M > (/3max . However . while the approximation may be far off for large u.2 If .0)u. These results suggest the approximation Vi(u) e6(0_. Of course. This follows since rl = 1/p . 7d The light traffic approximation As for heavy traffic . VIII). u * oo in such a way that (3max . and hence 2µ2B 1 .3) is reasonable for g being say 1020% and u being small or moderate. the term light traffic comes from queueing theory. the premiums are much larger than the expected claims . we shall represent this situation with a limit where 3 10 but B is fixed./3)u). We return to heavy traffic from a different point of view (diffusion approximations) in Chapter IV and give further references there .
( 3 J O B dx. En'=2 • • • = O(/32) so that only the first terms matters.Q limIP ( u) + Q lim z/'(u) Amax &0 amax ATAm. Again.T > u) = J o" B(x + u)/3eax dx .T > u).u)+. 0 u Notes and references Light traffic limit theory for queues was initiated by Bloomfield & Cox [69]. Sigman [347]. . Omax max m. ao n=1 00 n=1 (u) P) anllBBon(U) onPaBon(u) • Asymptotically. and hence 00 (U) /3pBBo (u) = 0 / B(x)dx. cf. u Note that heuristically the light traffic approximation in Proposition 7. by monotone time T of the first claim . the Poisson case is much easier than the renewal case.5) follow by integration by parts.3 is the same which comes out by saying that basically ruin can only occur at the F(U . 10 ( u The alternative expressions in (7. (7.= 1 aJ 1 a 0+ 1 = = p. 0(u) /3 J B(x)dx = /3E[U .u. Asmussen [19] and references there. Another way to understand that the present analysis is much simpler than in these references is the fact that in the queueing setting light traffic theory is much easier for virtual waiting times (the probability of the conditioning event {M > 0} is explicit) than for actual waiting times . U > u] = /3iE(U .(3 10. THE COMPOUND POISSON MODEL Proposition 7. 7e Interpolating between light and heavy traffic We shall now outline an idea of how the heavy and light traffic approximations can be combined.3 As . see Daley & Rolski [96].e. For a more comprehensive treatment.82 CHAPTER III. [97]. The crude idea of interpolating between light and heavy traffic leads to 0 (u) C1 . z/' (u) convergence P(U . Indeed.5) u Proof According to the PollaczeckKhinchine formula. i. Light traffic does not appear to have been studied in risk theory.
^ LT Q maxQ m"^ Qlo V LT) ( CHT(v) (say). with rate 1/µB = /3max. Substituting v = u(. we see that the following limits HT) (u'). one may hope that some correction of the heavy traffic approximation has been obtained. 8 Comparing the risks of different claim size distributions Given two claim size distributions B(1). (U).VHT) ( ax QmQ ) h (B) ( . Al . even if the safety loading is not very small. however.O0 M. [84].x ./3)) .O(E)(u) 1 (1 . _(E) (u). ) M. "/Qmex Cu) CLT(u ( /3max 0) + O16 CHT( U(Qmaz . COMPARISONS OF CLAIM SIZE DISTRIBUTIONS 83 which is clearly useless .3 and use similar notation for %(B) (u) = (u). the idea of interpolating between light and heavy traffic is due to Burman & Smith [83 ]. Another main queueing paper is Whitt [380]. that is. Instead. available. where further references can be found . The adaptation to risk theory is new. Thus . no empirical study of the fit of (7. ^IE) exist: 1 (B) HT QmsxQ hm J e e6" 2µE/µE2)'" = e(1 6)" =  Q1Qm.6) (1p) The particular features of this approximation is that it is exact for the exponential distribution and asymptotically correct both in light and heavy traffic.3). (7. z/i(E) (u) = pe(QmaxQ)u. Let OLT) (u) denote the light traffic approximation given by Proposition 7.6) is .3n. 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. to get nondegenerate limits . B(2). .8. Notes and references In the queueing setting . . 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).Wmax f(x ) dx + pee6mQ.
A weaker concept is increasing convex ordering: B(1) is said to be smaller than B(2) (in symbols. we have the convex ordering. we shall need various ordering properties of distributions. In particular (consider the convex functions x and x) the definition implies that B(1) and B(2) must have the same mean. for more detail and background on which we refer to Stoyan [352] or Shaked & Shantikumar [337]. Taking probabilities. In terms of the time to ruin. Proof According to the above characterization of stochastical ordering.2 If B(') <j. B(') <d B(2)) if B(1)(x) < B(2)(x) for all x. u Of course. then Bill = B(2). U(2) such that U(l) has distribution B('). an equivalent characterization is f f dB(') < f f dB (2) for any nondecreasing convex function f. 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)) in the increasing convex order if f BM (y) dy < f 00 Bi2i (y) dy x x for all x. the proof is complete. . Proposition 8. Recall that B(') is said to be stochastically smaller than B(2) (in symbols. Rather than measuring difference in size. B(') <i.1 is quite weak. XI. this ordering measures difference in variability.s.ill(u) < V)(2) (U) for all u. Finally. most often the term stoploss ordering is used instead of increasing convex ordering because for a given distribution B. one can interpret f x°° B(y) dy as the net stoploss premium in a stoploss or excessofloss reinsurance arrangement with retention limit x. then i. or the existence of random variables U(l).' 1)(u) < V)(2) (U) for all u. B(2)) if f fdB(1) < f fdB(2) for any convex function f. In the literature on risk theory. whereas (consider x2) B(2) has the larger variance. B(2) and PB(1) = µB(2). Here convex ordering is useful: Proposition 8. then . U(2) distribution B(2) and U(1) < U(2) a. Bill is said to be convexly smaller than B(2) (in symbols.1 If B(') <d B(2). equivalent characterizations are f f dB(') < f f dB (2) for any nondecreasing function f. we can assume that 1) < St 2l for all t. THE COMPOUND POISSON MODEL To this end. cf.84 CHAPTER III. 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). Proposition 8.6. B(' <.
B(2). The heavy traffic approximation (7. A first attempt would of course be to identify 'variation' with variance. say to p.. then /'(')(u) < 0(2)(u) for all u. and here is one more result of the same flavor: Corollary 8. from which the result immediately follows. (D) (u) < O(B) (U ) for all u. we have Bol) (x) f ' B(1) (y) dy < ' f' B(2) (y) dy = Bo2) (x)• µ 85 I.5 If '0(1)(u) < p(2) (U) for all u and a.1. Example 8.p ) E /3"µ"Bo2)* n(u) _ V(2) (u) n=1 = Corollary 8. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS Proof Since the means are equal. B.3 provides another instance of this. then B(1) <.1 and µB at 1 so that the safety loading 11 is 10%. Proof If f is convex.4 Let D refer to the distribution degenerate at 'LB . .3 If B(1) <. we have by Jensen 's inequality that E f (U) > f ( EU).(1) (. 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.4) certainly supports this view: noting that.u) = (1 _ P) E /3npnBo( 1):n(u) n=1 00 < (1.8. A general picture that emerges from these results and numerical studies like in Example 8. larger variance is paramount to larger second moment. it is seen that asymptotically in heavy traffic larger claim size variance leads to larger ruin probabilities. B(2). Then V. A partial converse to Proposition 8.6 below is that (in a rough formulation) increased variation in B increases the risk (assuming that we fix the mean). Bo1) <_d Bo2) which implies the same order relation for all convolution powers. Proof Consider the light traffic approximation in Proposition 7.e. and consider the following claim size distributions: B1: the standard exponential distribution with density ay.6 Fix /3 at 1/1. Corollary 8..2 is the following: Proposition 8. Hence by the PollaczeckKhinchine formula . This u implies that D <. with fixed mean. The problem is to specify what 'variation' means.
B3: the Erlang distribution with density 4xe2x. One then obtains the following table: U005 U0. Let ua denote the a fractile of the ruin function.) = a. Note to make the figures comparable. Pellerey [287] and (for the convex ordering) Makowski [ 252]. and consider a = 5%.4.1358. We return to ordering of ruin probabilities in a special problem in VI.e. For B1i B2. [166].86 CHAPTER III. with the hyperexponential distribution being more variable than the exponential distribution and the Erlang distribution less. i.01%. 0.4142. 1/)(u. A standard example from queueing theory is .9A2e'2r where A.. in comparison to B2 the effect on the ua does not show before a = 0. B4: the Pareto distribution with density 3/(1 + 2x)5/2.0' U0. B3 the comparison is as expected from the intutition concerning the variability of these distributions. which appears to be smaller than the range of interest in insurance risk (certainly not in queueing applications!).1%. However. 11 Notes and references Further relevant references are Goovaerts et al.01%. 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). we have 0r3 = 2 < or2 = 1 < 02 = 10 < 04 = 00 so that in this sense B4 is the most variable. 9 Sensitivity estimates In a broad setting. all distributions have mean 1. THE COMPOUND POISSON MODEL B2: the hyperexponential distribution with density 0. B. van Heerwarden [189]. and this is presumably a consequence of a heavier tail rather than larger variance. sensitivity analysis (or pertubation analysis) deals with the calculation of the derivative (the gradient in higher dimensions) of a performance measure s(O) of a stochastic or deterministic system. A2 = 3.e'\1x + 0.001 u0. = 0.000. the behaviour of which is governed by a parameter 9. 0.lA. Kluppelberg [234]. 1%. In terms of variances o2.
u Proposition 9.1. t].e. we may be interested in a'/ap for assesing the effects of a small change in the premium. say estimated from data. Thus at p = 1.19P a/ . 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. For example. Thus. and s(9) the expected sojourn time of a customer in the network. and hence a _ e(60)u + u e(60)u = ( i + which is of the order of magnitude uV. where the partial derivatives are evaluated at p = 1. Proof This is an easy time transformation argument in a similar way as in Proposition 1. a0 as ao 80 19P . where Q2 = fl ( l2 1113 / _ Ou2v)2. Then a p ao = 00 Qa/. Then ib = Pe(613)u. In particular .(u) for large u. or we may be interested in aV)/0/3 as a measure of the uncertainty on '0 if 0 is only approximatively known.3. and hence the effect of changing p from 1 to 1 + Ap corresponds to changing /3 to /3/(1 + Op) /3(1 . the standard deviation on the normalized estimate ^/1' (the relative error ) is approximatively . Similar conclusions will be found below. a/3 0 . it follows that ' is approximatively normal N(0.Ap). increasing in u. Example 9. i. Then if t is large . obtained say in the natural way as the empirical arrival rate Nt/t in [0. with 0 the vector of service rates at different nodes and routing probabilities. SENSITIVITY ESTIMATES 87 a queueing network.2 Consider a risk process { Rt} with a general premium rate p. Let R(P) = Rtli. a2/t).9.01/2u. while /3 = j3 is an estimate. In the present setting. the premium rate p and the claim size distribution B. Then the arrival rate /3(P) for { R(P) } is )31p.. Assume for example that 8 is known. the distribution of %3 0 is approximatively normal N(0„ Q/t).1 Consider the case of claims which are exponential with rate 8 (the premium rate is one). if = a e(6A)u.
Similar notation for partial derivatived are used below. x > 0 (9. In the case of the claim size distribution B.5) are similar.88 CHAPTER III. this intuition is indeed correct.()(0 +'0) ' (9 . Proposition 9. () Proof According to (9. /3 yields w e(e + Y. but must look for approximations for the sensitivities 0. However .2) (see Remark 9. u Now consider the ruin probability 0 = 0 (u) itself.Owe (9. it suffices to fix the premium at p = 1 and consider only the effects of changing .()^ 1 . Of course.r.3 70 = 'Ye = = 7 /3(1 we(e +'y. Consider first the case of 8/8/3: . and the proofs of (9.w(6. 3) ( 9 . and write yp = 8y/8/3 and so on .g. 9. e.6) As will be seen below.10) below.3 or/and B. (3+'y)PC (0+7.()wC(e.0 = t/'(u) and the CramerLundberg constant C.3. ^) . 5) (Q+'Y)[we(0+7. various parametric families of claim size distributions could be considered.()YC = 1 +y/ /3 \ Q2 From this (9. but we shall concentrate on a special structure covering a number of important cases.4).6 below for some discussion of this assumption). Differentiating w. 4) (9 .t.w(O. mathematically a proof is needed basically to show that two limits (u * oo and the differentiation as limit of finite differences) are interchangeable. so that heuristically we obtain '00 50ryu = Coe"u . for the ruin probabilities .(/3 + y)we(9 + 7. (. () = log(1 + y//3). namely that of a twoparameter exponential family of the form Bo.1 or Proposition 9.uypCe7u urypO. The most intuitive approach is to rely on the accuracy of the CramerLundberg approximation . Viei '0(. (9. (9. we cannot expect in general to find explicit expressions like in Example 9.((dx ) = exp {Ox + (t(x) .3. THE COMPOUND POISSON MODEL As a consequence. ()} p(dx) .3) follows by straightforward algebra.^)] 1(/3+y)we(9+'y. Consider first the adjustment coefficient y as function of 3. we can rewrite the Lundberg equation as w(9+ y.
w(O. Z(u)/u a C//3PF where PF is the mean of F.([a] = exp {w(9 + a. () . () exp {w(9 + a. () .4 As u oo. ()} .9.g.x)B(x)dx.x).(e"U = = wS(O. ()] exp {w (O + y.3(x) dx. But from the proof of Theorem 5. BarndorffNielsen [58]). Combining these estimates .10) (9. SENSITIVITY ESTIMATES Proposition 9.we(9 .C). we note the formulas Ee.3 (see in particular (5. ()} . u 0 Proceeding in a similar way as in the proof of the CramerLundberg approximation based upon (9.x)B(x) dx + J U W(u .x). we get p(u) = J "O B(x) dx + J U O(u . () .9) (9. 0(u) = /3 Ju"O B(x) dx + f 0 0(u .4t (U)e°`U = which are wellknown and easy to show (see e. z2(U) = e7" J u b(u . u 0 Then Z = z + F * Z and F is a proper probability distribution . w((9 + a. By dominated convergence.w(9. Hence by a variant of the key renewal theorem (Proposition A1.QB(x) dx.11) Ee.p)/C'y. the proof is complete.w(9.2 of the Appendix ). Be. O} (9. F(dx) = e'yy/3B(x)dx. PF = (1 .8) (Section 5).St (U) Ee. it holds that 89 a ue ryu a/3 Q(1 P) 7C2 Proof We shall use the renewal equation (3.3) for z/'(u). (9.8). () .12)). z2(u) _ 1 ^ e'ri`i7i( u . Further write de = [we (9 +'y. and alsoo zl(u) + 0 because of B['y ] < oo. we multiply by e7" and let Z(u) = elt" cp(u). Z= zl + z2 where zl (u) = e7u J m B(x)dx.8) Letting cp = e0/e/3 and differentiating (9.x) F(dx ) f u J C F(dx) = C as u 4 oo. 11 For the following.
8 8() 8( (9.wc (O.12) f exp {O y + (t(y) .x). z = zl + z2. F(dx) = e7x. 0 x Multiplying by e7" and letting Z(u) = e"uV(u).x)f3 f ^[t(y) .2) holds. ()]e7vB(dy) 'fCd 7 c .w(0. )}B(dy)• Letting cp it thus follows from (9.8) that cp(u) . 01 (i+) do = +'Y.w( (0.9)(9. ()]B(dy) dx.w(e. C)] (1 + 7 ) Proposition 9.6e7u f "o f[t(y) . ()](e7v . u Z2(U) = e7° f u ^/i(u .e7x/3 f 00 [t(y) .wc(O.5 Assume that (9. C) . ^)} [wc (0 + 7. By dominated convergence and (9.6C do 89 1p 8( 1p Proof By straightforward differentiation.1) B(dy) 'f '[t(y) .wc(9.lB(x) dx = e7uzl(u) + e7°zz(u) + V(u T where zl (u) = .we (0. ^) .w (9. THE COMPOUND POISSON MODEL [we(e+7.11). oo z2 (u) f C . ()]B(dy) dx x 0 0C T ON O .w( (0. 8^ ue7u.90 CHAPTER III. Then as u > oo. ())B(dy) dx. ()} 1z(dy) = f [t(y) . 2 z 07P N ue7u (3C de . this implies Z = z + F * Z.QB(x)dx.
yu/3C2do u86 89 1p' az/) = 8z/.. we (9.rye) S 5rya.15) (9.Sry a/32 + a/37 + /37 . ())B(dy) < oo. We get w( (0.9. . ( 9.6 Consider the gamma density b (x ) = Sa xa. () = log r(a) . U 7µF from which the second assertion of (9.pa+1 . by inserting in the above formulas.ry) 5a1 cry (5 .a log S = log r(c) .QS 1 . < = a.12) takes the form y) alp a . and also zj (u) 4 0 because of f Hence. ue_Yu 'C2d( 8a 8( 1 p .w((9.Y)a+1 ' (9. that C = a. w(e.16) (9. and the proof of the first one u is similar./35' a/i'y + aryl 625ry. SENSITIVITY ESTIMATES as u 3 oo. Here (9.14) de = d( 7!3 76 = 7e = log ( \ ( \5a_ / \SSry ) 72 .) log(9) = %F(a) logs where %1 = F'/]F is the Digamma function.2) holds with p(dx) = xldx.3ary tog('Finally.1 .C log(9). 9 = S.13) (9. () = C/9 = a/S. Z(u) /3C 91 o c'o e11(t (y) .(log r(a) a log S)} • r(a) 1.18) (05 + 57 _'3_y .1edz = 1 exp {Sx + a log x . It follows after some elementary calculus that p = a)3/5 and.17) (9.. Example 9.a/35a&y' ' (9. t(x) = logx.12) follows.. a /(S . () ='I'(t.
l3 of Section 6a needed for the existence of ry becomes e^Q > 1+62 / 2.2) with µ(dx) = 2x3zrdx.9) (() ."62 . C) .22. for a < a* = z (.1 = eXP {c(C . which we omit in part . THE COMPOUND POISSON MODEL Example 9. C = .w(9.2a) } Thus the condition B[a*] > 1 + a* /.3Ee. Be.7 Consider the inverse Gaussian density ( b(x) Zx37 exp This has the form (9.log c = 2 In particular.S[a] = exp {w (9 + a. t(x) _ .2 log (0.1 16 +ry c C22ry 2( = + 70 We (e. C) = B = Yc = de = do = . ()} = exp {c (C . 9 = .CZ try)} 1 C C2 try . further yield . () = Cc .92 CHAPTER III.21og 2.([Y] . w(e. Straightforward but tedious calculations .3.2 .
. To this end.12) takes the form a = a 93 ar. Comparatively less work seems to have been done in risk theory. B are assumed to be completely unknown.10. the exponent of the density in an exponential family has the form 01 tl (x) + • • • + 9ktk (x). or Ct(x). and hence explicit or asymptotic estimates are in general not possible. ESTIMATION OF THE ADJUSTMENT COEFFICIENT Finally. However. in u which case the extension just described applies. kT (a) = /T (BT [a] . we can just fix k .1) . the exponent is either Ox. to our knowledge.3C2de 1p' z a = c . by the LLN both F (NT = 0) and F (PT > 1) converge to 0 as T . the results presented here are new.cue_7u)3C P Remark 9.7 and references there. Notes and references The general area of sensitivity analysis (gradient estimation) is currently receiving considerable interest in queueing theory. That it is no restriction to assume k < 2 follows since if k > 2. queueing networks) are typically much more complicated than the one considered here. Thus.2) is motivated as follows. sj=1 and let YT be defined by IKT('ryT) = 0. thus. then ryT < 0. we have assumed k = 2 and ti (x) = x. and we estimate y by means of the empirical solution ryT to the Lundberg equation. [379] consider a special problem related to reinsurance. if 1 PT = /3TNT(U1+. Van Wouve et al.a. Also. in which case we can just let t(x) = 0. then BT and hence ryT is undefined.oo.+UNT) > 1. the main tool is simulation. for which we refer to X. BT [a]= NT ^` e"U. ae t 1lEY u S _ . However .2 of the parameters. .8 The specific form of (9. the models there (e. (9.. 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).. let NT 16T = ^T .g. Note that if NT = 0. 10 Estimation of the adjustment coefficient We consider a nonparametric setup where /3. Finally if k = 1. Thus. In general.
THE COMPOUND POISSON MODEL Theorem 10 . then (10.T y . B [7]2 (10.: N 0. a2 where a2 = /3r. Lemma 10 .1)2 + E[27] . If furthermore B[27] < oo. . B[27] . 7T a4' 7.B[7]2 V2 .1 As T 4 oo./^ B[27] . (10.: N ()3.94 CHAPTER III. For the proof.2) rT(7) N N (0.B[7]2 }) ( T 0 .2 As T * oo. it is easy to see that we can write \ V1 1 l _ .B[7]2 n Hence ( 10.1) .b[Yp'V21 T { (E[7] . since NT /T .i3)(B[7] 1) + (3(BT[7]  . Hence KT(7) = (F' + (OT a(B[7l 0))((BT [7] . B[2'Y]  /3T ) .7 + (.a BT[7] I B[7] I + .Q and Anscombe 's theorem.'s. we need a lemma.v. N ( n[7]. V2 are independent N (0.If .B[7]) + B [7] .B[7]) 0+ Iv/o(b[y].)vl+ N CO.3) Proof Since Var(eryU) = we have B[7]. 1) r.'Y .3T .1) 'YT .1) .3/ T). More generally..2) follows from NT/T a4' . vfoVFB[2y]. 16T where V1.(27)/K'(7)2.
(ry .4) and Lemma 10.'T(a) = 1 E Uie°U' a$' EUe "u = B'[a].E) < 0 < kT(7 + E) for all sufficiently large T .E) < 4T(7T) < 4T(7 + E).e) < 0 < r. ESTIMATION OF THE ADJUSTMENT COEFFICIENT which is the same as (10.E ) < 4T(7T) < (7 +0' which implies 'T(ry4) a$' r. OT a 95 u 4 /3. and the truth of this for all e > 0 implies ryT at 'y.(ry + e) and hence KT(7 .10. it follows that 7T7 KT(7T) .KT(7) kT(7) K'(7) .c'(7) N (0' T (2(7) / N (0. If ryT E (7  we have KT(7 .2. Proof of Theorem 10. Now write KT(7T)  kT(7) = 4T(7T)( 7T 7). BT[a] 3 B[a]. where ryT is some point between ryT and ry.e.. lcT(a ) 4 /c(a). 6"Y (10. Then r.4) + E). NT i =1 n'(a) for all a so that for all sufficiently large T K7 . 7T E (y . 0 are estimated from data . Combining ( 10.1 By the law of large numbers. y + E) eventually. NT BT [a] Hence r. Let 0 < E < ry. By the law of large numbers. first note that e7TU N (e7U u2e27Uo'2/T) 7 .3). °7IT) .Q.'(y). Theorem 10.e.1 can be used to obtain error bounds on the ruin probabilities when the parameters . To this end . I.
Asmussen [23]) can then be used to produce an estimate of ry.g. Deheuvels & Steinebach [102].Q. Wn). the nth busy cycle is then [Wn1. various alternatives have been developed. 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... A major restriction of the approach is the condition B[2ry] < oo which may be quite restrictive. [197]. Hipp [196].T VIT where r7ry. Further work on estimation of y with different methods can be found in Csorgo & Steinebach [94]. with a tail of the form P(Y > y) . . Herkenrath [192].96 CHAPTER III. it means 2 (8 . and the known fact that the Y„ = max Vt tE[W„1. if B is exponential with rate 8 so that ry = 8 .. Frees [146].e.1 is from Grandell [170].. > 0 for some t E [Wn_ 1. i. U2. For example .) = a (e. V.g. i .3 or equivalently p > 1/2 or 11 < 100%.96 if a = 2.e. t]}.d. 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.0) < 5. Mammitzsch [253] and Pitts.T = 3TKT ( 21T)IKT (^T)2 is the empirical estimate of vy and fc.f. satisfies b(. Csorgo & Teugels [95].. Griibel & Embrechts [292]. This approach in fact applies also for many models more general than the compound Poisson one. Notes and references Theorem 10. = 1. ft..1 : Vt = 0.ueryuU ". Vt = St .C1e"a ( see e.5%). 6 < 2. For this reason .. wn = inf{t > W. Embrechts & Mikosch [133].info< „< t S.i.Wn) are i .
generalizations to other models are either discussed in the Notes and References or in relevant chapters. defined as solution of c(ry) = 0 where ic(s) _ /3(B[s] .g. the Poisson intensity is 0 and the claim size distribution is B with m. 97 . Unless otherwise stated. The safety loading is q = 1/p . In particular. See Fig.1) . the premium rate is 1.f. Further let 'Yo be the unique point in (0. The notation is essentially as in Chapter III.1 (the role of ryy will be explained in Section 4b).s.1 where p = 13µB. it is assumed that i > 0 and that the adjustment coefficient (Lundberg exponent) y.Chapter IV The probability of ruin within finite time This chapter is concerned with the finite time ruin probabilities 0(u. 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. 0. exists. 'y) where c(a) attains it minimum value. B[•] and mean AB.
2) Proof Let as in Example 111.(4. By the likelihood identity III. 7. In particular. FL and independent of T(u).t.98 CHAPTER IV. (u) is exponential with rate 0 w.5 .u is the overshoot.9).1 In the compound Poisson model with exponential claims with rate S and safety loading 77 > 0.(U) < 00] = ELT(u)ke'YS.(. 1 Exponential claims Proposition 1.1) (1.1 The claims surplus is {St}. Var[T(u) I T(u) < 00] = VarL T( U) . E[T(u) I T(u) < 00 ] = ELT (U). the time of ruin is T(u) and ^(u) = ST(t&) . . EL refer to the exponentially tilted process 3 with arrival intensity S and exponential claims with rate / (thus .) = e7u ELe'Y^(u) ELT(U)k = e'Yu b ELT(u)k = O(u)ELT(u)k. 1 FL. using that the overshoot l.. PL = 6/0 = 1/p > 1). the conditional mean and variance of the time to ruin are given by E[r(u) I T (u) < oo] Var [T ( u) I T( u) < oo] /3u+1 J )3 _ 2/3Su+/3+S (S)3)3 (1. 2 that E [T(u)k.r. PROBABILITY OF RUIN IN FINITE TIME Figure 0. we have for k = 1.
1 /3u + 1 u + 1 //3 = 6/3 6/01 For (1. which leads to the quadratic 02 + (/3 . we have by Wald's identity that (note that ELSt = t(pL .1)) ELST(u) ELT(u) (PL . of (1.6 + a)0 .12 Thus the l. u + ELe(u) _ PL . is V1rLSr( u) +VarL ((PL .B = a.6a = 0 with solution 0 (the . EXPONENTIAL CLAIMS For (1 .s. This means that /3(6/(6 .1)T(u))2 = UL where = s. . where = eBu I 1 .2) is aLELT( u) .0) .1)T(u)) = VarLe(u) + (PL . T(u) < oo] fora > r. Wald's second moment identity yields 2 EL (Sr(u) .1.3) B = 0(a) = + (6/3a)2+4a6 2 and hence that the value of ic(yo) Proof It is readily checked that yo = 6 .s.s./3 .1)T(u) are independent with QL the same mean . the Laplace transform of the time to ruin is given by Eea7( u) = E [eaT (u). the 1.h.I (1. 0 Proposition 1.2).1//32 (6/)3 1)2 26(/3u + 1)/(6 .h. Let 0 > yo be determined by ^c(0 ) = a."(ry) = 26//32.1)2VarLT(u) + 2 Ca 1I VarLT(u).(PL . 1).6.1)ELT(u).1) .1 (6)3)2 which is the same as the r./3) .h.V/ is as asserted. Since Sr (u) and (PL .2 In the compound Poisson model with exponential claims with rate 6 and safety loading rl > 0.(yo) = 2V .
Y2.Y1 Y2 Figure 1. Using 5 = 6 . Cf. (1..3) we have E [e«T(u ). Fig.100 CHAPTER IV.0. PROBABILITY OF RUIN IN FINITE TIME sign of the square root is + because 0 > 0).v.4.... St Ti F.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..T+ Ti a t U T I 1 a i F.3 that we can write EeaT( u) = eeuEe 017(o). 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 . and M(u)+1 is the index of the ladder segment corresponding to T(u).. More precisely. T(u) < oo] = e. Note that it follows from Proposition 1. are the lengths of of the ladder segments 2..3. T(u) < oo] = EB [exp {aT(u ) . are the ladder heights which form a terminating sequence of exponential r. 1. the result follows. Ti.v.1 where Y1. ..1 .'s with rate 5. But by the fundamental likelihood ratio identity ( Theorem 111.9ST(u) +T(u)!c(0)} . T2 .4) The interpretation of this that T(u) can be written as the independent sum of T(0) plus a r.
then VT = U1. T) to be evaluated by numerical integration: Proposition 1.e.T. 2. EXPONENTIAL CLAIMS 101 For numerical purposes . the conditional distribution of VT given QT = N is that of EN where the r.1.I ex cos B cos j O dB fo " . . Corollary 11.v.i.T + • • • + UN. . where U1.i (u. Note that the case 6 # 1 is easily reduced to the case S = 1 via the formula V. Since U1 .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. Proof We use the formula .T are conditionally i. Let {Qt} be the queue length process of the queue (number in system.T. T..T) 1 I fl(O)h(0) fdO where (1.. including the customer being currently served). Then V(u. . [4]) 00 (x/2)2n+3 Ij (x) OnI(n+j)! ...4. UN.6(u) = Vfl/j l(Su.1.T is the residual service time of the customer being currently served and U2 .T.T the service times of the customers awaiting service .6) fl(9) f2(0) = = fexp {2iTcos9(1+/3)T+u(/cos91) cos (uisin9) . U2.1 )!.d.6.1(u.ST). 1). cf. If QT = N > 0.. UN.. 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 .3 Assume that claims are exponential with rate b = 1.. . let (cf.0.3 sin0 + 29) f3(0) = 1+/32/cos9. i.cos (u/. density xN lex/(N . . EN has an Erlang distribution with parameters (N.k + 1). the following formula is convenient by allowing t. For j = 0. and exponential with rate S = 1.
31 /2eie L 1)] 1 I/31/2eie .102 CHAPTER IV.i(k +1)e R [/3( klal/2e:0 (01 /2 e . k k2 + $k+1 E bj 00 t j . 8789) 00 E aj j= 00 = 1. f3(0) .1 00 ok+lR 00 j=k1 +1)/2e . 00 E '3j/2 cos(je) j=k+1 00 _ j=k+1 ^j/ zeij = .(31/2 cos (( k + 2)9) .ie .cos (( k + 1)0)] f3(9) Hence the integral expression in (1.8 ) yields F(QT > k + 1) . let I _ j (x) = Ij (x). Then (see Prabhu [294] pp.)3k+1 = e(1+0)T e201/2Tcos 7r 0 e )3(k +l)/2 [31/ 2 cos ( kO) . (1. in particular equations (1.44).112 l 1( k +1)/2 [ 31/ 2 cos(kO) .cos((k + 2)9)] d9.(31/2eie .13(k +l)/2ei(k +1)9 R E .1)] L _112 /(k+1)/2 [.3(k +1)/ 2ei(k + l)6 (. similar formulas are in [APQ] pp.k + 1) = 1 k +1 + bj j=00 j=00 00 j=kk+1 j=k1 By Euler 's formulas. and define tj = e(1+R)Taj/2Ij(2vT T).)3k +1 tj g'(QT >.1 R [. 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 . PROBABILITY OF RUIN IN FINITE TIME denote the modified Bessel function of order j.38).
F(x. k! k=O k0 i/z Co Uk ate" o'/z e .T) which. or. Seal [327] gives a different numerical integration fomula for 1 .. going back to Cramer.e = e' COS a cos(uf31/2 sin 0). t) = P . T). T) in terms of F(. is numerically unstable for large T.7) that _ [^ au ak+l (30 k L. The first formula. equivalently. however. We allow a general claim size distribution B and recall that we have the explicit formula z/i(0) _ P(7(0) Goo) = p. the numerical examples in [12] are correct). 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. Ui < x I / (note that P(St < x ) = F(x + t.3 was given in Asmussen [12] (as pointed out by BarndorffNielsen & Schmidli [59]. t )). from the accumulated claim distribution N. 2 The ruin probability with no initial reserve In this section . E Fk. We first prove two classical formulas which are remarkable by showing that the ruin probabilities can be reconstructed from the distributions of the St.2. we are concerned with describing the distribution of the ruin time T(0) in the case where the initial reserve is u = 0. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Since P(QOO > k + 1) = flk+1. oo (u)31/2e^e)k = )3k z cos(k9) = R k. expresses V)(0.0(u. Related formulas are in Takacs [359]. there are several misprints in the formula there. k=0 103 Cu) A further application of Euler's formulas yields cc k =0 k 'ese)k __ U #kJ2 cos((k + 2)9) = R eNO ^` (u^1 L k= = eup i/z L OI = =ateU161/2 e '0+2iO COS a cos(u(31/2 sin 9 + 20). u Notes and references Proposition 1. however. . it follows as in (1. The rest of the proof is easy algebra.
(6.T))dv.(.T)dx. .T)) 1 fT P(M(v.b (0. PROBABILITY OF RUIN IN FINITE TIME Theorem 2 . co ). we define a new claim surplus process St StM NJ Figure 2. T T o where the second equality follows from II.104 CHAPTER IV.0<w<t} St+v . ") } is at a minimum at time t. T]. T) = P(Tr(0) > T) = P(M(0. [v. Then 1 .(0. 2.T))dv E^T I(M(v.3) with A = (0.T) T F(x.1 In formulas. meaning that we interchange the two segments of the arrival process of {St}o<t<_T corresponding to the intervals [0.T]. f T lStv)} 0<t<T by a 'cyclic translation'. See Fig.S„ 0 <t<Tv STS„+St_T+v Tv<t<T as the event that IS.1. v]. 1 1 . Proof For any v E [0.T)) does not {Stv)} depend on v. and the third from the obvious fact (exchangeability properties of the Poisson process) that has the same distribution as St = { Si0)} so that P(M(v.t)= {Stv) < SM. Stv^ _ Define M(v.i. resp.
xdx. cf. letting w = inf It > 0 : St_ = mino<w<T Sw}.. we can write M(v. t).2.T) occurs. t) denote the density of F(•. v < t < T} n M(0.v<t<T} = {ST<StSv.Tt))f(u+t. For example.T) and Sv < 0 on M(0. We claim that if M(0. If ST < 0. T)) dv.T) occurs or not as long as ST < 0. T) as {ST<St+ vS. v). THE RUIN PROBABILITY WITH NO INITIAL RESERVE 105 Now consider the evaluation of fo I(M(v. T)). 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. v<t<T}n{ST<STSv+St. v). T) occurs. then M(v. ST > 0. Fig 2. T)) dv = TEST = T fP(ST < x) dx T T NT 1 f P(ST < x) dx = 1 f P Ui T . Obviously. T) = M(0. It is then clear from the cyclical nature of the problem that this holds irrespective of whether M(0. v) = M(0. T. . or it occurs.Sv. we can take v E (w E.. v). Proof The event {ST < u} = { Ei T Ui < u + T j can occur in two ways: either ruin does not occur in [0. then i fT I(M(v. It follows that if M(0 . T].2. there exist v such that M(v. v)) dv = ST T T o (note that the Lebesgue measure of the v for which {St} is at a minimum at v is exactly . 0<t <Tv}n{ST<ST Sv+St T+v. T T o i =1 Let f (•.ST on M (0. this integral is 0 if STv) .2 10(u. T) occurs. where the last equality follows from ST < St on M(0. T Theorem 2 .T)) dv f T I(M(0. Indeed.T)f(I z /)(0.T) = F(u+T. w) for some small E.t)dt.
which is independent of St and has the stationary excess distribution B0. Proof of Theorem 111.Tt))P(StE[u. Then P(T(0) E • I T(0) < oo) = P(T_ (Z) E •). ST_ _ z}. PROBABILITY OF RUIN IN FINITE TIME u Q II T Figure 2. Proposition 2.(0. O(T .2.z. Hence P(ST<u) = 1 .T) = C(z. {St > . 0 < t <T . 0 < t < T. ST_ _ z}. t + dt] occurs if and only if St E [u.b(u.p.u+dt]). {S t > z. For a fixed T > 0.2 Here o.3 Define r_ (z) = inf It > 0 : St = z}.106 CHAPTER IV.T)+ J0 T (1V.ST_ t_ and let A(z.t). E [t.v.2. 2. C*(z. The proof is combined with the proof of Theorem 111.T) = {St < 0.T) = . 0 < t <T. u + dt] and there is no upcrossing of level u after time t. ST_ _ z} . u which is the same as the assertion of the theorem. which occurs w. The following representation of T(0) will be used in the next section. z > 0. define St = ST .2 . Let Z be a r.
Hence integrating (2. T + dT] I S7(o)_ E [z.3). ST(o)_ E [z. z + dz]. T(0) < oo) = OR(z) dz in (2.2.2. T(0) < oo) B(y B(z) + z) f3B(z) dz = 3 f °^ B(y + z) dz = f3 + x v f B(z) dz. {St }o<t<T have the same distribution .1) z T .3 But by sample path inspection (cf. Proof of Proposition 2. z + dz]. .1) yields P(ST(o)_ E [z.3.T)). r(0) < oo) = 3R(z) dz JP(C(z.1) that P(T(0) E [T. (2. 2.T) = C*(z. z + dz].ST(o) >y.T))dT = Off(z) dz P(T_ (z ) < oo) = 3B(z) dz.T))f3B(z) dz dT.T + dT]. we therefore have P(A(z. It follows by division by P(ST(o)_ E [z. z + dz]) = P(A(z.T)) = P(Cx. 7( 0) < oo) = P (C(z)) dT. and since {St}o<t<T. Thus P(Sr(o)_>x. A(z.2. u which is the assertion of Theorem 111. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Then 107 P(r(0) E [T. z + dz]. Fig. Figure 2.T).T(0)<oc) = f x F(U > y + z U > z) P(Sr(o)_ E [z.
s. a martingale proof is in Delbaen & Haezendonck [103]. Theorem 2. (3. T(0) < oo) = dTP(T_(Z) E [T. who instead of the present direct proof gave two arguments. cf. r(a) denotes the solution < 'Yo of the equation a = ic(r (a)) = . some relevant references are Shtatland [338] and Gusak & Korolyuk [181].5 and one upon excursion theory for Markov processes (see IX.108 Hence CHAPTER IV.1) where a > r.3.(3(B[r( a)] . because of77>0. PROBABILITY OF RUIN IN FINITE TIME ]P(7(0) E [T. Lemma 3. z + dz].c(r(a)) l = l er( a)se+at } u yields 1 = eyr(a)Eear(y).1) . Notes and references For Theorems 2. I L Let ga(x) be the density of the measure E[ear(°). Let T_ (y) be defined as Proposition 2. see in addition to Prabhu [293] also Seal [326].2.(yo).1 Eear( y) = eyr(a).1 and the present proof is in the spirit of Ballot theorems.2 ga(x) = Qexr(a) f "o eyr(a)B(dy) x . r(0) < oo. z + dz]. ^(0) E dx] (recall that ^(0) = Sr(o)) and write ga[b] = f OD ebxga(x) dx. Tak'ecs [359].T+dT]). T(0) < oo) 0 = dT f 0 P(C(z))P(Z E [z. Note that T_ (y) < oo a. Proposition 2. 2.6.5a). [329]. Lemma 3 . Proof Optional stopping of the martingale I er (a) 9 t.1. 3 Laplace transforms Throughout in this section.3 was noted by Asmussen & Kl(ippelberg [36].T + dT] T(0) < oo) dT f ' P(C(z))P(Sr( o)_ E [z.r(a). In the setting of general Levy processes. one based upon a result of Asmussen & Schmidt [49] generalizing Theorem 11.
(u . LAPLACE TRANSFORMS 109 Proof Let Z be the surplus . Z = y] = EeaT.r(a) The result now follows by inserting /3B[s] = ic( s) +/3+ s and ic(r(a)) =a. the result follows after simple algebra. T(u) < oo] du = Proof Define Za(u) = E [eaT(" ). y + dy].°° ga(x)dx. u .4 E[eaT (o).3.x)ga (x) dx where za(u) = f.3.5 f 00 o a/r(a) .ga [b] 1 .ST(o)_ just before ruin . Hence eb"du E[eaT(").f.T(0) < oo] = 20[b] = za[b] (9a[b] 9a[0])/b 1 .1] evr(a)B(dy)[ b . Corollary 3. time T(u): u u Here is a classical result : the double m.2 P(Z E [y.ga [b] 0 TO Using Lemma 3.r(a) b . £(0) E dx) = /3B(x + dy) dx and hence ga(x) = f e r)/3B(x + dy) _ /3 f x e(v.2.g. r(u) < oo).3 ga[b] = c(b) Proof + b + a . E[ear (o) I T(0) < oo . Then by Proposition 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)) .x)(a) B(dy)• Lemma 3 . b .3. It is then easily seen that Za(u) is the solution of the renewal equation Za (u) = za (u) + fo Z.ic(b)/b x(b) + a eb"E[eaT(" ). (Laplace transform) of the ruin Corollary 3.r(a) = a [B[b] B[r(a)]] . rr(0) < oo) = 1_ r(a) Proof Let b = 0.(v) = ev''(a). Further by Theorem 111.
for any c > 0 P( Further.mu D 2 4 N(0.6. where _ 1 _ 1 1 C ML w(ry) 6B'[7J 1 .3). Then given r(u) < 00.mL > E T(u) < 00 ) 40. Proposition A1. For the second .r(u) = Er(u) • ES.s. t T(u) T(u) T(u) t m = lim = lim = lim Utioo u + Sr(u) u+oo S. By Proposition 111. T(u) a. note that by Wald's identity u + EC(u) = ES. the known results are even less explicit than for the exponential claims case. (u) t. P = /3µB > 1. Theorem 4 . we need the following auxiliary result: Proposition 4. St/t 1 1/m.s. This proves the first assertion of (4.h(u.UProof The assumption 11 < 0 ensures that P(T(u) < oo) = 1 and r(u) a4' oo. u 1 ET(u) 1 p1 u where Pw2 = 311B)m3• 7(u) .e.00 St = lim . The first main result of the present section is that the value umL..1 Assume 77 > 0. For the proof. Later results then deal with more precise and refined versions of this statement.. and hence a. uoo u using e.(u ) = o(u) a. i. = (p . Then as u * oo.1)Er(u) . (4.2 Assume ri < 0.3LELU 1 1p' is in some appropriate sense critical as the most 'likely' time of ruin (here C is the CramerLundberg constant). That is.1) i.2. mu ) ( 0 m < ML '(u) 1 m > rL.w ) v/. T(u)/u mL as u + oo. PROBABILITY OF RUIN IN FINITE TIME 4 When does ruin occur? For the general compound Poisson model. for any m T(u) u .1.110 CHAPTER IV. cf. and take basically the form of approximations and inequalities.
Thus.^ N (o. though it is not easy to attribute priority to any particular author. 4a Segerdahl's normal approximation We shall now prove a classical result due to Segerdahl.s.mu (2) '• m3/2 µB 7 . 1'r(U) . implying T(u) . T (u) < 00 J 0(u) e7'PL U \ I T u) .r(u)/m T(u) ti µB2) Z. WHEN DOES RUIN OCCUR? and that Ee(u)/u a 0. the same conclusion holds with t replaced by r(u). For (4 . Theorem 7.2) follows immediately from u (4. note first that ( Proposition 111.4.1 (by considering 0(u.2.1).N(0. of (4.2 of [86]) and (4. this can be rewritten as u + 1(u) .g. and (4. apB ) . again Proposition A1.3. If Z .7 6  11 Proof of Theorem 4. Notes and references Theorem 4. Tu) T( u) .1). . the result comes out not only by the present direct proof but also from any of the results in the following subsections. PL (•)+ 0. proving (4.h.5) St .6. 4).mu m . which may be viewed both as a refinement of Theorem 4.1) is T (u)  U mL P( T (u) < I > E.6µB2) Z v m (3µB2) Z.mL U > E.t/m D (2) 111 .1.1 is standard.mL >E By Proposition 4.1). T) for T which are close to the critical value umL).3). cf.1 The l. According to Anscombe' s theorem (e. and as a timedependent version of the CramerLundberg approximation. T(u) < oo f / 00) e7uE L [e_7 (t1).
(u. O .mul h(oo)Eg(Z). Using ( 4.4 (SIAM'S LEMMA) If 71 < 0.^(T(u')).112 CHAPTER IV.))I h(ul /4  ^(u)) I(6 (u') C ) f < ul /4 + f(e(u') . and similarly as above we get E[f(^(u)) I Fr(u.3 (SEGERDAHL [333]) Let C be the CramerLundberg constant and define wL = f3LELU2mL = f3B"[ry]mL where ML = 1/(pL1) = 1/($B'[ry]1).6) whenever f.ul/4.um.5) For the proof.e(u') oo w . (oo.L+YWLV'U) .6). e'°'/b (u.v. S( u ) < ul/4] < ET(ul / 4) = O(ul/4).a C4'(y )• ( 4.T ( u')] = E[ T ( ul /4 . P because of ^(u') . E9(Z) (4. then e(u) and r(u) are asymptotically independent in the sense that. we need the following auxiliary result: Proposition 4.VU T.) is readily seen to be degenerate at zero if ST(u•) > u and otherwise that of T(v) with v = u . oo ).4). and thus in (4.w2) r. Then for any y.)mu \ h(oo)Eg (r(ul) . Then h(u) 4 h(oo) = E f (6(oo)). using that ul/4 .f ( (oo)) . one has 9 (r(u)_rnu) Ef (^(u)) * E. we can replace T(u) by r(u'). we get E[ T (u) .3).l:(oo) (recall that rt < 0).T(u') given F. with w2 as in (4. letting Z be a N(0. Then the distribution of T(u) . oo). g are continuous and bounded on [0. Let h(u) = E f (^(u)). Hence Ef (Vu )) 9 (T(u.r. PROBABILITY OF RUIN IN FINITE TIME Corollary 4. Proof Define u' = u .t.u1/4)I(S(u') > u1 /4) h(oo) + 0. resp .ST( u') = u1/4 .
y > k'(7) .1. e7v" y < ^'(7) (4 .7) to be valid is that T varies with u in such a way that y(T) has a limit in (. Cf. For practical purposes . umL + ywL f) = e"P(T (u) < umL + ywL) = EL [e7V ").4) in the last. y u) < . T(u) < umL + ywL f. ELe7E (") .5 '(u .umL wI V"U u (4. in practice one would trust (4. that for the fit of (4.5) and solve for y = y(T). WHEN DOES RUIN OCCUR? Proof of Corollary 4. For refinements of Corollary 4. just substitute T = umL + ywL in (4. .yK(ay)• (4.(ay) = 17 7y = ay . oo ) as u * oo.8) Note that ay > 7o and that 7y > •y (unless for the critical value y = 1/ML). The present proof is basically that of Siegmund [342]. also Hoglund [204]. however . 10) '5(u) . CL Fig. see also von Bahr [55 ] and Gut [182].7) to be good. Thus .oo.3 ery"z/i(u .4.9) ( 4 .7) To arrive at this . The precise condition for (4. 3 is due to Segerdahl [333]. y u) < e 7v" . see Asmussen [12] and Malinovskii [254].z/)(u . Theorem 4. yy by 1 K.7) whenever u is large and ly(T)l moderate or small (numerical evidence presented in [12 ] indicates . Segerdahl 's result suggests the approximation b(u.dependent version of Lundberg's inequality For y > 0. define ay. 4b Gerber's time. PL(T(u ) < umL + ywL) 113 4 C4(y). u needs to be very large). 0.T) Ce7"4 (T . where we used Stain's lemma in the third step and (4.3 in terms of Edgeworth expansions . Notes and references Corollary 4 .
8). T(u) < yu] < eayu + yUr(ay) Y < eayuEav [ eT(u)K(av )L T(u) < yu} Similarly. yu ) = < eayuEay [eay^ ( u)+T(U)K ( ay). For a different proof. From the proof it is seen that this amounts to that a should maximize ayic(a).Y' (u.h(u. f Some urther discussion is given in XI. if y > 1/ic'(y). 5 is due to Gerber [156 ].5.t. yu < T (u) < oo 1 l e ayuEav [eT ( u)K(ay). 0. Numerical comparisons are in Grandell [172 ].3 yields easily the following sharpening of (4. In view of Theorem 4. yy is sometimes called the timedependent Lundberg exponent. However. An easy combination with the proof of Theorem 111. see MartinLM [257] . Then ic(ay) > 0 (see Fig .114 CHAPTER IV. which may be understood from Theorem 4.r. dy) Notes and references Theorem 4 .7 i. PROBABILITY OF RUIN IN FINITE TIME Proof Consider first the case y < 1/K'(y). u Differentiating w. the bound a7y° turns out to be rather crude .b (u. yu 11 < T(u) < oo j < eayu +Y UK(ay) Remark 4.6. .2. we arrive at the expression in (4.ay4(u)+ T(u)K(ay ). who used a martingale argument. the point is that we want to select an a which produces the largest possible exponent in the inequalities.8). yu) < C+(ay)e7a„ where l C+(ay) = sup f 00 eayR(xy)B( . which shows that the correct rate of decay of tp(u. a. and hence t.v"U.9): Proposition 4. Hoglund [203] treats the renewal case.yu ) = eayuEav [e .1).8 below .6 It may appear that the proof uses considerably less information on ay than is inherent in the definition (4. and generalizations to more general models are given in Chapter VI. we have rc(ay) < 0 and get (u) . yu) is e 'Yyu/ .
c(&) = ic(ay) is < 0. The traditional application of the saddlepoint method is to derive approximations. i. yu ) eaauEaye .ay a. then ay > 0..ay y 'Yay  ay .4.e.ayuEay f eay^ ( u)+T(u)K(ay). the choice of ay. the formula 0(u.yyu y l ay I 21ry/3B" [ay] V fU_ u + 00..(u.6 with P replaced by Pay and FL by Pay.11) ' If y > 1/ r . if we want EaT(u) . not inequalities.ayC() . (4. yu) = e.ay and get Ea e ayf (00) y _ 'Ya( ayKal lay C 1 .i(u. (4. We thereby obtain that T is 'in the center' of the Padistribution of T(u). (4.13) .12) < yu] Here the first expectation can be estimated similarly as in the proof of the CramerLundberg ' s approximation in Chapter III.^3 ]1/ Bay [lay .z. (0) r1 (a) ' I. For any a > yo. As a motivation.'(y ). yu ) ayay e ryyu ayay 27ry/3B"[ay] u Proof In view of Stam 's lemma. WHEN DOES RUIN OCCUR? 115 4c Arfwedson's saddlepoint approximation Our next objective is to strengthen the timedependent Lundberg inequalities to approximations. then the relevant choice is precisely a = ay where y = T/u.: T. T(u) < yu] . it is instructive to reinspect the choice of the change of measure in the proof.2 yields EaT(u) u u r. and in case of ruin probabilities the approach leads to the following result: Theorem 4 .8 If y < 1/ic'(ry). This idea is precisely what characterizes the saddlepoint method. Using Lemma 111. Ea . and b(u. [eT(u )K( ay). we have ryas = ay . T(u) suggests heuristically that l t/. then the solution &y < ay of . u 4 oo. Proposition 4.e.. and ii(u) .yu) c ay .5.
V < 01 Ir 00 er(ay)"1'2"'x eyur. Example 4.9 Assume that B(x) = eay. it seems tempting to apply the normal approxiyu + ul/2wV. 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).c'(a) _ /3a/(8 .7ruw2 Inserting these estimates in (4.ay) ay +.116 CHAPTER IV.13) rigorously.1) under Pay mation (4.(j (1 .I ay &y a ^c'(ay) a (1 +. (4. Writing r(u) and W2 = I3ay{.13). (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 .ay)K(ay) ay ayI&YI For the second term in (4. we get heuristically that Eay Ler (u)r(ay).a)2 .ay + ayl /BLay] .a. a nr=. T(U) < yu] = eyuk (ay)E''ay (ek(ay )"1/2WV.ay ) r.B[ay] /ay &y y(ay .12) is 0 entirely similar.1. The proof of (4. Then ic(a) = .1)3 = (jB"[ay]l (Pay .1)3 = y3/3B"[ay]. .1) .11) follows.1B[ay]1 ) y(ay .4). and the equation ic'(a) = 1/y is easily seen to have .c(ay)ul/2W p 2ir = eyu(ay) dz 1 rc(ay ) 2.a) . where V is normal(0.3(5/(S . i B[7ay . PROBABILITY OF RUIN IN FINITE TIME ry I i .13).(ay) _ y(ay . The difficulties in making the proof precise is in part to show (4.l'B)y /(Pay .
(5. 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.g.3+52 1+/351/y' sy 7 B ii[ay] 25 _ 251/2(1 + y)3/2 (5 .5. 0 Notes and references Theorem 4. A related result appears in BarndorffNielsen & Schmidli [59].8 is from Arfwedson [9]. is the drift and o.tcp) Lo {Wo ( t)}t>0 . 2 = Var(Si ) the variance.11) gives the expression '31/4 ( .ay)3 0 3/2 and (4. y) a''y" L '3 _ fl ) 51 /4(1 +1IY)3/4 \. . It follows that 5^y =5ay = /«y =f3+ay=l3+d 1+1/y' V 1+^1/y /35 1+1/y /3' ay ay =Qay say =.i )( v s vc ('3 + s _2 / .because the c..p. is undefined for a > 5).f. and next to note that such an approximation in particular implies that the first passage probabilities are close.. c a 00.. then { __ ./4 ^y for 1/i (u.1) . The mathematical result behind is Donsker's theorem for a simple random walk {Sn}n=o.= (s. in discrete time: if p = ES. DIFFUSION APPROXIMATIONS solution ay=5 117 V 1 (the sign of the square root is .1. yu) when y < 1/ic'('y) = p/1 .
Mathematically.t} _ {W_1(t)} .. a2 =/3µB2) Proof The first step is to note that { WC (St P) . we have o {i!t s: .tp). We want an approximation of the claim surplus process itself.p/c < St(p) < S((n+l)/ c + Pp/c. we shall represent this assumption on 77 by a family {StP) L of claim surplus processes indexed by the premium rate p.e.2) t>o where p = pp = p . oo)). 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. However.3. 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. of which a particular case is the claim surplus process (see the proof of Theorem 5. Letting c = a2/pp.p. Lemma 111.1) with S.tcpp) y = { WC (Sct) pct) } {Wo( t)}t>o (5.3) whenever c = cp f oo as p 1 p. St = EN` U= .z } {W_1(t )}t>o (5. n/c < t < (n + 1)/c.1. for the purpose of approximating ruin probabilities the centering around the mean (the tcp term in (5. (5.1)) is inconvenient. p. such that the claim size distribution B and the Poisson rate a are the same for all p (i. where p is the critical premium rate APBTheorem 5 . and this can be obtained under the assumption that the safety loading rt is small and positive. + {Wo(t ) .7c). and consider the limit p j p. 0 .3) takes the form LI S(P) { a2 to2/µ2 + t LI S (P) { a2 ta2/µ2 {W0(t)}. It is fairly straightforward to translate Donsker's theorem into a parallel statement for continuous time random walks (Levy processes)..118 CHAPTER IV.a = Snp) and the inequalities Sn )C . cf.1 As p J.1 below). Indeed . this is an easy consequence of (5.
r.(u) ti IG(oo. 1. 119 It is wellknown (Corollary XI. ('.h. is IG(T.Ta2 /p2). is 1/ip (ua2 /IpI. C. ulpl /a2) = e2"1µl / or2.2 suggests the approximation u 0(u.6) from Theorem 5.5.t. DIFFUSION APPROXIMATIONS Now let Tp(u) = inf{t>0: S?)>u}. u). and the r.6) This is the same as the heavy traffic approximation derived in III.h. ^ p2 Proof Since f 4 SUP0<t<T f (t) is continuous on D a. has a continuous distribution.s. 199.7c.. any probability measure concentrated on the continuous functions. (5. . we obtain formally the approximation V. For practical purposes .8 or [APQ] p.2 As p j p. u) is defective when < 0.e. u) =PIT( (u) < x) = 1 . (ua2 To2 op \ IPI > IG ( T . 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. TS(u)=inf{t>0: WW(t)>u}. Corollary 5. w.s. ulpI/a2). we omit the details .1. 196. and in fact some additional arguments are needed to justify (5.( ^ I + e2( \ I .s. since ti(u) has infinite horizon .4) Note that IG(.5) Note that letting T * oo in ( 5.5). the continuous mapping theorem yields sup W Sz2 to lP 4 sup Wi(t)• O<t<T O<t<T a2 Since the r. (5. [169] or [APQ] pp.h. see Grandell [ 168].f I \\\ J \ (5.1 .1. Corollary 5 . 263) that the distribution IG(•..u). (. Because of the direct argument in Chapter III.1 I 7= . u) of r( (u) (often referred to as the inverse Gaussian distribution) is given by IG(x. However.T) IG(Tp2/ a2). the continuity argument above does not generalize immediately..
0) { 2 StQ2 /µ2 D { W_ i(t)}t>o t>o D 2 where p = pe = pe . However.6 of [APQ].120 CHAPTER IV.5) combined with the fact that finite horizon ruin probabilities are so hard to deal with even for the compound Poisson model makes this approximation more appealing. In contrast. that 00 4090. The first application in risk theory is Iglehart [207]. the B9. . as 0 * 00 and that the U2 are uniformly integrable w.1. and which is much more precise. Michna & Weron [152] suggested an approximation by a stable Levy process rather than a Brownian motion. [169]. Theorem 5.3 Consider a family {Ste) } oc claim surplus processes indexed by a parameter 9. However. The proof is a straightforward combination of the proof of Theorem 5. We conclude this section by giving a more general triangular array version of Theorem 5.t. Assume further that 039µB6 < pe. for more general models it may be easier to generalize the diffusion approximation than the CramerLundberg approximation. For claims with infinite variance. B0 * Boo. pe . e. Furrer.Pe.Po = 09µB6 . the claim size distribution B9 and the premium rate p9 depends on 0.g. on the premium rule involving interest.00µB6 + 0. in particular for large u. All material of this section can be found in these references.6) therefore does not appear to of much practical relevance for the compound Poisson model. Further relevant references in this direction are Furrer [151] and Boxma & Cohen [75].. as an example of such a generalization we mention the paper [129] by Emanuel et al. Then as 0 _+ 90. we have ^A. The picture which emerges is that the approximations are not terribly precise. a2 = ae = 00µa6 Notes and references Diffusion approximations of random walks via Donsker's theorem is a classical topic of probability theory. PROBABILITY OF RUIN IN FINITE TIME Checks of the numerical fits of (5. such that the Poisson rate Oe.6) are presented.5) for the compound Poisson model which does not require much more computation.r.1 and Section VIII.5) and (5. In view of the excellent fit of the CramerLundberg approximation. (5. pt? 4 peo. See for example Billingsley [64]. in the next subsection we shall derive a refinement of (5. and two further standard references in the area are Grandell [168]. in Asmussen [12]. the simplicity of (5.
Let PO refer to the risk process with parameters e9oz Qo = QB[90].6. Then EOU' = Boki[0] = Biki[eo]/E[9o] and "(s) = k(sBo)k(9o).90] B(dx). B9(dx) =Bale] Bo(dx) e9z keo)z = B[9 . Since Brownian motion is skip free. whereas there we let the given 3B. it is more convenient here to use some value 9o < 0 and let 9 = 0 correspond to n = 0 (zero drift). In terms of the given risk process with Poisson intensity . this means the following: 1.1 > 0.ao (0) _ /c(s + 9 . The setup is the exponential family of compound risk processes with parameters ( B9 constructed in III. which we have seen to play an important role for example for the CramerLundberg approximation . let P9 refer to the risk process with parameters Q9 = QoB0[9] = QB[9 9o].4.1) .6. For each 9. claim size distribution B .T) = Peo(r(u) < T) for 90 < 0.'(yo) = 0 and let 90 = 'Yo. 2. and we are studying b(u. 0(0) = 0.90) and the given risk process corresponds to Poo where 90 = 'yo.90) . Bo(dx) = B[eo]B(dx).Q (B[s] .9(s) = Ico ( s + 9) . and we want to consider the limit 77 10 corresponding to Oo f 0. In this setup. this idea ignores (among other things) the presence of the overshoot e(u). Then r. However . Determine yo > 0 by r.c(s) = ./c(9 . The objective of the corrected diffusion approximation is to take this and other deficits into consideration. 77 is close to zero. P9(r (u) < oo) = 1 for 9 > 0. PB('r(u ) < oo) < 1 for 9 < 0. risk process with safety loading 77 > 0 correspond to 9 = 0 . 3. .s and p = /3µB < 1. . this is because in the regime of the diffusion approximation . 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. 77 = 1/p . 9o T 0.
0o to. (..(y) = 0. C. . 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. C .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.1) IG(x. One has (6. and Si = QoEoU2 = Q B"'['Yo Eo U3 ]..122 CHAPTER IV. tu2 ) i IG (t.Varo S1 = f30Eo U2 = S1. write r = T(u).2) .1) . () where h (A. the solution of r. The first step in the derivation is to note that µ = k (0) = r0 (00) . Theorem 5. . Vargo S.e. (01. _ ^(u) = ST . The corrected diffusion approximation to be derived is (u.. C) = 2A + (2 .T) 1+u2 (6. u) = euh(a . PROBABILITY OF RUIN IN FINITE TIME Recall that IG(x. (6.2' where as ususal ry > 0 is the adjustment coefficient for the given risk process. 9otc0" (0) = 0061 = ul. IGu+u2.7(u)/u2} eh(A.C. means up to o(u1) terms): .() The idea of the proof is to improve upon this by an O (u1) term (in the following. bl IG(t81. 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.S.3) this implies (take u = 1) Ego exp { . u) = IG(x/u2.u. (.. 1) • Since L eatIG (dt. i. for brevity.
4. of a (defective) r. distributed as Z . that the saddlepoint approximation of BarndorffNielsen & Schmidli [59] is a serious competitor and is in fact preferable if 77 is large] . it holds for any fixed A > 0 that Ego exp { Ab1rr(u)/u2} .v. just replace t by Tb1/u2. . In ( 1) and (2). 1. 1% in (2) and (4).z .h.7.5) according to (6. The initial reserve u has been selected such that the infinite horizon ruin probability b(u) is 10% in (1) and (3). u is Eeazead2/++ Eeaz[1 + ab2/u] where the last expression coincides with the r. p = 0.6.exp { h(A.ry2 . we get by formal Laplace transform inversion that C 2 u.f. .1 As u + oo. that whereas the proof of Proposition 6. Note. and the dotted line the corrected diffusion approximation (6. however . is the c.1 below is exact. 6 . which is based upon exponential claims with mean µB = 1.s.1 + 629. the formal Laplace transform inversion is heuristic: an additional argument would be required to infer that the remainder term in (6.v. however. A numerical illustration is given in Fig. CORRECTED DIFFUSION APPROXIMATIONS 123 Proposition 6.52/u where Z has distribution IG (•.2 ).3).5) Once this is established . . The solid line represents the exact value .s. we have p =.h. . bl I IG I t +2 . (6. of (6.d. 9o T 0 in such as way that C = Sou is fixed. calculated using numerical integration and Proposition 1.2). To arrive at (6. But the Laplace transform of such a r. the r.2) is indeed o(u1).1 + u2 I Indeed. in (3) and (4).3.3 = 0.'yu /2)(1 + b2/u)} + Aug 1I J . 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.
124 0.19)2 11 20 20 i0 T 1n0 Figure 6. A51 7(SAT 3 3 h(X.011 L1 60 T IM 11.02 I 90 120 160 2W A0 Z WT 40 80 120 160 100 240 280 T 111 WI.TI CHAPTER IV.(061 0. The proof of Proposition 6.0 0.W21 0.^) .199 0.1 W IU..114 0. and all of the numerical studies the author knows of indicate that its fit at p = 0.00 0. Similarly. the fit at p = 0.() Lemma 6.08 0. it gives the right order of magnitude and the ordinary diffusion approximation hopelessly fails for this value of p. see Asmussen [12]. PROBABILITY OF RUIN IN FINITE TIME 0. (Inc 0s 0.01 0.T) 0.EB 0 p ex p ( 7 S h ^)u .u2 2u3 (e .7.07 0. .7 or at values of Vi(u) like 1% is unsatisfying. Note that the ordinary diffusion approximation requires p to be close to 1 and '0 (u) to be not too small.OOIi O.T) 111 0..1 proceeds in several steps. For further numerical illustrations. BarndorffNielsen & Schmidli [59] and Asmussen & Hojgaard [34].1 It is seen that the numerical fit is extraordinary for p = 0.111 W(U.2 e. OM 0.aa1 .4 may not be outstanding but nevertheless.T1 00.05{ 0.08 a.
the result follows.r0 (00)) } Replacing B by 8/u and Bo by C/u yields e(B() = E eo exp { (e .(3)Eea LauT exp i 3J . (6. the formulas Po(C(0) > x) Po(C(co) > x) imply 1 °° Po(ST(o) > x) = EIU fIP0 (U>y)dy ..h.. + a1b2 + .co ((/u)) } Let 8 = (2a + (2)1/2 = h().61a2T (B3 .6) u U3 Lemma 6 .s. () + C and note that 2 KO (0) = 102.4) that the r.7) 2 2 .2u (B3 .00)(u +C)  'r (.3 EoU2 + 103OoEoU3 + " 2 6 Using d2 .co (e) . 1 = PB(T < oo) = Eo0 exp 125 {(B .C)C/u . exp ue } al 1J 3 exP I [2). 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 ..+ h (A. 3 lim Eof (u) = EoC(oo) = a2 Ep = 3EoU2 uroo Proof By partial integration . in Lemma 6.6. CORRECTED DIFFUSION APPROXIMATIONS Proof For a>0. 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.T (co (8/u) .(3) J t _ aa1T l + eh(A. () 62 Eeo exp u u2 J ..4 Ea. (6.2 behaves like C l Eeo eXp r _ ^81T 1 Sl u2 1 u 2u3 [1+h(AC) S .1) h(A.C2 = 2).
and inserting this and 9o 2 = S/u on the r.6  d h(A.\+ (2 (3 e 2u [ (2.x.1 (y/2 + Oo)u . yu/2) h(A.() I 1 + u2 ) y . Thus a2 y = 290 + O (u2). letting formally T * oo yields 7/)(u) C'e7u where C' = e7a2). we get h(A.h (A.S) d e 62 .\ + () 1 2 / .2. we get the correct asymptotic exponential decay parameter ^/ in the approximation ( 6. yields +90 62 0 + O(u 3) 2u2 +O(u 3)..s.126 CHAPTER IV. C) ( 1+ u2 The result follows by combining Lemma 6 .2) for O(u) (indeed .7) and using eh(a. yu/2) 11+ 62 I} S 1 \\\ u/11 l 62 (3 2u 2A Proof Use first (6.() . PROBABILITY OF RUIN IN FINITE TIME The last term is approximately (e 3 (3) 27. 2 + 00 = . () .e h(aS)h (^^ 262 exp {_h(. [2+ (2 . yu/2).2 (^/2 + 3y9o + 390) + O(u3). () by h(\. 0 The last step is to replace h(A. and the correction terms which need to be added cancels conveniently with some of the more complicated expressions in Lemma 6. l Lemma 6 . Thus by Taylor expansion around ( = 90u.6) and 7co (Oo) = ico('y + Bo) to get 0 = 21 (^/2 + 2y90) + 1112 (_Y3 + 3_Y200 + 3y9o) + O(u4). 5 exp { _h(A) (1 + / y u J)) exp 1.h.4. There are two reasons for this : in this way.(2A + ()1/21 exp S h(A.2u [2A+ (2 3 . 2 and (6.
His ideas were adapted by Asmussen & Binswanger [27] to derive approximations for the infinite horizon ruin probability 'i(u) when claims are heavytailed. the approach to the finite horizon case is in part different and uses local central limit theorems. the same as for the unconditional Lundberg process. () (i+a ) 2A + (2 . In Siegmund's book [346]. with the translation to risk processes being carried out by the author [12]. 7 How does ruin occur? We saw in Section 4 that given that ruin occurs. () I 1 + u2 )I 2u L 2A+C2_(2 exp { _h.T) has not been carried out and seems nontrivial. The answer is similar: the process behaved as if it changed its whole distribution to FL. We shall now generalize this question by asking what a sample path of the risk process looks like given it leads to ruin.(i+ 62 exP{ h(A. u Notes and references Corrected diffusion approximations were introduced by Siegmund [345] in a discrete random walk setting.1 (y/2 + Oo)u )} 1 (i + U ) [2+ C2 2u 62 S Pt^ exP { J 62(2 exp { h(A.4. ()} 3 h (A. and to the Markovmodulated model of Chapter VI in Asmussen [16]. Hogan [200] considered a variant of the corrected diffusion approximation which does not require exponential moments. the 'typical' value (say in sense of the conditional mean) was umL. 0 1 Proof of Proposition 6. Fuh [148] considers the closely related case of discrete time Markov additive processes. i. 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. that is.7.1: Just insert Lemma 6. The corrected diffusion approximation was extended to the renewal model in Asmussen & Hojgaard [34]. The adaptation to risk theory has not been carried out. HOW DOES RUIN OCCUR? exp { h (x.5 in Lemma 6. () I 1 + u 2 ) } S 1 . the analogous analysis of finite horizon ruin probabilities O(u. . 'yu/2) 127 ( i+ M pz^ exP { h (A. ()} .e.
ST(u)}t> o is just an independent copy of {St}t>o). Recall that .(u)_ and similarly the denominator is exactly equal to Ce7u.vi(u) Ce'Yu Corollary 7.1. and let M(u) be the index of the claim leading to ruin (thus T(u) = Ti + T2 + . F(u)c] P(r(u) < oo) ?P(U) < EL[e7u. Theorem 7 . {St}0< t<T(u)) Proof Write e'rsr(u ) = e'rue'r£(u).r. Recall that 13L = (3B[ry] and BL(dx) = e'rxB(dx)/B[7].FT(u) = o' (T(u ). P(u) and rate = aL w.(u) is not . then P(u) and FL coincide on .(u)_ is that i.TT(u) _measurable. r(u) < oo) . the Poisson rate changes from .(u).128 CHAPTER IV. the numerator becomes e'ruELe7^ (u)PL(F( u)t) = e7uCFL (F(u)°) when F(u) E . Note that basically the difference between FT(u) and .t. FL As example. . Then also P(u)(F(u)) + 1.2 If B is exponential. In the exponential case.(u)_ and ^(u) are independent . PROBABILITY OF RUIN IN FINITE TIME changed its arrival rate from 0 to /3L and its claim size distribution from B to BL.EL[e7S. stating roughly that under F(u). Proof P(u) (F(u)c) = F(flu)c. {ST(u)+t . so in the in the proof. ^(u) is exponential with rate 8 w.3 to ...T.t.3L and the claim size distribution from B to BL. + TMOO ). u * oo.1 Let {F(u)}u>0 be any family of events with F(u) E F.FT(u) is the stopping time oalgebra carrying all relevant information about r(u) and {St}o<t<T(u)• Define P(u) = P(•IT(u) < oo) as the distribution of the risk process given ruin with initial reserve u.F. In fact. we give a typical application of Theorem 7. .r.F. We are concerned with describing the F(u) distribution of {St}o<t<T(u) (note that the behaviour after rr(u) is trivial: by the strong Markov property. (u) and satifying PL(F(u)) * 1. . F(u)c] ti e' ru]PL (F(u)`) > 0.
the queueing results are of a somewhat different type because of the presence of reflection at 0.eALx) M(u) k=1 u The proof of the second is similar.7. 129 M(u) >2 I(Tk < x) M(tu) p(u) M(u) >2 I(Uk < x) BL(x). A somewhat similar study was carried out in the queueing setting by Anantharam [6].eaLx. however. Notes and references The results of the present section are part of a more general study carried out by the author [11].(1 . From a mathematical point of view. see further XI.3. Proof For the first assertion.3 M(u) pcu) 1 . who also treated the heavytailed case. HOW DOES RUIN OCCUR? Corollary 7. . This is currently a very active area of research. take I(Tk < x) . the subject treated in this section leads into the area of large deviations theory.
This page is intentionally left blank .
the claim sizes U1. Var(St) = 11Ba2A + I4AaB lim t goo t PA 131 .i. and M is the maximum of {St}.1.7). A different important possibility is Al to be the stationary delay distribution A° with density A(x)/µA. with the same distribution A (say) for T2.. the one corresponding to the stationary case by 00)(u). AA t*oo lim St = lim ESt t tioo t = p .1 (T1 = a1).. . . the distribution Al of T1 is A as well. U2. are i. see A.Then no matter the distribution Al of T1i B.Q. We use much of the same notation as in Chapter I. .Chapter V Renewal arrivals 1 Introduction The basic assumption of this chapter states that the arrival epochs O'1. with common distribution B. Proposition 1. with Nt = # {n: Un <t} the number of arrivals before t. Thus the premium rate is 1.(1.. T3.. and the one corresponding to T1 = s by 1/i8 (u).. D'2. Then the arrival process is stationary which could be a reasonable assumption in many cases (for these and further basic facts from renewal theory. .. the Tn are independent.1). of the risk process form a renewal process: letting Tn = Qn . r(u) the time to ruin.d.1 Define p = !µ. {St} is the claim surplus process given by I. The ruin probability corresponding to the zerodelayed case is denoted by 1/'(u). In the socalled zerodelayed case..
1 of the safety loading appears reasonable here as well. the definition 77 = 1/p . 2). This has a direct physical interpretation (a large portfolio with claims arising with small rates and independently). Nt + EVar U. The simplest case is of course the Poisson case where A and Al are both exponential with rate 0. Thus.t. the . Nt ESt = E E UI Nt t = ENt•pB . Sparre Andersen whose 1959 paper [7] was the first to treat renewal assumptions in risk theory in more depth.a is really the accumulated claims over a period u of length a. say at a. one could imagine that the claims are recorded only at discrete epochs (say each week or month) and thus each U. Here are two special cases of the renewal model with a similar direct interpretation: Example 1.1) ENt/t + 1/µA. OFF. RENEWAL ARRIVALS lim E [St+a .132 Furthermore for any a > 0.2 (DETERMINISTIC ARRIVALS) If A is degenerate. Nt = Var(PBNt) + E(4Nt) Q2 2 0`2 A tpB B + o(t). and (1 .St] = a(p . From this ( 1. by the elementary renewal theorem (cf. A.0 > 0. such that no arrivals occur in the off state. CHAPTER V. However . but the arrival rate in the ON state is .1). s + t µA PA 0 Of course. we get similarly by using known facts about ENt and Var Nt that Nt Var(St) = Var E Nt U. Proposition 1. For (1 .1) follows .3 (SWITCHED POISSON ARRIVALS) Assume that the process has a random environment with two states ON.1 gives the desired interpretation of the constant p as the expected claims per unit time. stating that E[Nt+a . 3) follows similarly by Blackwell 's renewal theorem. Example 1 . after E. t 4oo Proof Obviously. If the environment is Markovian with transition rate A from on to off and u from OFF to ON. The renewal model is often referedd to as the Sparre Andersen process.Nt] * a/PA.
(an arrival occurs necessarily in the ON state. we note that the ruin probabilities for the delayed case T1 = s can be expressed as in terms of the ones for the zerodelayed case as u+8 z/i8(u) = B(u + s) + '( u + s . For the stationary case.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. arrival times. we feel it reasonable to present at least some basic features of the model. A is phasetype (Example 1. INTRODUCTION 133 interarrival times become i. The following representation of the ruin probability will be a basic vehicle for studying the ruin probabilities: Proposition 1.t.s. 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^. (1.. in general the mechanism generating a renewal arrival process appears much harder to understand. u For later use..1.. However. and the present author agrees to a large extent to this criticism. if for nothing else then for the mathematical elegance of the subject..4) w.i. the fundamental connections to the theory of queues and random walks.2. the first term represents the probability F(U1 . and for historical reasons. Ao. and then the whole process repeats itself). Therefore.r.4) with phase space {oN. Proof The essence of the argument is that ruin can only occur at claim times.d. .. 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 .s > u) of ruin at the time s of the first claim whereas the second is P(r(u) < oo.1.oFF}. the relevance of the model has been questioned repeatedly. S o<t<oo n=0. initial vector (1 0) and phase generator 11 However.} with {S(d)} a discrete time random walk with increments distributed as the independent difference U . we have From this the result immediately follows. integrate (1.y)B(dy). U1 .s < u). More precisely.4) fo Indeed.T between a claim U and an interarrival time T..1.
(2. St = t . A typical sample path of {Rt } is illustrated in Fig.1.134 CHAPTER V. t. each of which receive a payment at constant rate during the lifetime . The initial reserve is obtained by prepayments from the policy holders.3* pB. with common distribution B* (say) concentrated on (0. The compound Poisson model with negative claims We first consider a variant of the compound Poisson model obtained essentially by signreversion . then 0 * (u) = 1 for all u > 0. 00). Theorem 2 .a*PB• > 1. That is . A simple sample path comparison will then provide us with the ruin probabilities for the renewal model with exponential claim size distribution.Ut. then 0*(u) = e 'r" where ry > 0 is the unique solution of 0 = k*(ry) = *(B*[ry] . If .3* (say ) and the U. U Figure 2. we shall be able to compute the ruin probabilities i(i* (u) for this model very quickly (.1) . 2. are independent of {Nt} and i. At the time of death .1) +ry. RENEWAL ARRIVALS 2 Exponential claims.1 If. b=1 !=1 where {Nt } is a Poisson process with rate .1 r* (u) One situation where this model is argued to be relevant is life annuities.0* (u) = P (rr* (u) < oo) where rr* (u) = inf It > 0: Rt < 0} ) . resp . Using Lundberg conjugation . < 1. the remaining part of the prepayment (if any ) is made available to the company. i.d. the claim surplus process are given by Nt Nt Rt = u+^U. the claims and the premium rate are negative so that the risk reserve process .
0) and has typically the shape on Fig.2(b).. Hence y exists and is unique. T_ (u) = inf { t > 0 : St = u 'r* (u).(u ) < oo) = E {e7sr_ (u). > 1 . EXPONENTIAL OR NEGATIVE CLAIMS [Note that r. Then the c. and thus 1 = P(T.3*. Then the function k* is defined on the whole of (oo.2 Assume now . then by Proposition 111. 0 Now return to the renewal model. and the Lundberg conjugate of {St} is { St } and vice versa. Since ic'(0) < 0. cf.2 sup St = inf St = 00 t>o t>o and hence 0* (u) = 1 follows.2(a).(a) 7 Figure 2. Proof Define 135 St =u .* (a) = log Ee'st I. . Hence T_(u) < oo a. Fig. of {St} is c(a) = is*(a7). B* [7] and let {St} be a compound Poisson risk process with parameters . B.1. Let B(dx) = ^e7x B*(dx). 2. Then { St } is the claim surplus process of a standard compound Poisson risk process with parameters 0 *.Rt.2.s. B*. (a) is*(a) (b) . If I3*pB* < 1. the safety loading of { St} is > 0. T_ ( u) < 001 e7"P(T_ (u) < oo) = e"V)* (u).0.g.f. Define T_ (u) = inf It > 0 : St = u} .UB. St=Rtu=St. 2.
Un } = max St = t>0 n=0.4... Taking m.. 1) means that 8(A[ry] .Y a I.Tr+.•.1.1.1.u+ and lr+.. 3* = 6.+Tn U1 Un. with the probability that a particular jump time is not followed by any later maximum values being 1 . To + max {Ul+•••+UnTI..)(u) _ 1r+e7" where ry > 0 is the unique solution of 1 = Ee'Y(uT ) = S 8 A[..7r+ 7r EeTo b/(Sa) + +. Then B* = A.1 means that M* is exponentially distributed with rate ry. 2. and (2 .Y] (2. A variant of the last part of the proof.. To + M(d) in the notation of Proposition 1. with rate S (say).. However. the failure rate of this process is y. Hence M* max {To + Ti + • • • + Tn . and 5PA > 1.. and hence the failure rate . T2 = U2. then .e. which has the advantage of avoiding transforms and leading up to the basic ideas of the study of the phasetype case in VIII.2) 7 and7r+=1Proof We can couple the renewal model { St} and the compound Poisson model {St*} with negative claims in such a way the interarrival times of { St*} are To ...1) + ry = 0 which is easily seen to be the same as (2.'s and noting that V)*(u) = P(M* > u) so that Theorem 2.1 it is seen that ruin is equivalent to one of these values being > u. RENEWAL ARRIVALS Theorem 2 . we get Ee'M(d) = Ee°M* _ Y/(.• • • .4 goes as follows: define 7r+ = P(M(d) > 0) and consider {St*} only when the process is at a maximum value. Now the value of {St*} just before the nth claim is To +T1* +.Ti = U1.136 CHAPTER V. alternatively termination occurs at a jump time (having rate 8)...Ui .f.a) = 1 .g.2). the distribution of M(d) is a mixture of an atom at zero and an exponential distribution with rate parameter ry with weights 1 . and from Fig . respectively.Tn} n=0.. According to Theorem 2. u Hence P(M(d) > u) _ 1r+e'r".2 If B is exponential.
. the relevant exponential change of measure corresponds to changing the distribution F(d) of Y = U .2. However. Thus a ladder step terminates at rate b and is followed by one more with probability 7r+.3. we have ] A[a )3] E«d'efl' = Bad> [a] A ad> [Q] = B[a +. resp.7r+) and hence r+ = 1. B^d) where Aad> (dt) = ^[ a] A(dt).3 A[a] B[a] F( d) [a +)3] F(d) [a] = Fad) [^] Letting M(u) = inf in = 1. 111. : S(d) > u} ..7r+) = ry and hence P(M(d) > u) = P(M(d) > 0)e7u = 7r+e'r". consider instead the failure rate of M(d) and decompose M(d) into ladder steps as in II. 3a The imbedded random walk The key steps have already been carried out in Corollary 11. Bads (dx) = .7r+). CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 137 is b(1. Furthermore.6.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] . we see that ry = 6(1.B(dx).2.y/b.. which states that for a given a.5. 0 3 Change of measure via exponential families We shall discuss two points of view. letting P(d) refer to the renewal risk model with these changed parameters . The probability that the first ladder step is finite is 7r+. Hence the failure rate of M(') is 6(1 . hence exponential with rate b. It only remains to note that this change of measure can be achieved by changing the interarrival distribution A and the service time distribution B to Aad^. Putting this equal to y.4. a ladder step is the overshoot of a claim size. This follow since.. the imbedded discrete time random walk and Markov additive processes.
In fact. to converge in distribution since p(yd) (r(0) < oo) = 1 because of r (d)' (y) > 0.2 p. Proof Proposition 3.. in the easiest nonexponential case where B is phasetype. 7µA . We have the following versions of Lundberg' s inequality and the CramerLundberg approximation: Theorem 3 .2 In the zerodelayed case.1 For any a such that k(d)' (a) > 0. E(d)e 1' (u). 00)(u) .r..C(°)eryu where C(O) = C0[7] . (a) '(u) < eryu.138 CHAPTER V. Corollary 3. provided the distribution F of U . cf. i . ik. we get: Proposition 3. For claim (b). 187) and thereby for ^(u) to be nonlattice w. the evaluation of C is at the same level of difficulty as the evaluation of i/i(u) in matrixexponential form.e.p)/($B'[7] . (b) V)(u) ..1) is explicit given 7.u the overshoot . VIII.Ce"u where C = limu. and claim (a) follows immediately from this and e (u) > 0. just note that F7(d) is nonlattice when F is so .1). RENEWAL ARRIVALS be the number of claims leading to ruin and ^(u) u = SM(u) . let 7 > 0 be the solution of r.4.C8e7u where Cs = Ce78B[7].t. (d) (7) _ 0. O(u) = eauE (d)ea{ (u)+M(u)K (d)(a) .T is nonlattice.(u) .1 implies Cu) = e«uE ( 7d)e«^(u) . It should be noted that the computation of the CramerLundberg constant C is much more complicated for the renewal case than for the compound Poisson case where C = (1 .3 For the delayed case Tl = s. This is known to be sufficient for ^(O) ]p (d) ([APQ] Proposition 3. Consider now the Lundberg case. For the stationary case.
St)} and h. The expressions are slightly more complicated and we omit the details. delayed version of Lundberg's inequality can be obtained in a e7u. 0) = ah (s) . h(s) = e(a +x( a))8 (3.Sdt] = Ee'uh(T) means 1 = f ' e°^B(dy) f ' h( s)A(ds).dt(ah ( s) + h'(s)) so that Gha ( s. y) = e°yh(s).(°) ( Ce8B[7] Ao(ds) similar manner. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES Proof Using (1. E8h0 (Jdt.a . For s > 0.h'(s). (s. Sdt) = h(s .5. (u + s . The underlying Markov process {Jt} for the Markov additive process {Xt} = {(Jt. we look for a function h(s) and a k (both depending on a) such that Gh. 0) = tc(a)h(s).3. we invoke the behavior at the 1 = h«(0. 3b Markov additive processes We take the Markov additive point of view of II.1) (normalizing by h(0) = 1).y) B(dy) 0 For the stationary case. IPA 0 Of course..St)} can be defined by taking Jt as the residual time until the next arrival. According to Remark 11.. Let P8f E8 refer to the case Jo = s. Here K. we get r u +8 e"8(u) 139 e7uB(u + s) + 4 0 + L 00 J e7(v8)e7(u+8v). B(x) = o(e7x) and dominated convergence.dt ) eadt = h ( s) . where G is the infinitesimal generator of {Xt} = {(Jt.5./c. To determine boundary 0. another use of dominated convergence combined with Ao[s] = (A[s] 1)/SPA yields 00 u) e7u iP8(u) Ao(ds) + f 0 = CB['Y](A[y] . 0 0 .0 ) = Eo[ha ( Jdt.9. Equating this to tch(s) and dividing by h(s) yields h'(s)/h(s) _ .4).1) = C(O).(s.
e(«+k(a))t esy A(dt). Remark 3 .. Proposition 3. An easy extension of the argument shows that U1. Ease AU1+6T2 [ AU1+6T2 = Ea a LT. . resp.rc(a)] B[a] A[a .4 The probability measure Pa. An important exception is.s is the probability measure governing a renewal risk process with Jo = s and the interarrival distribution A and the service time distribution B changed to Aa. ] = E. = J8 = T2..s governing {(Jt.8.5. [a + /3] A[b . . however.a . Note that the changed distributions of A and B are in general not the same for Pa. T2 are independent with distributions Ba. . Aa as asserted . T2. ..s(Jo = s) = 1 follows trivially from Lo = 1.S„):0<v< )be Lt = eaSt tK(a) h(Jt) = east . i.2). RENEWAL ARRIVALS B[a]A[a .13]A[b . St)}too by letting the likelihood ratio Lt restricted to Yt = a((J. Further. 5 For the compound Poisson case where A is exponential with rate . [e1U1 + 6T2ea ( U1s)stc ( a)e(a+K(a ))( T2s)I B[a +.tK( a)e.( a+r' (a))(Jt s) h(s) where c(a) is the solution of (3. (3.e. .. (3. Ba(dx) = B(dx). .rc(a)] = B = Ba[13]Aa[5]. the determination of the adjustment coefficient ry where the defining equations rc(d) (ry) = 0 and rc(ry) = 0 are the same. rc(a) = 0 (B[a] . since JT.s and P(d).rc(a)] = 1.140 CHAPTER V. A[a .c(a)] B[a] Proof Pa.2) means 1 = B[a]/3/(/3+a+rc (a)).c(a)] which shows that U1. resp. Ba where Aa (dt) . J n+1 u are independent with distribution Aa for the Tk and Ba for the Uk..a .1)a in agreement u with Chapter III..2) As in 11. U. we can now for each a define a new probability measure Pa.
yu ) < e7yu.6 Let y < let ay > 0 be the solution of ic'(ay) = 1/y.5.y yuAa y [ay + K(ay) .yu ) e7vu A[ay . or FCFS = first come first served) queueing discipline and renewal interarrival times.. T(u) < yu h(JT(u)) < eayu+yuk(ay ) ( Eia y Le(a(+k(ay))s v. Proof As in the proof of Theorem IV. the time from he arrives to the queue till he starts service.4.yx(ay).t. which is the same as the asserted inequality for 0. it is easily seen that ic(ay ) > 0.s eaysr(")+r(u ) K(ay) h (s) . defined as the single server queue with first in first out (FIFO. 2.r. Then J(rr(u)) = TM(u)+1 and hence Ws(u. yu). yu) = F'ay. that is. [APQ] Ch. 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. and U„ the service time of customer n.g. . that is. for the zerodelayed case zp8(u. The actual waiting time Wn of customer n is defined as his time spent in queue (excluding the service time). . and assume that T„ is the time between the arrivals of customers n . not after time T.ay+ray))TM(. the amount of . see e. The virtual waiting time Vt at time t is the residual amount of work at time t. For the approach via Markov additive processes. .1 and n. say finite horizon ruin probabilities where the approach via the imbedded random walk yields results on the probability of ruin after N claims.)+1 e J j e(ay+w(ay))8 e .5 Proposition 3. Then "^ e(ay+w(aY))8 Ys(u. A(ds).(u. XII. Let M(u) be the number of claims leading to ruin .4.. The claim for the zerodelayed case follows by integration w.rc( ay)] = e(aa+(r ))sb[a ]e7yu L y1 In particular. and define yy = ay .4. Label the customers 1. u The approach via the embedded random walk is standard. Using the Markov additive approach yields for example the following analogue of Theorem IV. Notes and references 4 The duality with queueing theory We first review some basic facts about the GI/G/1 queue. see in particular Dassios & Embrechts [98] and Asmussen & Rubinstein [45].
1 Wn+1 = (Wn + U.Tn)+ Proof The amount of residual work just before customer n arrives is VQ„ .4.1 and Corollary 11. since customer n arrives at time on. and we have P(V > u) = ?/iiol(u). we have Wn = Van(left limit).Tn)+.. Wn converges i n distribution to a random variable W. the proposition follows. Thus.1. Then P(r(u) < T) is the probability z/iiNi (u) of ruin after at most N claims. (4. Then: (a) as n + oo. (u).142 CHAPTER V.3. but we shall present a slightly different proof via the duality result given in Theorem II.+ Un. whereas in [On .2.4): Proposition 4. we get: Corollary 4. Also {Zt}o<t<T evolves like the leftcontinuous version of the virtual waiting time process up to just before the Nth arrival.4. (4. 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 11.3) Proof Part (a) is contained in Theorem 11. .4. and obviously z/'(u) = limN.2) (b) as t * oo. Vt converges in distribution to a random variable V.1. an+1) = [on. It then jumps to VQ„ . (4.1).• • • Tk ). then Wn v M. 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.. 0 Applying Theorem 11.3 Assume rl > 0 or. in which case {V} remains at zero until time on+1. equivalently. If W1 = 0. and we have P(W > u) = V. Thus Vos}1 _ = (Wn + Un .1) The following result shows that {Wn} is a Lindley process in the sense of II.2 Let Mnd) = maxk=o.4: Proposition 4. and combining with (4.n1 (U1 +• • •+Uk Tl . on + Tn) the residual work decreases linearly until possibly zero is hit. equivalently. but interchanging the set . Let the T there be the random time UN. p < 1. the waiting time a customer would have if he arrived at time t..."^ Vi(N) (u). The traffic intensity of the queue is p = EU/ET.
Ti) and similarly for the U. Letting n oo in Corollary 4.. hence (since the residual lifetime at 0 and the age at T have the same distribution . we obtain: Corollary 4. For part (b)..5 (LINDLEY'S INTEGRAL EQUATION) Let F(x) = P(U1T1 < x).e. (4.T*)+. which implies the convergence in distribution and (4...2). { 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). convergence in distribution hold for arbitrary initial conditions .T* = y yields K(x) = P ((W + U* ..T*)+ < x) = P(W + U* .. (4.1. It follows that P(WN > u) =.5) Proof Letting n . However. namely W1 = 0 in (a) and Vo = 0. we get W = (W + U* .T* < x) fK(x_y)F(dy) (x > 0 is crucial for the second equality!). Then the corresponding queue is M/G/1. In fact . K(x) = P(W < x). by an obvious reversibility argument this does not affect the distribution . and hence in particular ZT is distributed as the virtual waiting time just before the Nth arrival.4. we let T be deterministic . x > 0.2.. cf. u Now return to the Poisson case . but this requires some additional arguments (involving regeneration at 0 but not difficult) that we omit. A. T1 .4 The steadystate actual waiting time W has the same distribution as M(d). Hence for x > 0.oo in Proposition 4.Ao in (b)...le) the same is true for the timereversed point process which is the interarrival process for { Zt}o<t < T• Thus as before . T] form a stationary renewal process with interarrival distribution A. where U*.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. THE DUALITY WITH QUEUEING THEORY 143 (T1. i. Then the arrivals of {Rt} in [0. TN) with (TN.T* are independent and distributed as U1. conditioning upon U* .(N)(u) has the limit tp(u) for all u. Then K(x) = J x00K(x .y)F(dy). as WN. and we get: . Corollary 4. resp .. T1.
144 CHAPTER V. the zerodelayed and the stationary renewal processes are identical. Some early classical papers are Smith [350] and Lindley [246]. W v V. the actual and the virtual waiting time have the same distribution in the steady state. implying P(W > u) = P(V > u) for all u. That is. . 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. Proof For the Poisson case. The equation (4. RENEWAL ARRIVALS Corollary 4. 0 Notes and references The GI/G/1 queue is a favourite of almost any queueing book (see e .5) looks like the convolution equation K = F * K but is not the same (one would need (4. Note that (4. see e. Cohen [88] or [APQ] Ch.g.6 For the M/G/1 queue with p < 1. Asmussen [24] and references there. VIII).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).g.
and can be computed as the positive solution of WA = 0. The ruin probabilities with initial environment i are '+ki(u) = pi(T(u ) < oo) = Pi (M > u). i=1 0 and r(u) = inf It > 0: St > u}. M = supt>o St. t St = E Ui . • The premium rate when Jt = i is pi. dv.Chapter VI Risk theory in a Markovian environment 1 Model and examples We assume that arrivals are not homogeneous in time but determined by a Markov process {Jt}0<t<oo with a finite state space E as follows: • The arrival intensity is . • Claims arriving when Jt = i have distribution Bi. N. . As in Chapter I. Ire = 1.(3i when Jt = i. 145 Oj( u.f pi. {Jt} describes the environmental conditions for the risk process.)iJEE and its stationary limiting distribution by lr.. Thus. here it exists whenever A is irreducible which is assumed throughout. The intensity matrix governing {Jt} is denoted by A = (A. {St} denotes the claim surplus process.T) = Pi (T(u) < T).
say. and assume that weather conditions play a major role for the occurence of accidents. For example. An example of how such a mechanism could be relevant in risk theory follows. meaning that accidents occuring during icy road conditions lead to claim amounts which are different from the normal ones. P = E 7riPi. f3i and claim size distributions Bn. According to Theorem A5.5 below.2. we could distinguish between normal and icy road conditions. one expects that 3i > on and presumably also that Bn # Bi.1) iEE Then pi is the average amount of claims received per unit time when the environment is in state i. leading to E having two states n. Assume similarly that the sojourn time in the normal state has distribution A(n) which we approximate with a phasetype distribution with representation (E(').146 CHAPTER VI.11 below. = iii when j E E(i). assume that the sojourn time in the icy state has a more general distribution A(i). Example 1. respectively. cf. 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. the operational time argument given in Example 1. T(=)).. Thus. this is no restriction when studying infinite horizon ruin probabilities. t(i) = T(')e are the exit rates. a(i). u .2 (ALTERNATING RENEWAL ENVIRONMENT) The model of Example 1.1 Consider car insurance. and we have f3. Unless otherwise stated. We let p Pi = /ji/AB. Example 1 .4) with representation (E(i). we shall assume that pi = 1. MARKOVIAN ENVIRONMENT where as usual Pi refers to the case Jo = i.14. i and corresponding arrival intensities Qn. the intensity matrix is A OW) T(i) T(n) t(n)a(i) where t(n) = T(n)e.1 implicitly assumes that the sojourn times of the environment in the normal and the icy states are exponential. cf. and p is the overall average amount of claims per unit time. Proposition 1. we can approximate A(i) with a phasetype distribution (cf.a('). in blockpartitioned form.T(n)). Then the state space for the environment is the disjoint union of E(n) and E(i). /3 = Nn when j E E(n). with rates Ani and Ain. which is clearly unrealistic. say. Bi. r^ = P (1.
3i. dt. . but assume now that the arrival intensity changes during the icy period.1. let T 9(T) = f pi.p.Q. where W = (w. the state space E for the environment is { ('q. (9) where q = CHI... 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.. Example VIII. say.. resp.4) with states i1..>. we assume again pi = 1 so that the claim surplus is Nt St = ?Ui_t. and 1/ii(u) = t/ii(u). and similarly for the normal period. is the probability that a long icy period is followed by a short normal one. i8f n1. i E E(n) }.3i/pi. such that the icy period is of two types (long and short) each with their sojourn time distribution A('L). T(9) +wggt(9)0.a(n).T(n)). St = SB=(t). u From now on.. iq (visited in that order) and letfOil >. 1 . n8}.2.3.1. Qi = . one could for example have H = {i1. A('^). .j = . Then for example wi. the parameters are ^ij = aid/pi. 0 Then (by standard operational time arguments) {St } is a risk process in a Markovian environment with unit premium rate. This amounts to a family (A(")) ?CH Of sojourn time distributions.. such that a sojourn time of type rt is followed by one of type c w.. One way to model this would be to take A(') to be Coxian (cf. t(n) = T("i)e. w.3 Consider again the alternating renewal model for car insurance in Example 1. say it is larger initially. Indeed. u Example 1 . it = Jel(t). In the car insurance example.J017. depending only on 77. and . The simplest model for the arrival intensity amounts to . u Example 1. 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) .5 (MARKOVMODULATED PREMIUMS) Returning for a short while to the case of general premium rates pi depending on the environment i.tEH is a transition matrix... i ) : n E H...n. Approximating each A('?) by a phasetype distribution with representation (E('l). MODEL AND EXAMPLES 147 Example 1 .
7 The Pidistribution of T1 is phasetype with representation (ei.(Qi)diag)• More precisely. we put )3* = E 7fi/3i. Pi (Ti E dx.A . )3*. 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. o = 0.(3iBi(dx). In particular. one can associate in a natural way a standard Poisson one by averaging over the environment. More precisely. The key property for much of the analysis presented below is the following immediate observation: Proposition 1. the empirical distribution of the claims is B*. t l=1 Note that the last statement of the proposition just means that in the limit. dx. B* = 1 /^* Bi.e(A(Oi)d'sg)xe. 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. Next we note a semiMarkov structure of the arrival process: Proposition 1. . iEE iEE )3 These parameters are the ones which the statistician would estimate if he ignored the presence of Markovmodulation: Proposition 1. 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). qij = 0 in the notation of Chapter 11.148 CHAPTER VI. N > 1(Ul < x) a4 B*(x). Nt Nt a . MARKOVIAN ENVIRONMENT We now turn to some more mathematically oriented basic discussion.5. . JT1 = j) = Qj • e. the Markov additive structure will be used for exponential change of measure and thereby versions of Lundberg's inequality and the CramerLundberg approximation. vi(dx) = .8 As t oo.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.
^j 7riNi. Then {St)} + {St*} in D[0. MODEL AND EXAMPLES 149 Proof Let ti = f1 I(JJ = i) ds be the time spent in state i up to time t and Nti) the number of claim arrivals in state i . By Proposition A5. and let {St °i} refer to the one with parameters Pi.(/3i)aiag). A. y Ni) i Nti) t a.. aA. i. Bi(x). {St} to the compound Poisson model with parameters 0 *. given {Jt}0<t<0.1. we may view Nt`i as the number of events in a Poisson process where the accumulated intensity at time t is Niti. Proof According to Proposition 1.9 Consider a Markovmodulated risk process {St} with param eters Ni.. the limiting distribution of the first claim size U1 is B*. 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. the Fidistribution of T1 in {St(a ) } is phase type with representation (E. cf. B*. e.* (u) for all u.6. has distribution (7ri)3i //3*)iEE and is independent of Ti. In particular. Bi. zli( (u) . Conditioning . and furthermore in the limit JT. Hence Nt'> a .x) = Nt E > I (Uk) X) Nt Bi(x) 1=1 iEE k=1 iEE 1: t5 Bi( x) = B*(x). N + oo Hence 1 Nt 1 N`+) Nits Nt E I ( Ut <. iEE 13 A different interpretation of B* is as the Palm distribution of the claim size.7. However . Proposition 1. this converges to the exponential distribution with rate 0* as a * oo...4.aA . Nt a' t t iEE Also. In particular. Then it is standard that ti lt '4' iri as t > 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$. Example 11. oo) as a 4 oo. Bi.2..
shows similarly that in the limit (T2.g. From this the convergence in distribution follows by general facts on weak convergence in D[0. Claims of type E3 arrive with intensity 2 . Thus.. MARKOVIAN ENVIRONMENT upon FT..31µB 2 = 2 5 3 7 70 Thus in state 1 where p. the company even suffers an average loss. 9 . oo). state 1 appears as more dangerous than state 2. B1=3E3+2E7.2}.1. On Fig. 5 5 where E5 denotes the exponential distribution with intensity parameter 5 and a > 0 is arbitrary. A= (  a a ) \ a a 5 5 J 9 3 2 a1=2. U.. thick. e. 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 = . 132=2.2 +2 2 = 3.s = 1o in state 2. which also yield O(a) (u..150 CHAPTER VI. we first get that 3 (3* = 2. T) for all u and T.).. the overall drift is negative since it = (2 2) so that p = 71P1 + 112P2 = 7.2. T) + ?P* (u. B2=1E3+4E7.2. lines in the path of {St}. resp.l3* and U2 having distribution B*. > 1. from Theorem 3.=1.... since E3 is a more dangerous claim size distribution than E7 (the mean is larger and the tail is heavier).10 Let E_{1. those of type E7 with intensity z s = 5 in state 1 and with intensity z . Continuing in this manner shows that the limiting distribution of (T.1 of [145]. 1. Computing the parameters of the averaged compound Poisson model. U2) are independent of . and (at least when a is small such that state changes of the environment are infrequent). the paths of the surplus process will exhibit the type of behaviour in Fig. there are p = 2 background states of {Jt}.FT. 1. we may imagine that we have two types of claims such that the claim size distributions are E3 and E7.. with T2 being exponential with rate .1 with periods with positive drift alternating with periods with negative drift. 0 Example 1.s = o in state 1 and with intensity 1 . s 5 in state 2. That is. marked by thin. The fact that indeed 0(a) (U) 3 0* (u) follows. is as in {St }.
= P.1 a.3* = 3/4 of the claims occur in state 1 and the remaining fraction 1/4 in state 2.8. 0 The definition (1. For (b). t * oo.1) of the safety loading is (as for the renewal model in Chapter V) based upon an asymptotic consideration given by the following result: Proposition 1.1. note first that EN Uk')/N a4' µgi. the averaged compound Poisson model is the same as in III. a fraction r.11 (a ) ESt/t * p . Hence B* = 415E3+5E7/ +4 ( 51E3 +5 E7) = 1E 3 +2E7. Hence (i) Nti) 1 U(i) k' N(i)k=1 E t 4 St + t = iEE Nt t 1: 7ri Qi µs.1).(3.1. Proof In the notation of Proposition 1.. 01 /.s. (b) St/t * p . iEE . t + oo. MODEL AND EXAMPLES 151 Figure 1. we have E[St + t I (t(i))iE EI = E t(i)OW = iEE t(i)Pi• iEE Taking expectations and using the well known fact Et(i)/t * 7ri yields (a). That is.1 Thus.
[315].s. also + .2(a) p. . Proposition 1. and hence M = 00. Then by standard Markov process formulas (e. There seems still to be more to be done in this area. see the Notes to Section 7..Jt=i}.. and so on. X2 =SW2 So. Proof The case 77 < 0 is trivial since then the a.\ i and EiX1 Ei f 13 J... having the Pidistribution of X.s. and hence oscillates between 0o and oo so that also here M = oo.g. then M < oo a.. see [APQ] p.152 CHAPTER VI. with X2. EiX = 0. let some state i be fixed and define w=wl=inf{t >0:Jt_#i.. If 77 > 0. limit p . 38) Eiw1 = 1/ir. + Xn SWn ](1 a . See Meier [258] and Ryden [314]. In risk theory. X3. Eiw.12 If 77 < 0. and .0i(u) < 1 for all i and u. some early studies are in Janssen & Reinhard [211]. The proof of Proposition 1.1 of St / t is > 0. dt .a form a renewal process . and hence 1/ii(u ) = 1 for all i and u. and hence wn /n a4.1 and the Corollary are standard. then M = 00 a. 136 or A. 2 The ladder height distribution Our mathematical treatment of the ruin problem follows the model of Chapter III for the simple compound Poisson model. and a more comprehensive treatment in Asmussen [16].1(b) is essentially the same as the proof of the strong law of large numbers for cumulative processes. [212]. The mainstream of the present chapter follows [16]. s. Theorem II. The case 77 > 0 is similarly easy.. Now let r) = 0. Now obviously the w. [302]... 0 Notes and references The Markovmodulated Poisson process has become very popular in queueing theory during the last decade. Statistical aspects are not treated here. X 1 =Sty. and involves a version of the . Since the X„ are independent ..Eiw o'o Eiw • E ^ifjµs. [APQ].ld. n n Thus {SWn l is a discrete time random walk with mean zero. with some important improvements being obtained in Asmussen [17] in the queueing setting and being implemented numerically in Asmussen & Rolski [43]. w2=inf {t>w1:Jt_#i.1 jEE = (p .1)Eiw = 0. MARKOVIAN ENVIRONMENT Corollary 1.4.Jt=i}. . PB.
T R(i. However . only with the product of real numbers replaced by convolution of measures.3*B *(y)dy. 00 (2.A) = Pt(ST+ E A.a/i.IIG +II)e. For measurevalued matrices.(u) = Pi(M < u) = e' E G+ (u)(I . Thus. •) * G +(k. by specializing results for general stationary risk processes (Theorem II . dx)/jBj(y . 6. . B* in Section 1. we define the convolution operation by the same rule as for multiplication of realvalued matrices.A) =ZI(St E.1) 0 (b) G+ (y.2.Jt=j)dt. but is substantially more involved than for the compound Poisson case . n=0 (2. j. the definition of .. let G+(i.x.1 irG+(dy)e =. for i.j. j.2) R(dx)S((y . j.j E E.4) we obtain the following result . oo) = J ao 0 G+(i. •) II = JG+(i.Jr+ =j. IIG+ II denotes the matrix with ijth element IIG+(i. see also Example II. oo)) = f R(i. and S (dx) the measure valued diagonal matrix with /3 Bj(dx) as ith diagonal element. j. •)• kEE Also. e.2(a) below ) where the ladder height distribution is evaluated by a time reversion argument.6*. That is.x). we get the same ladder height distribution as for the averaged compound Poisson model. •). which represents a nice simplified form of the ladder height distribution G+ when taking certain averages : starting {Jt} stationary. Proposition 2.2 (a) The distribution of M is given by 00 1 . Proposition 2.EA. THE LADDER HEIGHT DISTRIBUTION 153 PollaczeckKhinchine formula (see Proposition 2. G+ is the matrix whose ijth element is E G +(i. oo)). (y.dx). Let further R denote the preT+ occupation kernel.g.j.5.i. k. The form of G+ turns out to be explicit (or at least computable). j. cf.6. Define the ladder epoch T+ by T+ = inf It : St > 0} = r(0). T+ < oo) and let G+ be the measurevalued matrix with ijth element G+(i.
3) We let {St*} be defined as {St}. and let further {my} be the Evalued process obtained by observing {Jt } only when {St*} is at a minimum value. we need to invoke the timereversed version {Jt } of {Jt} . III to bring R and G+ on a more explicit form . only with {Jt} replaced by {Jt } (the /3i and Bi are the same ). marked by thin. MARKOVIAN ENVIRONMENT Proof The probability that there are n proper ladder steps not exceeding x and (x)ej. . thick. That is. resp. lines in the path of {St}. 0  x Figure 2. hence uniquely specified by its intensity matrix Q (say). and that the environment is j at the nth when we start from i is e .IIG+II)e. we need as in Chapters II. 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.2 useful . see Figure 2. The u proof of (2. St < S* for u < t. G+ the probability that there are no further ladder steps starting from environment j is e^ ( I .3 When q > 0.1 The following observation is immediate: Proposition 2. mx = j when for some (necessarily unique) t we have St = x. To make Proposition 2.1) follows by summing over n and j. JJ = j.3.154 CHAPTER VI. From this (2. To this end .2) is just the same as the proof of Lemma 11.6.1 for an illustration in the case of p = 2 environmental states of {Jt}.
( Q( n)) converges monotonically to Q. and S(dx) is the diagonal matrix with the f3iBi(dx) on the diagonal.2 . Proof The argument relies on an interpretation in terms of excursions. For example the excursion of depth 2 has one subexcursion which is of depth 1.*. In general. If there are no jumps in (t. 0 mms1   ^O \ T.2. {S.(/3i)diag + T S(dx) eQx. we recursively define the depth of an excursion as 1 plus the maximal depth of a subexcursion.2. } is a minimum value at v = t. the sequence {Q(n)} A* defined by Q(O) = . Figure 2. corresponding to two subexcursions of depth 0. s]. 2. THE LADDER HEIGHT DISTRIBUTION 155 Proposition 2.2 where there are three excursions of depth 1.4 Q satisfies the nonlinear matrix equation Q = W(Q) where 0 co(Q) = n* . Q( n+l) _ ^.and a jump (claim arrival) occurs at time t. = x}.. we say that the excursion has depth 0.(/3i) diag. Furthermore. and the excursion ends at time s = inf {v > t : S. An excursion of {St*} above level x starts at time t if St = x. and the excursion is said to have depth 1 if each of these subexcursions have depth 0. The definitions are illustrated on Fig. Otherwise each jump at a minimum level during the excursion starts a subexcursion.0. 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).
i.s. or through an arrival starting an excursion terminating with J. Then a jump to j (i.j. of the definition to make U be concentrated on (co.4). Writing out in matrix notation . 0)).Qi + )%pij) Now just note that t pij and insert (2. p1^) Define a further kernel U by f U(i. Similarly by induction . = j.St EA. Fi(mh =i ) = 1 + =hflh+Qihpii+o(h) implies qii = 'iii /i +)3ipii..4) To show Q = cp(Q). Suppose mx = i. either due to a jump of {Jt } which occurs with intensity A= j. it becomes clear that pij = r [eQh] 0 ij Bi (dy) • (2. Now let {m ( n) } be {mx } killed at the first time i7n (say) a subexcursion of depth at least n occurs . the subintensity matrix of {min+i ) } is cp (Q(n)) = Q(n +l) which implies that qgj +1) = \!. It follows that qij = A.156 CHAPTER VI. (2. A) = L' U(j. StEA .T+>t) _ ^iF 7ri (JJ =i.(01)diag = Q. j. By considering minimum values within the excursion. A) = f Pi(mx = j) dx eie4xej dx A u (2..5) A (note that we use A = {x : x E Al on the r. Theorem 2 .u< t). e. A). Q = W(Q) follows. we first compute qij for i $ j. MARKOVIAN ENVIRONMENT Let p=7) be the probability that an excursion starting from Jt = i has depth at most n and terminates at J8 = j and pij the probability that an excursion starting from Jt = i terminates at J8 = j. It is clear that { mini } is a terminating Markov process and that { mio) } has subintensity matrix A* . mx+dx = j) occurs in two ways . h.St <S*. Similarly.6) . 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.j +/3ipij. 7rE Proof We shall show that Fi(Jt=j.
. THE LADDER HEIGHT DISTRIBUTION 157 from which the result immediately follows by integrating from 0 to oo w. 0 +1) = cp (K( n)) defined by K(o) = A . Jt = i. Remark 2..=StSt. 0<u<t). where A is the diagonal matrix with 7r on the diagonal: Corollary 2.StEA. and get irPi(Jt =j. and we let k be the corresponding right eigenvector normalized by Irk = 1. consider stationary versions of {Jt}. G+((z. it is readily checked that 7r is a left eigenvector of K corresponding to the eigenvalue 0 (when p < 1). S.(. (c) the matrix K satisfies the nonlinear matrix equation K = W(K) where W( K) = A ( i) diag + fi J "O eKx S(dx). and this immediately yields (2. dt.Qi)diag. K( n (d) the sequence converges monotonically to K. and to obtain a simple solution in the .z+>t) = P.0<u<t. `` {K(n)} [the W(•) here is of course not the same as in Proposition 2.S„<0. we shall see that nevertheless we have enough information to derive.g.. We may then assume Ju=Jtu. (Jo = j. {Jt }.r. St < St U.6 (a) R(dx) = eKxdx.0<u<t) = P. 0 < u < t) = 7rjPj(Jt =i. From Qe = 0.. u It is convenient at this stage to rewrite the above results in terms of the matrix K = 0'Q'A.7 It is instructive to see how Proposition 2. (b) for z > 0.t.4]. oo)) = f o' eIXS((x + z.(Jt=j. To this end.St EA..Jo=i. St EA.6 is hardly all that explicit in general.St <Su. x < 0.1 can be rederived using the more detailed form of G+ in Corollary 2. e. St E A. oo))dx. the CramerLundberg approximation (Section 3).2.6(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 2.6).
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
V)i(u. yu < r(u) < 00 h 4(u) < h(av)C+o)(y)eavuEav .i [eT(u)K(av ). we shall need the matrices G+ and R of Section 2.j) * coj)(u) _ f u ^Pj(u .9) For the proof. as in the classical case (3. av 'i [h.i [e*(u)K(av).3.. r(u) yu o)(y)eavuEav. exp {e() + r(u))} . We further write G(u) for the vector with ith component Gi(u) = EiEE G+(i. if y > 1lk'(ry). for a vector <p(u) = (cpi (u))iEE of functions . (3. we let G+ * W(u) be the vector with ith component E(G+(i. dy)• o iEE jEE . yu) f h(av)e v avuE«v.5) will produce the maximal ryy for which the argument works. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 165 hiav)e _avuE.7. 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 3.11 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). Chie ryu < Vi(u ) < C+hie 7u.00 su e7( ( 3. However. hj P .(ay)}.y)G+(z. 1 Similarly. we have ic(ay) < 0 and get 'i(u) .8 ) Then for all i E E and all u > 0. r(u) < yu] hiay)C+ h=av)C+ o) (y)eayu+yuw(av).i I (a) exp {aye(u) + r(u)r. 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 (3.5).j. (u. oo)) and.
166 CHAPTER VI. Proof Write UN = EN G+ .ery(&u+x)Bj (dy) Bj(u Bj (u .j.12 Assume sup1. Then iterating the defining equation ip(n+1) = G + G+ * V(n) we get W(N+1) = UN * G + G+N+1) * ^(o) However. °O . dx). dy) 00 C+ ijhj f R(i. dx ) Bj (u . U = U".x) jEE 00 u 0 //^^ C+E. dx) f e7( vu)Bj (dy . 0 G+(i. = Eo G+ G. MARKOVIAN ENVIRONMENT Lemma 3 .x ) = Gi(u). we have G *(N +1) * ^. dy) = aj f Bj(dy .13 For all i and u. Lemma 3 . and define W(n+1) (u) = G(u) + (G+ * tp(n))(u). 00 f C_ hj f e(Y)G+(i. dy) : 1(u) < C+ > hj u e(1tL)G+(i. jEE u 0 j. n > oo.3jhj // f 00 R(i. Then cpin)(u) sit (u) as n + oo.dy).u Iv 2°)(u)I Pi(rr+(N + 1) < oo) + 0. if r+ (n) is the nth ladder epoch. j. _ To see that the ith component of U * G(u) equals ?Pi (u). dx) 100 C .u IMP:°) (u) I < oo.7. j. j. 00 Thus C+ > hj f"o e7(Yu)G +(i.j. just note that the recursion <p(n+1) = G + G+ * (p(n) holds for the particular case where cpin)(u) is the probability of ruin after at most n ladder steps and that then obviously u cp2n) (u) + t.x ) R(i.& (u). Hence lim cp(n) exists and equals U * G.x) x) jEE 0 E Qj f jEE R(i. j.(0) ] (u) < sup Jt t.
Indeed.13 Let first cp=°)(u) = C_ hie"u in Lemma 3.11) C_e7u 57 O+[i. j. from which the lower inequality follows by letting n * oo. ST < u] < C+(yo)e7ouEi [h^7o)e70ST1 l T J = C h(7o)e7ou8T . we get Wo n +1) (u) = ? 7 i ( U ) + E J u gyp.(u) < T ) to 0i (u) which is different from Theorem 3. Then 0< Vi (u )  0i(u.ST). letting MT = maxo<t<T St.8) with y replaced by yo and hi by h=7o ). +i .M>u) = Ei [VGJT (u . j.13) Hence. 14 Let yo > 0 be the solution of 'c'(yo ) = 0. jEE estimating the first term in (3.13. T) = Pi (7. and the proof of the lower one is similar. and let 8 = e'(70).n) ( u . it follows that Vi(u) < C_(yo) h=70)e7ou. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES proving the upper inequality. dy) (3. 13 and the second by the induction hypothesis . let C+(yo) be as in (3. (3.T) = Pi(M > u) . j. Here is an estimate of the rate of convergence of the finite horizon ruin probabilities 'i (u. taking cps°) (u) = 0. and assuming it shown for n.tpi(u.y)G+(i.12) Proof We first note that just as in the proof of Theorem 3. and using Lemma 3 . T) < C+(')' o)hi7u)e7ou8T . 167 u Proof of Theorem 3. u The proof of the upper inequality is similar . this is obvious if n = 0. 9 for the last equality in (3. j.MT<u. we have Vii (u) . y]hj = C_ e7uhi. MT < u.10) C_ 1 f hje7(y.M > u) = Pi(ST<u.11). dy) jEE u U +C_ hje7( yu)G jEE"" +(i. We claim by induction that then cpin) (u) > C_ hie7u for all n. dy) jEE o (3.11.3. (3.u)G+(i.Pi(MT > u) = Pi(MT < u.10: Theorem 3 .10 ) by Lemma 3 .
o. but that in general the picture is more diverse. V)" if z/i'(u) <'c"" (u). this is not the case for (4. Occasionally we strengthen (4. where it has been observed repeatedly that Markovmodulation increases waiting times and in fact some partial results had been obtained. B2 <_s. ".2) (4. MARKOVIAN ENVIRONMENT Notes and references The results and proofs are from Asmussen and Rolski [44]. where o*(u) is the ruin probability for the averaged compound Poisson model defined in Section 1 and . we define the stochastic ordering by 0' < s.. The conditions which play a role in the following are: .31:5)32 ..2) alone just amounts to an ordering of the states. we also assume that there exist i # j such that either /3i <.. It was long conjectured that 0* Vi. (4. u > 0. this correponds to the usual stochastic ordering of the maxima M'.3) Bl <_s.1) Obviously. The results to be presented show that quite often this is so. [177]. <s..o. (4. 0"(u) = P(M" > u)) Now consider the risk process in a Markovian environment and define i' (u) _ >iEE irioi(u). The motivation that such a result should be true came in part from numerical studies.3) to B = Bi does not depend on i.33 or Bi 0 Bj. we refer to .3)..0.5) Note that whereas (4.. is the one for the Markovmodulated one in the stationary case (the distribution of J0 is 7r).o. Further related discussion is given in Grigelionis [176].3p.168 CHAPTER VI. The Markov process {Jt} is stochastically monotone (4. For the notion of monotone Markov processes. Bp. M" of the corresponding two claim surplus proceses (note that 0'(u) _ P(M' > u).4) To avoid trivialities.o.. in part from the folklore principle that any added stochastic variation increases the risk. < . 4 Comparisons with the compound Poisson model 4a Ordering of the ruin functions For two risk functions 0'. and finally in part from queueing theory. (4..
2 If al < . COMPARISONS WITH THE COMPOUND POISSON MODEL 169 Stoyan [352].6)..x)dx _ /3*B*(u) + f u / ^ t=1 > 3 * B* ( ) + f (4.4. The first is a standard result going back to Chebycheff and appearing in a more general form in Esary.10) and (4.2.1 Assume that conditions (4.2. then P P P 7rjbj.6) 7r= fl*B*(u) + p> s=1 +) fu 0 b (u  x)Bt (x) /pB. dx (4. . also Proposition 2. = aP or b1 = . Theorem 4 . where 7r2+) = QiµBilri/p. and it is in fact easy to show that Vii(u ) < t/j(u) (this is used in the derivation of (4.8) ^j Tri/iBd(x) .7) and Lemma 4.9) follows by considering the increasing functions 3iBi (x) and Oi (u . (4.. < bp and 7ri > 0 (i = 1. we need two lemmas..4) say basically that if i < j ..1) which with basically the same proof can be found in Asmussen & Schmidt [49].4) is automatic in some simple examples like birthdeath processes or p = 2 . ^i 7ri = 1. Conditioning upon the first ladder epoch. T(0) < oo) = Bi(x) dx/tcai . 0 Here (4..2)(4.3* f uB(x) z/^. p).2)(4.. Proschan & Walkup [140].3 for the second) *(u) _ /3 *B* (u) +. Proposition 2.r (JT(o) = i.. Comparing (4.9) (4.* For the proof. = b..5 (cf.. Then V.x) of i and using Lemma 4. we obtain (cf. Lemma 4 . b1 < .r(u x)dx.x)B*(x) dx. .1. Conditions (4.4) hold.3iBi(x)YPi(u .. note that (4.10) Q*B*(u)+. Proof of Theorem 4. the second follows from an extension of Theorem I1.1 for the first term in (4.9 ) below). it follows by a standard . 7(0) < oo) = pirf+).6.. (b) P. < a.13* J0 u 0*(u . Section 4. E 7r i Wi(u .. 3 (a) P.x) dx u o i =1 i=1 (4.r (Sr(o) E dx Jr(o) = i..7) 7ri. 1:7riaibi > E 7riai i=1 i=1 j=1 The equality holds if and only if a1 = . then j is the more risky state .. Lemma 4 .
3µi < 1 for all i. As is seen. let = ( 1/2 1/2 ) .s.r (u ) fails for all sufficiently small e > 0. they are at present not quite complete.h. (4. Rolski & Schmidt [32]..8) we get P P '*' (0) = 3* + /3*1* (0) _ > lri'3qqi • E 7i/ipBi .(0) = V.0. that P P /^ 1r1NiµBi /^2 /^ ^i/ji pBi < 1il3i i=1 i=1 (4. Notes and references The results are from Asmussen. Frey. it is sufficient to show that 0'..4) is essential (the present author conjectures it is).11) is of order 104 and the r. of (4..1 and Proposition 4. except possibly for a very special situation .11) i=1` and that A has the form eAo for some fixed intensity matrix A0..1 of [145] for a formal proof) that z/ii(u) converges to the ruin probability for the compound Poisson model with parameters . MARKOVIAN ENVIRONMENT argument from renewal theory that tk. Bi as e J. Proof Since 0.4 Assume that ..s. Using (4. For u = 0.6).6). dominates the solution 0* to the renewal equation (4.4 is the understanding of whether the stochastic monotonicity condition (4. u To see that Proposition 4. (u) may fail for some u. it will hold for all sufficiently large u. 0. µB2 = 104. u Here is a counterexample showing that the inequality tp* (u) < V). What is missing in relation to Theorem 4.h.170 CHAPTER VI.3i.2. Q2 = 1./3*. = 102. of order 101. 01 = 103. 4b Ordering of adjustment coefficients Despite the fact that V)* (u) < *. Then the l.4 is not vacuous.* (0). Then i/i*(u) < . this ruin probability is /3iPBi. i=1 i=1 7'r(0) _ EFioiwi(0) . µB. (u) is not in general true: Proposition 4.(0) < b *'(0) for e small enough.0*• i=1 But it is intuitively clear (see Theorem 3. Recall that . and from this the claim follows.
The adjustment coefficient y* for the averaged compound Poisson model is the solution > 0 of rc*(ry*) = 0 where rc*(a) _ 13*(B*[a] . Lemma 4.13) implies A(a) > 0 for all a.. and by Proposition II.)a. (4. Jt = i])' EE = vA+n(6.a.13) (4. in particular . e.(a) > 0 for all a 0 0.1) .1) . This implies that A is strictly convex. it follows by Proposition A1. Now we can view {Xt} as a cumulative process (see A. It is clear that the distribution of X.g.2 we have (Ei[e"X'.g. Then {(Jt. is nondegenerate unless bi does not depend on i E E. (4.6 Let (di)iEE be a given set of constants satisfying EiEE iribi = 0 and define A(a) as the eigenvalue with maximal real part of the matrix A + a(bi)diag• Then )t(a) > 0.5.7) )i is convex with A'(0) = lim EXt tioo t = iEE 70i = 0. (4.5 y < ry*.12) iEE Theorem 4.5. 0 .a = E irirci(a).4(b) that the limit in (4.14) is nonzero so that A"(0) > 0. with strict inequality unless rci (y*) does not depend on iEE. Xt)} is a Markov additive process (a socalled Markovian fluid model.. cf. Hence if 5i 54 0 for some i E E. Asmussen [20]) as discussed in 11. with strict inequality unless a = 0 or bi = 0 for all i E E.ld) with generic cycle w = inf{t>0: Jt_54 k. Proof Define X= f &ids.Jt=kI A (the return time of k) where k E E is some arbitrary but fixed state.14) A„(O) iioo varXt t t By convexity. which in view of EiEE 1ibi = 0 is only possible if Si = 0 for all i E E. 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] . Further (see Corollary 11.4.5.
a = 1 in Lemma 4. Further a(1) = rc(y*) by definition of A(. In the case of e. If rci(y* ) is not a constant function of i E E.Qi and Bi are fixed . we get rc (y*) > 0 which in a similar manner implies that u y < y*.15) once more and letting e = 0 we get .6. h(0) = e. h depend on the parameter (e or a). multiply the basic equation by a to obtain 0 = (A0 + e(r£i(y))diag)h. h'(0) = (Ao .5. Then > risi = 0 because of (4.eir)h'(0). this implies that the solution y > 0 of K(y) = 0 must satisfy y < y*. 0 = ((ri(Y))diag + ery (4{('Y))diag)h + (A0 + e(?i'Y))diag)h'.5 is from Asmussen & O'Cinneide [40]. (4. Here we put a = 1/e.e7r)1 (Ici(Y*))diage. The corresponding adjustment coefficient is denoted by ry(e). whereas the . Thus y(e) * y* as e 10. Notes and references Theorem 4.. y. and our aim is to compute the sensitivity ay e a E=O A dual result deals with the limit a 4 oo. (4. where A..15) Normalizing h by 7rh = 0. 4c Sensitivity estimates for the adjustment coefficient Now assume that the intensity matrix for the environment is Ae = Ao/ e. we have 7rh' = 0.15) yields 0 = (Ii(y*)) diage + Aoh'(0) = (rci('Y*)) diage + (Ao . the basic equation is (A + (rci(y))diag)h = 0. Frey.172 CHAPTER VI. Hence rc (y*) > 0. Hence letting e = 0 in (4.) and rc (•).p yi and compute 8y 8a a=0 In both cases. Let bi = rci(y*).12) and rc*(y*) = 0. Since ic is convex with rc'(0) < 0 . MARKOVIAN ENVIRONMENT Proof of Theorem 4.. Rolski & Schmidt [32]. note that y(a) + mins=1. improving upon more incomplete results from Asmussen..16) Differentiating (4.
0 = (Ao + ry'(ii(Y)) diag )h + (aAo + (Ki(7'))diag)h'. Frey. Rolski & Schmidt [32]. We get 0 = (aAo + ( lc&Y))diag)h.8 when ryi < 0 for some i is open. (4. We assume that 0 < y < 7i. Inserting (4. 5 The Markovian arrival process We shall here briefly survey an extension of the model.17) (4. (4..19) Then 'y ^ ryl as a ^ 0 and we may take h(0) = el (the first unit vector).19) holds.20) Letting a = 0 in (4. multiplying (4. THE MARKOVIAN ARRIVAL PROCESS 173 0 = 27'(0)(ri(`Y *)) diage + 2(ci('Y* )) diag h' (0) + Aoh" (0) . which has recently received much attention in the queueing literature. The analogue of Proposition 4.7 8ry aE = 1 7r(ci ('Y*))diag ( Ao e7r)1(Xi(Y*))diage *=0 P Now turn to the case of a. i = 2. The additional feature of the model is the following: • Certain transitions of {Jt} from state i to state j are accompanied by a claim with distribution Bid. . .16) yields Proposition 4.. (4.5.17) by 7r to the left to get (4.18). then 8a a=o All rci (0) Notes and references The results are from Asmussen.18) 0 = 27'(0)p+27r(rs. the intensity for such a transition (referred to as marked in the following) is denoted by Aii l and the remaining intensity . and we have proved: Proposition 4. p. and may have some relevance in risk theory as well (though this still remains to be implemented).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).i(7' *))diagh'(0).8 If (4.
T). Again . with common distribution B. The extension of the model can also be motivated via Markov additive processes: if {Nt} is the counting process of a point process. the definition of Bij is redundant for i i4 j.2 for details). is neither 0 or 1 is covered by letting Bij have an atom of size qij at 0. . This is the only way in which arrivals can occur.^) etc. Thus . and that are determined by A = A(l ) +A(2) where A is the intensity matrix the governing {Jt}. Indeed. the Markovmodulated compound Poisson model considered sofar corresponds to A(l) = (. We then let (see the Appendix for the Kronecker E = E(1) x E(2). For i = j.2). Note that the case that 0 < qij < 1. where qij is the probability that a transition i * j is accompanied by a claim with distribution. Bij = B. we may let {Jt} represent the phase processes of the individual interarrival times glued together (see further VIII.1 (PHASETYPE RENEWAL ARRIVALS) Consider a risk process where the claim sizes are i. that Bii = Bi . the claim surplus is a Markov additive process (cf. but the point process of arrivals is not Poisson but renewal with interclaim times having common distribution A of phasetype with representation (v. In the above setting.i. and thus 1i = 0. then {Nt} is a Markov additive process if and only if it corresponds to an arrival mechanism of the type just considered.174 CHAPTER VI. u Example 5 . A(l) = T. the definition of Bi is redundant because of f3i = 0. Jt = (Jtl).2 (SUPERPOSITIONS) A nice feature of the setup is that it is closed under superposition of independent arrival streams . j(2) } be two independent environmental processes and let E(k).(13i )diag. A ( 2) = A (2`1 ) ® A.4). Here are some main examples: Example 5 . II.6i ) diag. and the marked transitions are then the ones corresponding to arrivals. Bii = Bi .2) A(1) = A(' 1) ® A(1. let { Jt 1) }. A(1'k) A(2 k1). B. Jt2)) (2. MARKOVIAN ENVIRONMENT f o r a transition i + j by A . we use the convention that a1i = f3i where 3i is the Poisson rate in state i. A(1) = A . A(l) = tv.d. refer to notation) { Jt k) }.
iN.iN C17 AilO. E = { WORKING..1i2 . or.g.5. RETIRED.. where ik = 0 means that the kth policy has not yet expired and ik = 1 that it has expired.. the idea of arrivals at transition epochs can be found in Hermann [193] and Rudemo [313].1i2. with rate ai. Hermann [193 ] and Asmussen & Koole [37] showed that in some appropriate . Similarly. 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 . INVALIDIZED. u Example 5 . all Al i2.. Bilo.. E 10. say.iN are zero and all Bi are redundant. after which it starts afresh..}.4 (A SINGLE LIFE INSURANCE POLICY ) Consider the life insurance of a single policy holder which can be in one of several states.iN.. Thus. e.. In this way we can model.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). MARRIED. i2i . 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. the kth policy enters a recovering state.. and that the policy then expires. • upon a claim. WIDOWED. DEAD etc... DIVORCED.. claims occur only at state transitions for the environment so that AN2.iil.kj = Bik) B13 4k = Bak) 175  (the definition of the remaining Bij.. assume that there is a finite number N of policies...iN = C27 All other offdiagonal elements of A are zero so that all other Bii are redundant.. iN. Assume further that the ith policy leads to a claim having distribution Ci after a time which is exponential.kl is redundant). However . In fact . superpositions of renewal processes. possibly having a general phasetype sojourn time.iN. The versatility of the setup is even greater than for the Markovmodulated model..iN = a2.. THE MARKOVIAN ARRIVAL PROCESS Bij. more recently. Example 5 . iN = all BOi2.. as the Markovian arrival process ( MAP).. 11. Easy modifications apply to allow for • the time until expiration of the kth policy is general phasetype rather than exponential.iil. This means that the environmental states are of the form i1i2 • • • iN with il..
p)/p. a claim arrives with rate /3(s + t) and is distributed according to B(8+0 . 0 < t < 1. [248]. Lucantoni [248]. one needs to assume also (as a minimum) that they are measurable in t. one limitation for approximation purposes is the inequality Var Nt > ENt which needs not hold for all arrival streams.176 CHAPTER VI. Obviously. continuity would hold in presumably all reasonable examples. .2) Note that p is the average net claim amount per unit time and µ* = p//3* the average mean claim size. we may assume that the functions /3(t). 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 . • The premium rate at time t of the year is p(t). p(t) and B(t) are defined also for t t [0. Lucantoni et at. Without loss of generality. For the Markovmodulated model. but now exhibiting (deterministic) periodic fluctuations rather than (random ) Markovian ones. we talk of s as the 'time of the year'. Thus at time t the premium rate is p(s + t).3*µs • p = f /3(v) dv 0 0 (6. 6 Risk theory in a periodic environment 6a The model We assume as in the previous part of the chapter that the arrival mechanism has a certain timeinhomogeneity. where i f00 xB(°) (dx) _ . )3 t 1 J (6. B* = J f B(t) ((*) dt.1) Then the average arrival rate is /3* and the safety loading rt is 77 = (p* . • Claims arriving at time t of the year have distribution B(t). for s E E = [0. 1). The basic assumptions are as follows: • The arrival intensity at time t of the year is 3(t) for a certain function /3(t). from an application point of view. Let 1 1 /3* _ f /3(t) dt. Some main queueing references using the MAP are Ramaswami [298]. Neuts [271] and Asmussen & Perry [42]. p * = 0 p(t) dt. We denote throughout the initial season by s and by P(8) the corresponding governing probability measure for the risk process. 1). Sengupta [336]. By periodic extension. let the period be 1.
it turns out that they have the same adjustment coefficient. Section 4b).10. St = SeI(t). the conditional distribution .w(t)) dt).(3.3(t) = 3A(1 + sin 27rt). 0 Then (by standard operational time arguments ) {St} is a periodic risk process with unit premium rate and the same infinite horizon ruin probabilities. let .1. since the added variation is deterministic. (6.9). and we recall from there that the ruin probability is 24 1 *(u) _ 3 5eu + 35e6u. It is easily seen that .t. or. of the periodic model as arising from the compound Poisson model by adding some extra variability. RISK THEORY IN A PERIODIC ENVIRONMENT 177 In a similar manner as in Proposition 1. Many of the results given below indicate that the averaged and the periodic model share a number of main features.2 Define T 6(T) = p(t ) dt.3) Note that A enters just as a scaling factor of the time axis. one may think of the standard compound Poisson model with parameters 3*. equivalently. B*. and thus the averaged standard compound Poisson models have the same risk for all A. In contrast. 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 . Thus. the discussion in 111. p* as an averaged version of the periodic model. The behaviour of the periodic model needs not to be seen as a violation of this principle. in agreement with the general principle of added variation increasing the risk (cf. the average compound Poisson model is the same as in III. Thus . The arrival process {Nt}t>0 is a timeinhomogeneous Poisson process with intensity function {/3(s + t)}t>0 . The claim surplus process {St } two is defined in the obvious way as St = ^N° Ui . 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. for Markovmodulated model typically the adjustment coefficient is larger than for the averaged model (cf. In contrast. In particular. respectively. u Remark 6 . We u assume in the rest of this section that p(t) .8. Example 6 .1) and Example 1. not random.3* = 3A.6.1 As an example to be used for numerical illustration throughout this section.w(t). we shall see that for the periodic model increasing A increases the effect of the periodic fluctuations.
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. Dassios & Embrechts [98] and Asmussen & Rolski [43].f.. Daykin et.al. Notes and references The model has been studied in risk theory by. of the averaged compound Poisson model (the last expression is independent of s by periodicity).a be the c.3(v)(B(vl [a] .a) = exp { .4) At a first sight this point of view may appear quite artificial..T) = P(8)(r(u) <T). and define h(s. e. of the claim surplus process.tc* (a)] dv then h (.1) dv .1) . a) etw*(a) h(s+t. we start by deriving formulas giving the m.1]) . The claim surplus process {St} may be seen as a Markov additive process.. 6b Lundberg conjugation Motivated by the discussion in Chapter II.g. [44] (the literature in the mathematical equivalent setting of queueing theory is somewhat more extensive. As usual. we obtain E. t + dt] or not.5.(3(s + t)dt)e«St adt + /3(s + t)dt . The exposition of the present chapter is basically an extract from [44].a. [101] .1) a = J8 . To this end.^8 [.a) Proof Conditioning upon whether a claim occurs in [t. .(1 .a .e.5 (see in particular Remark 11. a) is periodic on R. i.178 CHAPTER VL MARKOVIAN ENVIRONMENT of U. with some variants in the proofs. J Theorem 6 .f.3(s + t)dt[B(8+t)[a] .(8) [eaSt+dt I7t] = = (1 . with the underlying Markov process {Jt} being deterministic period motion on E = [0. r(u) _ inf It > 0 : St > u} is the time to ruin .Q(v) (B(„) [a] . 3 E(8)eaSt = h(s. let f 8+1 tc *(a) _ (B* [a] .g.8). (6. see the Notes to Section 7). given that the ith claim occurs at time t is B(8+t). 0 (5)(u.s .g. 1).east B(8+t) [a] east .adt +. Jt = (s + t) mod 1 P(8) . and the ruin probabilities are 0(8) (U) = P(s )(r(u) < 00).
so that obviously {Lo. u Remark 6.1].t = 1 by Theorem 6.3(s + t)[D(8 +t)[a] .1)dv  o h(t. RISK THEORY IN A PERIODIC ENVIRONMENT E(8)east+ dt d Et.5 The formula for h(s) = h(s. a) as well as the fact that rc = k` (a) is the correct exponential growth rate of Eeast can be derived via Remark 11.t}t>o = h(s. we can write Lo Jt.6. 0) exist and are finite.2.adt +. a) h(s + t.6 . a) et. a) = exp I f t3(v)(kv)[a) .. E (8)east (a +.(8)east 179 = = = = = E(8)east (1 .4 For each 0 such that the integrals in the definition of h(t .t} is a multiplicative functional for the Markov process { (Jt. St)} . B) eoSt t.3.1)dv l og E(8) et where atetk•(a) h(t.log h(s. h(s + t. a) .9 as follows.9) eastt.(e) Let = h( h(Jo. St)} and . a).1]) . at + f log h(s + t. a) = h(s.c* (e) {Le.5.t. 9) is a P ( 8)martingale with mean one.s. According to Remark 11. a) Corollary 6.4).* (a) h(s.0(s + t)dt[B(8+t)[a] . a) Thus E(8)east = h(s + t. it then suffices to note that E(8)Le.1]) . With g the infinitesimal generator of {Xt} = {(Jt. 0) P(8)a. dt log E(8)east a + f3(s + t) [B(8+t) [a] . + v)(B([a] . Proof In the Markov additive sense of (6.
3(v)( Bi"i [a] . Lemma 6 . J s [. the restrictions of Plsi and Pest to Ft are equivalent with likelihood ratio Le.3(s)dt • B(s)[a]h(s) = gha(s. Sdt) = h(s + dt) eadt (1 .g. That is.6 ( s ) exp { 0( s )&s) [a] + tc . (ii) use Markovmodulated approximations (Section 6c). MARKOVIAN ENVIRONMENT ha(s.3(s)ks)[a]h(s)} ah(s) 13(s)h(s) + h'(s) +. Proof (i) Check that m.60(t) = a(t)B(t)[0].f.5 that we can define a new Markov process {(Jt.1. St)} with governing probability measures Fes).180 CHAPTER VI.7 When a > yo. such that for any s and T < oo. Proposition 6. That is.3.6 The P(s). of St is as for the asserted periodic risk model. However. . correspond to a new periodic risk model with parameters ex . (iv) finally. When a = y. we put for short h(s) = h(s.(3(s)dt) +. it follows by Theorem II.0) = Kh(s).'y). B(s). Now define 'y as the positive solution of the Lundberg equation for the averaged model.3(s)B(s) [a]h(s). yo is determined by 0 = k* (70) = QB*.2. Equating this to rch (s) and dividing by h(s) yields h(s ) = h(s) = a + . see [44] for 11 a formal proof. as above E (s) ha(Jdt.1) . 0 < s < 1. For each 0 satisfying the conditions of Corollary 6. [70] .3(s)h(s) + h'(s) +.tc] dv} (normalizing by h(0) = 1). Proposition 6. A further important constant is the value yo (located in (0.T. P(s) (T(u) < oo) = 1 for all u > 0.a . the requirement is cha(i. ( iii) use approximations with piecewiese constant /3(s).y) = eayh(s). ry)) at which n* (a) attains its minimum.4. cf. y solves n* (y) = 0. 0) = h(s) + dt {ah(s) . say. That rc = is*(a) then follows by noting that h(1) _ u h(0) by periodicity. Bet)(dx) = ^ B(t ) (dx).
9) 0(')(u) = h(s.7) h(B(u). J C R+ such that the B(8).8 The ruin probabilities can be computed as (u)+T(u)k'(a) ^/i(8) (u.6(v) dv Jo ' xe«xB (°) (dx) r^ xe«xB'(dx) = Q'B' [ a] = ^' J 0 = ^c"'(a) + 1. the Markov process {(^(u).4. have components with densities b(8)(x) satisfying inf sEI. a) a > ry0 (6. the mean number of claims per unit time is p« 181 = Jo 1.1.2. considered with governing probability measures { E(8) }E[ . 9(u)) for any bounded continuous function (e. a) TI h(9(u). The proof involves machinery from the ergodic theory of Markov chains on a general state space. B(oo)). a) e«uE(8 ) e «^ . Corollary 6. f (x. xEJ 0 (s)b(8)(x) > 0. which is not used elsewhere in the book. The relevant likelihood ratio representation of the ruin probabilities now follows immediately from Corollary 11.9) and noting that weak convergence entails convergence of E f (^(u). say s0. 1). has a unique stationary distribution. RISK THEORY IN A PERIODIC ENVIRONMENT Proof According to (6.9(u))} u>0. Here and in the following. s E I.10) Then for each a. and we refer to [44]. Wu). (6. a)e«uE (a iP(s) (u) = h( s)e7uE(` ) h(O(u)) To obtain the CramerLundberg approximation from Corollary 3. u which is > 1 by convexity.g. 0(u)) * (b(oo). ^(u) = ST(u) . and no matter what is the initial season s. we need the following auxiliary result . T) = h(s.u is the overshoot and 9(u) = (T(u) + s) mod 1 the season at the time of ruin. q) = eryx/h(q)).2).8) (6.6. we get: . Lemma 6 . e(cc)) Letting u > oo in (6.1) the distribution of (l: (oo).9 Assume that there exist open intervals I C [0. T(u) < (6.
it does not seem within the range of our methods to compute C explicitly.10 Under the condition (6. Theorem 6 .9).11) Note that ( 6. 6.ir) } Plots of h for different values of A are given in Fig.W.182 CHAPTER VI. Among other things. illustrating that the effect of seasonality increases with A. 1. which may provide one among many motivations for the Markovmodulated approximation procedure to be considered in Section 6c.1. where C(o) = 1 + info < t<i h(t) . 11 7/'O (u) < C+°)h(s) ery".10) of Lemma 3. (6. For our basic Example 6 .6 for the Markovmodulated model: Theorem 6 . where e. 10 shows that certainly ry is the correct Lundberg exponent. At this stage . we obtain immediately the following version of Lundberg ' s inequality which is a direct parallel of the result given in Corollary 3. MARKOVIAN ENVIRONMENT Theorem 6.Ch(s)ery". this provides an algorithm for computing C as a limit. Vi(8) (u) . Noting that ^(u) > 0 in ( 6.1 In contrast to h. elementary calculus yields h(s) = exp { A C 2^ cos 2irs  4^ sin 21rs + 11 cos 41rs . A=1/4 A=1 A=4 0 Figure 6.) C = E1 h(B(oo)) u + oo.11) gives an interpretation of h(s ) as a measure of how the risks of different initial seasons s vary.16.1.
w)e4u dx 9w + 7(1 .13) Elementary convexity arguments show that we always have ryy > Y and ay > ry. Consider first the timedependent version of Lundberg's inequality. RISK THEORY IN A PERIODIC ENVIRONMENT Thus. ay) • (6. Theorem 6.42 • exp {J_ cos 27rs . Just as in IV. 1 (6. whereas ay < y.12) As for the Markovian environment model.14) < C+)(y)h(s) e7yu. 1 ) and all u > 0. (ay).(8) (u.0(8) (u+ yu) (6.13 to our basic example.w)e4u .3x + (1 . Theorem 6 .(ay) > 0 when y < 1/ic' (7). where ay is the unique solution of W(ay) =y• (6. C_h(s)e7u < V. Lundberg's inequality can be con siderably sharpened and extended..7e .167r I Cu. we substitute T = yu in 0(u.yr. .16 In order to apply Theorem 6.17) (6. We state the results below. e7 ( yx)B(t)(dy) > Then for all $ E [0. we obtain Co) = 1.w ) • 7e u{w • 3e3x + ( 1 . we first note that the function fu° exu {w • 3e . yu) 000 (u) .7x j dx _7x } _ 6w + 6(1 .11 as well as it supplements with a lower bound.g. in our basic example with A = 1. #c( ay) < 0 when y > 1/tc'('y). T) and replace the Lundberg exponent ry by ryy = ay .13 Let = 1 B(t) C o<tf i h(t) 2no f °O e'r(Yx)B( t) (dy)' (x) x 1 B(t) (x) C+ = sup sup o<t<i h ( t) xo J. r.12 Let 00)(y) 1 Then info < t<i h(t. (6.w) .6.(s)(u) < C+h(s)e7". the proofs are basically the same as in Section 3 and we refer to [44] for details.47r sin 27rs + 167r cos 47rs .42 so that 183 tp(8) (u) < 1. e.4.15) The next result improves upon the constant C+) in front of eryu in Theorem 6.
but thereby also slightly longer.'Yo)e (6.181 s(u) < 1. such a deterministic periodic environment may be seen as a special case of a Markovian one (allowing a continuous state space E = [0. This observation motivates to look for a more formal connection between the periodic model and the one evolving in a finite Markovian environment. Then 7oudT . n}.66.013. much of the analysis of the preceding section is modelled after the techniques developed in the preceding sections for the case of a finite E. with s the initial season.. Of course.9 3 0<8<1 p 27r 47r 167r 161r 2 _ _e. 1/i18 1 s (u) > 0. .(8)(u. . MARKOVIAN ENVIRONMENT attains its minimum 2 /3 for u = oo and its maximum 6 /(7 + 2w) for u = 0.I eu.\ 3 C+ = sup 6 exp { A (. C+ = 1.19 I eu. yo)..L sin 27rs + 1 I cos 47rs .1 sin 27rs + 1 cos 47rs .20 •exp { 2n cos 27rs . exp 2^ cos 21rs . where the environment at time t is (s + t) mod 1 E [0. 0 <'p(8)(u ) . for A = 1 (where 3 e0.T) < C+('Yo)h( s. . completing a cycle . 6c Markovmodulated approximations A periodic risk model may be seen as a varying environment model.16. Some of the present proofs are more elementary by avoiding the general point process machinery of [44].184 CHAPTER VI.0. and let 8 = er' (Y0). the nth Markovian environmental process {Jt} moves cyclically on {1. Thus C_ = 2 inf ex cos 2irs .19 } 0 <8<1 8 + cos 21rs Thus e.18) Notes and references The material is from Asmussen & Rolski [44]. Finally.16) with 'y replaced by yo and h(t) by h(t. Thus.g. 1).4^ sin 2irs + 16^ cos 41rs . 14 Let C+('yo) be as in (6.20).1 sin 2irs + 16_ cos 47rs . we have the following result: Theorem 6 . and in fact.\ = 0 . The idea is basically to approximate the (deterministic) continuous clock by a discrete (random) Markovian one with n 'months'.cos 27rs .013.66. 1) for the environment).
and the ruin probability corresponding to the initial state i of the environment is then Y'yn)(t) = F (M(n) > t). one simple choice is Oni = 0( i . Notes and references See Rolski [306]. it is desirable to have formulas permitting freely to translate from one setting into the other. M(n) = Supt>o Stn). We want to choose the /3ni and Bni in order to achieve good convergence to the periodic model. (6. Bi. This queue is commonly denoted as the Markovmodulated M/G/1 queue and has received considerable attention in the last decade. T) can be expressed in a simple way in terms of the waiting time probabilities of a queueing system with the input being the timereversed input of the risk process. Thus.20) be the claim surplus process of t>o the nth approximating Markovmodulated 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.21) which serves as an approximation to 0(1)(u) whenever n is large and i/n s.7. Let 0j. but others are also possible. We let {Stn)} (6. 7 Dual queueing models The essence of the results of the present section is that the ruin probabilities i/ (u).19) n 0 0 ••• n Arrivals occur at rate /3ni and their claim sizes are distributed according to Bni if the governing Markov process is in state i. DUAL QUEUEING MODELS 185 within one unit of time on the average . To this end. since the settings are equivalent from a mathematical point of view. .jEE of the risk process. AE= Aii'r?/7ri• The arrival intensity is /3i when Jt = i. so that the intensity matrix is A(n) given by n n 0 ••• 0 0 n n ••• 0 A(n) _ (6. z/'i (u.1 ((i 1)/n) ) and Bni = B .
{Jt }o<t<T• Then we may assume that Jt = JTt.4) where 0* = >jEE 7rj/3j. Proof Consider stationary versions of {Jt}o<t<T. MARKOVIAN ENVIRONMENT • Customers arriving when Jt = i have service time distribution Bi. just sum (7.3) 7ri where (V. In particular.T(V > u I J* = i).2) Oi(u) = 1.. JJ = i). (7. For (7. and the virtual waiting time (workload) process {Vt}too are defined exactly as for the renewal model in Chapter V. (7.3).2). J*) is the steadystate limit of (Vt.. Now let In denote the environment when customer n arrives and I* the steadystate limit.oo in u (7. Then Pi(T(u) < T. T) = 7ri 1 P.3.n(VT > u. Proposition 7. (7. ii (u) = it /3 P(W > u. and for (7. J* = i) = P. 2 . JT = i) = P(V > u.0i (u . (VT > u I JT = 2). Jt ). Taking probabilities and using the stationarity yields 7riPi(T(u) < T. Proposition 7. The actual waiting time process 1W1.1) over j. let T . Jo = i.186 CHAPTER VI.2) and use that limF (VT > u. and (7. Jo = j. . J* = i) for all j. JT = Z).1 Assume V0 = 0.2 The relation between the steadystate distributions of the actual and the virtual waiting time distribution is given by F(W > u. I* )3i P(V > u.=1 . JT = j} and {VT > u.1) follows. J* = i). JT = i} coincide. The first conclusion of that result then states that the events {T(u) < T.1). I* = i). .1) 7ri In particular. (7.. JT = i) = 'P.P(V > u. 0 < t < T and that the risk process {Rt}o<t<T is coupled to the virtual waiting process {Vt}o<t<T as in the basic dualitylemma (Theorem 11. • The queueing discipline is FIFO. JT = j) = LjPj (VT > u. JT = j) = 7rjPj(VT > u.
see Regterschot & van Doorn [123]. on average /32TP(V > u. A more probabilistic treatment was given by Asmussen [17]. >u.4) and (7. a general formalism allowing this type of conclusion is 'conditional PASTA'. P(1')(r(u) < oo) = P(')(00) > u). a paper relying heavily on classical complex plane methods. DUAL QUEUEING MODELS 187 Proof Identifying the distribution of (W. and (7. and one has PI'>(rr(u) < T) = P('_T)(VT > u). P(. (7.I.4) can be found in Regterschot & de Smit [301].g.l. that /3(t).=i) a4. and Rolski [306]. and further references (to which we add Prabhu & Zhu [296]) can be found there. and of these. T].7) (7.3) improving somewhat upon (2.7) of that paper.o. we have 1: I(W. see in particular Harrison & Lemoine [186]. N * oo.T)(T(u) <T) = P(8)(VT > u). on average 0*T customers arrive in [0..5) follows from (7. J* = i) see W > u. I*) with the timeaverage . n=1 N However. In the setting of the periodic model of Section 6. With {Vt} denoting the workload process of the periodic queue. [243]. u Notes and references One of the earliest papers drawing attention to the Markovmodulated M/G/1 queue is Burman & Smith [84]. The relation (7. Taking the ratio yields (7. the dual queueing model is a periodic M/G/1 queue with arrival rate 0(t) and service time distribution B(') at time t of the year (assuming w.7. if T is large.4). B(t) have been periodically extended to negative t). P(W >u. I* = i.3).1 is from Asmussen [16]. Proposition 7.6) (7.8) For treatments of periodic M/G/1 queue. with (7. The first comprehensive solution of the waiting time problem is Regterschot & de Smit [301].I *=i). Lemoine [242]. p < 1 then ensures that V(*) = limNloo VN+9 exists in distribution. .
This page is intentionally left blank .
Thus in between jumps. i&(u. {Rt} moves according to the differential equation R = p(R).T) = FloinfTRt< OIRo=u1 denote the ruin probabilities with/initial reserve u and infinite. resp . finite horizon. and the evolution of the reserve may be described by the equation Rt = u . and T(u) = inf {t > 0 : Rt < u} is the time to ruin starting from Ro = u so that '(u) = F(T(u) < oo).1) (other terms are accumulated claims or total claims).. with common distribution B and independent of {Nt}. However . .2) tk(u. are i.Chapter VII Premiums depending on the current reserve 1 Introduction We assume as in Chapter III that the claim arrival process {Nt} is Poisson with rate .6.i. t] are Nt At = Ui (1. and that the claim sizes U1. z/i(u) = F IinffRt< 0IRo=u 1 (1. the premium charged is assumed to depend upon the current reserve Rt so that the premium rate is p(r) when Rt = r.. the aggregate claims in [0.T) = F(T(u) < T). Zt As earlier. 189 . U2.d.At + p(R8) ds. Thus.
when x > p/S. 1 . pi > p2 and p(r) = One reason could be competition. say e. i. Proof Obviously '(u) < ilb(v) when u > v. If Ro = v < u. Thus at deficit x > 0 (meaning Rt = x). P(r) _ p + e(r .'(u)) > 0 so that V'(v) < 1. oo) is given by i (u + p/S). However. it seems reasonable to assume monotonicity (p(r) is u . where one would try to attract new customers as soon as the business has become reasonably safe. In this situation. No tractable necessary and sufficient condition is known in complete generality of the model.Vi(v) u > e(1 . dividends are paid out at rate pi . the payout rate of interest is Sx and absolute ruin occurs when this exceeds the premium inflow p.i(u) = 1 for all u. there is positive probability. Proposition 1.3 (ABSOLUTE RUIN) Consider the same situation as in Example 1. and the probability of absolute ruin with initial reserve u E [p/S. Assume 0(u) < 1 for some u.2. Now return to the general model. Hence in terms of survival probabilities. but assume now that the company borrows the deficit in the bank when the reserve goes negative. Another could be the payout of dividends: here the premium paid by the policy holders is the same for all r.p/S) r > p/S p5(p/5r) 0<r<p/5 Then the ruin problem for {Rt } is of the type defined above. Example 1.1 Assume that the company reduces the premium rate from pi to p2 when the reserve comes above some critical value v.190 CHAPTER VII. A basic question is thus which premium rules p(r) ensure that 'O(u) < 1. or o(u) < 1 for all u.2 (INTEREST) If the company charges a constant premium rate p u but invests its money at interest rate e.e. that {Rt} will reach level u before the first claim arrives. we can put Rt = Rt + p/S.4 Either i. That is. say at interest rate b. RESERVEDEPENDENT PREMIUMS The following examples provide some main motivation for studying the model: Example 1 . we get p(r) = p + er.p2. rather than when the reserve itself becomes negative. Example 1. but when the reserve comes above v.
Starting from Ro = uo. 296297): Theorem 1. we have z/i(u) <p(u . appealing to Proposition 111. B and (constant) premium rate p. { Vt} remains at 0 until the next arrival). then ?(u) = 1 for all u.o(uo) = 1 so that t/'(u) = 1 for all u by Proposition 1. which was proved in 11.4) 0 and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0.3µB for all sufficiently large r. then l/i(u) < 1 for all u.6) . In case (a). Proposition I1I. Then if u > no.f p(Vs) ds. Let Op(u) refer to the compound Poisson model with the same 0. instead of (1.2) we have t Vt = At . (1. That is.1. This is basically covered by the following result (but note that the case p(r) .2(d)).b(u.I3IB requires a more detailed analysis and that µB < oo is not always necessary for O(u) < 1 when p(r) 4 oo. We next recall the following results. INTRODUCTION 191 decreasing in Example 1.1. that u zPp(u . if and only if V)(u) < 1 for all u. let uo be chosen such that p(r) > p = 0ILB + e for r > uo. V = p(V)).3.. hence Rt < uo also for a whole sequence of is converging to oo.2 once more. T] i n such a way that the events {r(u) <T} and {VT > u} coincide. (1. {Vt} decreases at rate p(v) when Vt = v (i.e. Proof This follows by a simple comparison with the compound Poisson model.6 For any T < oo. Theorem 1. obviously infu<uo z/'(u) > 0. However.T) = P(VT > u).uo) < 1.1 and increasing in Example 1. and P(Rt + oo) > 0. one can couple the risk process and the storage process on [0. say V. Here {Vt}two is a storage process which has reflection at zero and initial condition Vo = 0. . In particular.1.+ p(r) exists. (1.2) for r sufficiently large so that p(oo) = limr.1.5 (a) If p(r) < /. Hence ik(u) < 1 for all u by Proposition III. let uo be chosen such that p(r) < p = /3µB for r > uo. Then 0(u) = P(V > u). cf. and hence by a geometric trials argument.2(d). the probability that Rt < uo for some t is at least tp(0) = 1 (cf. (b) If p(r) > /3µB + e for all sufficiently large r and some e > 0. In between jumps.4. In case (b).5) and the process {Vt} has a proper limit in distribution . [APQ] pp.uo) and.
h.8) as the rate of upcrossings.7 p(x)g(x) = tofB (x) + a f (x .8) as g(x) = p 1 {yo13e_6x +. Oeax f x e'Yg (y) dy } = p) eaxa(x) .h.y. of (1. u Define ^x 1 w(x) Jo p(t) dt. In view of the path structure of {V t }. where g(x) = p( ^ exp {.8 Assume that B is exponential with rate b. and is succesful if the jump size is larger than x . Corollary 1. (1. yo ^ 1 + oo Q exp {.y)g(y) dy.Sx} dx. It follows in particular that 0(u) = fg(Y)dy. Now obviously. we thus need to look more into the stationary distribution G. Jo AX) (1.192 CHAPTER VII. oo) must be the same as the flow the other way. and the other being given by a density g(x) on (0.9) Proof We may rewrite (1. the l.s. Considering the cases y = 0 and 0 < y < x separately.6w(x) . x] to (x. we arrive at the desired interpretation of the r. one having an atom at 0 of size 'yo. An attempt of an upcrossing occurs as result of an arrival. t + dt] if and only if Vt E [x. Then the ruin probability is tp (u) = f' g(y)dy. 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. RESERVEDEPENDENT PREMIUMS In order to make Theorem 1. say when {Vt} is in state y. say. x + p(x)dt]). the flow of mass from [0. Note that it may happen that w (x) = oo for all x > 0.6 applicable.6x and that w(x) < oo for all x > 0. B(x) = e.8) Proof In stationarity. this means that the rate of upcrossings of level x must be the same as the rate of downcrossings. It is intuitively obvious and not too hard to prove that G is a mixture of two components. oo). say.7) Proposition 1. say if p(r) goes to 0 at rate 1 /r or faster as r j 0.Qw(x) .Sx}.8) is the rate of downcrossings (the event of an arrival in [t. for the storage process {Vt}. of (1. t + dt] can be neglected so that a path of {Vt} corresponds to a downcrossing in [t. (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
. From ic(a) = .b P/E dx /' P/ ' (p/  x)p/e 150/f I' (/3/E) (6P1'E.e.3. with density x(3/e1aQ/e fV (x) _ e 6X ' x > 0.2y +µ ) dy .pa.3 is also valid if {Rt} is obtained by adding interest to a more general process {Wt} with stationary independent increments./3 log(b + a)] = log ePa/f (a + a ) e which shows that Z is distributed as p/E .2) follows by elementary algebra. The process {St} corresponds to {Wt} so that c(a) or2a2/2 . 13/E).200 CHAPTER VII.g.5 The analysis leading to Theorem 2.13 /E) r (.f. of Z is IogH[a] = f ytc(y)dy = e fa (0. RESERVEDEPENDENT PREMIUMS u from which (2. i. H(u) = P(Z r < u) = P(V > u + p/E) = (8(p + Eu)/E. r (j3/E) In particular. Proof 2 We use Theorem 2.3/E) By the memoryless property of the exponential distribution. assume that {Wt} is Brownian motion with drift µ and variance v2.a) . As an example.pa.01'E) + (p/E)al aO l febP/E } IF (0 /0 jF From this (2. and the c. where V is Gamma(b. 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 . RT(u) has an exponential distribution with rate (S) and hence E [H(RT(u))I r(u) < oo] L Pe6'r (P/C .2) follows by elementary algebra.V < x)]0 + f P(V > p/E ) + eby fv (p/E . /^ u Example 2 .3a/ (5 . then {Rt} is the diffusion with drift function p+Ex and constant variance a2.x) dx e.V.
[129] and Harrison [185].e. Some of these references also go into a stochastic interest rate. it follows that the ruin probability is Cu) H(u) H(0) 11 Notes and references Theorem 2. of the form Ei° p"X" with the X„ i. [357]. see e. Gerber [155].2 is a special case of a perpetuity. 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). It must be noted. write y* for the solution of the Lundberg equation f3(B[ry *] . [129]. Further studies of the model with interest can be found in Boogaert & Crijns [71]. Paulsen & Gjessing [286] and Sundt & Teugels [356].3. Q2/2E).3 is from Harrison [185]. Goldie & Griibel [167]. A r. Logarithmic asymptotics For the classical risk model with constant premium rule p(x) . for a martingale proof. THE LOCAL ADJUSTMENT COEFFICIENT _ Q2a2 pa 4e E 201 I.Y*p* W*(u) < ery*u = 0. Gerber [157] p. or to nonlinear premium rules p(•).i. [283]. 134 (the time scale there is discrete but the argument is easily adapted to the continuous case).g. 3 The local adjustment coefficient. Z is normal (p/E. not even Erlang(3) or H3..v. that the analysis does not seem to carry over to general phasetype distributions.p*. however.d.1) .. it is also used as basis for a diffusion approximation by these authors. and since RT = 0 by the continuity of Brownian motion.g. Paulsen & Gjessing [286] found some remarkable explicit formulas for 0(u) beyond the exponential case in Corollary 1.3) was derived by Emanuel et at. Paulsen [281].8. write Vi* (u) for the ruin probability etc. Emanuel et at. [282]. and recall Lundberg 's inequality . Delbaen & Haezendonck [104]. as in the proof of Proposition 2.4 is classical. The formula (2. se e. Corollary 2.
let p* be a in (3.2) such that p(x) < c(. as will hold under the steepness assumption of Theorem 3. and (for simplicity) that inf p(x) > (3µs ..'y ( x)) = 0 where r. which in turn by Lundberg's inequality can be bounded by ery*(1E)" Hence limsup„. choose c(.>o 7(x) > 0. Let y* < So. 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.w (u) J dt > c(3)eeu v 1 p(u+ t) . 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).1) .1 ). Proof of Theorem 3. e(1o+e)2 (x ) u > 00.log '(u)/u < ry*(1 .e. a first step is the following: Theorem 3 . The steepness assumption and p(x) + oo ensure 'y(x) * So. Then lim sup u>oo u and e E''p(r) + 0. If 60 s f 6o. i. (3. as solution of the equation n(x. 1) and for a given E > 0.E). a) = f3(B[a] . i.e.i)eex.ap(x). x * oo. 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 .1.1 Assume that for some 0 < 5o < oo. Letting first E * 0 and next ry * T 5o yields the first statement of the theorem.1.*(u) . and that p(x) * oo. For the last asssertion . c(. 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) .4) we assume existence of y(x) for all x. (x. (3. B(x) > C(2)e(ao+f)x for all x. choose uo such that p( x) > p* when x > u0E.3) When trying to extend these results to the model of this chapter where p(x) depends on x. When u > uo. it holds that f3[s] T oo..C*ef*". oo for all E > 0.202 CHAPTER VII.5) which implies inf. then log u (u) In the proof as well as in the remaining part of the section . log ?i(u) < < 00 JO . we will use the local adjustment coefficient 'y(x). RESERVEDEPENDENT PREMIUMS and the CramerLundberg approximation V. x>0 (3.
2).' (u) < eI("). or (b) Condition 3. 3. For e > 0. 2 Assume that p(x) is a nondecreasing function of x and let I(u) = fo ry(x)dx. the limit is not u + oo but the slow Markov walk limit in large deviations theory (see e. the result is not very informative if bo = oo. let 0e (u) be evaluated for the process only with 3 replaced by /0/e and U.3 Assume that either (a) p(r) is a non decreasing function of r. {Rte)} defined as in (1. If p(x) = pis constant .1 only presents a first step. ruin will occur if the claim is at least u + v.e. noting that in many cases the constant C is close to 1. Theorem 3.3 to be reasonably precise and use e` (u) as approximation to 0 (u). Condition 3. which essentially says that an overshoot r. THE LOCAL ADJUSTMENT COEFFICIENT 203 where c. u Obviously.4)eE" Given such an arrival. (u) = O(u/e).ea°/(ecf1)). 3) = (1 . by cU2. then Rte) = CRtie for all t so that V). . the asymptotics u * oo and c . (3. The rest of this section deals with tail estimates involving the local adjustment coefficient.13 is a technical condition on the claim size distribution B. The first main result in this direction is the following version of Lundberg's inequality: Theorem 3 . The form of the result is superficially similar to the CramerLundberg approximation. Theorem 3 .v.6) The second main result to be derived states that the bound in Theorem 3.(u) > so.0 are the same. UJU > x cannot have a much heavier tail than the claim U itself. Therefore the probability that a claim arrives ( before the reserve has reached level u + v is at least c(. I.13 below holds. (3.2 is also an approximation under appropriate conditions. one can then assume that e = 1 is small enough for Theorem 3.g.. and in particular.3. and hence '(u) > c(4)eeuc( 2)e(do+e)u The truth of this for all e > 0 implies lim inf log V. Then limelog l/ie (u) = I(u). Bucklew [81]). Then .7) CIO Remarks: 1. However. 2. 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.
4 Consider again the exponential case B(x) = eax as in Corollary 1.8. One would expect the behaviour in 2) to be important for the quantitative performance of the Lundberg inequality (3.3. rather than eI(u)). u .(3/p(x).8 in terms of I(u) when the claims are exponential: Example 3 . say.2. However.bx}]o + b /' oo exp low (x) .bx} dx .bu}.bx} dx fo 00 1 + [exp {/(3w(x) . RESERVEDEPENDENT PREMIUMS 4.bx} dx = 1 + J0 dodx(x) exp {. it is formally needed only for Theorem 3.204 CHAPTER VII.3.7) is only captures 'the main term in the exponent' but is not precise to describe the asymptotic form of O(u) in terms of ratio limit theorems (the precise asymptotics could be logI(u)e1(U) or I(u)"e_I(u).exp {/33w(u) . we consider some simple examples. As typical in large deviations theory.(x) dx. Then y(x) = b .(iw(x) . J0 ^oo g(x ) dx f AX) lexp IOW (X ) bx + b u 1 exp low(x) . 3. we show how to rewrite the explicit solution for ti(u) in Corollary 1. and r j 1 'Yo v(x)dx = bu  a J0 p(x)ldx = Integrating by parts. we get = 1+ J" AX) exp {(3w (x) .1 + b f e. the logaritmic form of (3. First.bx} dx 1+0. 5. 3a Examples Before giving the proofs of Theorems 3.6).bx} dx oo exp low(x) bx dx 70 Ju r oo = b J exp low (x) .
u .fa 7(x+u)dx/Edy o The analogue of (3. BE * 0.f y(x)dxd y If 7(x) is increasing . (3.9 ) 11000 eI(v)dy f000 e. (3. Similarly. (u) = I(u) + AE .2(X) = ev2(x) so that 7e(x) = 7(x)/e.7) follows.9) yields e log . THE LOCAL ADJUSTMENT COEFFICIENT and hence 205 f°° eI(v )dy . note first that the appropriate slow Markov walk assumption amounts to u. In particular. 191195) that 1P (U) = fu0 eI(v)dy = eI(u) follo e.I ( v )dy fo +u) dxdy .2.0.fo 7(x)dx/Edy f . 0.3. and (3. Choosing yo... we get r 00 e.BE. oo) with drift µ(x) and variance a2 (x) > 0 at x. (X) = µ(x).e.1/8 .fory(x+u)dxdy ( 3.3. and (3.8) 7(x)dxdy 11 We next give a direct derivations of Theorems 3.2.fo 7(x) dx /E dy > av 'yo /Edy = E (1 .1.3 in the particularly simple case of diffusions: Example 3.I ( u) fool.5 Assume that {Rt} is a diffusion on [0. IE(u) = I(u)/e. 3. in the definition of AE converges to 0. Be = e log U000 e. ry(x /b I u o e f0 °° e  e.10 or Karlin & Taylor [222] pp. It is well known that (see Theorem XI. the integral is bounded by 1 eventually and hence lim sup AE < lim sup a log 1 = 0.10) where AE = e log 000 e. applying the inequality 7(x + u) > 7(x) yields immediately the conclusion of Theorem 3. The appropriate definition of the local adjustment coefficient 7(x) is then as the one 2p(x)la2(x) for the locally approximating Brownian motion. 70 > 0 such that 7(x) < 7o for y < yo.5) is infx>o 7(x) > 0 which implies that f °O . For Theorem 3.ev 0 O /E) J0 70 70 Yo This implies lim inf A. > lime log e = 0 and AE * 0.
10) holds if we redefine AE as AE = flog (j °° efo 7(x)dx/edy _ E/5 I and similarly for B. this leads to (3.206 CHAPTER VII. the slow Markov limit a * 0 and the limit u walk approximation deteriorates as x becomes large.4. Further. Then the solution of the Lundberg equation is y* = b . the results are suggestive in their form and much more explicit than anything else in the literature. however . we have 5 > 7o and get lim inf AE > lime log e .0) = 0. that the interchange of the slow Markov walk oo is not justified and in fact. I(u ) = G1(u) + . E+o e*O By (3.6/p* so that u 1 I (U) = bu . .+1 (u) = o( 1). 0.0.5 for risk processes with exponential claims is as follows: Example 3 .5. G. lim sup Af < lim sup c log(1 . ) Note that this expression shows up also in the explicit formula for lk(u) in the form given in Example 3.. G.5.0/e.5) and 7* = 5 . As in Example 3. > . Of course. + Gq(u) + o(G9(u))• Gi (u) It should be noted .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./3 1 AX dx. . Ignoring 1/5 in the formula there.e. Nevertheless . G. so our approach is to determine standard functions Gl (u). Thus 7e(x) _7(x)/e and (3. I. the slow Markov walk assumption means 5E = b/c.1 3. . 0 Now (3. _ . .7o C 15 I I.6) exactly as in Example 3...5. RESERVEDEPENDENT PREMIUMS The analogue of Example 3. (u) representing the first few terms in the asymptotic expansion of I(u) as u + oo. .6 Assume that B is exponential with rate S. 7(x) is typically not explicit.(u) oo.Q/p*..
1) leads to .c2 Su a dx ) Su a<1 Su .11) that b[s] * co as s f S and hence 7* T S as p* + oo.ry*°p*. u Example 3 . 1. phasetype distributions (Example 1.Y . It follows from (3. This covers mixtures or convolutions of exponentials or. if the phase generator is irreducible ( Proposition VIII. B[s] = 1 + s exB(x)dx = 1 +c1SF(a) ('+o(')) (S . 2. in the phasetype case . ry* loge*+ g7loglogp*.. say 1 is the upper limit and B(x) .1/k). the typical case is a = 1 which holds . I(u) Pt. .1/a).y/s)dy sn 1 1 f ' evy'7ldy = cse8r(T7) as s T oc.s)C' f "o o as s T S.cs(1 . c4 = c2b 1/'/(1 .C2p* C2 = (3clr( a))11'. THE LOCAL ADJUSTMENT COEFFICIENT Example 3 . . and hence (3. more generally.12) with y > 1. e.1 =$ f cse8 Sn f e"B(x)dx = e8 Jo s eIB ( 1 . For example. 1) and 17 = k + 1 if B is the convolution of k uniforms on (0.3. x T 1. fu I(u) Su .7 Assume that B(x) . u(logu + r7loglogu).x)n1. Here B[s] is defined for all s and B[s] .8). Hence (3. y = 2 if B is uniform on (0.:.4) or gamma distributions.1) leads to (S7T N Ocp a. (3.clxale5x 207 (3. 77 = 1 if B is degenerate at 1. More precisely.c3 logu a= 1 J 0 a + bx 1/ ( c4ul 1/° a > 1 where c3 = c2 /b.11) with a > 0.g.8 Assume next that B has bounded support.3cse7*I7(77) .
15) Proposition 3. Hence for u<V. of the increment in a small time interval [0.14) for the m . one could also have considered the increment ru (T1) .10 Assume that p(x) is a nondecreasing function of x. RESERVEDEPENDENT PREMIUMS Example 3 . By convexity of the m .(t))dt. The assumption implies that ru(t) .4).3 (B[s] . 1 0 3e.log p*.Ote7o( u)(u. h 10.sp(u). (3.9 As a case intermediate between (3.g. 7 * .f..13) We get b[s] .B[7o (u)] .4) is the formula h logEues ( Rhu) . g. (3.c8 log .208 CHAPTER VII.11) and (3.2 We first remark that the definition (3.(T1)) > Ee7o(u)(ul+vr»(Ti)). 3b Proof of Theorem 3. x f oo .Ul up to the first claim (here ru (•) denotes the solution of i = p (r) starting from ru(0) = u).css 2%rc7eC782/2.12). I (u) c8u log u 0 where c8 = 2/c7.e7o ( u)(ul+u r.u . h].u is a nondecreasing function of u. this is only possible if 7o(v) 2 7o(u)• .4) of the local adjustment coefficient is not the only possible one: whereas the motivation for (3. (3.f. 1 = E. . of U1 + v .r^.1) . (b) 'y(x) <'Yo(x)• Proof That 7(x) is nondecreasing follows easily by inspection of (3.ru(TI)) .r„(Ti). assume that B(x) CO x2/2c7. ec78)2/2c7 dx C7 .•. Then: (a) y(x) and 7o(x) are also nondecreasing functions of x.1 Cgs o"O 0 esxex2/2c7 dx = cgsec782/2 f . This leads to an alternative local adjustment coefficient 7o(u) defined as solution of 1 = Ee''o(u)(vi+u .
(n+l) (u) 1 .(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)] . Then (u) < efo Yo(x)dx.1) .11 Assume that p(x) is a nondecreasing function of x. this is only possible if yo(u) > 7(u). Assume (3. We shall show by induction that (' Y'(n) (u) < e fo 'Yo(x)dx (3.17) shown for n and let Fu(x) = P(U1 + u . we obtain „I. The case n = 0 is clear since here To = 0 so that ik(°)(u) = 0.10(b): Theorem 3. Separating after whether ruin occurs at the first claim or not.u > tp(u).Fu(u ) + J ^(n)(u .u[70(u)] fo eyo(x)dx .3.7o (u)p(u)• Since (3. Also. it is easily seen that fu x7o(y)dy < xyo (u). note that the assumption implies that ru(t) .es'Yo(u)Fu(dx)} o0 e fo yo( x)dx j. (3. Hence 1 = EeYo(u)(U1+uru(T1)) < E. THE LOCAL ADJUSTMENT COEFFICIENT For (b).2 in terms of 7o. Hence „/.x)Fu(dx) 00 U efo J = o (y) dYF (dx) )+f I 11 /' / 00 e f oFu fu dx) + of u :7o(Y)dYFu(dx) 00 J u 1 l` Considering the cases x > 0 and x < 0 separately.16) Proof Define 411(n)(u) = P('r(u) < on) as the ruin probability after at most n claims (on = TI + • • • + Tn). u We prove Theorem 3. the case of 7 then follows immediately by Proposition 3.17) from which the theorem follows by letting n + oo.4) considered as function of 7 is convex and 0 for y = 0. fa 7o(y)dy < u7o(u) < xyo (u) for x > u.ru(T1) < x).
3/e and U. 0 It follows from Proposition 3. P k.10(a) for some of the inequalities.2.E (u/ n) Y'E (un . However. we have chosen to work with y(u) as the fundamental local adjustment coefficient.n inf n uk1.n = ku.2).n u k}1. in .n.n. yo(u) appears more difficult to evaluate than y(u). 3.n.3 is required. The probability of this is at least n n. for either of Theorems 3. Further.11 be reasonably tight something like the slow Markov walk conditions in Theorem 3.E (u) denote the ruin probability for the classical model with 0 replaced by . let Op*.n) pn niE (u /n) n n_1 n. we used also Proposition 3.3). in accordance with the notation i/iE (u).E (u/n) Now as e . For these reasons.. and.n) must first downcross unl.11 is sharper than the one given by Theorem 3.x/n. RESERVEDEPENDENT PREMIUMS where the last identity immediately follows from (3. To this end. Let Ck.: y(u).n. C*e where the first equality follows by an easy scaling argument and the approximation by (3. given downcrossing occurs. Proof For ruin to occur.n = sup p(x). 0. {RtE)} (starting from u = un.10(b ) that the bound provided by Theorem 3.. the probability that ruin occurs in the CramerLundberg model with p* = pn.n (starting from u/n) without that 2u/n is upcrossed before ruin.2. by €U=.n) > k =1 II v ^k n.E (u/n).n <Z auk}l. 3c Proof of Theorem 3.15).12 lim sup4^o f log O. (3.E ( u/n) ^•e.. x + x/n] by two classical risk processes with a constant p and appeal to the classical results (3. 0.3). y* evaluated for p* = Pk.3 The idea of the proof is to bound { R( f) } above and below in a small interval [x .n so that n. Also. and here it is easily seen that yo(u) . pk n = uk_l. define uk. W O ..210 CHAPTER VII.I.. ryk. op*.e (u) = v'. resp.nbe C*. Lemma 3.n AX). the value of {R(E)} at the time of downcrossing is < unl. (u) < I(u).. (u). (un2. Y*u /E.
. v. y > 0 it holds that F(U>x +yIU>x) B(x + y) < F (V > y). (3.nu /fn( 1 Ck  e.. also ryk.13 There exists a r.nu/en(1 + where o(1) refers to the limit e . B(x) (3. uk_1. k=1 k=1 n u _ nE7 k.2 gives 7PE (u) < eIi"i/f = lim inf Clog 0E (u) > I (u). ne7k.log Ck. In case (b). . 211 Clearly. *p•. 3 now follows easily in case (a). 40 Combining with the upper bound of Lemma 3. It follows that n log V'C (u) k =1 log Ypk.E (urn) < \ *I.E (u/n) OP +^p•.n .. so that Theorem 3. 0 with n and u fixed.n + 0(1).. (u) CIO < Letting n 4 oo and using a Riemann sum approximation completes the proof.18) (ii) the family of claim overshoot distributions is stochastically dominated by V. i. /' (u/n) 'T nk.nu /En) o(1)). we need the following condition: Condition 3. in obvious notation one has tC (x) = y(x)/e.e.F (2u/n). THE LOCAL ADJUSTMENT COEFFICIENT particular.i.n = sup ?'(x).nk=1 limsupelogv). for all x . since ry' is an increasing function of p'.n. 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. Indeed .n cE (2u/n) Ck ne7k.a( u)z.12 completes the proof.19) .n <X<Uk.7k.3.E (u/n) Op•.! (u/n) n n m 7k. 11 Theorem 3.
u/n) < oo) EV). u/n)) I T(E) (u..nu /EnE [e71. (R.E(E) (u. u/n)) . u/n) < oo) .v.E (0) cf.18) for the last equality).EV) = e.EV) = EiI 1 ..5) and the standard formula for b(0).^(E) (u. Ei + E2 < e71. u/n) < oo] . let v < u and define T(E) (u. V < u/En] + P(V > u/En) (u/En . P (T(E) (u. For E2.2y 1 ' . where El is the contribution from the event that the process does not reach level 2u/n before ruin and E2 is the rest.n < ery1. (3. infx>2u /n P(x) .R<) (u v).( .nu/En0(1) .EV) = El + E2.1 n. T(E) (u.of:>2 in n(x). The probability of ruin in between two downcrossings is bounded by Epp . N with EN < 1 = infx>2u/nA(x) = 0(1). (u/n . Write EO. .E (2u/n .212 CHAPTER VII. RESERVEDEPENDENT PREMIUMS To complete the proof.QEU 1 . u /n) < oo] l = = < E [OE (u/n . v ) = v .E (u/n . ) (u u /n)) . we first note that the number of downcrossings of 2u/n starting from RoE) = 2u/n is bounded by a geometric r. (u/n . v ) = inf { t > 0 : R(c ) < v R) = u } . T() (u. u /en 0(i) _n so that E2 < e2ryl nu/En0(1).^'' = E [ .. Then Y'E (u) ^(E) (u.n V. u/n) < oo] E [OE (u/n .. Then the standard Lundberg inequality yields El < E?.V) = e71 nu/Eno(l) (using (3.eV) • P (T(E) (u.
)ds + Y(R2. Djehiche [122] gives an approximation for tp(u. Comparing these references with the present work shows that in the slow Markov walk setup. (u) 40 213 lim inf e log(Ei +E2) + logP (r(`) (u.7) then comes out (at least heuristically) by analytical manipulations with the action integral. . they also discuss simulation based upon 'local exponential change of measure' for which the likelihood ratio is ( /'t /'t Ns Lt = exp S . where the key mathematical tool is the deep WentzellFreidlin theory of slow Markov walks (see e .1.g.21) (the initial condition is r(0) = u in both cases). Whereas the result of [122] is given in terms of an action integral which does not look very explicit.3EU) (3. 0 and b T 00 are interchangeable in the setting of [89].20) (with ic(x. Similarly. one can in fact arrive at the optimal path by showing that the approximation for 0(u.)Ui } . it might be possible to show that the limits e .=1 J An approximation similar to (3.T) = P „(info<t <T Rt < 0) via related large deviations techniques. THE LOCAL ADJUSTMENT COEFFICIENT Hence lim inf e log Ali. T) is maximized over T by taking T as the time for (3.7) for ruin probabilities in the presence of an upper barrier b appears in Cottrell et al.4) and the prime meaning differentiation w.t. the risk process itself is close to the solution of the differential equation r(x) _ r (x. u/n) < oo { 40 )I U nryl n+liminfelogP (T(')(u.J y(Rs)dR.r.21) to pass from u to 0. 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 Notes and references With the exception of Theorem 3.u/n) < oo) CI  > u n n ryi n' i=1 Another Riemann sum approximation completes the proof. the approximation (3.) = exp . l o JJJ o . s). Typically. s) as in (3. the results are from Asmussen & Nielsen [39].3.7(x)) (3.13. Bucklew [81]).J r(Rs)p(R. whereas the most probable path leading to ruin is the solution of r(x) _ k (x. 0 ) (= p(x) .
. RESERVEDEPENDENT PREMIUMS the simplest being to require b[s] to be defined for all s > 0 (thus excluding . the exponential distribution ). e.214 CHAPTER VII.g. For different types of applications of large deviations to ruin probabilities . 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. however.3. . see XI. We should like.
if a problem can be solved explicitly when the relevant distributions are exponentials. if v = (vi)iEE is a probability distribution. on Eo = E U {A} where A is some extra state which is absorbing.1) is 'Here as usual . We often write p for the number of elements of E. F (Jt = A eventually) = 1 for all i E E 1 and where all states i E E are transient. This implies in particular that the intensity matrix for { it } can be written in blockpartitioned form as T 0 0 . then the problem may admit an algorithmic solution involving a reasonable degree of computational effort if one allows for the more general assumption of phasetype structure. A proper knowledge of phasetype distributions seems therefore a must for anyone working in an applied probability area like risk theory. and not in other cases. we write Pv for the case where Jo has distribution v so that Pv = KER viPi• 215 . P.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. Typically. refers to the case Jo = i. a terminating Markov process {Jt} with state space E and intensity matrix T is defined as the restriction to E of a Markov process {Jt}o<t<. More precisely. 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. A distribution B on (0. that is.
T) (or sometimes just (a. A convenient graphical representation is the phase diagram in terms of the entrance probabilities ai. (1. j.216 CHAPTER VIII. k}. an exponential distribution with rate parameter . a. B(t) = Fa(^ < t ).3.1 The phase diagram of a phasetype distribution with 3 phases. T is a subintensity matrix2. that is. the ith component ti gives the intensity in state i for leaving E and going to the absorbing state A.0 = t11. and we have t = Te. We now say that B is of phasetype with representation (E. Thus the phasetype distributions with p = 1 is exactly the class of exponential distributions. and the phasetype distribution is the lifetime of a particle with constant failure rate /3. Then a = a1 = 1. Equivalently.e. the rows sum to one which in matrix notation can be rewritten as t + Te = 0 where e is the column Evector with all components equal to one. t1 = /3.1 Suppose that p = 1 and write . MATRIXANALYTIC METHODS the intensity matrix of a nonterminating Markov process. E = {i.2) The interpretation of the column vector t is as the exit rate vector. C is the lifetime sup It > 0 : Jt E E} of {Jt}. i. 0 2this means that tii < 0. i. the exit rates ti and the transition rates (intensities) tij: tj 3 aj ai i ti tk tjk FkJ ak Figure 1. tij > 0 for i 54 j and EjEE tij < 0 .T)) if B is the Padistribution of the absorption time C = inf{t > 0 : it = A}. Here are some important special cases: Example 1 . The initial vector a is written as a row vector.e. In particular.
.2 The Erlang distribution EP with p phases is defined Gamma distribution with integer parameter p and density bp XP1 6x (p.... p}. 6. 0 SP 0 and the phase diagram is (p = 2) . 0 0 0 0 S 6 . 0 S 6 Example 1. . ..x i=1 Thus E _ Si 0 T 0 S2 0 0 .. so that the density is P E ai6ie6..... .1)!e Since this corresponds to a convolution of p exponential densities with the same rate S. 00)) S s o ..2 corresponding to E = {1.. PHASETYPE DISTRIBUTIONS 217 Example 1. . the EP distribution may be represented by the phase diagram (p = 3) Figure 1. . 0 ••• 0 0 Sp1 0 0 t= 0 0 00 •.1. a = (1 0 0 .3 The hyperexponential distribution HP with p parallel channels is defined as a mixture of p exponential distributions with rates 51.. 0 0 0 T= t= 0 ••• S S 0 0 0 0 0 0 .. 0 •. .
4 (COXIAN DISTRIBUTIONS) This class of distributions is popular in much of the applied literature.g.1)"n! aT"e. the backwards equation for {Jt} (e. Theorem 1 . (b) the density is b(x ) = B'(x) = aeTxt. .3 . Then: (a) the c. [APQ ] p. The basic analytical properties of phase type distributions are given by the following result .g. j E E. p:. Then for i . 36) yields s d.t2 yt bP.e. T). and is defined as the class of phasetype distributions with a phase diagram of the following form: 1 617 ti t2 2 b2. E t ikp kj = kEE kEE 3For a number of additional important properties of matrixexponentials and discussion of computational aspects . Recall that the matrixexponential eK is defined by the standard series expansion Eo K"/n! 3.4 For example.218 CHAPTER VIII.d. the restriction of P8 to E. 5 Let B be phasetype with representation (E. dp. (c) the m. a. Proof Let P8 = (p ^) be the sstep EA x EA transition matrix for {Jt } and P8 the sstep E x Etransition matrix for {Jt} . MATRIXANALYTIC METHODS Figure 1. see A. i.aeTxe.3 0 Example 1 . the Erlang distribution is a special case of a Coxian distribution.f is B (x) = 1 . B[s] = f0°O esxB (dx) is a(sI T)lt (d) the nth moment f0°O xnB(dx) is (. ds^ = ds' = ttlaj + tikpkj.1 tP1 1 Figure 1.f.
tii we go to A.s j# tii i (1. (Jx E E) = this proves (a). = aPxe. B(n)[0] = _ Alternatively..12) for integrating matrixexponentials yields B[s] = J esxaeTxt dx = a ( f°°e(81+T)dx ) t a(sI . this means in vector notation that (T + sI)h = t. in which case the time to absorption is 0 with m .g. hj . PHASETYPE DISTRIBUTIONS 219 That is. Part (d) follows by differentiating the m. or w. 1.T) 't = (. After that. of the initial sojourn in state i. Alternatively.B(x) = 1'a (( > x) = P. i. .f.p.6) .e. j#i jEE tijhj + his = ti. Then h tit ti + ti3 h j .tii and have an additional time to absorption either go to state j which has m . Rewriting ( 1.s) is the m . for n = 1 we may put ki = Ei( and get as in (1. tij / .1 ) n +l n ! a (s I + T ) . d8 P8 = TP8. and (b) then follows from 1: aipF.T) 1t.f.g.n lt .f.n1t = (1)nn!aTn1Te (1)nn! aTne. we i w. the rule (A. we arrive once more at the stated expression for B[s].jEE B'(x) _ cx Pxe = aeTxTe = aeTxt (since T and eTx commute).tii tii . d" dsn a (.g.5) Indeed . and since b[s] = ah. define hi = Eie8S.tii is the rate of the exponential holding time of state i and hence (tii)/(tii . Since 1 . For (c). and since obviously P° = I.f.p. the solution is P8 = eT8.5) ki = 1 + tii L jj:Ai tii (1..s I . (1) n+1n!aT .g.1. i. ti/ .5) as hi(tii + s) = ti  t ij hj. h = (T + sI)1t.
Consider for example 3 9 a= (2 2). One obvious instance is the hyperexponential distribution. Example A3. we get the density as 9 9 6 (1 1) 10 7 1 0 10 2 aeTyt = e x .h. T= 2 111 so that 2 2 Then (cf.6 Though typically the evaluation of matrixexponentials is most conveniently carried out on a computer. MATRIXANALYTIC METHODS which is solved as above to get k = aTle. are idempotent."n! ( ( l 2 2 ) 17 9 0 \ 1 / 10 10 32 n! 35 6" +n!353 Similarly. see the Appendix. another the case p = 2 where explicit diagonalization formulas are always available.220 CHAPTER VIII. making the problem trivial.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.s. 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. there are some examples where it is appealing to write T on diagonal form. 0 Example 1.
4. There are two ways to interpret this: • The phasetype distribution B is defective. 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 a mixture of a phasetype distribution with representation (a/llall.f.4b for definitions and basic rules): Proposition 1. 0 Sometimes it is relevant also to consider phasetype distributions. T) is then defined to be oo on a set of probability 1. or one just lets U be undefined on this additional set. i.T). where the initial vector a is substochastic. 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 If B is phasetype with representation (v.g.11aDD. B[Q] of B is f3[Q] = J e'1zB(dx) _ (v (9 I)(T ® Q)1(t ® I). then the matrix m. (1. and in fact one also most often there allows a to have a component ao at A.e 11BIJ = 1laDD < 1. hail = E=EE a.7) Proof According to (A.1. 5 and serves at this stage to introduce Kronecker notation and calculus (see A. • The phasetype distribution B is zeromodified.T) with weight hall and an atom at zero with weight 1 .hall.29) and Proposition A4. a random variable U having a defective phasetype distribution with representation (a. This is the traditional choice in the literature. . i. < 1.
In older literature.. and we have eTx .8). Lipsky [247]. Example A5. (or Laplace transform) are often used where one would now work instead with phasetype distributions. let v.Ce7'.. Schmidli. Notes and references The idea behind using phasetype distributions goes back to Erlang. (1. x * oo. See in particular the notes to Section 6. In Proposition A5. 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.hve7x.222 CHAPTER VIII.f. but the relevant T is not irreducible. 77 > 0 and k = 0. h be the corresponding left and right eigenvectors normalized by vh = 1 and define C = ah • ve . it is easily seen that the asymptotic form of the tail of a general phasetype distribution has the form B(x) _ Cxkenx. h can be chosen with strictly positive component. Schmidt & Teugels [307] and Wolff [384]. the Erlang case). the conditions of Proposition 1. here k = p1). we give a criterion for asymptotical exponentiality of a phasetype distribution B.8 Let B be phasetype with representation (a. i is real and positive. cf.8) Proof By PerronFrobenius theory (A. distributions with a rational m. . B(x) . let . Rolski.g. O'Cinneide [276] gave a necessary and sufficient for a distribution B with a rational m. the result follows (with C = (ah)(ve)). Here is a sufficient condition: Proposition 1. The Erlang distribution gives an example where k > 0 (in fact. No satisfying . but in many practical cases. Then the tail B(x) is asymptotically exponential. one has k = 0. but todays interest in the topic was largely initiated by M. not only in the tail but in the whole distribution.g. MATRIXANALYTIC METHODS la Asymptotic exponentiality Writing T on the Jordan canonical form. assume that T is irreducible . Using B(x) = aeTxe . 1.1 of the Appendix. T).F. 0 Of course. see his book [269] (a historical important intermediate step is Jensen [214]). 2.f.8 are far from necessary ( a mixture of phasetype distributions with the respective T(') irreducible has obviously an asymptotically exponential tail. v. cf. Neuts.q be the eigenvalue of largest real part of T. Other expositions of the basic theory of phasetype distributions can be found in [APQ]. the text is essentially identical to Section 2 of Asmussen [26]. All material of the present section is standard.4c).
t. n=O We may think of the U. but is in part repeated below... (2. U2. +UnEA} 00 = EEI(U1 +. as the lifetimes of items (say electrical bulbs) which are replaced upon failure.T).+UnEA).1) Proof Let {Jtk)} be the governing phase process for Uk and define {Jt} by piecing the { J(k) } together. or the density is available ) is. known. but nevertheless. is Markov and has two types of jumps . we refer to U as the renewal measure..1 Consider a renewal process with interarrivals which are phasetype with representation (cr. The explicit calculation of the renewal density (or the renewal measure) is often thought of as infeasible for other distributions. Jt={Jt?ul}. . For this reason.1. RENEWAL THEORY 223 algorithm for finding a phase representation of a distribution B (which is known to be phasetype and for which the m. if U is absolutely continuous on (0.. + U0 is 0 . If B is exponential with rate 0.. Then the renewal density exists and is given by u(x) = ae(T+ta)xt. U1<t < U1+U2. A related important unsolved problem deals with minimal representations: given a phasetype distribution .. oo) w.. with common distribution B and define4 U(A) = E# {n = 0.. . however....g. and U(A) is then the expected number of replacements (renewals) in A.1 of the Appendix.r.. 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 +. the renewals form a Poisson process and we have u(x) = 0.2.i.f. the problem has an algorithmically tractable solution if B is phasetype: Theorem 2. we denote the density by u(x) and refer to u as the renewal density. Lebesgue measure. 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. JtJt1) Then { 0<t<U1 . .d..: U1 + . Let U1. be i.
u Returning to nonterminating renewal processes . Then the lifetime is zeromodified phase type with representation (a.224 CHAPTER VIII. this is welldefined. B is defective . Proof Just note that { it } is a governing phase process for the lifetime. and the jumps of the first type are governed by T.1) remains valid for that case. the lifetime of the renewal process. Equivalently. the phasetype assumptions also yield the distribution of a further quantity of fundamental importance in later parts of this chapter . fi(t) U2 U1 .1. as the time of the last renewal. Corollary 2.. T). IIafl < 1.1) follows by the law of total probability. i.3 Consider a renewal process with interarrivals which are phasetype with representation (a.e. that is.T).IIBII which is > 0 in the defective case. Hence ( 2. (b) £(t) has a limiting distribution as t * oo. is the first k with Uk = 00. define the excess life e(t) at time t as the time until the next renewal following t.. However. This is defined as U1 + .T + ta). MATRIXANALYTIC METHODS of the last type from i to j occurs at rate tiaj .1 Corollary 2. + Uit_1 where s. and the distribution of Jx is ae ( T+t«)x. the density is veTxt = B(x)/µB. The renewal density at x is now just the rate of jumps of the second type. and hence ( 2. 2. . since Uk = oo with probability 1 .T) where vt = ae (T+ta)t . and let µB = aTle be the mean of B.e. Then: (a) the excess life t(t) at time t is phasetype with representation ( vt.2 Consider a terminating renewal process with interarrivals which are defective phasetype with representation (a.U1 U3 U2 U3 U4 Figure 2. u The argument goes through without change if the renewal process is terminating. .T) where v = aT1 /µB. i. Hence the intensity matrix is T + ta. see Fig. which is phase type with representation (v. which is ti in state i.
we get B(x) aeTxe aT1eTxTe µB µB PB = veTxt. Hence in (b) it is immediate that v exists and is the stationary limiting distribution of it. i. hence e(t) is phasetype with representation (vt. RENEWAL THEORY 225 Proof Consider again the process { Jt } in the proof of Theorem 2.2) v(T + ta) = 0.2): aT1 e = AB = 1 µB µB a + aT'Tea aT1(T + ta) µB PB a + aea a + a µB µB =0. we first compute the stationary distribution of Q.4 Consider a nonterminating renewal process with two phases. Next appeal to the standard fact from renewal theory that the limiting distribution of e(x) has density B(x)/µB. the unique positive solution of ve = 1.e. (2.2.) ( t2 ) .1. u Example 2 .q2. = qz ql (x1 xz) = ql + qz ql + q ' and the nonzero eigenvalue A = ql .T) where vt is the distribution of it which is obviously given by the expression in (a).6. (ii) First check the asserted identity for the density: since T. cf.e. Here are two different arguments that this yields the asserted expression: (i) Just check that aT1/µB satisfies (2. The formulas involve the matrixexponential of the intensity matrix Q = T + to = ( tll + tlal t12 + t2al tlz + tlaz _ q1 ql t22 + t2a2 q2 q2 (say). T1 and eTx commute. Al. According to Example A3. The time of the next renewal after t is the time of the next jump of the second type. The renewal density is then aeQtt = (al a2) ( 7i 7"2.
5 Let B be Erlang(2).`t (al a2) + C 11 172 ir12 / \ t 2 ) r1 (7r1 7r2) ( t2 7rltl + J + eAt (al a2) ( 71(t2 . A = 25.a27r1) (t1 .4 yields the renewal density as u(t) = 2 (1 . t1B 0 Example 2 .52) 25152 51x2+5251 51a2+5251 Notes and references Renewal theory for phasetype distributions is treated in Neuts [268] and Kao [221].6 Let B be hyperexponential. The present treatment is somewhat more probabilistic. Then _ Q Hence 51 0 0 52 + 51 52 _ 5152 51a2 ) (al a2) 52a1 62a1 Slat + 52a1 51a2 51a2+52a1 A = 51a2 .e2bt) 13 Example 2 . and Example 2. MATRIXANALYTIC METHODS e.(biaz + aza.tl) 7r2t2 + eat (a17r2 .a27rl) (tl . .t2) 1 + eat (a17r2 .226 CHAPTER VIII. )t (51 .4 yields the renewal density as u(t) = 5152 e.t2) . and Example 2. Then Q= 0 55 )+(1o)=( j ad ).52a1. Hence 7r = (1/2 1/2).
T + to+). Considering the first. add a more selfcontained explanation of why of the phasetype structure is preserved.p. Next. T). r(u) the time of ruin with initial reserve u. and rewriting in matrix form yields the phase generator of {my} as T + ta+. Then: (a) G+ is defective phasetype with representation (a+.(u) = a+e(T+tQ+)u Note in particular that p = IIG+II = a+e. a+j. (b) V. The stars represent the ladder points ST+(k).3. represent the maximum M as the lifetime of a terminating renewal process and use Corollary 2. however. 3. Since the results is so basic.f3aT1. Within ladder steps. T(0) < oo) the ladder height distribution and M = supt>o St. Proof The result follows immediately by combining the PollaczeckKhinchine formula by general results on phasetype distributions: for (a). Corollary 2. B the claim size distribution.e. 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. and if there is a subsequent ladder step starting in j whic occurs w. the Markov processes representing ladder steps can be pieced together to one {my}. T).2. G+(. For (b). we shall. use the phasetype representation of Bo. marked by thin and thick lines on the figure. Thus the total rate is tip + tia+. Now just observe that the initial vector of {mx} is a+ and that the lifelength is M.) = F(ST(o) E •.3. 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. . which occurs at rate ti. Corollary 3. We asssume that B is phasetype with representation (a. Here we have taken the terminating Markov process underlying B with two states. cf. T) where a+ is given by a+ = . Then each claim (jump) corresponds to one (finite) sample path of the Markov process. The essence is contained in Fig. 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. and M is zeromodified phasetype with representation (a+. itself phasetype with the same phase generator T and the initial vector a+ being the distribution of the upcrossing Markov process at time ST+_.i. i. {St} the claim surplus process.1 on the next page. with 0 denoting the Poisson intensity.1 Assume that the claim size distribution B is phasetype with representation (a.
 Figure 3.1 This derivation is a complete proof except for the identification of a+ with . T = (3 .QaT1. This is in fact a simple consequence of the form of the excess distribution B0. MATRIXANALYTIC METHODS t . see Corollary 2.1 . 7e7x 2 2 Thus b is hyperexponential (a mixture of exponential distributions) with a (2 2 ).t t d kkt S.2 Assume that ... 3e3x + .Q = 3 and b(x) = .228 CHAPTER VIII.7)diag so that a+ = QaT 1 = 3 ( 3 2 2) 0 3 9 2 14 7 2 11 2 T+ta+ = 3 0 07/+( 7I \ 2 14 . 0 Example 3.M {mx} ST+(2)  S .3..
the duality result given in Corollary 11.and Markovmodulated models. the discussion around Fig. where a+ is the (defective) .4.4. 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 . with A denoting the interarrival distribution and B the service time distribution.T). see Section 6. his derivation of +'(u) is different. It is notable that the phasetype assumption does not seem to simplify the computation of finite horizon ruin probabilities substantially. but there the vector a+ is not explicit but needs to be calculated (typically by an iteration). The parameters of Example 3. we encounter similar expressions for the ruin probabilities in the renewal. T). For the compound Poisson model.6.2 are taken from Gerber [157]. see Stanford & Stroinski [351] . We shall derive phasetype representations of the ruin probabilities V) (u).1): Proposition 4. see Shin [340]. 3. (a) G+ is of phasetype with representation (a+. Fig.6). this was obtained in Section 3. and the argument for the renewal case starts in just the same way (cf. We assume p = PB/µA < 1 and that B is phasetype with representation (a. T) for some vector a+ = (a+.1 In the zerodelayed case. cf. For further more or less explicit computations of ruin probabilities. 3.^(u) = a+e( T+ta+)ue = 24eu + 1 e6u 35 35 0 Notes and references Corollary 3. but that such a simple and general solution exists does not appear to have been well known to the risk theoretic community. That is. THE RENEWAL MODEL This is the same matrix as is Example 1. 4 The renewal model We consider the renewal model in the notation of Chapter V. The result carries over to B being matrixexponential.1 which does not use that A is exponential) by noting that the distribution G+ of the ascending ladder height ST+ is necessarily (defective) phasetype with representation (a+.1 can be found in Neuts [269] (in the setting of M/G/1 queues. if we define {mz} just as for the Poisson case (cf. In the next sections. For an attempt.j). 0(8) (u) (recall that z/i(u) refers to the zerodelayed case and iY(8) (u) to the stationary case).
4. where a(8) = aT1/PA. Since the conditional distribution of my given T1 = y is ae4y. CHAPTER VIII.T).230 distribution of mo.5. The key difference from the Poisson case is that it is more difficult to evaluate a+. where B0 is the stationary excess life distribution corresponding to B. where u w(a +) = aA[T + to+) = a J0 e(T+t+)1A(dy). but with initial distribution a rather than a+. Also. In fact.T).1. Then {m. the calculation of the first ladder height is simple in the stationary case: Proposition 4. We have now almost collected all pieces of the main result of this section: Theorem 4 . Fig. the form in which we derive a+ for the renewal model is as the unique solution of a fixpoint problem a+ = cp(a+). obviously mo = m.*} from {St+y . Then . But by Corollary 2. Proof Obviously. MATRIXANALYTIC METHODS (b) The maximum claim surplus M is the lifetime of {mx}. cf.4 Consider the renewal model with interarrival distribution A and the claim size distribution B being of phasetype with representation (a.3. with intensity matrix Q given by Q = T + to+.1). (c) {mx } is a (terminating) Markov process on E.*'} is Markov with the same transition intensities as {mx}.3 a+ satisfies a+ = V(a+).Sy} in the same way as {mx} is defined from {St}. the Palm distribution of the claim size is just B. it follows by integrating y out that the distribution a+ u of mo is given by the final expression in (4.2 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). G(') = pBo. which for numerical purposes can be solved by iteration. B0 is phasetype with representation (aT1/µa. Hence by Theorem 11. (4.T)• Proposition 4.6.1) Proof We condition upon T1 = y and define {m. Nevertheless.
... It remains to prove convergence of the iteration scheme (4. The second follows in a similar way by noting that only the first ladder step has a different distribution in the stationary case. (4.•.0.^(u) = a+e ( T+ta+)xe. a+) > 0 = a+o) implies a+) _ (a+) > W (a+)) = a+) .4. (4.0) is an increasing function of /3.1/pA.^(8)(u) = a ( 8)e(T+ta +) xe.. I {mx} . thus .3). Furthermore . the maximum claim surplus for the stationary case has a similar representation as in Proposition 4.1) and a(8) _ aT.1(b). .3) (defined on the domain of subprobability vectors .M. a+ can be computed by iteration of (4.2) where a+ satisfies (4. The term tf3 in cp(i3) represents feedback with rate vector t and feedback probability vector (3. i.2 ) follows from Proposition 4. a+l ) = cp (a+°)) . only with initial distribution a(*) for mo. i y ^ T1= y `•r Figure 4. In particular . Hence ^p(.3) Proof The first expression in (4. and that this is given by Proposition 4. a+2) = ^p (a+l)) ..1).1 by noting that the distribution of mo is a+.e.1 . THE RENEWAL MODEL 231 . by a+ = lim a +n) where a+°) .2.
1 arrivals (n arrivals are excluded because of the initial arrival at time T1 ). 0 0 We next give an alternative algorithm.4) whenever EeR(S)U < oo.5 Let s be some complex number with k(s) > 0. limn4oo a ) < a+.232 CHAPTER VIII. s ¢ sp(T). and hence we may assume that h has been normalized such that ahA[s] = 1. Thus . Thus by (4. the corresponding right eigenvector may be taken as (sI . (4. let F be the distribution of U1 . Then (4. Theorem 4. MATRIXANALYTIC METHODS and (by induction ) that { a+ n) } is an increasing sequence such that limn. For n = 0. the normalization is equivalent to F(s) = 1. Obviously. Fn ).g.P[s] = A[s]B[s].T)'t • A[s] (4. To this end.1.T)It.T1. F[s] being interpreted in the sense of the analytical continuation of the m. However. Proof Suppose first Qh = sh. Then F[s] = a(sI .4). Let Fn = {T1 + • • • + Tn+1 > r+}be the event that {my} has at most n arrivals in [T1.f.4) makes sense and provides an analytic continuation of F[•] as long as s ¢ sp(T). To prove the converse inequality. this implies that ahA[s] # 0. 7+ ]. n) &+n) T a+.} can contain at most n . Similarly. .2. we use an argument similar to the proof of Proposition VI. Then each subexcursion of {St+Tl .T)1t. which links together the phasetype setting and the classical complex plane approach to the renewal model (see further the notes). a+ ) exists .4. 0 = a+) < a+ yields a+) _ (a+0)) (a+) = a+ (n and by induction that a(n) < a+ for all n . (4. Then e4'h = e82h and hence sh = Qh = (T + taA[Q])h = Th + A[s]tah.ST.5) yields h = (sI .5) Since s $ sp(T). Then s is an eigenvalue of Q = T + ta+ if and only if 1 =.) = P(mTl = i. Assume the assertion shown for n . so to complete the proof it suffices to show that &+ < a+) for all n. and let &+".. both quantities are just 0 . with B[s]. It follows that n1) so that on Fn the feedback to {mz} after each ladder step cannot exceed &+ a+ n) < a f ^ e(T+ t&+ 1))YA(dy) o < a is e(T+t«+1')YA(dy) _ w (a+1 )) = a+n). In that case.
As in Corollary 4....p1i .4.' that the equation F(s) = 1 has d distinct roots p1. the matrix Q in Theorem 2. THE RENEWAL MODEL 233 Suppose next F(s) = 1. Given T has been computed...T)lt + t = s(sI . .. Pd in the domain ER(s) > 0 . 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. we get at a(Q .6 Suppose u < 0. Notes and references Results like those of the present section have a long history.. Q = CD1 where C is the matrix with columns hl. the classical algorithm starts by looking for roots in the complex plane of the equation f3[y]A[ry] = 1. t(ry) > 0. we have IG_ [s] I < 1 .lt we get Qh = (T + to+)h = T(sI . .T) = 1 ata+ = a+. in turn. and the solution is ..9) we have G+[s] = 1 which according to Theorem 1.T)lt = sh.. 0).5. This gives d roots 'y. T) with a+ = a(QT)/at.. pdhd. . W v M(d) in the notation of Chapter V). (4.. Pd with corresponding eigenvectors hl. hd.. .. In older literature . Corollary 4. Then G+ is phase..5(c) means that a+(sI T)1t = 1. D that with columns p1 hl. hd. yd satisfying R(ryi) > 0.. and hence by the WienerHopf factorization identity (A. .1 has the d distinct eigenvalues . Further.. .T)It. Hence with h = (sI T).type with representation (a+.6) i=1 i=1 Proof Appealing to Theorem 4. letting vi be the left eigenvector of Q corresponding to pi and normalised by vihi = 1 . .. This immediately implies that Q has the form CD1 and the last assertion on the diagonal form . Let d denote the number of phases. Q has diagonal form d d Q = dpivi®hi = dpihivi. and define hi = (piI . explicit expressions for the ruin/ queueing probabilities are most often derived under the slightly more general assumption that b is rational (say with degree d of the polynomial in the denominator) as discussed in Section 6.6. 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 ). Since R(s) > 0 and G _ is concentrated on (oo.
contained derivation).. with representation say (a(' ). the background Markov process with p states is {Jt}. involving . starting around in 1975.) d (see.g. MATRIXANALYTIC METHODS d F 1 + a J e°" ip(u) du = Ee°w = 11(t. which contains somewhat stronger results concerning the fixpoint problem and the iteration scheme. the fixpoint problems look like R=Ao+RAI+R2A2+ . Neuts and his students. [270] and Latouche & Ramaswami [241]. a pioneering paper in this direction is Tacklind [373]. The distribution of W comes out from the approach but in a rather complicated form . The exposition here is based upon [18].4. That is . In queueing theory. and the distribution of an arrival claim is B. Numerical examples appear in Asmussen & Rolski [43]. see Dickson & Hipp [118].type assumptions are basic. Asmussen & O'Cinneide [ 41] for a short self. 5 Markovmodulated input We consider a risk process {St } in a Markovian environment in the notation of Chapter VI. see Neuts [269].. e. It turns out that subject to the phase. The matrix. The number of elements of El=> is denoted by q. but the models solved are basically Markov chains and processes with countably many states ( for example queue length processes ). where R is an unknown matrix. an alternative approach (the matrixgeometric method ) has been developed largely by M. The solutions are based upon iterations schemes like in Theorem 4.exponential form of the distribution was found by Sengupta [335] and the phasetype form by the author [18]. the intensity matrix is A and the stationary row vector is ir . whereas the approach was introduced in queueing theory by Smith [350]. T(').234 then in transform terms CHAPTER VIII. Here phase. and appears already in some early work by Wallace [377]. the ruin probability can be found in matrixexponential form just as for the renewal model. The arrival rate in background state i is a.type assumption . [119]. 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 surveys . is phasetype. In risk theory. similar discussion appears in Kemperman [227] and much of the queueing literature like Cohen [88].. E(t)). We assume that each B. For further explicit computations of ruin probabilities in the phasetype renewal case .F.
and the one E(•) for B. 5a Calculations via fluid models. (a) 0 0 ♦ o ° tl ♦ • 0 0 o } o o (b) 0 } ♦ • 0 o f o Figure 5. say with slope r(i) on intervals where It = i.1. Section 5b then gives a representation along the lines of Theorem 4.Vt)} obtained by time reversing the I component. p = ql = Q2 = 2. O. The two environmental states are denoted o. Vt)}t>o such that {It} is a Markov process with a finite state space F and {Vt} has piecewiese linear paths. The stationary distribution is obtained by finding the maximum of the Vcomponent of the version of {(It. The connection between the two models is a fluid representation of the Markovmodulated risk process given in Fig. MARKOVMODULATED INPUT 235 some parameters like the ones T or a+ for the renewal model which need to be determined by similar algorithms.2. for which the relevant fixpoint problem and iteration scheme has already been studied in VI. the phase space E(°) for B. The key unknown is the matrix K.6. However. has states o. 5. states .5.4. Diagonalization Consider a process {(It.1 In Fig. 5. We start in Section 5a with an algorithm involving roots in a similar manner as Corollary 4. the analysis involves new features like an equivalence with first passage problems for Markovian fluids and the use of martingales (these ideas also apply to phasetype renewal models though we have not given the details).1. 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. •. This calculation in a special case gives also the ruin probabilities for the Markovmodulated risk process with phasetype claims.
A claim in state i can then be represented by an E()valued Markov process as on Fig. This implies that in the fluid context. resp.236 CHAPTER VIII. The intensity matrix for { It} is (taking p = 3 for simplicity) I A .92a(2) 0 0 T(2) 0 0 0 f33a(3) 0 0 T(3) with the four blocks denoted by Ei„ i. we have more martingales at our disposal.1) if and only if s is an eigenvalue of E.31a(l) (/3i)diag . '31a(1) 0 0 f32a(2) 0 0 AI = t(1) 0 0 0 t(2) 0 0 0 t(3) 0 T1 0 0 0 0 T(2) 0 '33a(3) 0 0 T(3) The reasons for using the fluid representation are twofold. a) = 1. of E into components indexed by E.1(b) {(It . Second. a) : i E E. 5. MATRIXANALYTIC METHODS 4. < oo for all s.Vt)} is then obtained by changing the vertical jumps to segments with slope 1.1))diag ) a = sa and the eigenvector b = . First. 2. j = 1. Recall that in the phasetype case. a E E(i) } .1 A complex number s satisfies 'A+ (f3i(Bi[s] . the probability in the Markovmodulated model of upcrossing level u in state i of {Jt} and phase a E Eli) is the same as the probability that the fluid model upcrosses level u in state (i. 4.(Ni)diag r(i. F = E U { (i. Bi[s] = a(i)(T(i) + sI)it('). t.A 0 Or 1A/ _ t(i) 0 t(2) 0 0 0 0 0 t(3) 0 T1 0 0 0 . •. V. o.1(a).1))diag + sII = 0 (5. In the general formulation . If s is such a number. whereas Ee8s' = oo for all t and all s > so where so < oo. corresponding to the partitioning + Epp). 4. consider the vector a satisfying (A + (13i(Bi[ s] . 4}. i E E. Let E denote the matrix . Thus F = {o. r(i) _ 1. in the fluid model Eel'. 5. Eli) + Proposition 5. The fluid model on Fig . a) of {It}. F is the disjoint union of E and the Eli).
E12E22 E21 I .sI.32a(2) (/3i)diag .A .(sI . with Eii replaced by Eii .1). it follows that Ell E12 ( E 21 E22) (d) = s 1 d I .sI 0 0 0 T(2) .sI + E12 (sI .E21a + sd = sd. it follows that if Qla(1) 0 0 .E22 .sI 0 0 0 T(3) . d correspond to the partitioning of b into components Proof Using the wellknown determinant identity Ell E12 E21 E22 E22 I ' I Ell .E22)1 E21a. resp . and let d = (sI .sI+ ((3ia(i)(T(i) . Noting that E11c + E12d = se by definition. E(1) + + E(P).sI ()3i)diag .A . Then E21c+E22d = E21a .Nla(1) 0 0 T 1. d = (sI E22)1E21a = E ai(sI .T('))1t(i) . where c. For the assertions on the eigenvectors. t(1) 0 0 then also 0 t(2) 0 .sI 0 0 t(3) 0 0 = 0. MARKOVMODULATED INPUT 237 indexed by E. Then (up to a constant) c = a. assume that a is chosen as asserted which means (Ell . iEE (a> of 0* 1 AI.E22)1 E21a E21a .E22)1 E21) a = 0. 0 .sI) (sI . c = a.sI)1t)) iag I = 0 which is the same as (5.5.
q. the result u follows.. a )d(a + e8 °vpi (u . it follows by Proposition II. a)).v) = = p i( u . .5. .sv)b(v) = 0. w(u. . a)). Example 5 .pi(u.v) = v) I.. j.upi(u. s2 are the negative eigenvalues of Al +01 A1 E _ A 2 b1 0 52 A2 +32 0 . v.. Letting v ^ oo and using Rsv < 0 yields e8'u = Epi(u.4 that {e"1b(v) is a martingale . . We can take a = c = 1 and get d = (s + b)16 = 5/(3 = 1/p. v > 0. u) Iw(u. v = 1. a) and noting that i1 (u) = >I j.. j. For u. ... pi(u.a)d^ ). v.a Solving for the pi(u./' u = e' (esiuc ( 1) . MATRIXANALYTIC METHODS Theorem 5. j. j.4 Assume that E has two states and that B1. a). ..j)c v .j.. Proof Writing Or'Alb( v) = svb( v) as (AI .. v... c j.Q.3 Consider the Poisson model with exponential claims with rate 5.. we first look for the negative eigenvalue s of E = I 0 I which is s = ry with yy = b . j.O. To determine 0 (u).v)=inf{t >0:Vtu orVt=.( u. < 0 and let b(v) = I d(„)) be the right eigenvector corresponding to s. define w(u.v) = Optional stopping at time w (u. Iw(u. a) = (j. Thus 0(u) = esu/d = pe7 ° as u should be. d("))1 e.v) = j).238 CHAPTER VIII.2 Assume that E = Or 'Al has q = ql + + qp distinct eigenvalues si. a) = Pi (Vw(u. I' i( V P2 (w (u) < oo.v.v}.. v) yields C{V) = e8 . Then we get V)i (u) as sum of two exponential terms where the rates s1.v) = (j. Here E has one state only.. Then . B2 are both exponential with rates 51 i b2. sq with $2s. Example 5 . j) pi( u . w(u)=inf{t >O:Vtu}. e89uc(e)) (d(1) .
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. (y.h. oo)) j)ye. Proof We must show that G+ (i. U) where t(j) + t(j)O(j j = k uja. 8^')IT(j)) where e 3^') =.( 2. 9('). j.2) the l.y = to B k7 j # k In particular. •) is phasetype with representation (E(i).5.xxej • a 00 oo el . MARKOVMODULATED INPUT 239 5b Computations via K Recall the definition of the matrix K from VI. In terms of K.5 G+(i.6 For i E E. 0 Theorem 5 . j.k. (5.3) . is 0 /3 f R(i .s.3j eye.33(e = 0 a(j))(K ®T ( j))(ej (9 I).2. j. according to VI. the Pidistribution of M is phasetype with representation (E(1) + + E(P).x) 00 f ° (') (j) eT (yy)edx . dx)Bj(y . i. (') a T( However .b (u) = Pi(M > u) = 9(i)euue.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.
g.. Then b*[0] is rational if and only b(x) is matrixexponential. a) is obviously chosen according to e(`). oo) and b* [0] = f °O eBxb(x) dx the Laplace transform. Starting from Jo = i. some square matrix T and some column vector t (the triple (a. which occurs at rate t^^7.. bn1 bn). Piecing together these phase processes yields a terminating Markov process with state space EiEE E(')..f.) which is rational. (6. . +aii10+anI then a matrixexponential representation is given by b(x) = aeTxt where a = (b1 b2 . For j = k.. For a transition from (j. and it just remains to check that U has the asserted form. 6 Matrixexponential distributions When deriving explicit or algorithmically tractable expressions for the ruin probability.p. we have sofar concentrated on a claim size distribution B of phasetype...1 Let b(x) be an integrable function on [0. 0 1)'. the ratio between two polynomials (for the form of the density. we have the additional possibility of a phase change from a to ry within the ladder step. see Example 1. t = (0 0 .k y. i.240 CHAPTER VIII. t) is the representation of the matrixexponential distribution/density): Proposition 6. 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. say. and a new ladder step of type k must start in phase y. which occurs w.. This yields the asserted form of uja. which occurs at rate t(i). a) to (k. if b* [0] = b1 +b20+b302 +. i. the initial value of (i.5). that the density b(x) can be written as aeTxt for some row vector a. and lifelength M. Furthermore. u Notes and references Section 5a is based upon Asmussen [21] and Section 5b upon Asmussen [17]. +bn0i1 0n +a10n1 +. T. An alternative characterization is that such a distribution is matrixexponential. intensity matrix U. Bk7 .e. a m. equivalently. MATRIXANALYTIC METHODS Proof We decompose M in the familiar way as sum of ladder steps .y) to occur when j # k.e.2) .2. However. with phase space EU> whenever the corresponding arrival occurs in environmental state j (the ladder step is of type j). Associated with each ladder step is a phase process. Numerical illustrations are given in Asmussen & Rolski [43].. the current ladder step of type j must terminate.
. of (6. .. 0 1 an an1 an _2 .2). . The converse follows from the last statement of the theorem. Namely..(6.h. t= 0 .3) that we can take 0 1 0 0 a= (1 + 47r2 0 0). Thus. One of his elementary criteria. then b*[0] = a(0I T)1t which is rational since each element of (01 .. S. 1 . personal communication).T)1 is so. where c = 1 + 1/47r 2. s = c/ 2 .an_3 an _ 4 .3) was suggested by Colm O'Cinneide. t).. a2 a1 Proof If b(x) = aeTxt. matrixexponentiality implies a rational transform. shows that the distribution B with density b(x) = c(1 cos(21r x))ex.. For a proof..3 A set of necessary and sufficient conditions for a distribution to be phasetype are given in O'Cinneide [276]. since 1 + 4ir2 03 + 302 + (3 + 47x2)0 + 1 + 47r2 it follows by (6.e(tai1)x/2 + e'T) it follows that a matrixexponential representation ()3. u giving b(x) = E 1 ciebiz/bY.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. but as follows from Proposition 6. cannot be phasetype.47r2 3 ./(0 + bi). . b(x) > 0 for x > 0..1 0 0 )3 = (111). T. s) is given by 27r i . (6. namely to asssume the roots 6l.s. bn of the denominator to be distinct and expand the r..2 A remarkable feature of Proposition 6. Writing b(x) = c(e( 2ni1 ) y/2 . 0 0 0 0 1 0 0 . (6.2). 0 0 .3) 0 0 0 0 0 . we can always obtain a real one (a. T= 0 0 1 .1.47x2 3 1 0 .6. Example 6 . MATRIXEXPONENTIAL DISTRIBUTIONS 241 T = 0 1 0 0 0 . u Remark 6.4) 0 0 1 c This representation is complex.. (6.1 0 . see Asmussen & Bladt [29] (the representation (6. S = f c/2 0 21ri . .1) as E 1 c.
Corollary 111. t) of b(x). 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. q are polynomials without common roots.242 CHAPTER VIII. Consider the distribution with density = 15 ((2e2x .4) the Laplace transform of the ruin probability is /g(e)PO 0*[e] _ /' eeu^G(u)dU = 0 9(/3a0p(9)ap (9)/q(9)) .6) The remarkable fact is. we have represented ti* [0] as ratio between polynomials (note that 0 must necessarily be a root of the numerator and cancels).6) in Section 3 seems to use the probabilistic interpretation of phasetype distribution in an essential way. Then (cf. and present two algorithms for calculating '(u) in that setting. and that the minimal number of phases in a phasetype representation increases to 0o as 5 .1 to get i (u) = f3esus. . As for the role of matrixexponential distributions in ruin probability calculations. T the phase generator and t = Te.5) Thus. t) a phasetype representation with a the initial vector. T. that despite that the proof of (6. we shall only consider the compound Poisson model with arrival rate 0 and a matrixexponential claim size distribution B. then: Proposition 6. (6. For the second algorithm. then 5(u) = a+e(T+t+)uTle where a+ = /3aT1.3.1)2 + 6). and can use this to invert by the method of Proposition 6. We recall (see Section 3. we take as starting point a representation of b* [0] as p( O)/q(9) where p. recall that t = Te) that if B is phasetype and (a. MATRIXANALYTIC METHODS Example 6 .5 (6. 7 + 155ex b(x) Then it is known from O'Cinneide [276] that b is phasetype when 6 > 0. For the first. 0. leading to matrix calculus in high dimensions when b is small.4 This example shows why it is sometimes useful to work with matrixexponential distributions instead of phasetype distributions: for dimension reasons .6) holds true also in the matrixexponential case.1 shows that a matrixexponential representation can always be u obtained in dimension only 3 independently of J. (6. we use a representation (a. T.
.7) 9( cf.1BVA1. b+ = a+(9I .T)1t.1UB(B + BVA1UB). Presumably. U =.T .5 ). (91. Now.T)1 (91.T)1 J0 00 b(x) dx = f aT1t.T)1t ( l .A .T .6). the assertion is equivalent to a+(BI .T)1T 2 = and 1 = AB IT2 + 82T .T)1t)1a +(9I .6.1 + 82 (9I . with A = 91T.1 = A1 . Then in Laplace transform formulation .T .to+)1 = (BI .1 = ^(T1 + ( 91T)1).T)1 + (6I . 519) (A + UBV ).1 + b+ = b++ 1 .T)1T1t.b* (6.6b* .T)1 so that b* b** b** a+(9I . but we shall give an algebraic proof.T). we get (91. (6.t.to+)1T .B=land V=a+. this can be verified by analytic continuation from the phasetype domain to the matrixexponential domain .to+)1T .T)1 + 1 ib* (91.T)1ta+(OI . (6. From the general matrix identity ([331] p.'t. since (91T)1T .1t = f3a (0I T)1T1t . MATRIXEXPONENTIAL DISTRIBUTIONS 243 Proof Write b* = a(9I . we get b+ = 0aT1(9I T).1t du = .1t = b* .a+(9I . b+ = a +(BI . xb(x) dx = aT2t.
1. /3aT1(0I .s.T)1)t = 8 (1 .8. but the argument of [286] does not apply in any reasonable generality). a key early paper is Cox [90] (from where the distribution in Example 6. (for some remarkable explicit formulas due to Paulsen & Gjessing [286]. VII.3a (1 0 T 2 + 1 T 102 (9I + 02 1 T)1) t P + 7. premium rate p(r) at level r of the reserve {Rt} and claim size distribution B which we assume to be of phasetype with representation (E.1. 7a Computing O(u) via differential equations The representation we use is essentially the same as the ones used in Sections 3 and 4. to piece together the phases at downcrossing times of {Rt} (upcrossing times of {St}) to a Markov process {mx} with state space E. 3.h. see the Notes to VII.la.5 is similar to arguments used in [29] for formulas in renewal theory. which is selfexplanatory given Fig. 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.244 CHAPTER VIII. some key early references using distributions with a rational transform for applied probability calculations are Tacklind [373] (ruin probabilities) and Smith [350] (queueing theory). See Fig. 7.b*). 7 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate 0. cf. of (6. . T). a. A key tool is identifying poles and zeroes of transforms via WienerHopf factorization.8 a(T1 + (01.1t = /3a (9I . From this it is straightforward to check that b**/(b+ . 0 Notes and references As noted in the references to section 4.1) is the same as the r.T)1T.7).82b*. In Corollary VII. see Asmussen & Bladt [29]. For expositions on the general theory of matrixexponential distributions. the ruin probability(u) was found in explicit form for the case of B being exponential.1. Lipsky [247] and Asmussen & O'Cinneide [41].3 is taken). Much of the flavor of this classical approach and many examples are in Cohen [88].1.T)1T2t . MATRIXANALYTIC METHODS . The proof of Proposition 6.
7. Ai(t) = P(mt = i). 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.tl < t2 < u. we obtain V)(u) = P(m„ E E) = v(u)P(0.t)).I] I 80 { tq f Q(v) dvl t1 1 .1 The difference from the case p(r) = p is that {m2}. RESERVEDEPENDENT PREMIUMS 245 Rt l0 u . Let P(tl. >iEE Vi (U) is the ruin probability for a risk process with initial reserve 0 and premium function p(u + •). Given the v(t) have been computed. By general results on timeinhomogeneous Markov processes.1 A(0) = v(u) and A'(t) = A(t)(T + tv(u . In fact. Note that in general >iEE Vi (U) < 1. Also. i. 0 < t < u. the A(t) and hence Vi(u) is available by solving differential equations: Proposition 7. Proof The first statement is clear by definition. t + s) .u)e = A(u)e (7. in contrast to Section 3. t) is the vector of state probabilities for mt. P(tl. t2) = exp where Q(t) = ds [P(t. is no longer timehomogeneous.1) where A(t) = v(u)P(0.1z I. Since v(u) = (vi(u))iEE is the (defective) initial probability vector for {m8}.t2) be the matrix with ijth element P (mt2 =j I mtl = i).e. the definition of {m8} depends on the initial reserve u = Ro. O<. Figure 7. though still Markov.
A'(t) = A(t)Q(t) = A(t)(T + tv(u . the probability of downcrossing level u in phase i for the first time is E vj (u + p(u)dt) (Sji + p( u)dt • tji + p(u)dt • tjvi(u)) jEE vi(u) + vi' (u)p(u)dt + p(u) dt E {tji + tjvi(u)} jEE Collecting terms. In the first case. MATRIXANALYTIC METHODS However. 0 < t < u. two things can happen: either the current jump continues from u + p(u)dt to u. those corresponding to state changes in the underlying phase process and those corresponding to the present jump of {Rt} being terminated at level u .2 For i E E. Given this occurs.t and being followed by a downcrossing. dt]. the probability of downcrossing level u in phase i is 8ji(1 + p(u)dt • tii) + (1 .246 CHAPTER VIII. Hence Q(t) _ T + tv(u .Qdt) vi(u) + vi'(u)p(u)dt + p(u) dt E{tji+tjvi(u)}. jEE . given A. Given A'.t)). the interpretation of Q(t) as the intensity matrix of {my} at time t shows that Q(t) is made up of two terms: obviously.t) for the second. (7.4) jEE jEE Proof Consider the event A that there are no arrivals in the interval [0. {mx} has jumps of two types. or it stops between level u + p(u)dt and u. Given A.3dt. Proposition 7. the probability of which is 1 . Thus. vi.Sj i)p(u)dt • tji = Sji + p(u)tji dt.(u) p ( u) = . the probability that level u + p(u)dt is downcrossed for the first time in phase j is vj (u + p(u)dt).t).(tai + vi(u) E vj(u)tjp (u)  Q + vj (u)tjip ( u). whereas in the second case the probability is p(u)dt • tjvi(u). 0 Thus. we get vi(u) = aidt + (1 . The intensity of a jump from i to j is tij for jumps of the first type and tivj(u . from a computational point of view the remaining problem is to evaluate the v(t). the probability that level u is downcrossed for the first time in phase i is ai.
This yields v.^ 0. say. P u which implies that v. Then pv(r) p(r) r < v p r>v ' and (no matter how p is chosen) we have: Lemma 7.i7rT1/p. starting from v"(v) = . we face the difficulty that no boundary conditions is immediately available.) + 0 as v + oo. .4) backwards for {va (t)}v>t>o. vi (U) = lim v= (u). after a certain level v... <oo.. consider a modification of the original process {Rt} by linearizing the process with some rate p. denotes the time of downcrossing level u . When solving the differential equation in Proposition 7.0 as v + 00 we have P(A) P"(A) = P(AnBv)+P(AnBv) P"(AnB.2.00 Proof Let A be the event that the process downcrosses level u in phase i given that it starts at u and let B" be the event By={o.) P"(AnBv) = P(AnB. (u) on both side and dividing by dt yields the asserted differential u equation. refer to the modified process. From Section 3. of (7.. Thus.s. (u) for any values of u and v such that u < v. Since the processes Rt and Rt coincide under level B. To deal with this. Rt . say. F" etc. and similarly P"(Bv) . we can first for a given v solve (7.3 For any fixed u > 0. then P(A n Bv) _ P"(A n BV').7. V . we have p(r) = p = vi (u) 0aTe. supRt>v l t<7 I where o.5).h.) is the tail of a (defective) random variable so that P(Bv) + 0 as v 4 oo. Let p" (t). Now since both P(A n Bv) 3 0 and P"(A n Bv) . (v) is given by the r. RESERVEDEPENDENT PREMIUMS 247 Subtracting v. Then P(B.)P"(AnB.
0 < u < v. To evaluate p1(u). cf. Let ii'( u) = a+'ie(T+ta +^)"e denote the ruin probability for R't where a+ = a+i) = laT1/pi.V" M 0 . Recall that q(w) is the probability of upcrossing level v before ruin given the process starts at w < v. the evaluation of Vi(u) requires q(u) = 1 .x) dx f v(u)eT xt dx . typically the complexity in n is 0(n2) for integral equations but 0(n) for integral equations. We recall from Propositon VII.1.9. where. MATRIXANALYTIC METHODS Next consider a sequence of solutions obtained from a sequence of initial values {v. v = u. The trapezoidal rule used in [288] gives a precision of 0(n 3).6) We may think of process Rt as pieced together of two standard risk processes RI and Rte with constant premiums p1.. r<v r > v. Then v(u) is the initial distribution of the undershoot when downcrossing level v given that the process starts at u.q(v dx +( ) ) = ( ) ( q( )) vueTva (7. as well p1(u). the probability of ruin between a and the next upcrossing of v. Corollary 3. < 0}). which is available since the z/i'(.. 2/n. The f iin in (7. The algorithm based upon numerical solution of a Volterra integral equation (Remark VII. assuming u > v for the moment. > 0} and the last term the contribu tion from {R. 1/n.. 2u.7) f o (the integral is the contribution from {R.z51(v)). 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. p(r) P. 7b Twostep premium rules We now assume the premium function to be constant in two levels as in VII. i. for u > v the distribution of v . 3u etc.1. while the fourthorder RungeKutta method implemented in [30] gives 0(n5). Therefore u pl(vvueTa t 1. numerically implemented in Schock Petersen [288]) and the present one based upon differential equations require both discretization along a discrete grid 0.v v(u)eTat 1 1 .248 CHAPTER VIII. T). let v(u) = a+2ieiT +ta+>)(uv).1a.1.. However. where v = inf It > 0 : Rt < v}. (7.10 that in addition to the O'(•).zp1(u)/(1 . (u)}. The precision depends on the particular quadrature rule being employed.e. such that Rt coincide with RI under level v and with Rt above level v.) are so. p2.RQ (defined for or < oo only) is defective phasetype with representation (v(u)..7) equals 01 (v . say.
x) dx} V 1 1(v) f V v(u) eTxt.v(u)eTve).v(u ) eTV e J v(u)eTxtz/)l (v .24e.e6u 35 .01 (v . B is hyperexponential corresponding to 3 0 3 a(2 2)' T= ( 0 7 t.21 = ? yields 0(u) = 1. RESERVEDEPENDENT PREMIUMS 1 . Example 7. Then one gets X20 20 21 f 1ea1(u v) + 1 3 3 ^ A 2(u e .4 Let {Rt } be as in Example 3. so we consider the nontrivial case example p2 = 4 and p1 = 1.v) + (2^ + 3v2 ea'(u " .4) can be written as (Y(u) ®a+)e(T+t°+>)°1 (T ® (T .(7 The arrival rate is (i = 3.be the eigenvalues of T + to( 2 ). From Example 3. 01(u) _ 24 u + 35 e6u 1 35 e 4(u) _ 35 .8) The integral in (7.1 from which we see that pl (u) = 1 + 1 249  1 .e. p2 < 3.to+))11 {e{T®(Ttoy+ ))}„ .8) equals v v(u)eTxta+2) e(T+ta +))( vx)edx which using Kronecker calculus (see A.and A2 = 3 .jl (t ®e) Thus.24ev .x) dx 1 ^(v) ( 1 . I.v) 1eai(u v) + 7 7 1 e\2(u v) 1 3 ^') eA2 (u.2. all quantities involved in the computation of b(u) have been found in matrix form. Since µB = 5/21.^1(v) 1 .2.v(u)eTVe .e6v Let Al = 3 + 2V'2..2V"2. (7.7.u .
The analysis and the example are from Asmussen & Bladt .250 CHAPTER VIII. MATRIXANALYTIC METHODS From (7./2) ea 1(u .1 Thus. ) e sv + ( 2v/2.b(v) = 192esv +8 35e6v + 168esv + 7* Thus all terms involved in the formulae for the ruin probability have been exu plicitly derived. 192esv + 8 P1 ./2 ea1(u") .1.1)' ?. pi (u) = p12(u)/p1 l(u) where p1i(u) p12(u) 35e6v .24e5v . and one gets 12e5" .. 21 3 In particular.+ it (3 4'I 1 ea2(uv e1\2(u") 7 + ( 32 +4.21(35e6v .v)esv + 7 4_ 2.24es" . Notes and references [30].24e5v .1 V2 = 4e5"+6 35e6v .7) we see that we can write pi (u) = v(u)V2 where V2 depends only on v.24es" .2 35e6v .
L(tx)/L(x) 4 1.2b. (b) the lognormal distribution (the distribution of eu where U . B(x) = ex0 with 0<0<1. for all t > 0.B(x). For example.4.f. For further examples. The definition b[s] = oo for all s > 0 of heavy tails is too general to allow for a general nontrivial results on ruin probabilities.N(µ. x 4 oo. x 2iror2 (c) the Weibull distribution with decreasing failure rate .g. and instead we shall work within the class S of subexponential distributions . Some main cases where this lighttail criterion are violated are (a) distributions with a regularly varying tail. B[s] = f e8x B(dx) is finite for some s > 0 in the lighttailed case and infinite for all s > 0 in the heavytailed case. oo ) and say then that B is subexponential (B E S) if 251 . III. the exponential change of measure techniques discussed in II.46 and at numerous later occasions require a light tail. A rough distinction between light and heavy tails is that the m. see I. 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. we require that B is concentrated on (0. For the definition . a2)) with density 1 e(logyFh) 2/2az .
Proof By the inclusionexclusion formula. given X1 + X2 > x. That is. In terms of r. HEAVY TAILS B*2\ 2.2.v.252 CHAPTER IX. X2 with distribution B. then P(X1>xI X1+X2>x)* 2. We later show: Proposition 1. As contrast to Proposition 1. 1).'s. the behaviour in the lighttailed case is illustrated in the following example: Example 1. (b) liminf BB(() ) > 2. oo). X2 but none of them exceeds x. B(x) Here B*2 is the convolution square. That is.B(x)2 . the distribution of independent r. X1 is w.2 If B E S.F(X1 > x. and thus the lim inf in (b) is at least lim inf P(max(Xi. Since B is concentrated on (0. one can check that x x where U is uniform on (0. (1. To capture the intuition behind this definition. note first the following fact: Proposition 1. B(x) ax. Thus . if X1 + X2 is large .1 Let B be any distribution on (0. Thus the liminf in Proposition 1. the r.v.p. then (with high probau bility) so are both of X1. X2) > x} C {X1 + X2 > x}. Then: (a) P(max(Xi. x 3 00.v. 1/2 it has the distribution of X1I X1 > x.1(b) is oo. that is.3 Consider the standard exponential distribution. X2) > x) ^' 2B(x). X2) > x}.2B(x). Then X1 +X2 has an Erlang(2) distribution with density yeY so that B*2(x) xex. X2) > u x)/B(x) = 2. oo). P(max(Xi.p. In contrast. . we have {max(Xi. X2 > x) = 2B(x) . X2) > x) is P(X1 > x) + P(X2 > x) . in the subexponential case the only way X1 + X2 can get large is by one of the Xi becoming large. proving (a).1) then means P(X1 +X2 > x) 2P(Xi > x). P(Xi <yI Xi+X2>x) 1B(y). The proof shows that the condition for B E S is that the probability of the set {X1 + X2 > x} is asymptotically the same as the probability of its subset {max(Xi. 1/2 'typical' (with distribution B) and w.'s X1.
S)x + B(Sx)2 < lim sup B(x) xaoo B(x) lim sup 2L((1 x^oo .5)x)' + 0 _ 2 L(x)l xa (16) Letting S 10.'s: if X . we get limsupB*2(x)/B(x) < 2.xIX > x converges in distribution tooo. x]. B( 0 .B(y)) . Hence lim sup a+oo B*2(x) 2B((1 . y] and (y. we therefore get lim sup B*2(x)/B(x) > 1+B(y)+ 1 . We now turn to the mathematical theory of subexponential distributions.v. Proposition 1. This follows since the probability of the overshoot to exceed y is B (x + y)/B(x ) which has limit 1. [In terms of r. then B(B(x)y) * 1 uniformly in y E [0. If X1 + X2 > x.z B(x) . we get BZ(x)) > 1 + B(y) + B(B()y) (B(x) . The uniformity now follows from what has been shown for y = yo and the obvious inequality y E [0. then the overshoot X .B*n(x .2) B(x) B(x ) B(x) Jo with n = 1 and splitting the integral into two corresponding to the intervals [0. a contradiction. or they both exceed Sx.] Proof Consider first a fixed y.yo].B*(n ) B(dz) (1. then either one of the Xi exceeds (1 .4 Any B with a regularly varying tail is subexponential. 253 Proof Assume B(x) = L(x)/xa with L slowly varying and a > 0. SUBEXPONENTIAL DISTRIBUTIONS Here is the simplest example of subexponentiality: Proposition 1. Finally lim inf B(x . and combining with Proposition u 1.y)/B(x) > 1.y)/B(x) > 1 since y > 0.1.5 If B E S. yo] as X + 00. Using the identity B*(n+1)(x) = 1+ + 1)(x) 1+ 2 1 .B E S.B(y) = 2.S)x. Let 0 < 5 < 1/2.6)x)/((1 . If lim sup B(x .1(b) we get B*2(x)/B(x) * 2. 1 < B(x ) B( x) Y) < B( 0).
z) B(dz) (n + O(e)) ^x JO B(x) (n + 0(0) I B (x) . P(X1 > xIX1 + X2 > x) _ P(Xi > x) _ B(x) 1 P(X1 + X2 > x) B2(x) 2 1 y P(X1<y X1 + X2 > x) B(x . its intuitive content is the same as discussed in the case n = 2 above.7 If B E S.z ) (x ) = 1 + (^ B(x . 0 Proof of Proposition 1.z) B(x) Here the second integral can be bounded by B*n(y) B(x) . Then by (1. and this immediately yields the desired conclusions. Proof For 0 < 5 < e. then for any n B*n(x)/B(x) * n.B(x .254 CHAPTER IX.2).5 that B(n) > e6B(n .5 and the induction hypothesis.6 If B E 8.z) B(dz) 2B(x) o rv 2 0 2 using Proposition 1. HEAVY TAILS Corollary 1.z) B(dz).B*2 (x) B(x) (x . x oo. b[c] = oo for all e > 0.y) sup v>o B(v) B(x) which converges to 0 by Proposition 1. we have by Proposition 1. choose y such that IB*n(x)/B(x) . The case n = 2 is just the definition. so assume the proposition has been shown for n.. O The following result is extremely important and is often taken as definition of the class S. The first integral is y B(x .1) for all large n so that B(n) > cle6n for all n. This implies B(x) > c2e5x for all x. Proposition 1.z) B(dz) _y B(x) 111 Lx B . B(x) \Jo _ B(x .2. then e"R(x) * oo. B*(n+1) (x I xy + Jxx y) W.nI < e for x > y. Given e > 0.5 and dominated convergence. Proof We use induction.
z) B(dz) < 1 + A + an(1 + d) .2).4) . Proof Let X1.z) B(dz) x .ala2B(x)2 which can be neglected. oo) such that Ai (x) _ aiB(x) for some B E S and some constants al.y))/B(x). 0 Proposition 1.9 Let A1. Xi <= v Ai (x . Then by (1.z) B(x) < 1 + A + an sup f x B(x . Proof Define 5 > 0 by (1+5)2 = 1+e.ajB(x)Ai(v) = ajB( x)(1+o„(1)) (j = 3 .X2 > xv) < A1(xv)A2(x v) .5 easily yields P(X1 + X2 > x. A2 be distributions on (0. Proposition 1.'s such that Xi has distribution Ai. Since P(X1+X2 > x. an = supx>o B*n(x)/B(x).y)Ai(dy) = (x)o(1) (1.B(x .(al + a2)B(x).z) B(x . then there exists a constant K = KE such that B*n(x) < K(1 + e)nB(x) for all n and x. x>T o B(x) The truth of this for all n together with al = 1 implies an < K(1 + 5)2n where K = (1 + A)/e. Then Al * A2(x) = P(X1 + X2 > x). For any fixed v.i). choose T such that (B(x)B*2(x))/B(x) < 1 + b for x > T and let A = 1/B(T). X2 be independent r.X1 > xv.3) Using the necessity part in the case Al = A2 = B yields f xv B(x .0 completes the proof. e > 0. Combining these estimates and letting a 4. an+1 fX B*n( *n(x . SUBEXPONENTIAL DISTRIBUTIONS 255 Here the first term in {•} converges to 1 (by the definition of B E S) and the second to 0 since it is bounded by (B(x) .y)Ai(dy) v) f o .8 If B E S. a2 with a1 + a2 > 0.1. it follows that it is necessary and sufficient for the assertion to be true that JX_VA (x .v. 0 Lemma 1.z) B(dz ) + sup < 1 + sup f x<T B ( x) x>T 0 B(x . Then Al * A2 (x) .y)B(dy) = B(x)ov (1)• v (1.
y)Ai(dy) = B(x)o„(1). u Corollary 1. A(x) = o(B(x)). f " By a change of variables. u It is tempting to conjecture that S is closed under convolution.y)B(dy). it should hold that B1 * B2 E S and B1 * B2 (x) . L2 slowly varying. i = 1.B(x) Proof Take Al = A.2aB(x) . u Corollary 1. However.12 Assume that Bi(x) = Li(x)lxa. L2 are slowly varying. Proof Taking Al = A2 = A. it is easy to see that if L1. B1 * B2 (x) . .5) Here approximately the last term is B(x)o„(1) by ( 1.11 Let B E S and let A be any distribution with a ligther tail. Hence Corollary 1.10 The class S is closed under tailequivalence.Bl (x) + B2 (x) follows precisely as in the proof of Proposition 1.v)Ai(v) . HEAVY TAILS 'VV B(x .aiB(v)) = B(x)o„(1). then A E S. That is.4).v)B(v) + _'U Aq(x .2A(x).9).13 Let B have density b and failure rate A(x) such that . the l.h.(x) is decreasing for x > x0 with limit 0 at oo. whereas the two first yield B(x)(Ai(v) .Ai(x . B2 E S. In the regularly varying case. a1 = a2 = a yields A*2(x) . Then L = L1 + L2 is slowly varying and B1 * B2(x) sim L(x)/x«. V (1. with a > 0 and L1. Then A * B E S and A * B(x) .5) becomes x B(x .2. if q(x) aB(x) for some B E S and some constant a > 0.Bl (x) + B2 (x) when B1. a2 = 1. Recall that the failure rate A(x) of a distribution B with density b is A(x) = b(x)/B(x) Proposition 1.256 Now (1. then so is L = L1 + L2.3) follows if CHAPTER LX. Then B E S provided fo "O exA(x) b(x) dx < oo. We next give a classical sufficient (and close to necessary) condition for subexponentiality due to Pitman [290]. B1 * B2 E S does not hold in full generality (but once B1 * B2 E S has been shown.s. A2 = B so that a1 = 0. of (1. That is.
1.(x .`(x)b(x) is integrable.1 B(x) eA( x)A(xv )A(y)A(y) dy f B(x . the first integral has limit 1 . and exa(x)b(x) = (3x01e(10)x9 is integrable. 0 A(x) . Then B(x) = eA(x). replace B by a tail equivalent distribution with a failure rate which is everywhere decreasing). elementary but tedious calculations (which we omit) show that A(x) is ultimately decreasing. Example 1.1 has limit 1 + 0.12. Goldie & Teugels [66]): Proposition 1. Since ) (x . Thus. we first quote Karamata's theorem (Bingham. To illustrate how Proposition 1. an integrable function by assumption. a(x) = ax01. x .. subexponentiality has alrady been proved in Corollary 1.y ) b(y)dy = B (x) o ox _ J = ox/2 eA( x)A(xy ). Then b(x) = Ox0lexp. the DFR Weibull distriu bution is subexponential.y) < yA(x . SUBEXPONENTIAL DISTRIBUTIONS 257 Proof We may assume that A(x) is everywhere decreasing (otherwise.1)xcl1 . we can use the same domination for the second integral but now the integrand has limit 0 . Jo For y < x/2.y) < A (y) for y < x/2.14 Consider the DFR Weibull case B(x) = ex0 with 0 <.. B*2(x) .A(x .A(y)a(y ) = ev'(y) b(y).U) /or) v 2x This yields easily that ex. Example 1.y) dy.16 For L(x) slowly varying and a > 1.y) y\(y)• The rightmost bound shows that the integrand in the first integral is bounded by ey"(v).A(xy)A ( y). Thus by dominated convergence .3 < 1. By (1. L(x) y° (a . The middle bound shows that it converges to b(y) for any fixed y since \ (x .y) * 0.A(y)\(y) dy + fox/ 2 eA(x ). Further.e009xv)2/2a2/(x 2irv2) logx ( ) 't ((logx .(y) dy.13 works in this setting. Define A(x) = fo .15 In the lognormal distribution. Thus A(x) is everywhere decreasing. the u lognormal distribution is subexponential. proving B E S.2). Thus. f ' L(y) dy . Thus B*2(x )/ B(x) . In the regularly varying case.
13 may present a problem in some cases so that the direct proof in Proposition 1. We conclude with a property of subexponential distributions which is often extremely important: under some mild smoothness assumptions.6) EX(x) . the overshoot properly normalized has a limit which is Pareto if B is regularly varying and exponential for distributions like the lognormal or Weibull.1)])a 1 1 .L(x)/x" and )t(x) .1/A(x) and P(X ixil'Y (x) > y) * e'. . Proof ( a): Using Karamata's theorem.ea b(x) is integrable. yo] .1)X(x)/x > y) = P(X > x[1 + y/(a . However.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) = EXix>. then B(x) .y(x)B(x).E(X . HEAVY TAILS Proposition 1.)/Y(x) > y) (1 + y/(a . Then 7(x) . the monotonicity condition in Proposition 1. (c) Under the assumptions of either ( a) or (b).1)] I X > x) L(x[1 + y/(a .x)+ _ 1 °° P(X > x) P(X>x )J L PX >y)dy 1 x L(y)/ydy L(x)/((a1)x'1) x )l ° J °° ( ()l a x a1 Further P ((a . then 7(x) x/(a . Thus exa(x)b(x) .258 From this we get CHAPTER IX. More precisely.1))a .18 (a) If B has a density of the form b(x) = aL(x)/xa with L(x) slowly varying and a > 1.xjX > x. we get (1. f O B(y) dy . let X W = X .1))^ ' (b) Assume that for any yo )t(x + y/A(x)) 1 A(x) uniformly for y E (0. Then: Proposition 1.1) and P(X (. (1 + y/(a .4 is necessary in full generality.1)]) xa L(x) (x[1 + y/(a .a/x.
2 Let Y1. Examples 1. r(u) = inf it > 0. The remaining statement (1. Recall that B0 denotes the stationary excess distribution.n0 1•P(K= n)•n = EK.1/. Lemma 2.1 If Bo E S. be i. and that for each Proof Recall from Section 1 that G*n (u) z > 1 there is a D < oo such that G*n(u) < G(u)Dzn for all u..nn. .(x).15. 0 G(u) L G(u) . 1. Bo(x) = f0 B(y) dy / µB. St > u}.t be the claim surplus at time t and M = sups>0 St.EK G(u). Then P(Y1 + • • • + YK > u) . with common distribution G E S and let K be an independent integervalued r.A(x) I X > x) = exp {A(x) . 2 The compound Poisson model Consider the compound Poisson model with arrival intensity /3 and claim size distribution B.14.v.. Theorem 2 . We assume p = /3µB < 1 and are interested in the ruin probability V)(u) = P(M > u) = P(r(u) < oo). THE COMPOUND POISSON MODEL 259 We omit the proof of (c) and that EX (x) .7) is referred to as 1/A(x) being selfneglecting. . Let St = Ei ` Ui .8) in (b) then follows from P (A(x)X (x) > y) = F(X > x + y/. It is trivially verified to hold for the Weibull. with EzK < oo for some z > 1. i. d. We get p(yl+. Y2.A(x + y/A(x))} =a(x) a(x + x) dx = ex p ex P .2. then Vi(u) P Bo(u).f yl 0 0 = exp {y (1 + 0(1))} 0 fY A( x + u /A( x)) a(x) du } The property (1. u a oo.. Kliippelberg & Mikosch [134].and lognormal distributions . Notes and references A good general reference for subexponential distribution is Embrechts.+YK> u) = ^•P(K = n)G* n(u ) .. P The proof is based upon the following lemma (stated slightly more generally than needed at present). nG(u). cf.
the result follows immediately from Lemma 2. Since EK = p/(1. . Bo E S. In general: Proposition 2. Note that in these examples . The PollaczeckKhinchine formula states that (in the setup of Lemma 2.2. Weibull) one has Bo(x ( B(x) .2) M = Yl + • • • +YK where the Yt have distribution Bo and K is geometric with parameter p. in our three main examples (regular variation . mathematically one must note that there exist (quite intricate) examples where B E S. then Bo(x)/B(x) + 00. Bo is more heavytailed than B .1 is essentially due to von Bahr [56].?(xµ 8 (x). HEAVY TAILS u using dominated convergence with >2 P(K = n) Dz" as majorant. we have fx B(y)dy = a B0 (x) > lim inf lim inf x+oo B(x) .260 CHAPTER IX. (2. The problem is a very slow rate of convergence as u ^ oo. The approximation in Theorem 2. see Abate. Proof Since B(x + y)/B(x) * 1 uniformly in y E [0. _ B(x^sx Bo(x) µ8 I aoB(y )dy = (^) .13).18. and for the lognormal and Weibull cases it can be verified using Pitman 's criterion (Proposition 1. Bo ¢ S. r(1/Q) xlQexp B(x) = ex' From this ..µ J ) . P(K = k) = (1. as well as examples where B ¢ S.1) In particular . The tail of Bo is easily expressed in terms of the tail of B and the function y(x) in Proposition 1. See also Embrechts & Veraverbeeke [136]. However.3 If B E S.. a].x^ ) B(x) _ f or ( lox . lognormal .x400 PBB(x) PB Leta+oo. Proof of Theorem 2.1 is notoriously not very accurate. u x+a Notes and references Theorem 2. Bo E S is immediate in the regularly varying case.1.p) and EzK < oo whenever pz < 1. u The condition Bo E S is for all practical purposes equivalent to B E S.p)p'. Borovkov [73] and Pakes [280]. x 4 00.µB(01 . For some numerical studies.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 ) .
Then K M=EY.3.. The main result is: Theorem 3 . T1 the ith interarrival time and Xi = U.1.. E. also a second order term is introduced but unfortunately it does not present a great improvement. Based upon ideas of Hogan [200]. in [219] p.1) this end .] The proof is based upon the observation that also in the renewal setting. [279]. there is a representation of M similar to the PollaczeckKhinchine formula.. Thus G+ is the ascending ladder height distribution (which is defective because of PB < PA). 1 Assume that (a) the stationary excess distribution Bo of B is subexponential and that (b) B itself satisfies B(x . Somewhat related work is in Omey & Willekens [278].y + as usual denotes the first ascending ladder epoch of the continuous time claim surplus process {St}.1 when u is small or moderately large. Define further 0 = IIG+II = P(r9+ < oo). M = sup s$ .g. t9(u) = inf {n : Snd> > u} .y)/B (x) > 1 uniformly on compact y internals. one may have to go out to values of 1/'(u) which are unrealistically small before the fit is reasonable. In [1].1 gives 1010. + Xn.Ti. We assume positive safety loading. THE RENEWAL MODEL 261 Choudhury & Whitt [1].. To Bo(u) u + 00. Snd) = Xl +. Asmussen & Binswanger [27] suggested an approximation which is substantially better than Theorem 2. Kalashnikov [219] and Asmussen & Binswanger [27]. (3. Then l/i(u) 1 P P [Note that (b) in particular holds if B E S. G+ (A) = P(Sq+ E A. T+ < oo) where r+ = T1 + • • • + T. i.} Then ik(u) = F ( M > u) = P(i9 (u) < oo).+ E A. i=1 . p = iB /µA < 1. 195 there are numerical examples where tp(u) is of order 105 but Theorem 2. .9+ < oo) = P(S... This shows that even the approximation is asymptotically correct in the tail. 3 The renewal model We consider the renewal model with claim size distribution B and interarrival distribution A as in Chapter V.e. Let U= be the ith claim . {n= 0. let t9+ = i9(0) be the first ascending ladder epoch of {Snd> }..
Let further 19_ _ inf {n > 0: S^d^ < 0} be the first descending ladder epoch.. A(dy) = 1. As for the compound Poisson model. 0] normalized by IPG_ I so that we should have to G+(x) .Y2. Then 0 0 F( x . HEAVY TAILS where K is geometric with parameter 9. Lemma 3 .9)9'' and Y1. x > 0.d.y) dy = 1 Pi (X) oo IPG_ I .g+ > x.y+ given r+ < oo). cf. Lemma 3 .PBBo(x).FI(u). u a 00. Proof Let R+(A) = E E'+ ' I(S. d+ < oo). The heuristics is now that because of (b). G_(A) = P(S.y_ E A) the descending ladder height distribution (IIG II = 1 because of PB < P A) and let PG_ be the mean of G_. 0 The lemma implies that (3. U_ (dy) is close to Lebesgue measure on (. P(K = k) = (1 . (b) and does not rely on the structure Xi = Ui .Ti). 0] to the integral is O(F(x)) = o(FI(x)). are independent of K and i.B(x).2 F(x) .3 G+ (x) . x * oo. oo) = F(S.FI(x) /IPG_I. whereas for large y .1 IPG_ I / F(x .2).. this representation will be our basic vehicle to derive tail asymptotics of M but we face the added difficulties that neither the constant 9 nor the distribution of the Yi are explicit.(.oo.d)) E A) denote the pre19+ occupation measure and let and U_ = Eo G'_" be the renewal measure corresponding to G_.N. we will use the fact that the proof of (3.1) is equivalent to P(M > u) " . Proof By dominated convergence and (b).3) and we will prove it in this form (in the next Section.i. x + oo. Let F denote the distribution of the Xi and F1 the integrated tail.. with distribution G+/9 (the distribution of S. (3. FI (x) _ fz ° F(y) dy. B(x) _ J O° B(B(x)y) A(dy) f 1 . Write G+( x) = G+ ( x.1) holds for a general random walk satisfying the analogues of (a).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_.262 CHAPTER IX. and hence FI(x) . the contribution from the interval (. A.
3.1. n] + 1/I µG_ I. the proof is complete.1 I .9) 1 . n] F1 ( n=N _1 1+e E F(x+n) 0 + limsup xr00 FI(x) FAG.y) U.O[s])(1 .I n=N (1 E)2 r00 F(x + y) dy + e) lim sup .2). and that U_(n .e) z lim inf G+(x)  FI (x) Ip G_ I Letting a 10.1.(dy) fN FI ( x) + lim sup ZY00 N F(x .3. > (1 . u Proof of Theorem 3. If G_ is nonlattice. F(Y= > x) FI(x)/(OIp _ 1). (3. we can assume that the span is 1 and then the same conclusion holds since then U(n . THE RENEWAL MODEL 263 We now make this precise.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 0.1. n] is just the probability of a renewal at n. By Lemma 3.1.F[s] = (1 .=1 BIp G_ I (1.G+[s]) . Given e. n] < (1 + e)/1µc_ I for n > N. In the lattice case.y) U_ (dy) 00 FI (x) < lim sup F(x) U(N.0)0k k I(u) A.UG_ I x. Similarly.1)/F(n) < 1 + e for n > N (this is possible by (b) and Lemma 3.1.2) yields 00 F F I (u) P(M > u) _ E(1 . We then get lim sup G+(x) xro0 Fj(x) < lim sup X)00 o F(x . 0] x+00 FI(x) 00 + lim up 1 x) E F(x + n) U_ (n . Hence using dominated convergence precisely as for the compound Poisson model. then by Blackwell 's renewal theorem U_ (n .9)IpG_ I Differentiating the WienerHopf factorization identity (A. choose N such that F(n . and in the last that FI is asymptotically proportional to Bo E S.
HEAVY TAILS µF = (1 . Note that substantially sharper statements than Lemma 3. see Asmussen & Kliippelberg [36].a.a. 4 Models with dependent input We now generalize one step further and consider risk processes with dependent interclaim times.a) N P(M > u). allowing also for possible dependence between the arrival process and the claim sizes. S+q(u) .l. .IIG+II)µc_ = (1 .AB iP We conclude by a lemma needed in the next section: Lemma 3 .SS(u)}n=o.4 on the joint distribution of (S.a.1)6+[0] .(1 . Proof Let w(u) = inf {n : Sid) E (u . u)).. u)) > P(w(u) < oo)(i lp (0))• On the other hand.1 is due to Embrechts & Veraverbeke [136].264 and letting s = 0 yields CHAPTER IX. on the set {M > u.2.So(u)) are available. Then P(M E (u .u)) = o(P (M > u)) = o(FI(u)).yiui_1. Notes and references Theorem 3. with roots in von Bahr [56] and Pakes [280].So( u)_1 < a) < P (w(u) < oo)j/i(0) < 0(0) P(M E (u . u). In view of the `one large claim' heuristics it seems reasonable to expect that similar results as for the compound Poisson and renewal models should hold in great generality even when allowing for such dependence.Sty(u)_I < a} we have w(u) < oo. Mn < u}. S+9(u) . 10(0) But since P(M > u . must attain a maximum > 0 so that P(M > u. FJ(u) UBBO(U) PBo(u) N = (10)Ipc_I JUA .. Sty(u) . we have P(M E (u .0)ua_ ..4 For any a < oo.a. Therefore by Lemma 3. and {Su.Se(u)_1 < a) = o(Fj(u)). P(M > u.(u)+n .
see [47]. and the distribution of {Sxk+t . such that {SXo+t  SXo}0<t< X 1Xo .1 based upon a regenerative assumption.... and apply it to the Markovmodulated model of Chapter VI.... The zerodelayed case corresponds to Xo = Xl = 0 and we write then F0. t>0 S. Theorem 4. +1..1) . 2. 0o(u) etc.. (viewed as random elements of the space of Dfunctions with finite lifelengths) are i... 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 <.1 except for the first one) is a random walk. (corresponding to the filled circles on Fig.1 = max k=0. MODELS WITH DEPENDENT INPUT 265 Various criteria for this to be true were recently given by Asmussen. 4. examples and counterexamples. ..Sxi}0<t<x2Xl .d. {SX1+t . E0. We let F* denote the Podistribution of Si.n n=0.1.F*(X) = P0(Si > x) .. We give here one of them..1. M* = max S.. 4.1) is assumed.. = Sx.4. 4... M = sup St. M. assume pp. The idea is now to observe that in the zerodelayed case.. Schmidli & Schmidt [47]. {Sn}n=o.i.X2 < . Thus the assumption ..X1 is the generic cycle. See Fig.4 below..Sxk}o<t<xk+1xk is the same for all k = 1.. G(x) (4. For further approaches. We return to this point in Example 4. Figure 4. < 0 and EoX < oo where X = X2 .1 where the filled circles symbolize a regeneration in the path.1 Note that no specific sample path structure of {St} (like in Fig. Define S.
the assumption means that Mix) and Sl are not too far away.2. Then '00 (u) = Fo(M > u) .4) liminf u>oo F(M > u) .S. jF11 F* (U).3) hold.. HEAVY TAILS for some G such that both G E S and Go E S makes (3. (4.266 CHAPTER IX. The one we focus on is Fo (Mix) > x) . Sxn +t . N N Xi=0 N Figure 4. See Fig.Sxn = sup Sxn+t . 4. Since clearly M(x) > Sl ..2) to show F(M* > u) > 1. Fo(Si > X).1 Assume that (4. (4.* i o<t<xn+1x. it suffices by (4. u p 00...3) applicable so that F(M* > u) 141 F*(u).3) where Mnx) = sup o<t<xn +1 X. Proof Since M > M*.1) and (4.2) Imposing suitable conditions on the behaviour of {St} within a cycle will then ensure that M and M* are sufficiently close to be tail equivalent.2 Theorem 4. (4.
(1 .4)..5) which follows since Po (M > u.: Sn > u} . E (u .Mn +1 >aV(u n=1 00 S.1 = limti00 St/t.2.a..: S.Sn 0<t<x„+j ( 1 . We shall use the estimate Po(M > u) Miu^+ 1 < a) = o(Po (M > u)) (4. Under suitable conditions . Po(M* > u) .. Let a > 0 be fixed. 0 yields (4. M^xu)+l > a) . . MW O(u)+1 < a) IN ( U n=1 A1.e)Po (MMX> > x).4.1 can be rewritten as 00 (U) (4. Then by Lemma 3.a... u)} < P(M* E (u . u))/P(M* = 0) = o(Po(M* > u)). )) > (1 . To this end.Po (M* > u..6) 1 p pBo(u) u where B is the Palm distribution of claims and p . 2. assume the path structure Nt St = EUit+Zt i=1 . Letting first u + oo and next e . MODELS WITH DEPENDENT INPUT Define 79* (u) = inf {n = 1 .e)Po (M > U).( u)1 > a) 00 1: Po(Mn<u..(u) . x > a.+Mn+1>u} 267 (note that {M> u} = {3(u) < oo}). /3(u) = inf{n=1.S. Theorem 4.E) Po ( n max St u. choose a such that Po(Si > x ) > (1 .4. Given e > 0. S.e)Po (M > u.Sn+1Sn>aV(uSn*)) n=1 00 > (1E)EPo(Mn<u. Mn+l > a V (u .
v.} and satisfying Zt/t N. a4' 0. HEAVY TAILS N` U.X both have tails of sup Zt. Then the Palm distribution of claims is B(x) = E N Eo 0 I( U1 < x) . i=1 (4.1 is in force.268 with {Zt} continuous.2 Assume that {St} is regenerative and satisfies (4.3PB.8) x Write . cf. Mix) < > UE + i=1 o<t<x Thus Theorem 4. since the tail of Zx is lighter than B(x) by (iv). are Fmeasurable and NX Po J:U=>x i=1 (iv) Po sup Zt > x / (0:5t<x o(B(x)) Then (4.6 below. X and N. Assume further that (i) both B and Bo are subexponential.7).4). (iii) For some o field Y. and also for Mix) since Nx FNX U. we get 00 (u) 1 IPF. The same is true for Sl. Corollary 4. the proof of Lemma 4.'s order EoNx • B(x). Proof It is easily seen that the r.6) holds with p = . and ENX Ui . (ii) EozNX < oo for some z > 1.p) Ju P Bo(u) 1p 0 . and the rest is just rewriting of constants: since p = 1+tlim St = 1+ .I u J Po(Sl > x) dx 1 EoNxB(x) dx EoX(1 .Q = EoNx/EoX. independent of {> CHAPTER IX. oX (see Proposition A1.
More precisely. The number N. X2 = 1. 1) is Poisson with rate /3 = fo /3(s) ds so that (ii) holds. X2 = 1. Example 4 . (iii) is obvious.5 Consider the Markovmodulated risk model with claim size distributions satisfying (4.t} is standard compound Poisson and {Zt} an independent Brownian motion with mean zero and variance constant a2.6) holds. Taking again Xo = Xi = 0.6) holds. Bo E S.t + EN'I Ui where {>N`1 Ui . i. (i) holds. Zt . .3 As a first quick application.. in particular lighttailed. Bo E S. . we conclude just as in Example 4.. Thus we conclude that (4. We now return to the Markovmodulated risk model of Chapter VI with background Markov process {Jt} with p < oo states and stationary distribution 7r. The regenerative assumption is satisfied if we take Xo = Xi = 0. > 0.4 Assume that St = Zt . and for some constants ci < oo such that cl + • • • + c. X3 = 2. Assume that B E S. The arrival rate is /3i and the claim size distribution Bi when Jt = i. of claims arriving in [0. We consider the case where one or more of the claim size distributions Bi are heavytailed. .e.. Again .0 (thus (iv) is trivial).6) u holds..9). we assume that B E S. consider the periodic model of VI. X3 = 2.4. i=1 B = >2 7riaiBi i=1 and we assume p = 014 B = Ep ri/3ipB. 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 .3 that (4. The key step of the proof is the following lemma.6 with arrival rate /3(t) at time t (periodic with period 1) and claims with distribution B (independent of the time at which they arrive).. Then (4. Theorem 4. 3 The average arrival rate / and the Palm distribution B of the claim sizes are given by P P Q = ir i/i.(NX). In particular. then (iv) holds since the distribution of supo<t<i Z(t) is the same as that of I Zl 1. . < 1. note that the asymptotics of i/io( u) is the same irrespective of whether the Brownian term Zt u in St is present or not. and taking F = o. MODELS WITH DEPENDENT INPUT 269 Example 4 .
. X > 0 a r. Then P P(Yx > x) . 6 Let (N1.F) < CG(x)zn'1+. NP ) and X are . oo) and define p Ni Yx = EEX'i . and that for some + cp distribution G on [0. Markovmodulation typically decreases the adjustment coefficient y and thereby changes the order of magnitude of the ruin . as x a oo.Fmeasurable.270 CHAPTER IX.2.. 1."+Np .ciG(x). u Proof of Theorem 4. i=1 Proof Consider first the case X = 0. HEAVY TAILS Lemma 4 . cp with cl + > 0 it holds that Fi(x) .... NP ) be a random vector in {0.v.X i=1 j=1 where conditionally upon F the Xi. i=1 P(Yx > x ^) < P(Y0 > x I. . The same dominated convergence argument completes the proof. It follows by a slight extension of results from Section 1 that P P(Yo > x I Y) G( x) ci Ni. i =1 P(Yo > x I ^ ) < CG(x)zN1+ +Np for some C = C(z) < oo. and the rest of the argument is then just as the proof of Corollary 4. are independent with distribution Fi for Xij. . If Jo = i. Let {Fi}t=1 P be a family of distributions on [0.5. i1 = E\ G(x) In the general case. 2 . .}P. oo) such that G E S and some c1.F) = P(Yo > X+x I •^) G (x +x)>2ciNi i=1 . and F a aalgebra such that (N1.c'(x) where c = ciENi .. Thus dominated convergence yields ( P(Yo>x P(Yo>x . . .G( x ) > ciNi .. An easy conditioning argument then yields the result when Jo is u random.. P P P(YX and > x I. For lighttailed distributions.^•) G(x) P ^ E ciNi = C. . we can define the regenerations points as the times of returns to i. Assume EzN1+"'+Np < oo for some z > 1 and all i..
> 0) matter for determining the order of magnitude of the ruin probabilities in the heavytailed case. this then easily yields approximations for the finite horizon ruin probabilities (Corollary 5..1.5 that the effect of Markovmodulation is in some sense less dramatical for heavytailed distributions: the order of magnitude of the ruin probabilities remains ft°° B(x) dx. for lighttailed distributions the value of the adjustment coefficient y is given by a delicate interaction between all B. and the final reduction by Jelenkovic & Lazar [213].7).2 and Example 4. Theorem 4.p)Bo(u).. i. 5 Finitehorizon ruin probabilities We consider the compound Poisson model with p = /3pB < 1 and the stationary excess distribution Bo subexponential. Floe Henriksen & Kliippelberg [31] by a lengthy argument which did not provide the constant in front of Bo(u) in final form. 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)..5. Theorem 2.4. the discussion provides an alternative point of view to some results in Chapter IV. this is applied for example to risk processes with Poisson cluster arrivals. ).d.1 is from Asmussen.4. in particular Proposition 2. Schmidli & Schmidt [47]. As usual. we let PN"N = P(. I T(u) < oo).. ) form a general stationary sequence and the U.i.T2. IV. That paper also contains further criteria for regenerative input (in particular also a treatment of the delayed case which we have omitted here). It follows from Theorem 4. Then O(u) . FINITEHORIZON RUIN PROBABILITIES 271 probabilities for large u. this should be compared with the normal limit for the lighttailed case.5 shows that basically only the tail dominant claim size distributions (those with c. In contrast. i.4.pl(1 . m is a (orfinite) . r(u) is the time of ruin and as in IV.. Notes and references Theorem 4.e. states that under mild additional conditions. Within the class of risk processes in a Markovian environment.. as well as a condition for (4.T2. cf. We start by reviewing some general facts which are fundamental for the analysis.3.6) to hold in a situation where the interclaim times (T1. Essentially. cf. VI. cf.4.7. 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. For further studies of perturbations like in Corollary 4. The main result of this section. 5 was first proved by Asmussen. The present approach via Theorem 4. An improvement was given in Asmussen & Hojgaard [33]. and independent of (T1. see Schlegel [316]. Combined with the approximation for O(u). Theorem 5.
the whole of R and not as usual impose the restrictions x > 0. We let QS be the corresponding distribution and Qx.272 CHAPTER IX. follows by the substitution y = x .y = Qx (. say. y to vary in. resp. u For F C E. Proof Starting from Ro = x.2) means ffh(a.h.= y. to consider only the case Px(w(F`) = 0) 0. j.2) for all bounded measurable functions h. .s. Then (5. The simplest example is a discrete time discrete state space chain.z) dx G(dz) = ffh(y + z) k(y)dy G(dz).1 A compound Poisson risk process {Rt} and its associated claim surplus process {St} are in classical duality w . The equality of the l.t. a familiar case is time reversion (here m is the stationary distribution). a main difficulty is to make sense to such excursions also when Px(w(F°) = 0) = 1.t.s. St is distributed as y . t. w(Fc) < oo ) 'In general Markov process theory. for states i. Let G denote the distribution of ENt U.00). Say {St} is reflected Brownian motion on [0. (note that we allow x. to the r. Rt is distributed as x + t .)k(x . x = 0+ and F = (0. oo). HEAVY TAILS measure on E such that L for all measurable A C E and all t > 0. and starting from So = y. but the example of relevance for us is the following: Proposition 5. in the terminology of general Markov process theory. 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} .rij = mjsji where r13.). Sw(F. Thus. . m.2) with t = 1 means m. r. Lebesgue measure. and (5. y = 0).r.h.s=j are the transition probabilities for {St}. where we can take h. For the present purposes it suffices . k on E.t + EI U. Then there is a Markov process {Rt} on E such that fE m(dx)h(x)Exk(Rt) = Lm(dy)k(y)Eyh(St) (5. k as indicator functions.z.>N` Ui. however . {St} and {Rt} are in classical duality w. {Rt}.
We can then view Qy. in with i0. .. 0]. i1. QR and QRy are defined similarly.1 for the case F = (oo... and we let Qy y refer to the time reversed excursion .2 Qy. Thus. /^s x (S1 = Z1. That is. . Qx y is the distribution of an excursion of {St} conditioned to start in x E F and terminate in y E F. in = y. The theorem is illustrated in Fig . x = 0. this simply means the distribution of the path of {Rt} starting from y and stopped when 0 is hit. Sn = in = y. in E F.y = Qy Q.5.y() = P ({SW(F`)t} 0<t<w(F °) E So = x. . w(0.itt) = P Px(w(Fc) < 00. Sn+1 E Fc) nx. 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). In particular: Corollary 5. But in the risk theory example (corresponding to which the sample paths are drawn). Qx.SS(F.y(2p21 .. z > 0.1 The sample path in (a) is the excursion of {St} conditioned to start in x = 0 and to end in y > 0. when p = . Sw(F)1 = y) .).(0)_ = y < 0 is the same as the distribution of w(y) where w(z) = inf It > 0 : Rt = z}. the one in (b) is the time reversed path. [note that w(z) < oo a. oo) = r(0) x= St y (a) Figure 5.s.3 The distribution of r(0) given r(0) < oo. FINITEHORIZON RUIN PROBABILITIES 273 y E F (in discrete time.y as a measure on all strings of the form i0i1 . 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. io = x.. 5. ... Sw(Fo)_ should be interpreted as Sw(F^)_1).2. S.13AB < 1] Proof of Theorem 5..= y) Theorem 5 .
.. in with 20.274 note that Fx(w(Fc) < 00. note first that Pt' (R l = il..y(inin _ E SYj jEF` 00 Sxik _1 .. 2n) = Qx. t' y and Qy x are measures on all strings of the form ipi l .TI( 2n2n _1 ...... (Fc)1 = y) 00 CHAPTER IX. Si l io E mjSjx.in1 ...ik_1EF . Si1y 00 jEF° E E 5xik_ 1 . To show Q y x (i0 i 1 .. S. .. R Qy x(2p21 . .. in) = Qx. Sn+1 E Fc) n=1 i1. ..ik_1EF Sxin_1 .. .rin_1in E Txj jEFC m21 s2120 m2252221 m in Ssn n1 mjSjx Mx m2p mil min1 jEF` 1 Sinin _ 1 ..ii .. R ... ... . in = x.(F<)1 = Y) S S and Qx y( ipil .. Rn = in = x... Rn+1 E FC) TioilTili2....... = in = y.i„_iEF Similarly. Rn = in = x...ik1EF Similarly but easier Sxin_1 .. .... in E F.gilt' k=1 ii . in = x. 2p).. Si1y k=1 i1 . 20 = y. S.in E F. i0) Q x... Si11 S 1 .. Silt' E Sxik_1 .... 21 ..J (i. MY Thus Qx(ioii . i0 = y.. in)  Pt' (R1 = ii. Rn+1 E F`) F (w(Fc) < 00..... HEAVY TAILS E E Px (Si = 21i ...... 2p) when 20. Silt' E SO k=1 i1. . .... in) = oo jEF^ Sxin1 ..
P(") = P(. Y > y} . S. the P(u.2. 7(0) < oo. that is.2. U T(O) = T (u) Y Figure 5. the distribution w. Now the P(u. ST(o) > y. Z = Zl = ST+( 1)_ the value just before the first ladder epoch (these r.2 The distribution of (Y. 5. y > u. Let Y = Yl = Sr+( 1) be the value of the claim surplus process just after the first ladder epoch .B(a) +a PBBo(u) . FINITEHORIZON RUIN PROBABILITIES 275 5b The time to ruin Our approach to the study of the asymptotic distribution of the ruin time is to decompose the path of { St} in ladder segments . the case r (O) < oo. To clarify the ideas we first consider the case where ruin occurs already in the first ladder segment . see Fig.'s are defined w. P(o) ).t. Bo") is also the P(u. Y > u).')distribution of Yu is Bo"). That is.r.UBBo(u)]. We are interested in the conditional distribution of T(u) = T(0) given {T(0) < oo. Z follows the excess distribution B(Y) given by B(Y) (x) _ B(y + x)/B(y).t. Z) is described in Theorem 111. 1 w . The formulation relevant for the present purposes states that Y has distribution Bo and that conditionally upon Y = y. that is.r.(o) > y} = {T(0) < oo.')density of Y is B(y)/[.p.2.')distribution of Z since P(Z>aIY>u) = 1 °° B(y) B(y + a) dy FLBBo(u) B (y) J°° (z) dy .v.5.
. Yn_1 'typical'. it therefore follows that T(u)/Z converges in Pi"'')probability to 1/(1 . > u with high probability. Z). more precisely. It is straightforward that under the conditions of Proposition 1.p).276 CHAPTER IX.i. We now turn to the general case and will see that this conclusion also is true in P(")distribution: Theorem 5 .1. . let r+(1) = T(0). Z/'y(u) * W in Pi "' ')distribution .18(c) Bo")(yY (u)) + P(W > y) ( 5.. However. 2. the duration T+ (n) . Then 7(u)/y(u) ^ W/(1 . conditionally upon r+ (n) < oo.. a slight rewriting may be more appealing. (Y.o be defined by w(z) = inf It > 0: Rt = z} where {Rt} is is independent of {St}. i.. are i. .: r+ (n) < oo. then by the subexponential property Yn must be large. Since w(z)/z a$. HEAVY TAILS Let {w(z)}Z^... and distributed as (Y.. 4 Assume that Bo E S and that (5. Then Corollary 5. In the proof.p) in Pi"'')distribution. and YI. Zn_1 'typical' which implies that the first n1 ladder segment must be short and the last long. denote the ladder epochs and let Yk.3. Hence Z.. r(u)/Z 4 1/(1 . this in principle determines the asymptotic behaviour of r(u)..d.T+(2). must be large and Z1.r+ (n .. in particular of Z.. .. Y1 + • • • + Yn > u} denote the number of ladder steps leading to ruin and P("'n) = P(• I r(u) < oo. we get the same asymptotics as when n = 1.e.3) holds.p) in F(u) distribution. . cf.e. That is . Z = ZI but relative to the kth ladder segment. i.1) of the last ladder segment can be estimated by the same approach as we used above when n = 1. 5. Zn). . 1/(1 . the random vectors (YI.p) then yields the final result T(u)/y(u) + W/(1 . z ^ oo. K(u) = n). Recall the definition of the auxiliary function y(x) in Section 1.. The idea is now to observe that if K(u) = n. Z1).p). Zk be defined similarly as Y = Y1... . Fig. Then..3 implies that the P("'1)distribution of T(u) = r(0) is that of w(Z). P(Z < a I Y > u) 3 0. Bo") ). Since the conditional distribution of Z is known (viz. We let K(u) = inf In = 1. Now Bo E S implies that the Bo ")(a) + 0 for any fixed a.3) where the distribution of W is Pareto with mean one in case ( a) and exponential with mean one in case (b). and since its dominates the first n .
A"(u) are events such that P(A'(u) AA"(u)) = o(F (A'(u)) (A = symmetrical difference of events). I A'(u)) = P(u.2.. Proof We shall use the easily proved fact that if A'(u). Y„1. A"(u) _ {K(u)=n} = {Y1+ P(. I A"(u ))II + 0..5.5 Ilp(u. II ' II denotes the total variation norm between probability measures and ® product measure.Yn1iYn . .. .Bo (ri1) ®B( . suitably adapted). . P(.. then IIP( I A'(u)) Taking A'(u) = {Y. Further.Yl+ +Yf1>u}.3 In the following. +Yn1<u.n).. FINITEHORIZON RUIN PROBABILITIES 277 16 Z3 Z1 r+(1) T+(1) T+(1) Figure 5.. the condition on A'(u) A A"(u) follows from Bo being subexponential (Proposition 1.. > u}.u) E • I A'(u)) = Bo (n1) ®Bou) .n) (y1.u) II 0. .u) E •) . P (Yj. Lemma 5. Yn .
wk(Zk) has a proper limit distribution as u + oo for k < n. .1). then 11P(Z E •) . .. in particular his Proposition (2. see Fitzsimmons [144]).1 and Y„ .6 IIPIu'n ) CHAPTER IX. + Y" > u) Flul (K (u ) = n) _ Cu) P"F(1'i +. .. Let {wl(z)}. n .. HEAVY TAILS ((Z1'. Thus F(u'n)(T(u) /7(u) > y) = F(u'n)((wl (Z1) + .).y(u)T) ..u has distribution Bout That is..d.. the density of Yn is B(y)/[IBBO(u)]. the discussion just before the statement of Theorem 5.. Proof of Theorem 5. Z' are arbitrary random vectors. (Y. copies of {w(z)}. .i.P(Z' E •)II > 0 (here Y. and that Yk has marginal distribution B0 for k = 1.. Z11). ..4. whereas wn(Zn) has the same limit behaviour as when n = 1 (cf.6.278 Lemma 5 .Bo (n1) ®Bo' 0.t. +wn(Z n))l7( u ) > 1y) ^' P(u'n)(wn (Zn)/7(u) > y) 4 NW/(1 .. Notes and references Excursion theory for general Markov processes is a fairly abstract and advanced topic..2. . By Lemma 5.+y 1 p"F(Yn > u) P)Pn1 P/(1 . k = 1.4). y > u. Then according to Section 5a. . {wn(z)} be i. Zn) E •) . The first step is to observe that K(u) has a proper limit distribution w.1. .... Now use that if the conditional distribution of Z' given Y' is the same as the conditional distribution of Z given Y and JIF(Y E •) .. be independent random vectors such that the conditional distribution of Zk given Y. Proof Let (Y11.. The same calculation as given above when n = 1 shows then that the marginal distribution of Zn is Bou).. and clearly Zi. the F'distribution of r(u) is the same as the P'distribution of w1(Zl) + • • • + wn(Zn). P(u) since by Theorem 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 .. . For Theorem 5.p) < y). Y1 +. Z.P) > y) Corollary 5.P) Bo(u) for n = 1. the marginal distribution of Zk is Bo for k < n. Similarly (replace u by 0). in our example Y = (Y1....r.P(Y' E •)II * 0...' = y is BM. . n_1 < u. Zn are independent. 2.. Y") u etc. .. Y'.. n.7 O (u.. Zn).
We will show that the stationary density f (x) of {Vt} satisfies f (x) /B(x) r(x) We then get V. RESERVEDEPENDENT PREMIUMS 279 The results of Section 5b are from Asmussen & Kluppelberg [36] who also treated the renewal model and gave a sharp total variation limit result . claim size distribution B. Theorem 6 . The heuristic motivation is the usual in the heavytailed area.B(u). cf. and define the cycle as a = inf{t>0: Vt=0. Then 0 (u) Qf "O ^) dy.(3 u u J B(y) dy . . the results only cover the regularly varying case.y) .1) The key step in the proof is the following lemma on the cycle maximum of the associated storage process {Vt}. Then P(MT > u) . V. that MQ becomes large as consequence of one big jump. The rigorous proof is.6. Assume for simplicity that {Vt} regenerates in state 0 . Extensions to the Markovmodulated model of Chapter VI are in Asmussen & Hojgaard [33]. x > oo.2 Define M. 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). More precisely.(u) = P(V > u) = f f (y) dy . Corollary II. that fo p(x)1 dx < oo. p(Y) and the result follows. T) when T + oo with u fixed.1. Proof of Theorem 6.. the probability that is exceeds u is then B(u . however. = supo<t<0. and premium rate p(x) at level x of the reserve. 3.1 Assume that B is subexponential and that p(x) > 00.e./3Ea B(u). nontrivial and we refer to Asmussen [22]. i. one expects the level y form which the big jump occurs to be 0(1). Asmussen & Teugels [53] studied approximations of i (u. max VB>0I Vo=0^ o<s<t J11JJJ Lemma 6 .2. u (6. 6 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate /3.
Then D(u) = f(u)p(u) and. there exist constants c(u) 4 0 such that the limiting distribution of r(u)/c(u) given r(u) < oo is exponential. D(u) = DQ(u)/µ.q ( u)) 1 . u Notes and references The results are from Asmussen [22]. Hence f (u)r(u) = D(u) = Do(u) .q(u) Now just use that p(x) * oo implies q (x) + 0. HEAVY TAILS Define D(u) as the steadystate rate of downcrossings of {Vt} of level u and Da (u) as the expected number of downcrossings of level u during a cycle. by regenerative process theory. .280 CHAPTER IX. It is also shown in that paper that typically. 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).P(MT > u) $B(u) Ft µ(1 . where also the (easier) case of p(x) having a finite limit is treated .
d. 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.. The crude Monte Carlo ( CMC) method then amounts to simulating i. Rubinstein [310] or Rubinstein & Melamed [311] for more detail . where a2 = Var(Z ).2) is an asymptotic 95% confidence interval . ZN.z) 4 N(0. z) 2 = Zit NE ii ii According to standard central limit theory . We shall be brief concerning general aspects and refer to standard textbooks like Bratley. topics of direct relevance for the study of ruin probabilities are treated in more depth. Hence 1. vrN(z . 281 . 4Z)..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.. estimating z by the empirical mean (Z1 + • • + ZN)/N and the variance of Z by the empirical variance N s2 = E(Z{  N 2.i. replicates Zl. Ripley [304]. Fox & Schrage [77]. .
variance reduction is hardly worthwhile. conditional Monte Carlo and importance sampling. However. typically by modifying Z to an alternative estimator Z' with EZ' = EZ = z and (hopefully) Var(Z') < Var(Z). an added programming effort. Letting Z' = E[Z I Y]. Conditional Monte Carlo Let Z be a CMC estimator and Y some other r . The situation is more intricate for the infinite horizon ruin probability 0(u). and a longer CPU time to produce one replication. This is a classical area of the simulation literature. one can argue that unless Var(Z') is considerable smaller than Var(Z). We mention in particular ( regression adjusted) control variates and common random numbers. there are others which are widely used in other areas and potentially useful also for ruin probabilities. Therefore.282 CHAPTER X. 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). generated at the same time as Z. and many sophisticated ideas have been developed. We survey two methods which are used below to study ruin probabilities. we then have EZ = EZ = z. 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. 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'. it is straightforward to use the CMC method to simulate the finite horizon ruin probability z = i.b(u. Further. lb Variance reduction techniques The purpose of the techniques we study is to reduce the variance on a CMC estimator Z of z. Sections 24 deal with alternative representations of Vi(u) allowing to overcome this difficulty. so that Z' is a candidate for a Monte Carlo estimator of z. SIMULATION METHODOLOGY In the setting of ruin probabilities. Say that Var(Z') = Var(Z)/2. and in most cases this modest increase of N is totally unproblematic. v. writing Var(Z) = Var(E [Z I Y]) + E(Var[Z I Y]) . 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.
3). i.z2 = 0. This may also be difficult to assess .1.v. even if the optimal change of measure is not practical. However.96 sis v^ N 2 1 where srs = N j(LiZi . Variance reduction may or may not be obtained: it depends on the choice of the alternative measure P.zrs. To this end. it may often be impossible to describe P in such a way that it is straightforward to simulate from P). the argument cheats because we are simulating since z is not avaliable analytically.3) Thus. it gives a guidance: choose P such that dP/dP is as proportional to Z as possible. (1. (ZN. . and the problem is to make an efficient choice.E [Z Z]2 = z2 . Nevertheless. a crucial observation is that there is an optimal choice of P: define P by dP/dP = Z/EZ = Z/z.. 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. . GENERALITIES 283 and ignoring the last term shows that Var(Z') < Var(Z) so that conditional Monte Carlo always leads to variance reduction. LN) from P and uses the estimator N zrs = N > L:Zj i=1 and the confidence interval zrs f 1. but tentatively.e.[E(LZ)] = E Z2 Zz . L such that z = EZ = E[LZ]. Thus we cannot compute L = Z/z (further. Thus. Then z Var(LZ) = E(LZ)2 . using the CMC method one generates (Z1.zrs) = 2 1 N 2 2 2 i=1 i=1 N > Lt Zi . it appears that we have produced an estimator with variance zero. one would try to choose P to make large values of Z more likely. . the obvious possibility is to take F and P mutually equivalent and L = dP/dP as the likelihood ratio. In order to achieve (1. L = z/Z (the event {Z = 0} is not a concern because P(Z = 0) = 0). L1)..
e.z) 1 > 00. 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. The CMC method leads to a variance of oZ = z(1 .B = iP(AB) = P(BIA). To introduce these.e. u + oo.0. For each u. say 10%. as is the case of typical interest. and let Z(u) be a Monte Carlo estimator of z(u).96 2 z2 z increases like z1 as z . it does not help telling whether z is of the magnitude 104.96oz /(zV) = 0. say of the order 103 or less. SIMULATION METHODOLOGY 1c Rare events simulation The problem is to estimate z = P(A) when z is small .1. Two established efficiency criteria in rare events simulation are bounded relative error and logarithmic efficiency.z) 1001. in terms of the halfwidth of the confidence interval.5 or even much smaller . However. Again. I. In ruin probability theory.e. large sample sizes are required. This leads to the equation 1.96 2Z ( 1 . assume that the rare event A = A(u) depends on a parameter u (say A = {r(u) < oo}). let z(u) = P(A(u)). we may try to make P look as much like P(•IA) as possible. Thus. i. the optimal P is the conditional distribution given A. Z = I(A) and A is a rare event. Another way to illustrate the problem is in terms of the sample size N needed to acquire a given relative precision . a confidence interval of width 10 4 may look small.z) which tends to zero as z ^ 0. the issue is not so much that the precision is good as that relative precision is bad: oZ z(1 . An example where this works out nicely is given in Section 3. 10 . assume that the A(u) are rare in the sense that z(u) * 0. and further it is usually not practicable to simulate from P(•IA). We shall focuse on importance sampling as a potential (though not the only) way to overcome this problem. N .. if z is small. However.. A = {T(u) < T} or A = {r(u) < oo} and the rare events assumption amount to u being large. The optimal change of measure ( as discussed above) is given by P(B) = E [ Z] i.284 CHAPTER X.100 . z I. but if the point estimate z is of the order 105. Z z V5 In other words .1. We then .
2.1) may be written as V) (u) = P(M > u). and in practice. Notes and references For surveys on rare events simulation.e. Otherwise. O (u) = z = EZ.i. i. F(K = k) = (1 . but as a CMC method . 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. Therefore . which gives a logarithmically efficient estimator . logarithmic efficiency is almost as good as bounded relative error.log Var(Z(u)) lim inf > 2 u+oo . 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. 2.0. where X1. XK from the density bo(x). We shall here present an algorithm developed by Asmussen & Binswanger [ 271. let Z +.4) for any e > 0. where M = X1 + • • • + XK.(2. This allows Var(Z(u)) to decrease slightly slower than z(u)2. the PollaczeckKhinchine formula III.d. the mathematical definition puts certain restrictions on this growth rate. . Var(Z(u)) hm sup U+00 z (u) 2E < oo (1.. let Z +. X2. Generate K as geometric. this means that the sample size N = NE(u) required to obtain a given fixed relative precision (say a =10%) remains bounded. . The algorithm gives a solution to the infinite horizon problem . According to the above discussion.. so that NE (u) may go to infinity.p)pk. The term logarithmic comes from the equivalent form . . P(K = k) = (1 . 3. are i.1. it is not efficient for large u . it is appealing to combine with some variance reduction method . If M > u. see Asmussen & Rubinstein [45] and Heidelberger [190]. Generate X1. However.. where Z = I(M > u) may be generated as follows: 1.4).. Thus. with common density bo(x) = B(x)/µB and K is geometric with parameter p. 2 Simulation via the PollaczeckKhinchine formula For the compound Poisson model. Let M .X1 + + XK.p)pk.log z(u) of (1.
To see this....XK_1). and considering only the remaining ones.XK_1 and let Z( 1)(u) = Bo (Y) (if K = 0.286 CHAPTER X...SK1)2. note first that To check the formula for the P(X(n) > x I X(1). y < 0).b(u) = P (Xl +•••+XK>u) = EF[Xl + .p/(l . For the simulation. we generate only X1.. Xl > x.. So. asymptotically it presents no improvement : the variance is of the same order of magnitude F(x).. Then (cf.S( K_1)) V X(K1)) / Bo(X(K 1)) where S(K_l) = X(1) + X(2) + • • • + X(K_1). XK1.X(2)... However.... form the order statistics X(1) < X(2) < . . Z(1) (u) has a smaller variance than Zl (x). XK. Theorem IX..X1 .. The idea of [27] is to avoid this problem by discarding the largest X.p)Bo(x).. X1 + + XK_ 1 > x when X1 > x.. Thus.. SIMULATION METHODOLOGY when the claim size distribution B (and hence Bo) has a regularly varying tail. As a conditional Monte Carlo estimator . Z(1)(u) is defined as 0). 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. and that Bo(y) = 1. .L(x)/x`' with a > 0 and L(x) slowly varying.2. . and let Z(2)(u) = _ P (SK B0((u > u I X(l). compute Y = u .Xl . ..X(2). just note that EZ(1)(u ) 2 > E[Bo (x . we thus generate K and X1i ..X(K1)) .. A first obvious idea is to use conditional Monte Carlo: write i.+XK > uIXl...X(n_1)) Bo(X(„_l) V X) Bo(X(n1)) .1) V)(u) . K > 2] = P2p(Xl > x) = P2Bo(x) (here we used that by positivity of the X. assume in the following that Bo(x) ..XK1] = EBo(uX1 . < X(K) throw away the largest one X(K). conditional probability... This calculation shows that the reason that this algorithm does not work well is that the probability of one single Xi to become large is too big...
1) .5). and that paper contains one more logarithmically efficient algorithm for the compound Poisson model using the Pollaczeck.modulated model P(r+ < oo) and G+ are not explicit ). 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. Asmussen .. . l)) BO(X(n1)) Theorem 2 . Compute y > 0 as solution of the Lundberg equation 0 = K(y) = )3(B[y] . Thus. using 13L. Binswanger and HOjgaard of [28] give a general survey of rare events simulation for heavy tailed distributions . and define )3L.S(n_1) I X(1). However . . l)) . for the purpose of recording Z(u) = erysr(u). BL(dx) = e7sB(dx)/B[y].1 is elementary but lengty... and simulate from FL. 3 Importance sampling via Lundberg conjugation We consider again the compound Poisson model and assume the conditions of Ce7". X(n1)) P(X(TZ) + S(.3.y. X (.Khinchine formula and importance sampling .u is the representation 0(u) = e7sr(u) overshoot (cf._1) > P(X(n) > _ X X(1). the continuoustime process {St} is simulated by considering it at the discrete epochs {Qk} corresponding to claim arrivals. 1 Assume that Bo (x) = L(x)/x° with L(x) slowly varying. B. Notes and references The proof of Theorem 2. . Also in other respects the findings of [28] are quite negative: the large deviations ideas which are the main approach to rare events simulation in the lighttailed case do not seem to work for heavy tails.. and we refer to [27]. X(2).S (n1)) V X (. Then the algorithm given by { Z (2) (u) } is logarithmically efficient. BL by I3L = /3B[y]. The algorithm is sofar the only efficient one which has been developed for the heavytailed case. X(2). X(n1)) Bo((x . that is. the algorithm for generating Z = Z(u) is: 1.. . IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 287 We then get P(S" > x I X( 1). 111. X .. . BL instead of 0. For practical purposes. use the the CramerLundberg approximation so that z(u) = '(u) = e7"ELe7E(") where ^(u) = ST(") . X(2). in the renewal or Markov. .
to deal with the infinite horizon problem . cf. . 4. let S.2ryu _ z (u)2/C2.l3 and U from B. The proof is given below as a corollary to Theorem 3. Generate T as being exponential with parameter . In detail . Let S . namely ELery£(").2 The estimator (3. BL).F.(u)) are asymptotically coincide on {r(u)} < oo.r(u) < oo) = 1. Let Sf0 CHAPTER X. Xi = U. More precisely.. and avoid simulating the known part e7". with distribution F. b different from . SIMULATION METHODOLOGY 3. X2. + X. We may expect a small variance since we have used our knowledge of the form of 0(u) to isolate what is really unknown. and we have: Theorem 3. the results of IV. B) is not logarithmically efficient when (/3. It resolves the infinite horizon problem since FL(. Proof Just note that EZ(u)2 < e .T. = X1 + .d. If S > u.3.. one must restrict attention to the case 4µB > 1.... A > AL as in Chapter V. so that changing the measure to FL is close to the optimal scheme for importance sampling .7 tell that P(.. and the change of measure F r FL corresponds to B > BL.1) (simulated with parameters ^3. The algorithm generalizes easily to the renewal model . M(u) = inf {n : S„ > u}. b) # (/3L. The answer is no. the discussion at the end of Section 1b. . BL could improve the variance of the estimator . return to 3. and assume that µF < 0 and that F[y] = 1. We formulate this in a slightly more general random walk setting '. P'[y] < oo for some ry > 0. Otherwise.Q.i.1 The estimator Z(u) = e'rs* "u) (simulated from FL) has bounded relative error.S+U . Let FL (dx) = 'For the renewal model. 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. In fact: Theorem 3. r(u) < oo) and FL (both measures restricted to.QL. let Z F e_'s. There are various intuitive reasons that this should be a good algorithm. be i.. u It is tempting to ask whether choosing importance sampling parameters .288 2. Ti.
where e' = EL Iog dFL (Xi) > 0 by the information inequality.3 The estimator (3.2'X1 .. More generally. When F # FL.2) dF Theorem 3. = c'. Proof The first statement is proved exactly as Theorem 3 . 1. K2. + KM(u))} = exp {ELM(u)(E .yu+elu u +oo etry' 1 > lim up C2e2. Since K1... By the chain rule for RadonNikodym derivatives. write W(F IF) _ F(XI). For the second.+KM(u)}. let F be an importance sampling distribution equivalent to F and M(u) dF Z(u) _ I (Xi) . are i.2ryELXi)} . IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 289 e7yF(dx). Jensen's inequality and Wald's identity yield EpZ(u)2 > exp {EL(K1 + . (3.2 > 0.yu = G.P = FL. The importance sampling estimator is then Z( u) = e'rSM( ). F(XM(u)).3. EFZ(u)2 EFZ(u)2 lim sup z(u)2eeU = lim cop C2e2.2ryELXi. is not logarithmically efficient. Here ELK.. it thus follows that for 0 < e < e'/ELXi. .2) (simulated with distribution F of the X3 has bounded relative error when .. Since ELM(u)/u + 1/ELXi. EFZ(u)2 = EeW2(FIF) = Ep [W2(FIFL)W2(FLIF)] = EL [W2 ( FIFL)w(FLIF)] = ELexp {Kl+.d. .... where Kl og (X) (j) 2 ) = log dFL (Xi) .i..
i. In fact: Proposition 4. 4 Importance sampling for the finite horizon case The problem is to produce efficient simulation estimators for '0 (u.'(y) or y > 1/r.T". we write T = yu.T" has a left exponential tail with rate /3'. U' . B" and generic claim sizes U'. Consider compound Poisson risk process with intensities /3'.3"eQ x 0 J eQ zB (z) dz x (x > 0) and /3' = /3".T" > x) J /3"e0 yB (x + y) dy = .'(y).T' has a left exponential tail with rate /3' and U" .T' = U" . U".4.g. T". all that needs to be shown is that if U' . In [13].T".e.290 which completes the proof.B'=B". with the present (shorter and more elementary) proof taken from Asmussen & Rubinstein [45]. . Next.T' D U" . /3'eQ'YR'( x + y) dy = . then /3' B' = B". generic interarrival times T' . The easy case is y > 1/k'(y) where O(u. U' . then the estimator Z(u) = e7Sr(°)I(r(u) < yu) (simulated with parameters /3L. As in IV.1 is from Lehtonen & Nyrhinen [244].3'eO'x f f P (U" . optimality is discussed in a heavy traffic limit y 10 rather than when u + oo. The queueing literature on related algorithms is extensive . the references in Asmussen & Rubinstein [45] and Heidelberger [190].4 indicate that we can expect a major difference according to whether y < 1/r. Then according to Theorem 3. The extension to the Markovian environment model is straightforward and was suggested in Asmussen [ 16].2. we conclude by differentiation that Bo(x)=B' (x)forallx > 0. The optimality result Theorem 3. CHAPTER X. T) with T < oo. 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 . yu) is close to zk(u). so that one would expect the change of measure F 4 FL to produce close to optimal results.3. SIMULATION METHODOLOGY u Proof of Theorem 3. from 3' P(U'T'>x) ^ = ^ eQ'zB (z) dz. Further discussion is in Lehtonen & Nyrhinen [245]. see e. This immediately yields / = 3".1 If y > 1/ic'('y). /3". BL) has bounded relative error. The results of IV. First by the memoryless distribution of the exponential distribution . claim size distributions B'.
. that ryy = ay . We recall that ay is defined as the solution of a'(a) = 1/y. lim inf uoo 27yu .2) Since the definition of ay is equivalent to Eay r(u) .4. T( u) < yu] e2ryyuEay le 2ay^(u).1) so that z(u) = zP(u.3) and we have: Theorem 4.to g x ( u ) u u so that (1.1).4.8). We next consider the case y < 1/r. (4. The corresponding estimator is Z(u) = eavS' ( u)+T(u)K (ay)I(T( u) < yu). and in fact. the result follows as in the proof of Theorem 3.log Var(Z(u)) l im of . Bounding u ELZ(u)2 above by a7u.4.3) (simulated with parameters /gay. yu) = eayu Eay Leay^(u)+r(u)K(ay). Bay) is logarithmically efficient. Remark 4 . Proof Since ryy > y. Let Qy2 = . one would expect that the change of measure F Pay is in some sense optimal. yu)/z. T(u) < yu] .(u) * 1 (Theorem IV. 3 Theorem IV. we have ic(ay ) > 0 and get Eay Z(u)2 = Eay [e  2aySr( u)+2r(u )r. and that ryy > ry.2 The estimator (4.4. yu) in the sense that .1) which is all that is needed here can be showed much easier .8 has a stronger conclusion than (4.yk(ay) determines the order of magnitude of z'(u.yu.log 4')u) 4 u (Theorem IV.'(7).1.O(u.5) follows. IMPORTANCE SAMPLING FOR THE FINITE HORIZON CASE 291 Proof The assumption y > 1/n'(y) ensures that 1fi(u. yu) is of order of magnitude a71. T(u) < yu] e Hence by (4. 7y (4. (4.1).yy> 2 . Further .(ay).log Var(Z(u)) _ .
(av)Eav l e. there exists a dual process { V t} such that i.1) where the identity for Vi(u) requires that Vt has a limit in distribution V.u1/2 < r(u) < yu Le ] l = e..2) . the object of interest is {Vt} rather than {Rt}.o . related discussion is given in a heavy traffic limit q J.a vt(u).4. we believe that there are examples also in risk theory where (5. (5.3: for many risk processes {Rt }.1) is used to study Voo by simulating {Rt} (for example.T) = P O<t<T inf Rt < 0 = P(VT > u).7y x(u) > hm inf u+Oo U . In most of the simulation literature (say in queueing applications).4).1): then by Proposition A1. Hence lira inf log ryyu + vyu 1/2 tc(ay) . Then z(u) = Eay Z(u) > Eay avS'(u)+T( u)k(av 1 ). 5 Regenerative simulation Our starting point is the duality representations in 11.1) (see Proposition IV. and (5.4. In Asmussen [13].a yu +l/ur' (av)Ei`av reav^(u)+(T(++)(U) yu . yu . zi(u) = INV. 0 rather than when u 3 oo.uaoo U That lim sup < follows similarly but easier as when estimating En. SIMULATION METHODOLOGY Vara„ (r(u))/u so that (T(u) .b(u. One main example is {Vt} being regenerative (see A. yu . the algorithm in Section 3 produces simulation estimates for the tail P(W > u) of the GI/G/1 waiting time W).3.1) may be useful.o.Qyu1/2 < T(u) C yu e. > u) = E f I(VV > u) dt 0 (5.yu)/(uyu1/2) .ryyu +oy u1/2K'(av)Eo l v 1/2) where the last step follows by Stam's lemma (Proposition IV. However.yu1/2 <1 T(u) < yu l r > e7vu +avul/ 2r. Z (u)2 above. '%(u) = P I info Rt < 0) = P(VV > u). 0 Notes and references The results of the present section are new. N(0.2).292 CHAPTER X.
Z2 a4* z2.d. Z1 = (Zl1i +. )).. Z2 .+Z(N) z 1.... The method of regenerative simulation. i. Therefore . EZ2'i = z2 = E Thus. > u) (and more general expectations Eg(V. oh) for h : R2 ^ R and Ch = VhEVh.. z2)) > N(O. Thus. 02) (5. and Z2'>) where Zi'i = w.. To derive confidence intervals . + Z1N>) . Z2'> the time during the cycle where { Vt} exceeds u and zj = EZJ'). which we survey below . REGENERATIVE SIMULATION 293 where w is the generic cycle for {Vt}. Zl the LLN yields Z1 a$' Z(1) +. . Z2 = N (X21' + . Vh = (8h/8z1 8h/ 8z2).. Then (Z1z1i Z2z2 ) 4 N2(0. let E denote the 2 x 2 covariance matrix of Z('). z2) z2/z1 yields Vh = (z2/z2 1/zl). consider first the case of independent cycles . . Simulate a zerodelayed version of {V t } until a large number N of cycles have been completed. letting J0 'o I (Vt > u) dt .. Z(N) are i . For the ith cycle.h (zl. i (^(u) . the regenerative estimator z%(u) is consistent. Taking h(zl.. is the cycle length.3) . (u) ?2 = E fo I(Vt > u) dt = 0( u ) zl Ew as N > oo. EZ1'i = z1 = Ew.t(u)) 4 N(0.5. Thus the method provides one answer on to how to avoid to simulate { Rt} for an infinitely long time period. j = 1. Then Z(1)...E). a standard transformation technique (sometimes called the delta method) yields 1 V 2 (h (Zi. provides estimates for F ( V. For details . 2. record Zi'i = (Z1'). + Z2N)) .
consider an extremely simple example . Notes and references The literature on regenerative simulation is extensive.294 where 01 2 CHAPTER X.v.z^ i=1 so a2 can be estimated by 2 2 = 72 S11+ 12 S22 .g S12 (5. () depending on C. Z of the form Z = ^p(X) where X is a r . In 111. Let X have a density f (x. say risk processes with a complicated structure of the point process of claim arrivals and heavy tailed claims . 9. in some situations it may be the only one resolving the infinite horizon problem . 6 Sensitivity analysis We return to the problem of 111 .9. There is potential also for combining with some variance reduction method. However . () dx f Ax) (dl d()f (x' () f ( z. () dx = E[SZ] f(X. with distribution depending on a parameter (. asymptotic estimates were derived using the renewal equation for z /i(u).g. Before going into the complications of ruin probabilities . We here consider simulation algorithms which have the potential of applying to substantially more complex situations. v. Then z(() = f cp(x) f (x.Z) ^Z(=) .C)dx = f w(x) d( f ( x.2 E1 2 z1 z1 Z2 The natural estimator for E is the empirical covariance matrix N S = N 1 12 (ZW . The regenerative method is not likely to be efficient for large u but rather a brute force one. to evaluate the sensitivity z/i( (u ) = (d/d() 0(u) where ( is some parameter governing the risk process .96s/v"N. Rubinstein [310] and Rubinstein & Melamed [311]. () dx so that differentiation yields zS d( fco(x)f(x.2. the expectation z = EZ of a single r.5) Z1 Z1 Z1 and the 95% confidence interval is z1 (u) ± 1. SIMULATION METHODOLOGY 2 Eli = Z2 z1 + 2 E22 . see e. Here are the ideas of the two main appfoaches in today 's simulation literature: The score function ( SF) method .0 .
So assume that a r. 11 /3oeOoT. For the SF method. The likelihood ratio up to r(u) for two Poisson processes with rates /3. ()).6. () can be generated as h(U. C)). zc = E [d co(h(U. IPA will estimate zS by 0 which is obviously not correct. the Poisson rate /3 in the compound Poisson model. ()) is 0 for C < Co and 1 for C > Co so that the sample path derivative cp'(h(U. SZ is an unbiased Monte Carlo estimator of z(. Thus .r. /3 is 0.log U/(. cp' (h(U. () = . one can take h (U. The following example demonstrates how the SF method handles this situation. r(u) = Tl + • • • +TM(u)). nonpathological examples where sample path derivatives fail to produce estimators with the correct expectation. this phenomenon is particularly unpleasant since indicators occur widely in the CMC estimators .t. C) f(X. just take cp as an indicator function . = E [`d (h(U. cp(h(U. () Thus. () = (8/8()h (u. For example . C).r. giving h( (U. for some Co = (o(U).() d( is the score function familiar from statistics . Let M(u) be the number of claims up to the time r(u) of ruin (thus. A related difficulty occurs in situations involving the Poisson number Nt of claims: also here the sample path derivative w.t. if f (x.1). The derivations of these two estimators is heuristic in that both use an interchange of expectation and differentiation that needs to be justified. Thus.1 Consider the sensitivity tka(u) w. /3o is M(u) Oe (3T: < oo) . To see this. say W(x) = I(x > xo) and assume that h(U. () is increasing in C. I(r(u) . Then . one. this is usually unproblematic and involves some application of dominated convergence . Then z(() = Ecp(h(U. In the setting of ruin probabilities . ()) is 0 w .v. Example 6 . () = log U/(2. ()) d( hc(U. () _ (eSx. () = d log f (X. ( where h( (u. For IPA there are. Infinitesimal perturbation analysis (IPA) uses sample path derivatives. ()) h((U. p. with density f (x. () where U is uniform(0. () is an unbiased Monte Carlo estimator of zS. SENSITIVITY ANALYSIS where 295 S = (d/d()f (X. however .
differentiating w.296 CHAPTER X. the risk process should be simulated with parameters . j3 and letting flo = 0. in part for different measures of risk than ruin probabilities.T(u)) I(T(u) < co) ] . To resolve the infinite horizon problem . different parameters.t. Thus. There have been much work on resolving the difficulties associated with IPA pointed out above. we get 1 M(u) 00(u) = E (_Ti)I(T(U)<) E [(M(u) . We recall (Proposition 111.9 . We then arrive at the estimator ZZ(u) = (M(u) .T(u)) e7uerVu) for ?P.3 (u) is of the order of magnitude ue7u.r. . the estimation of z(ip(u) is subject to the same problem concerning relative precision as in rare events simulation . for different models and for the sensitivities w. Example 6. However. change the measure to FL as when simulating tp(u). whereas for the SF method we refer to Rubinstein & Shapiro [312]. ) we have VarL(ZQ(u)) ZO(u)2 O(u2)e2 u2e2ryu yu . SIMULATION METHODOLOGY Taking expectation. In the setting of ruin probabilities.3L. BL).r.0(1) so that in fact the estimator Zf(u) has bounded relative error.3 (u) (to generate Zp (u).t.1 is from Asmussen & Rubinstein [46] who also work out a number of similar sensitivity estimators. 4) that V5. 0 Notes and references A survey of IPA and references is given by Glasserman [161] (see also Suri [358] for a tutorial). a relevant reference is VazquezAbad [374]. since ELZp(u)2 < (M(U) _T(u) \ 1 2 a2ryu = O(u2)e27u.
T(u.. and {1.. 297 . X2. Besides its intrinsic interest . '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 ). in most cases ... as e. ..+• • •+X.i. with P(Xk = 1) = 9. in the Bernoulli random walk example below. Consider first a Bernoulli random walk. Oa(U ) can also be a useful vehicle for computing t/i(u) by letting a * oo. defined as Ro = u (with u E {0.g...(u) = 0 ) = 0) or it is trivial to translate from one setup to the other.(u) is defined as the probability of being ruined (starting from u) before the reserve reaches level a > u. }).. The twobarrier ruin problem The twobarrier ruin probability 0. a) = r(u)).P(•r(u.. wherel T(u) = inf {t > 0 : Rt < 0} . a) = r(u) A T+(a).1.. That is. either this makes no difference (P(R.d. T+(a) = inf It > 0 : Rt > al. R„ = u+X. are i. where X1.Chapter XI Miscellaneous topics 1 The ruin problem for Bernoulli random walk and Brownian motion. Y'a(U) = P(T (u) = r+(a)) = 1 .1}valued .
.1 For a Bernoulli random walk with 0 0 1/2. i. one elementary but difficult to generalize to other models.. then 'Oa(u) _ au a We give two proofs . tba(2) _ (1 .. = (1 . zu = EzRO = EzRT(u.1..a) Y.1) is solution. = z°Va(u) + za(1  .0)/0.(u) I\ e = 1 oa ' ()i a = u.+Xn) F[ a]n n=0. Wald's exponential martingale is defined as in 11.o)'t/1a(a . z and the solution is z = (1 . 7/la(a . and insertion shows that ( 1. The Lundberg equation becomes 1=F[ry]=(19)+9z.. By optional stopping. u + 1.e.2).a) = 0) + zap ( R..(4.2) Oa(a . Proof 1.298 CHAPTER XI.. u Proof 2. C1_0\a. The martingale is then {zuzXl+•••+X„ } = {zR° }. MISCELLANEOUS TOPICS Proposition 1. (1.o» = z°P (RT ( u. and in view of the discrete nature of a Bernoulli random walk we write z = e7.a(u)).1) o If 0 = 1/ 2. and the other more advanced but applicable also in some other settings.. Conditioning upon X1 yields immediately the recursion 'a(1) = 19+00a(2).y] = 1.. where a is any number such that Ee°X = F[a] <oo.(1B)u oJ 0. the solution of F[. We choose a = ry where ry is the Lundberg exponent.r(u.1) = (19)4/'0(a3)+9ba(a1).o)T/la (1) + 8z/'u(3).4) by ea(u+Xl+. In a general random walk setting .
then Proof Since 'Oa (U)  au a Eea(R°.zu)/(za . 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)).1. BROWNIAN MOTION. 1h (u) = a el u \1 If 9 < 1/ 2.1) for p # 0. u Proposition 1.2) is trivial (z = 1). Then for p 0 0. .1).e7u)/(e7° . and solving for 9/la(u) yields Z/)a(u) = (e 76 .2 For a Bernoulli random walk with 9 > 1/2.u) = et(a2 /2 +aµ) the Lundberg equation is rye/2'yp = 0 with solution y = 2p.1). Corollary 1.1 If p = 0.u)/u.ba(u) = e2µa . thenz1 (u)=1. TWO BARRIERS 299 and solving for 4/la(u) yields t/ia(u) = (za . Proof Let a+ oo in (1. {R.• a2µa e2µu . pa( u) _ u Corollary 1. Applying optional stopping to the exponential martingale {e7R. {Rt} is itself a martingale and just the same calculation as in the u proof of Proposition 1. If p<0. then Vi(u) = 1. (1.3 Let {Rt} be Brownian motion starting from u and with drift p and unit variance . } yields e7u = Ee7R° = e°Wa(u) + e7a(1 .4 For a Brownian motion with drift u > 0. (1. If p = 0.1 yields 't/la(u) = (a . RANDOM WALK.0a(u)).5) . However. If 9 = 1/2..
616). 0.a ) < 0) + e 7aF ( R (u. 7/'(u) = 1). Here is one more case where this is feasible: Example 1.vi(a) Proof By the upwards skipfree property. valid if p < 1 (otherwise . Ic 5ry 'pa(u) Using y = 6 . However. 1 . implying R(u. however. and thus one encounters the problem of controlling the undershoot under level 0.a) = a ) + e ' ° ( 1 . CHAPTER XI.+^a(u))^(a) If 7k(a) < 1.7). passing to even more general cases the method quickly becomes unfeasible (see. the paths are upwards skipfree but not downwards. For most standard risk processes . say. Here the undershoot under 0 is exponential with rate 5.a) = a on {r (u. . It may then be easier to first compute the onebarrier ruin probability O(u): Proposition 1.5a). (1. we obtain 'Oa a7u . a) I R(u a ) < 0] P (R(u . letting a * oo yields the standard expression pe7u for .4).0 (u) (where u p =. and hence e7u = Ee7Ro E [e7R(.300 Proof Let a * oo in (1.e7a Again .a) = r+ (a)} and similarly for the boundary 0. MISCELLANEOUS TOPICS u The reason that the calculations work out so smoothly for Bernoulli random walks and Brownian motion is the skipfree nature of the paths.7) .7/la(u)).3. VIII. 5). 7O(u) = 7/la(u) + (1 ..a) < 0) + e7°P (R(u.a) = a) = 5 y = P (R (u.5 Consider the compound Poisson model with exponential claims (with rate.6 If the paths of {Rt} are upwards skipfree and 7//(a) < 1. (u) _ O(u) .0(a) 0 < u < a.e7a (u) = 6 /0 . this immediately yields (1.
8 ).11) VIT ) Proof For p = 0.. = 1 . ( 1.T) P(MT > u) where MT = maxo<t<T St.. Here {St } is Brownian motion with drift 0 (starting from 0).8 Let {Rt} be Brownian motion with drift . We now return to Bernoulli random walk and Brownian motion to consider finite horizon ruin probabilities. MT > u) = P (ST > u) + P (ST > u.f. For µ # 0. the density dPµ / dP0 of St is eµsttµ2/2. MT > U) = P(ST > u) + P(ST > u) (1.10) Pµ (T(u) < T) !. For the symmetric (drift 0) case these are easily computable by means of the reflection principle: Proposition 1..ST<u) = P(MT>u. (1.2 .µ T I + e2µ"4) ( . T ) = P(T(u) < T ) = 241. 10) follows then by straightforward differentiation. Hence P(MT>u. TWO BARRIERS 301 Note thas this argument has already been used in VII.)_ _( u)µ2 /2.ST>U).1a for computing ruin probabilities for a twostep premium function. Corollary 1. and (1 .4) I = . of r(u) are ( U2 Pµ (T(u ) E dT) = 2^T 3/2 exp µu .Rt}. we have ili(u. Then the density and c.1. BROWNIAN MOTION. P(MT > u) = P(ST > u) + P(ST < u.µ%T (1. and hence Pµ('r(u) E dT) = Eo [e µsr(. RANDOM WALK. 0(u.µ so that {St} is Brownian motion with drift µ . + µ2T) } .9) = 2P(ST > u). (1. (i).d.r(u)...8) Proof In terms of the claim surplus process { St} = {u . T(u) E dT.7 For Brownian motion with drift 0.11 ) is the same as (1. = eµuTµ2/2Po (T( u) E dT) 2 eµuTµ2/2 u T3/2 ex p u 27r p 12 T . 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 .
12) P(ST = v) = 0 otherwise.12) is the same as ( 1. is (1. Proof The argument leading to ( 1. S(x) = f x s(y)dy.13) The following results gives a complete solution of the ruin problem for the diffusion subject to the assumption that S(x).g.11) then follows by checking that the derivative of the r.302 CHAPTER XI. MISCELLANEOUS TOPICS which is the same as (1.. is zero for all u > 0 but that nevertheless Rt ^4 0 (the problem leads into the complicated area of boundary classification of diffusions. is finite for all x > 0.9) goes through unchanged. as defined in (1. Thus. e.9).. Vi(u. the behaviour at the boundary 0 is more complicated and it may happen. 226).10). The expression for F ( ST = v) is just a standard formula for the u binomial distribution. that 0(u). (1.3 we can define the local adjustment coefficient y(x) as the one 2µ(x)/a2(x) for the locally approximating Brownian motion.T2. and in a similar spirit as in VII.T+2.13) with 0 as lower limit of integration. see e.8 also applies to the case 9 54 1/2.g.T (1.T) = P(ST = u) + 2P (ST > u).s. 0 0 (1. If this assumption fails. Here {2T( (v}TT)/2) v=T. We assume that u(x) and a2 (x) are continuous with a2 (x) > 0 for x > 0.10) and that the value at 0 is 0. The same argument as used for Corollary 1. such that the drift µ(x) and the variance a2(x) are continuous functions of x and that a2(x) > 0 . We finally consider a general diffusion {Rt} on [0. Breiman [78] or Karlin & Taylor [222] p. as defined above as the probability of actually hitting 0. u Small modifications also apply to Bernoulli random walks: Proposition 1. Let s(y) = ef0 ry(. T are integervalued and nonnegative.h.. Theorem 1.9 For Bernoulli random walk with 9 = 1/2. oo).T)dx.. oo) with drift µ(x) and variance a2 (x) at x.10 Consider a diffusion process {Rt} on [0. whenever u. but we omit the details. and (1. close to x {Rt} behaves as Brownian motion with drift µ = u(x) and variance a2 = a2(x). S(oo) = f c s(y)dy.
10. .16).14) S(oo) < 00. b = 0. A good introduction to diffusions is in Karlin & Taylor [222]. if (1. Wa. (1 . where Lq(u) = 0'22u) q "(u) + p(u)q(u) is the differential operator associated with the diffusion. then 0 < 2l. the function S(x) is .(u) < 1 for all u > 0 and ^ S^ Conversely. The obvious boundary conditions '0a. [117]. 0 Proof of Theorem 1.b('u) = Eu &0. 1'. E„ q(Rdt) = q(u)+Lq(u)dt.0(u) = 1 for all u > 0.10.16) S(a) . and we get Wo.14) fails.1.ba. BROWNIAN MOTION.S(u) (1. Letting a T oo and considering the cases S(oo) = oo. TWO BARRIERS 303 for x > 0. Lemma 1.e. RANDOM WALK.b(Rdt). elementary calculus shows that we can rewrite L as Lq(u) d 1a2 (u)s(u)d [ s (u) ? ] .17) Hence L. Assume further that S (x) as defined in (1.b('u) = Eu .S(u)/S(a). see in particular pp. If b < u < a. 0 in (1. A classical reference for further aspects of Bernoulli random walks is Feller [142]. see Asmussen & Perry [42]. Letting b J.b(u)dt. Notes and references All material of the present section is standard. Then YIa. Using s'/ s = 2p/a2.b (Rdt) = Oa. For generalizations of Proposition 1.b(u) be the probability that {Rt} hits b before a starting from u.b(b) = 1. then. we can ignore the possibility of ruin or hitting the upper barrier a before dt.6 to Markovmodulated models . so that Y)n.11 Let 0 < b < u < a and let t&0.13) is finite for all x > 0.S(b) Proof Recall that under mild conditions on q.b(u) + L. 15) i. i.b = 0 implies that VQ b/s is constant. Further references on twobarrier ruin problems include Dickson & Gray [116].b(a) = 0 then yield the result. S(oo) < oo separately u completes the proof. O. 191195 for material related to Theorem 1. If (1.16) yields 4b (u) = 1 .b(u) = S(a) .ba.e LVa.b = a+/3S. (1. In view of (1.
(7) .(a) (2.1 ) was given already in II. y > .2. They all use the fact that ( tx(a) l ( eaRt = eau + aSttx(a) < e7yu. yu) where W (ay) = y. is currently an extremely active area of research. variance 0. defined by the density 1/va(u)s(u) showing up in (1.6) .3..ytc (ay). (2. 111 . Another basic quantity is the speed measure M . equivalently.304 CHAPTER XI.9 ) and optional stopping applied to the stopping time r(u) A T. which is motivated from the study of modern ATM (asynchronous transfer mode ) technology in telecommunications.1) (2.1.6. with the drift and the variance depending on an underlying Markov process .5. Remark 11. (2.2) C_e7u < t(u) < C+e _7u.aR.17).)AT . yielding eau = Ee. MISCELLANEOUS TOPICS referred to as the natural scale in the general theory of diffusions (in case of integrability problems at 0. See Asmussen [20] and Rogers [305] for some recent treatments and references to the vast literature.13)). but by duality. Lo is a martingale (cf. where C_ = B(x) _ B(x) sup 2no fy° e7(Y )B(dy)' f2e7(Y2)B(dy)' C+ i/i(u. 2 Further applications of martingales Consider the compound Poisson model with adjustment coefficient ry and the following versions of Lundberg 's inequality (see Theorems 111.4) I. IV.5) A martingale proof of (2. Lo I.(. correponding to piecewise linear paths or . and here are alternative martingale proofs of the rest .(T(u)AT) r. (2. (2.4.aRo . 7y = ay . information on ruin probabilities can be obtained . 1 y < k (y).3) < e 7yu.o•K(a) = Ee . The emphasis is often on stationary distributions .t&(u. yu) '+/1(u) . Markovmodulated Brownian models . much of the literature dels with the pure drift case. one works instead with a lower limit 5 > 0 of integration in (1.5): _ z/'(u) < e 7u.4.
1.d. Let H(dt.1 .4): We take a = ay in (2. Proof of ( 2. we have ic(ay ) < 0 and use the lower bound E [e7Rr („). dr JO Zoo ) f e7'B(r + dy) B(r) Jo ^00 ^00 H(dt.yu))• b(u. RT(u)_) given r(u) < oo.1.6) below by 1 E Le7Rr(.3).T)  V.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).yuk (ay)(u&(u.T(u)K(ay) I yu < r(u) < T] F(yu < r(u) < T) > e. (2. yu))• Letting T + oo yield e_ayu > eyur4ay)(0(u)  Notes and references See II.6) with = 'y that eyu . and the proof of the lower inequality is similar. Hence E [e7Rr (u) Jr(u) < ool ^00 H( dt.2.(ay)I T(u) < yu] P(r(u) < yu) (using RT(u) < 0).B(r))/B(r). For (2. when Rt_ = r. Rt has distribution B(r + dy)/B(r).E [e. Equivalently. (B(y) . eyuk (ay) = e7yu e > eyu"(ay ) ij(u.)r(u)r. u Proof of (2. so that i/1(uL yu) < eayu .3). y > r. it follows easily from (2.4). ..7R. A claim leading to ruin at time t has c.f. dr) denote the conditional distribution of (T(u).2): As noted in Proposition II.yu) Y Similarly for (2. dr) 1 = 1 I0 /o C+ C+ From this the upper inequality follows. we have tc(ay) > 0 and we can bound (2. dr) e 7( yr)B(dy) B(r) f oo o 0 r > H(dt.(u.( u ) I T(U) < 00] .
the parameter will be u rather than n).2). Cramer considered a random walk Sn = X1 + . Example 3. The classical result in the area is Cramer's theorem..nn or C2e. we will write fn 1.1 We will go into some more detail concerning (3.1). however . Thus .gn if nioo lim 109 fn = 1 log gn (later in this section. gn 4 0. . og For sequences fn. Accordingly. ri. then P C S. logarithmic asymptotics . cle .1) amounts to the weaker statement lim 1 log P I Sn > x I = 17. 1) but only the dominant exponential term ./n E I) for intervals I C R. For example. v2 later. nroo n n /// Note in particular that (3. .306 CHAPTER XI. The limit result (3.?n typically only give the dominant term in an asymptotic expression .3na with a < 1. not quite so much in insurance risk. in being capable of treating many models beyond simple random walks which are not easily treated by other models . Thus. and that a considerable body of theory has been developed.means (as at other places in the book) that the ratio is one in the limit (here n * oo).(B) = log EeOX 1 is defined for sufficiently many 0. The advantage of the large deviations approach is.. + X.g. large deviations results been.2) can be rewritten as F (Sn/n > x) 1g a'fin. (3. such that the cumulant generating function r.1) does not capture the \ in (3.. e. gn with fn + 0 .1) where we return to the values of 0.. logarithmic asymptotics is usually much easier to derive than sharp asymptotics but also less informative . (3. its generality.1) is an example of sharp asymptotics : . MISCELLANEOUS TOPICS 3 Large deviations The area of large deviations is a set of asymptotic results on rare event probabilities and a set of methods to derive such results.^ e nn 1 > x n 0o 2xn (3. However . if x > EX1. which in the setting of (3. and gave sharp asymptotics for probabilities of the form P (S. The last decades have seen a boom in the area and a considerable body of applications in queueing theory. large deviations results have usually a weaker form.the correct sharp asymptotics might as well have +.
425.sseo f which in conjunction with (3. of P(X1 E dx) = E[e9X1K.4) immediately yields (3.960/) * 0.r. S rtn > x 1. 2 where o2 = o2(x) = rc"(0).9S„+n' ( 9).(e)i XI E dx].2).3.4 enn +1. Define .q = rc* (x).(0)) e 307 (other names are the entropy. we get P(Sn/n > x) E [e9nx +nK(9)9" '. P with mean nx and variance no.tin f o') o e9o^y 1 1 ey2/2 dy 21r = etin 1 Bo 27rn . i.the mean rc'(0) of the distribution of X1 exponentially tilted with 0. V > 0 e. More precisely.4) n Next. the sup in the definition of rc* can be evaluated by differentiation: rc*(x) = Ox .96o /] > 0. (3. we have P(nx < Sn < nx + 1.rc(0) where 0 = 0(x) is the solution of x = rc'(0). In fact. which is a saddlepoint equation . replacing Sn in the exponent and ignoring the indicator yields the Chernoff bound P Sn > x 1 < e°n (3.3) is put equal to x. since Sn is asymptotically normal w. nx < Sn < nx + 1. Since P nn > x) = E {e_8 ' ( 9). the LegendreFenchel transform or just the Legendre transform or the large deviations rate function). rc*(x) = sup(Ox .r. exponential change of measure is a key tool in large deviations methods.e.t. LARGE DEVIATIONS Define rc* as the convex conjugate of rc. if we replace Sn by nx + o / V where V is N(0. and hence for large n P(Sn/n > x) > E [e. Most often.1).
commonly denoted as is the saddlepoint approximation.. We shall need: Lemma 3 . .3. MISCELLANEOUS TOPICS which is the same as (3.2 (GLYNN & WHITT [163]) Let X1.s. Further main results in large deviations theory are the GartnerEllis theorem. is differentiable at ry with 0 < K'(y) < 00.dxn) = 05nKn(7)Fn(dx1. we shall concentrate on a result which give asymptotics under conditions similar to the GartnerEllis theorem: Theorem 3 . Then i/.. Ee9X n < oo for e < 0 < e. In the application of large deviations to ruin probabilities.e < 8 < y + e. r(u) = inf {n : Sn > u} and o(u) = P('r(u) < oo).. which is of similar spirit as the dicussion in VII. there exists z E (0. which is a version of Cramer's theorem where independence is weakened to the existence of c(O) = limn... (iii) #c (8) = limn.. to be made rigorous. that is..308 CHAPTER XI. see Jensen u [215] or [APQ] p.'(u) )Ng a"u. .. For the proof. (ii) lim supn. We further write µ = tc'(ry). we introduce a change of measure for X1. (iv) tc(ry) = 0 and r.p > 7 < zn.'s.3 For each i > 0. Assume that there exists 'y. Mogulskii's theorem which gives path asymptotics.. X2.. and write Sn = X1 + • • • + Xn. asymptotics for probabilities of the form P ({S[nti/n}o<t<l E r) for a suitable set r of functions on [0. e > 0 such that (i) Kn (0) = log Ee°Sn is welldefined and finite for 'y . however. integrates to 1 by the definition of Icn). Xn) and sn = x1 + • • • + xn (note that the r.v./^ >7 < zn n for n n0. .dxn) where Fn is the distribution of (X1i .o log Ee9Sn /n. 1]. . and the WentzellFreidlin theory of slow Markov walks. .. n Icn(0) exists and is finite for ry ...e < 8 < y + e. 1) and no such that Sn . The substitution by V needs. Sanov's theorem which give rare events asymptotics for empirical distributions..1).h.. 260 for details.. Xn given by Fn(dxl. be a sequence of r. Pn Sn1 ..
71 < e and jq9j < e. We first show that lim inf„_.y) .. This establishes the first claim of the lemma . S.s.+r.YS. LARGE DEVIATIONS Proof Let 0 < 9 < e where a is as in Theorem 3..ne(p+ 17).Kn(7)e'n (p(O +7))/p I Ee geXn]1/q where we used Holder's inequality with 1/p+ 1/q = 1 and p chosen so close to 1 and 0 so close to 0 that j p(0 +. log zl'(u)/u > 'y. S.2. Let r7 > 0 be given and let m = m(77) = [u(1 + 77)/µ] + 1. Clearly. in particular the r.r (7) n = e.h. can be chosen strictly negative by taking 9 small enough. can be chosen strictly negative by taking p close enough to 1 and 0 close enough to 0.µ?7 .s. This proves the existence of z < 1 and no such that Pn (Sn/n > µ. > 1 +17] m(7).ne(µ limsup 1 log Pn (Sn/n > µ + 17) < ic(9 + ry) .. The corresponding claim for Pn(Sn/n < µ .W. 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. h. P n(Sn/n > {c+77) < e no(µ 309 +n)Enees n +n)elcn(B +7).s.77) follows by symmetry (note that the argument did not use µ > 0).077 n^oo n and by Taylor expansion and (iv ). ( U) P(S. Then V.m(7).n e(µ +o)w"(7) [Eep(B +7)Sn]1 /p [EegoX. For Sn1i we have Fn(Sn 1/n > µ+r7) < ene(µ+ 1?)EneeS„1 = ene ( µ+n)EneeSneX„ eno(µ +n) Ee(e+7)Sn ex„ wn (7) < e. for Sn.]1/q = e.n > u ) = [ Em [em Em 1e.> . 0.3. The rest of the argument is as before.+r7) < zn for n > no. h. u Proof of Theorem 3.2.n m µ 1 + rl .91) + o(O ) as 0 J... mµ Sm > u] km e7Sm+n. Since I EeqOX „ ] 1/q is bounded for large n by (ii). the r . it is easy to see that the r.Bµ . is of order .
(Y). logO(u)/u > ry.3.YS +^c CHAPTER XI. 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). 14 = = E Lu(16)/aJ+1 Lu(1+6)/µJ+l = n) and n(S) is chosen such that icn('y )/n < 6 A (. n=1 n=n(b)+1 00 Lu(1 +6) /µJ 13 F( T (u) = P(T(u) n).+wn(7).0 log i'(u )/u < 'y..310 ]Em I e... MISCELLANEOUS TOPICS (7). P(T(u) = n) < P(Sn > u) = En [e7S.. I > IL exp `S. we get lum inf z/i(u) 1 +12r7 >_ ry + 77 Letting r7 J. Pn \ > la+ 8 I < zn (3. I2 = F(T(u) = n). n=1 . 0 yields liminfu __.I < µl1 1+77 I M 1_ 1+277 S. Sn > u] < eYu+Kn(7)pn(Sn > u) (3. For lim supu. this is possible by (iii). and since Ic..n Yµ 1 + m + r ('Y) } U n \ 77 m µ µ7 1 < 1+ 77 ) Here E..log z) /2 and Sn Fn\ n >lb+S) <Zn..6) for some z < 1 and all n > n(E).(•) goes to 1 by Lemma 3.n(ry)/u 4 0andm/u* (1 + r7)/µ. 3. (iv) and Lemma 3. Obviously.7) so that n(b) I1 < e'Yu E en.
C 26u `p / +1 I e6u(1+6)/µ (3. Sn1 C U. u .3.zl/z en6 [u(1 +6)/µJ 1u (1 +6) /µJ ekn(7) < e' 13 < C" E Yu l u(16)/lij+1 Lu(16)/µJ+l1 < e7U Finally. eryu en logz/2p n nt n. 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 . 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. > u) Lu(1+6) /µJ +l 00 )^n 'YSn+kn (7) .10) 00 I4 < E F(Sn_1 < u. S. µ n=n(6)+1 \ 1u(16)/µ1 00 1 zn < e7u E Z n/2 < e(U xn/2 E n=n(6)+1 n=0 eYu = 1 . we get lim sup log u/00 O (U) < y + b(1 + b) U Letbl0.11) [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).
For 12. 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. 13 = P(T (u) E (u(1 b)l^ (7). MISCELLANEOUS TOPICS The following corollary shows that given that ruin occurs. 2. For I. it holds for each b > 0 that 0(u) 1' g F(T(u) E (u(1 .2.xl/2 to give the desired conclusion. 4 there is an aj > 0 and a cj < oo such that Ij < c3e.312 CHAPTER XI.. the typical time is u/rc'(7) just as for the compound Poisson model.4. the last steps of (3.u(1+b)/rc'(7)).8) by P(S. ryue«iu . this is straightforward since the last inequality in (3. Corollary 3. IV. I2. Then for n large.('+'Y). u(1 + b)/i(7)) Proof Since V. u .z 1/z For I1.4/3rc'(y) > 0. e'.b)/i(7).9) can then be sharpened to x LQuJ /2 I2 < e7u 1 .7' a"ju. > u) < e"' E eIsn = ectueKn (a+'Y)Kn(7) where 0 < a < e and a is so small that r. we have rcn (a + 7) < 2n^c(7 + a) < 4narc' (7). we need to redefine n(b) as L. (7 + a) < 2arc'(7). say n n1.11 ) can be sharpened to x 4 [u(1+6)/µJ /2 1 ..3ui where . Letting c11 = maxn<n.4 Under the assumptions of Theorem 3. cf. For 14.(u) = I1+I2+I3+I4'^ ery( u).Q is so small that w = 1 . it suffices to show that for j = 1. we replace the bound P(Sn > u ) < 1 used in (3.
f. 11 Inspection of the proof of Theorem 3. The reader not satisfied by this gap in the argument can easily construct a discrete time version of the models! The following formula (3..v.14) is needed in both examples . but nevertheless. If {St}t> 0 is the claims surplus process.5 Assume the Xn form a stationary Gaussian sequence with mean p < 0.. Theorem 3.2 then immediately yields the estimate log F( sup Skh > u) a7u (3. 09(9).. The problem is whether this is also the correct logarithmic asymptotics for the (larger) ruin probability O(u) of the whole process.3.1.g.2 shows that the discrete time structure is used in an essential way. Let {Nt}t>0 be a possibly inhomogeneous Poisson process with arrival rate . Assuming that the further regularity conditions can be verified.. Hence z z\ 2 z nrn(9) _ n Cn0p+BZn/ * .... and in fact. 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. for the ruin probability z/'h(u) of any discrete skeleton {Skh}k=0.LARGE DEVIATIONS 313 Example 3 . An event occuring at time s is rewarded by a r.'(y) > 0. i. we shall give two continuous time examples and tacitly assume that this can be done.e.. V(s) with m. whether P ( sup St > u ltg a ^" 0<t<oo // (3. r.1.12) k=0.13) One would expect this to hold in considerable generality. the key condition similar to (iii).3(s) at time s.(O) = 9µ+02 for all 9 E R. Obviously many of the most interesting examples have a continuous time scale. criteria are given in Duffield & O'Connell [124]. and we conclude that Theorem 3 . t] is Rt = E V (Un) n: o„ <t . Thus the total reward in the interval [0. 2 is in force with y = 2p/wz. (iv) becomes existence of a limit tc(9) of tct(9) _ log Ee8S° It and a y > 0 with a(y) = 0. Xk+l) k=1 00 naoo n provided the sum converges absolutely. To verify these in concrete examples may well present considerable difficulties.
Thus. Kt (0) t (Ee9U"it8i J0 .1) ds rt (3. Since the remaining conditions of Theorem 3. Un represents the total payment for the nth claim).1) . O'n +S] is a r . = U„ ( t .Lundberg model has the larger ruin probability.d.g. e.2 are trivial to verify.1) ds . .t. (9) < oo for 9 < 'y + C. 0 Example 3 .14) (to see this . the CramerLundberg model implicitly assumes that the Poisson intensity /3 and the claim size distribution B (or at least its mean µB) are known. It is interesting and intuitively reasonable to note that the adjustment coefficient ry for the shot . Then logEeOR° = J0 /3(s)(^8(9) .314 where the an CHAPTER XI. e > 0 such that ic('y) = 0 and that r. we conclude that Cu) log e7 u (cf. i. Thus by (3. then the payments from the company in [on. n: o. 0 and since EeOUn(8) + Ee°U^ as s * oo.14). Thus. derive .. nondecreasing and with finite limits Un as s T oo ( thus.1) ds . We let ic (9) = 3(EeWU° . (3. leading to St = At(1+77) Joo t S8 ds. this is not realistic . but that a claim is not settled immediately. it contributes to St by the amount Un(t .s). one would take p(t) = (1 + rt)At/ t. where Ft = a(A8 : 0 < s < t).v.'`1 U. If the nth claim arrives at time Qn = s.noise model is the same as the one for the Cramer Lundberg model where a claim is immediately settled by the amount Un. we have rct (9)/t 4 ic (9). the above discussion of discrete skeletons). Of course.. if the nth claim arrives at time a. 7 Given the safety loading 77. we have S.. We further assume that the processes {U1(s)}8>0 are i. An apparent solution to this problem is to calculate the premium rate p = p(t) at time t based upon claims statistics . Of course . Un(s).Q„) . <t which is a shotnoise process. More precisely. the best estimator of /3µB based upon Ft.9t = /3 J t (Ee8U° i8l .15) .It. At = . is At .. Most obviously. MISCELLANEOUS TOPICS are the event times.9t. assuming a continuous premium inflow at unit rate. the Cramer. a differential equation in t). Example 3.6 We assume that claims arrive according to a homogeneous Poisson process with intensity 0 .0 and assume there are y.
(1 + r7) log t (3./3.1) or . 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 . and since the remaining conditions are trivial to verify.d. one has y > y' (3. rewrite first rc as te(a) _ /3E 1 1 +(1+77)aUJ eau 1 .b(u) IN a'Yu (cf.16) i=1 o i=1 Let ict (a) = log Eeast .2 hold. Indeed. To see this . we have Nt t N.14) that rt _ 13 Jo _ (a [1_( i+77)log]) ds_flt = t (a) (3. (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. uniform (0. typically the adaptive premium rule leads to a ruin probability which is asymptotically smaller than for the CramerLundberg model .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 . i. Ui Nt / t 01i 315 St = Ui . the solution of /3(Eelu .1) .i.19) with equality if and only if U is degenerate.log Oi are i. equivalently. we conclude that t.17) K(a) f o 1 O (a[I + (1 + 77) log u]) du )3.e. It then follows from (3.3.(1 + 17)0µB = 0.d.i. (3. the Vi = .(1 +i) f > i= 1 s ds = E Ui 1 . Thus. standard exponential . LARGE DEVIATIONS With the Qi the arrival times.20) (3.18) Thus (iii) of Theorem 3.
k'(0) < 0.i. though we do not always spell this out.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. the function k(x) = e7*x . For Example 3. assuming that the U. 4 The distribution of the aggregate claims We study the distribution of the aggregate claims A = ^N' U. 11 Notes and references Some standard textbooks on large deviations are Bucklew [81]. so there exists a unique zero xo = xo(r7) > 0 such that k(x) > 0. [245]. a* (s) are convex with tc'(0) < 0 . This implies n(y*) < 0. MartinL6f [256]. and since tc(s). x > x0. In addition to Glynn & Whitt [163]. Further. at time t. see Nyrhinen [275] and Asmussen [25].2 expressing the finite horizon ruin probabilities in terms of the distribution of A. rc*' (0 ) < 0. with common distribution B and independent of Nt. This is a topic of practical importance in the insurance business for assessing the probability of a great loss in a period of length t. using that Ek(U) = 0 because of (3. the proof of (3. Lehtonen & Nyrhinen [244].(1 + ri)y*x is convex with k(oo) = 00.2.d. For notational simplicity. Further. Dembo & Zeitouni [105] and Shwartz & Weiss [339]. much of the analysis carries over to more general cases.xo. the study is motivated from the formulas in IV. The main example is Nt being Poisson with rate fit. . Therefore e7'U _ k(U) E [1+(1+77)y*U] . we then take t = 1 so that p. and k(x) < 0. say one year. 0 < x < x0. k(0) = 0.20) is due to Tatyana Turova. [257] and Nyrhinen [275]. = P(N = n) = e(3an However.316 CHAPTER XI. MISCELLANEOUS TOPICS Next. we are interested in estimating P(A > x) for large x. this in turn yields y > y*. are i. Further applications of large deviations idea in risk theory occur in Djehiche [122].7. see also Nyrhinen [275] for Theorem 3. In particular.1 . y = y* can only occur if U .19)..
The exponential family generated by A is given by Pe(A E dx) = E [eeA K(9). A E dx] . The analysis largely follows Example 3. no(a) = logE9e'A = rc(a + 9) .3B[9] and Be is the distribution given by eox B9(dx) = B [9] B(dx). e9x+K(°) P(A > x) B 2ir /3 B" [9] Proof Since EBA = x.1 Assume that lim8T8. 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 .4.x)//3B"[9] is standard normal. K'(0) _ ic'(9) = x. we define the saddlepoint 9 = 9(x) by EBA = x. only with 0 replaced by a9 and B by B9. B"[s] = oo. 818' where s' = sup{s : B[s] < oo}. (4.1).e. For a given x. This shows that the Pedistribution of A has a similar compound Poisson form as the Fdistribution.1. Proposition 4.[s])3/2 = 0. Then as x * oo. i.2) implies that the limiting Pedistribution of (A ."(0) = . Hence P(A > x) = E e [e9A+ ic(9).9(Ax). THE DISTRIBUTION OF THE AGGREGATE CLAIMS 317 4a The saddlepoint approximation We impose the Poisson assumption (4. Vare(A) = s.3e(bo[a] .1) where )30 = . A > x)] = eex+K( e)E9 [e .1).3B"[9]. B"' [s] lim (B". Then Ee"A = e'(") where x(a) _ 0(B[a] . In particular.ic(9) = .
Jensen [215] and references therein. bounded with b(x) . just the same dominated convergence argument as in the proof of Theorem 2. it is quite questionable to use (4.Q{AB (4. Var(A) _ ^3p. For a rigorous proof.e. 1 . b is logconcave.4) .3) The result to be surveyed below improve upon this and related approximations by taking into account second order terms from the Edgeworth expansion. b(x) = q(x)eh(z). Remark 4 . For example. b is gammalike. under the Poisson assumption (4.ycix °ie6x B. In fact. either of the following is sufficient: A.2) is often referred to as the Esscher approximation. it holds that EA = .2i and that (A .1 yields: Proposition 4. and (4.(D X . In particular. 4b The NP approximation In many cases . B covers distributions with finite support or with a density not too far from ax° with a > 1.x') where x' = sup {x : b(x) > 0}. see Embrechts et al. For example.1). the distribution of A is approximately normal . MISCELLANEOUS TOPICS It should be noted that the heavytailed asymptotics is much more straightforward. Thus . then P(A > x) . 3 A word of warning should be said right away : the CLT (and the Edgeworth expansion) can only be expected to provide a good fit in the center of the distribution . The present proof is somewhat heuristical in the CLT steps. Furthermore 00 b(x)Sdx < oo for some ( E (1.1 goes all the way back to Esscher [141]. i. where q(x) is bounded away from 0 and oo and h (x) is convex on an interval of the form [xo. Notes and references Proposition 4.(3µB)/(0µB^))1/2 has a limiting standard normal distribution as Q ^ oo. leading to P(A > x) :.2 If B is subexponential and EzN < oo for some z > 1.318 CHAPTER XI. large x. or. A covers the exponential distribution and phasetype distributions. some regularity of the density b(x) of B is required. 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. 2).EN B(x).l3pB. [138]. more generally. For details.v.3) and related results u for the case of main interest .
the standard normal distribution. f °o 9(y) = 1 e'uye u2/2(1 + iSu3) du 27r _ cc(y) .5) follows by integration.2 ^ \1 .5) is obtained by noting that by Fourier inversion.1). Let Y = (A . the NP (normal power) approximation deals with the quantile al_E.. one expects the u3 term to dominate the terms of order u4.s. If the distribution of Y is close to N(0. which is often denoted VaR (the Value at Risk)..99. u5.2X2 . are small. (4. . Remark 4.EA)/ Var(A) and let yl_E..2 2 . however.. the density of Y is 1 °° _ eiuy f(u) du 2x _.i 3 K3 } Pt^ exp .5) may be negative and is not necessarily an increasing function of y for jyj large. (4.2K3 + 4i 64 + .5). ylE should be close to zl_E (cf.4.. in particular.5 (y3 . then P(Y < y) 4(y) . A particular case is a. are the cumulants .y2)^P(y)• 319 Note as a further warning that the r.e. K2 = Var (Y).c2i. . Var(Y) = 1 as above .l = EY.6(1 .. THE DISTRIBUTION OF THE AGGREGATE CLAIMS where b is a small parameter. In concrete examples . as u2 u3 u4 9(u) = Ee'uY = exp {iuci .h. If this holds . so that 1(u) 3 exp { . . Thus if EY = 0. the CLT for Y = Y6 is usually derived via expanding the ch. of (4.equantile in the distribution of Y.: EA + zl_E Var(A) . Rather than with the tail probabilities F(A > x). K3 = E(Y ... zl_e be the 1 . K4 .EY)3.. one needs to show that 163. and so as a first approximation we obtain a1_E = EA + yle Var(A) .3& (y).3!). where Kl .i 6 r 1 3 so that we should take b = ic3/6 in (4.f. resp. s. Heuristically. defined as the the solution of P(A < yle) = 1 .6) . and from this (4.
3ni /3 .5) by noting that the 4.. In particular .(y) terms dominate the S(1 . k3 is small for large /3 but dominates 1c4. n = 1.zl E)^o(zl E) . this yields the NP approximation 6(Z1 _E .y2)cp( y) term.zlE)W(zlE) 1 .. as required .6pBki) d/2. b = /3 for the Poisson distribution with rate /3 since Pn = Pn1 n! n (n .yi. 21 . this holds with a = 0. MISCELLANEOUS TOPICS A correction term may be computed from (4. b such that EN 1 U%.E . This leads to t( yl E) . K5 .E + (yl. the kth cumulant of A is /3PBk' and so s. We can rewrite (4.E(/3PB^1 )1^2 + s(z1E .zi.E = z1E + S(zi_E . that [101] distinguishes between the NP and Edgeworth approximations.320 CHAPTER XI.zl E)V(zl_E) ..S(1 .k = /3µB^1 / (.EA ) / Var(A).1). and assume that there exist n ) Pn_i . 4c Panjer 's recursion Consider A = constants a.5(1 .E)A1 l E) 1 E 4)(yl E) ^' . Using Y = (A .S(1 ..1)! n ^eQ . however.7) as 1 (3) a1E = Qµa +z1 . . Note..1)^ 2) µ'E Notes and references We have followed largely Sundt [354].EA)3 a1_E = EA + z1_E(Var (A))1/2 + 1 Var(A) Under the Poisson assumption (4.1) E (A .E )Azl E) 4(z1E) + ( ylE .1)EY3.6 (1 . let pn Pn = (a+ = P(N = n). For example. [101].zlE )w(zl _E) = which combined with S = EY3/6 leads to q^ 1 Y1 . Another main reference is Daykin et at...
.12) we get for j > 0 that fj n a b + n p nlgj *n 00 U I n 1 *n = E a+bUi=j pn19j n=1 j i=1 CC) n Ui EE n=1 Ia +b Ul i=1 =j pn_1 . . . 2. . fj = E (a+ b k =1 )9kfi_k .11) Remark 4. 2.} and write gj = 2 .4. u Proof of Proposition 4. (4. . and calculating the gj*n recursively by 9*1 = 9j..4 is that the algorithm is much faster than the naive method.12). THE DISTRIBUTION OF THE AGGREGATE CLAIMS 321 Proposition 4.. .4.12) where g*n is the nth convolution power of g.1. fj = P(A = j)..14) is independent of i = 1. n = k=n1 9k(n1 )9j k • (4. n. Hence by (4.. .10) f o = po. then j (a + b!) 1ag k_1 3 gkfj. 2. the complexity (number of arithmetic operations required) is O(j3) for (4. j = 1. .4 Assume that B is concentrated on {0.9).13) Namely.13) but only O(j2) for Proposition 4. if go = 0. E[a +bU=I >Ui =j l i=1 J (4... .5 The crux of Proposition 4. j = 1. (4.4.k . which would consist in noting that (in the case go = 0) fj = pn9jn n=1 (4... Since the sum over i is na + b. By symmetry. the value of (4... j = 0. (4. (4. j1 g. 1.. Then fo = >20 9onpn and fi = 1 E In particular.14) is therefore a + b/n. The expression for fo is obvious. .
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).
h(Ui)] .6.h(Ui)] .P I = EC [7]N. ' ii (6.4).. . appealing to (6.4).4).d. we get N 1 = Eexp ry E[Ui ii . this implies 7* > 7.b)+ with b determined by E(U . expectations like those in (6.h(Ui)] . N E X . it suffices to show that Eexp {ry ii 'UiAb.6 Assume the Ui are i.3).d. Local reinsurance with h(u) = (u . (6.i.P]}. This follows by taking Xl = U A b.h * (X) . (6. y = Ei [Ui .P. however.b)+ is referred to as excessofloss reinsurance and plays a particular role: Proposition 6. that 01[ry] < 0[y] where 0[y] = Ee'r(U^') .4. i.P } < 1 = Eexp E[Ui. REINSURANCE Proof Define N 329 h* (x) = E > h(Ui) X = x .b)+ = Eh(U) (and the same P) satisfies 71 > ry. The arrangement used in practice is. as often local as global. we get EX = EN • EU.6) u where C[ry] = Ee'r(u4(u)).h(U) (as in the proof of Proposition 6. Applying the inequality Ecp(Y ) > EW(E (YIX )) (with W convex ) to W(y ) = eryy. Assuming for simplicity that the Ui are i.h(U)].P > EexP{7[X .5 Because of the independence assumptions . Remark 6.h(Ui)P JJJ l:='l {ry ] or.5) holds trivially. Proof As in the proof of Proposition 6.h(u) and any P satisfying (6. Then for any local retention function u .6).5) reduce quite a lot. then (6. the excess ofloss rule hl (u) = (u .4) and u g(x) = e7x in Ohlin's lemma. X2 = U .h( UU) = EN • E[U . ry* > 0 because of (6. u But since ry > 0. Eexp 7 [E [Ui . and so on.
e.330 CHAPTER XI. Bowers et at.g. see also Sundt [354]. The original reference for Ohlin's lemma is Ohlin [277]. See further Hesselager [194] and Dickson & Waters [120]. .many texts on insurance mathematics. MISCELLANEOUS TOPICS Notes and references The theory exposed is standard and can be found in. The present proof is from van Dawen [99]. [76]. Heilman [191] and Sundt [354].
. some condition is needed: that F is nonlattice. note in particular that U({0}) = 1.. The associated renewal measure U is defined by U = u F*" where F*" is the nth convolution power of F.e. = T„ . the renewal process is called zerodelayed. i. That is.. Y2. .U(t) is the expected number of renewals in (t. are independent and Y1. all have the same distribution. not concentrated on {h. . t]) so that U(t + a) . The mathematical representation is either the ordered set 0 < To < T1 < . Lebesgue measure for some n > 1). the distribution of Yo is called the delay distribution. The renewal theorem asserts that U(dt) is close to dt/µ.t.. of interarrival times and the time Yo = To of the first arrival (that is. If Yo = 0. when t is large. . + U2 where U1 is a finite measure and U2(dt) = u(t)dt where 331 . Y1. 2h.r. Lebesgue measure dt normalized by the mean to of F. of epochs or the set Y1.1) (here U(t) = U([0.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] is denoted by Nt. The number max k : Tk_j < t of renewals in [0. Technically. Y2... Y. If F satisfies the stronger condition of being spreadout (F*' is nonsingular w . The point process is called a renewal process if Yo.. then Stone 's decomposition holds : U = U. t 00 (A. Then Blackwell 's renewal theorem holds.. denoted by F in the following and referred to as the interarrival distribution.. stating that U(t+a)U (t) ^ a.} for any h > 0.T„_1). t +a]).. U(A) is the expected number of renewals in A C R in a zerodelayed renewal process. .
wee shall need the following less standard parallel to the key renewal theorem: Proposition A1.2). oo). that z(u) has a limit z(oo) (say) as u 4 oo.9. the asymptotic behavior of Z(u) is given by the key renewal theorem: Proposition A1.2) has the unique solution Z = U * z. z(x) = 0. and that F has a bounded density2. (A.a. z(u) a known function. Under weak regularity conditions (see [APQJ Ch. and F(dx) a known probability measure . see [APQ] Ch. Then Z(u) 4 z(oo). stating that U(t)/t > 1/p. Equivalently.x)F(dx).332 APPENDIX u(t) has limit 1/µ as t 4 oo. the statements being EN(t + a) . IV).5) 2This condition can be weakened considerably . A weaker (and much easier to prove) statement than Blackwell's renewal theorem is the elementary renewal theorem.e. (A. but suffices for the present purposes . In 111.4) If F is spread.. Both result are valid for delayed renewal processes. (A.EN(t) . in convolution notation Z = z + F * Z.out. Note in particular that F is spreadout if F has a density f. then it suffices for (A.4) that z is Lebesgue integrable with limZ. i.i.3) Further.1 if F is nonlattice and z (u) is directly Riemann integrable (d.2 Assume that Z solves the renewal equation (A. resp. IV). µF (A.R.2) Z(u) = J0 u z(x)U(dx). U Z(u . (A. ENt 4 1 lb Renewal equations and the key renewal theorem The renewal equation is the convolution equation Z(u) = z(u) + f where Z(u) is an unknown function of u E [0 . then Z(u) i f0 z(x)dx .i". u u PF 4 00.
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). The simplest case is when {Xt} has i.. • . However. results from the case fo F(dx) = 1 can then be used to study Z and thereby Z. Tk and {Xt }o<t<Tk • For example.d. . . Here the relevant F does not have mass one (F is defective). To this end. Assuming that y can be chosen such that f °° Ox F(dx) = 1.5a. this expression is to be interpreted as a random element of the space of all Evalued sequences with finite lifelengths.x)u(x) dx = z(u( 1 . cycles.. A stochastic process {Xt}t>0 with a general state space E is called regenerative w.t))u(ut) dt 0 0 J f z(oo) • 1 dt = z(OO). a basic reason that renewal theory is relevant is the renewal equation II... {Tn} if for any k.. We let FO. i.2) by e7x to obtain Z = z +P * Z where Z(x) = e'Y'Z(x). Z(u) U = 1 u 1 u f z(u . Hence by dominated convergence. that the existence of y may fail for heavytailed F. Tk (or. F(dx) = e7xF(dx). equivalently.APPENDIX 333 Proof The condition on F implies that U(dx) has a bounded density u(x) with limit 1/µF as x * oo. Yk ). however.i. z(x) = e7xz(x).3) satisfied by the ruin probability for the compound Poisson model. is called the cycle length distribution and as before. The kth cycle is defined as {XTk+t}o<t<Yk . or many queueing processes. 0 PF µF 11 In risk theory.. Y2.. we let µ denote its mean. . The property of independent cycles is equivalent to the postTk process {XTk+t}t>0 being independent of To. this covers discrete Markov chains where we can take the Tn as the instants with Xt = i for some arbitrary but fixed state i. refer to the zerodelayed case.} be a renewal process..k+t }t>o is independent of To. of Yo. . 1c Regenerative processes Let {T.t..e. . . the postTk process {XT. Eo etc. Y1 . that F is a probability measure. Note.(3. multiply (A. T1. However. asymptotic properties can easily be obtained from the key renewal equation by an exponential transformation also when F(dx) does not integrate to one. T1. This program has been carried out in III.r. the present more general definition is needed to deal with say Harris recurrent Markov chains. A regenerative process converges in distribution under very mild conditions: . The distribution F of Y1.
under the condition of Blackwell's renewal theorem. {Tn} if the processes {ZT +t . {i7(t)} are Markov with state spaces (0... oo).t. cycles (we allow a different distribution of the first cycle).d. oo). Otherwise . just the same proof as there carries over to show: Proposition A1. {Tn}.6) id Cumulative processes Let {Tn} be a renewal process with i. (b) If in addition Var(Ul ) < oo.r.t..0 is called cumulative w.. in total variation. Then {e(t)}. where the distribution of X.'s by e. and q(t) = sup It .r.i. {Tn}.334 APPENDIX Proposition A1. C).e. C(t) and ij (t) both have a limiting stationary distribution F0 given by the density F (x)/p. but in fact.. then (Zt . We denote the limiting r.r.oo (i.e.v. Y1) le Residual and past lifetime Consider a renewal process and define e ( t) as the residual lifetime of the renewal interval straddling t.3 Consider a regenerative process such that the cycle length distribution is nonlattice with p < oo. If p = oo. for n = 1. is given by Eg(Xoo) = 1 E0 f Ylg (Xt)dt. [0.t. fi (t) = inf {Tk .ZTOI < 00..0 be cumulative w.Tk : t < Tk} as the age.. then e (t) . e(t )) .. µ 0 If F is spreadout.i.3. r.t : t < Tk}. Then {Zt}t^. and we have: holds more generally that (rl(t). i.ZT Then: (a) If E sup I ZTo+t . 0<t<Yi then Zt /t a$• EU1/µ. An example is Zt = fo f (X8) ds where {Xt} is regenerative w. P(C ( t) < a) 4 0 for any a < oo) and ij (t) * oo. are i.4 Let {Zt}t^. assume that p < 00 and define Un = ZT}1 . 2. Then Xt Di X.. (A.+ X. resp .. Then it (ii.ZT }0<t<Y„+..d.. This is the case considered in [APQ] V.tEU1/µ)/f has a limiting normal distribution with mean 0 and variance Var(Ui) + (!)2Var (Yi)_ 2EU1 Cov(U1. then Xt .
U(x) < U( 1)). . Proof The number Nt of renewal before t satisfies Nt/t a4' p. we can bound e(t) by M(t) = max {Yk : k < 2t/p}.d. Since z ( k) < E[Yi . but governed by a Markov chain {Jn} (we . assume first the renewal process is zerodelayed. 1) and W has distribution Fw given by dFw/dF(x) = x/pF.6 Consider a renewal process with µ < oo. are not i. For the second. and the conditional distribution of ri given l. In the general case.5 Under the condition of Blackwell's renewal theorem.. (d) the marginal distribution of ^ is FO. use t E^(t)/t = E[Yo .i. the joint distribution of (rl. (c) the marginal distribution of q is FO.. the first statement follows. r.t. Y1 > t] 4 0. In IV. U(x + 1) .i. Yl > t].. Y1i Y2.U(x) (c < oo because it is easily seen that U(x + 1) . W are independent. Hence for t large enough.^(t))} as a regenerative process. V is uniform on (0.t. Yo > 0] + f Eo^ (t . Then fi(t)/t a4' 0 and.v. (1 V)W) where V. = z is Foz) The proof of (a) is straightforward by viewing {(r.y)P(Yo E dy) . the sum is o(t) so that Eo£(t)/t + 0 . l:) is the same as the distribution of (VW. and the conditional distribution of given 17 = y is the overshoot distribution R0(Y) given by FO(Y) (z) = Fo (y+z)/Fo(y).dy )z(y) < c ^ l z(k) Eoe(t 0 0 k=o where c = sup.d. Hence t t lt ) = f U(dy)z(t . ^) is given by the following four equivalent statements: (a) P (77 > x.(t).APPENDIX 335 Theorem A1. 0 If Markov renewal theory By a Markov renewal process we understand a point process where the interarrival times Yo . we used: Proposition A1. ^ > y) = 1 f +Y (z)dz. Then Eo^(t) satisfies a renewal equation with z(t) _ E[Y1 .4. and the equivalence of (a) with (b)(d) is an easy exercise. (b) the joint distribution of (ri. if in addition EYo < oo.'s with finite mean satisfies Mn/n a$• 0 (BorelCantelli).y) = f U(t . Since the maximum Mn of n i. EC(t)/t + 0.
A2 WienerHopf factorization Let F be a distribution which is not concentrated on (oo..i .}.. oo).7 Consider a nonlattice semiregenerative process..T_ < oo). 0] or (0 . e. Y1.. Jn = i is the same as the P. We call r+ (T_) the strict ascending (weak descending) ladder epoch and G+ (G_) the corresponding ladder height distributions. be i. ..336 APPENDIX assume here that /the state space E is// finite) in the sense that P(Y.t.t. in [APQ]. Jn_1. X2. = io for some arbitrary but fixed reference state io E E. IT. These facts allow many definitions and results to be reduced to ordinary renewal. namely {Twk } where {Wk } is the sequence of instants w where Jo. and define r+=inf{n>0: Sn>0}.. A stochastic process {Xt}t>o is called semiregenerative w. A Markov renewal process {Tn} contains an imbedded renewal process. .. . 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. The semiregenerative process is then regenerative w. Then Xt 4 Xo. the conditional distribution of {XT„+t}t>o given Yo.and regenerative processes.d. the Markov renewal process if for any n. T_=inf{n>0: Sn<0}.. . < yIJ) = Fij( y) on {Jn= i.r. .+ < x.g. Let X1. .r...jEE is a family of distributions on (0. For example. Alsmeyer [5] and Thorisson [372].. distribution ofjXt}t>o itself where Pi refers to the case Jo = i. where the distribution of X. Jo. Notes and references Renewal theory and regenerative processes are treated. .) and (Fij )i. with common distribution F. r+ < oo). Jn +1=j} where J = a(JO. is given by Eg(X00) = 1 YO vjEj f g(Xt) dt µ jEE o where p = ujEEViAj. J1 i . the semiregenerative process is called nonlattice if {T.} is nonlattice (it is easily seen that this definition does not depend on i).. Yn. oo). Further: Proposition A1.. . . . G_(x) = P(ST_ < x. G+(x) = P(S.
7) (A. (d) R+ = U_.1 (a) F = G+ + G_ . 0).S. 0<j<m. A C (oo.8) (e.. (c) G+(A) = f °. More rigorously. we may rewrite (a) as G_ (A) = G+(A) = F(A) + (G+ * G_)(A). .T_=n} = {S. define w as the time where the preT_ path S1. >0. .7) follows since G+(A) = 0 when A C (oo. n=0 n=0 00 00 and the T+. S.x)R+(dx). A C (oo. n 0 R_(A) = E I(Sn E A). On {T_ > 2}. A C (0. In (A. 0]). the renewal measures U+=>G+.r.. (e) R_ = U+. Proof Considering the restrictions of measures to (oc. m<j<n}. G+.g.APPENDIX 337 Probabilistic WienerHopf theory deals with the relation between F. 0] and (0. Sr_ _1 is at its minimum .and r_ preoccupation measures T+1 r_1 R+(A) = E E I(Sn E A).G+ * G_: (b) G_ (A) = f °° F(A . 0]. oo). U. we consider the last such time (to make w unique) so that {w=m.x)R_ (dx). F(A) is the contribution from the event {T_ = 1} = {X1 < 0}. u .7). oo). A C (0. oo) (A. n=0 The basic identities are the following: Theorem A2.1 .=n w=m i Figure A. (A. G_..=EGn. F(A) + (G+ * G_)(A). F(A .S.>0.
7) follows._ E A) n1 f P(r_=nw=m Sm EduSrEA) m=1 n1 F(r+=mSr+Edu)..3 8 APPENDIX Reversing the time points 0.+ E du) E P(S. Sr_ E Adu) (s ee again Fig . m it follows (see Fig. A.3. S. ST+Edu)...+ E du)P(S.1. Aso.m.XnEAx) 00 f 0 f 0 00 00 1: F(A ..du) (G+ * G)(A)• C llecting terms._ = n . . .. SmEdu) = P(T+=m. clearly (Sj Sm>0.>0.x)P(Sk < 0. and the proof of (A.. and reversing the order of summation yields P(T_ > 2.x)R+(dx).1) that P(Sj Sn. (A. A.0<k<ri . 0<j<m. ST_ E A) P(T+ = m. m=1 f S mming over n = 2. (b) follows from 00 G+ (A) _ E F(Sn E A..= n. . m < j <n._ E A .. 0 < k < n. E du) = P(T_=nm.8) is similar. SnEAIS. r+ = n) n=1 n=1 0  C0 E fF(Sk< 0. It follows that for n > 2 F (7. ST_ E A .Sn_1Edx.F(r_n_mSrEA_u). S. Sn1 E dx) n=1  F(A .1).u) f0m m=1 n=m+1 00 J0 OO P(S.
which is basic for the PollaczeckKhinchine formula.T+> n) = P(Sk < O.'s.G_ [s] is defined and bounded in the halfplane is : ERs > 01 and nonzero in Is : ERs > 0}. Since G+ is concentrated on (0.. then T+ = inf It > 0 : St = 0} is 0 a. Summing over n yields R+ (A) = U_ (A). The present proof of Theorem A2. there are direct analogues of Theorem A2. G_ [s] are defined at the same time.1. and similarly H_ (s) = 1 . consider a fixed n and let Xk = Xn_k+l.. Again. WienerHopf theory is only used at a few places in this book.0+[s])(1 .APPENDIX 339 and the proof of (c) is similar.O<k<n.s. being concentrated at 0. Then for A C (oo.g. see e.SnEA) = P(SnSn_ k. and using timereversion as in (d) to obtain the explicit form of R+ (Lebesgue measure).P as a product H+H_ of functions with such properties. see for example Bingham [65]. For (d). G_ are trivial. cf.4). there is no direct analogue of Theorem A2. P(SnEA .0<k<n.f. 11.F[s] = (1 . and the proof of (e) is similar. the analogue of a random walk is a process with stationary independent increments (a Levy process.SnEA) = P(Sn<Sk. However. oo). . is based upon representing G+ as in (b). we can rewrite (a) as 1 . the survey [15] by the author and the extensive list of references there. In discrete time. and sometimes in a larger strip.1. if {St} is Brownian motion.0<k<n.g.O<k<n.1(a) is from Kennedy [228]. and G+. In this generality of. u Remark A2.2 In terms of m. For example. In continuous time. The classical analytical form of the WienerHopf problem is to write 1 . 6+ [s]. such developments motivate the approach in Chapter VI on the Markovian environment model.g.1).SnEA) = P(Sn<Sk. Nevertheless.G_[s]) (A. H+ (s) = 1G+[s] is defined and bounded in the halfplane Is : ERs < 0} and nonzero in Is: Rs < 01 (because IIG+lI _< 1).SnEA) is the probability that n is a weak descending ladder point with Sn E A. 0]. a number of related identities can be derived. u Notes and references In its above discrete time version. it serves as model and motivation for a number of results and arguments in continuous time.9) whenever F[s].6. the derivation of the form of G+ for the compound Poisson model (Theorem 11. E. Another main extension of the theory deals with Markov dependence. Sk = X1 + • • • + Xk = Sn . this holds always on the line its = 0.Sn_k.
Here it is standard to compute matrixinverses by GaussJordan el imination with full pivoting . To circumvent this. Eo Kn/n! converges rapidly and can be evaluated without p oblems. Some fundamental properties are the following: sp(eA) = {e' : A E sp(A)} (A. It is seen from Theorem VIII.11) A f eAtdt = eA. 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).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. 0 . if m is s fficiently large. Thus.13) henever A is a diagonal matrix with all diagonal elements nonzero. whereas there is no similar single established a proach in the case of matrix exponentials. Here are.10) d dteAt = AeAt = eAtA (A. write eQ = (eK)m where = Q/m for some suitable integer m (this is the scaling step).12) eA'AO = Ale AA (A. one needs to compute matrix inverses Q1 and matrix exponentials eQt ( r just eQ ). ere A is the eigenvalue of largest absolute value.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. and eQ can then be computed as the mth power (by squaring if = 2). _I 0 (A. 1. three of the c rrently most widely used ones: xample A3. however .5 that when handling phase type distributi ons. JAI = max {Jjt : µ E sp(A)} and sp(A) is the set of all eigenvalues of A (the spectrum).
letting P = I + Q/i and truncating the series in the identity = e17t 00 Pn(. condition upon the number n of Poisson events in [Olt])  Example A3. . i. In practice.2 (UNIFORMIZATION) Formally.15) Then it is easily checked that P is a transition matrix . 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 . The approach is in particular convenient if one wants eQt for many different u values of t. p different eigenvalues Aj i . Ap.3 i (A. Zo = a (Z = QZ. the procedure consists in choosing some suitable i > 0.e. The probabilistic reason that (A.7t) n=0 n! u °O n Pn (to see this.. assume that Q is the intensity matrix for {Xt} and choose q with rt > max J%J = max qii• 1.]t)n (A.3 (DIFFERENTIAL EQUATIONS) Letting Kt = eQt.14) holds is therefore that the tstep transition matrix for {fft} is eQt = E ent (.APPENDIX 341 Example A3. some jumps are dummy in the sense that no state transition occurs ).. One then can reduce to p linear differential equations by noting that k = ZQ.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. 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. vp be the corresponding left .e.. i.. Here is a further method which appears quite appealing at a first sight: Example A3 . Zo = h). what is needed is quite often only Zt = TreQt (or eQth) with it (h) a given row (column) vector. However .4 (DIAGONALIZATION) Assume that Q has diagonal form. 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. construction of {Xt} by realizing the jump times as a thinning of a Poisson process {Nt } with constant intensity 77. . Let vi..
Complex calculus : Typically.17) eQt = E e\`thivi = E ea:thi ® vi.. Everything is nice and explicit here: 411+q2+D' )12_g11+q2^^ where (411422z + 4412421. hi have been computed.18) Namely. Nevertheless. this last step is equivalent to finding a matrix H such that H1QH is a diagonal matrix. however. In view of this phenomenon alone care should be taken when using diagonalization as a general tool for computing matrixexponentials. say Al.18) contains terms which almost cancel and the loss of digits may be disasterous..342 APPENDIX (row) eigenvectors and hl. we have an explicit formula for eQt once the A j.16) (A. There are. (A. v5Q = Aivi. not all ai are real. the eigenvalue.5 If Q= ( 411 ( q21 q12 q22 is 2 x 2. i=1 i=1 Thus. and vihi ¢ 0.. some cases remain where diagonalization may still be appealing.g H1. (A. D = ) 2 2 . we can take H as the matrix with columns hl. Example A3. hp.. Then P P Q = > Aihivi = E Aihi (9 vi.. i= 1 i=1 P P (A. and we may adapt some normalization convention ensuring vihi = 1. and hence A2 is so because of A2 = tr(Q). hp the corresponding right (column) eigenvectors. and writing eQt as eQt = He°tH1 = H (e\it)di.. of largest real part is often real (say. i # j. Then vihj = 0. under the conditions of the PerronFrobenius theorem).. say A = (Ai)diag. Qhi = vihi. two serious drawbacks of this approach: u Numerical instability : If the A5 are too close.. The phenomenon occurs not least when the dimension p is large. vi. and we need to have access to software permitting calculations with complex numbers or to perform the cumbersome translation into real and imaginary parts.
21) Here the first term is the stationary limit and the second term thus describes the rate of convergence to stationarity. v2 and h2 can be computed in just the same way.Q2i and after some trivial calculus one gets eQt = 7r 1 112 + eat 7r1 7r2 / (7fl 7r2) = ( 7r2 1r2 7r1 IF. b are any constants ensuring//Irk = 1. Then Al = 0 and the corresponding left and right eigenvectors are the stationary probability distribution 7r and e. k  C k2 ) =b ( A1 q 1 Q11 / where a . replacing ai by A2.6 A particular important case arises when Q = q1 qi ) q2 q2 J is an intensity matrix. where (A. Then 7r = (ir1 7r2 ) = a (q21 Al . However. i. The other eigenvalue is A = A2 = q1 .7 Let 3 9 2 14 7 11 2 2 . l ab (g12g21 + (A1  411) 2) = 1.e.k1). u Example A3.19) Example A3 . it is easier to note that 7rh2 = 0 and v2k = 1 implies v2 = (k2 .APPENDIX 343 Write 7r (= v1) for the left eigenvector corresponding to a1 and k (= hl) for the right eigenvector. Of course.20) ir = q2 ql qi +q 2 9l +q2 (A. 1) . h2 = Thus. eqt = eNlt ( ir1ki i2k1 \ ir1 k2 72 k2 + e azt 7r2k2 i2k1 7ri k2 7r1 k1 (A.q.
5 . and a generalized inverse may not unique.satisfying AAA = A. (A+A)' = A+A. 2 2 1=ab(142+(1+2)2 ) = tab.23) .344 Then D= 2+ 11)' 7 T4 2 =52. (AA+)' = AA+. Generalized inverses play an important role in statistics. A+AA+ = A+... A2 = 3/2 . APPENDIX x1 3/2 . (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_.11/2 + 5 1. They are most often constructed by imposing some additional properties . but only that dimensions match .6.11/2 . for example AA+A = A. e_6u A4 Some linear algebra 4a Generalized inverses A generalized inverse of a matrix A is defined as any matrix A.22) Note that in this generality it is not assumed that A is necessarily square. (A.
if A is a possibly singular covariance matrix (nonnegative definite).1 goes under the name fundamental matrix of the Markov chain). Here is a typical result on the role of such matrices in applied probability: Proposition A4.25) .. E.APPENDIX 345 A matrix A+ satisfying (A. Then for some b > 0. Am+1 = .eir )1. (A.P + e7r ).eir ).24) = te7r . are ordered such that Al > 0. 0 01 In applied probability..ew. . and exists and is unique (see for example Rao [300]).. one is also faced with singular matrices .g. Assume that a unique stationary distribution w exists .D + O(ebt).eir)1 = I .1 Let A be an irreducible intensity matrix with stationary row vector it. Am > 0. = 0 where m < p is the rank of A. ( Q .. (I .23) is called the MoorePenrose inverse of A. These matrices are not generalized inverses but act roughly as inverses except that 7r and e play a particular role .e.= (I .g. Rather than with generalized inverses .P + e7r)1 (here ( I .. most often either an intensity matrix Q or a matrix of the form IP where P is a transition matrix. lt o eAx dx = te7r + D(eAt .I) (A. and can define /ail 0 0 0 0 0 0 A+ = C A' 0 0 0 C' . _ A. then there exists an orthogonal matrix C such that A = CDC' where 0 0 D = AP Here we can assume that the A .. one then works with Q = (Q .e ® 7r)1. and define D = (A .P).1Q = Q(Q . .
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. For example.3 Let 2 A= 4 3 Vf' N7 5 )' B= ( 8 ). .346 t APPENDIX 2 xe Ax dx = eir + t(D + e7r) + D(eAt . it follows that h ® it is the k x m matrix with ijth element hi7rj . Then A(O) _ B(O) = 0. resp.26) 2 = 2 e7r + tD .h. I. B(t) denote the l. see below. o Finally.91a(2) . h as 1 x m and k x 1 matrices.27) Proof Let A(t).s.eir)eAt = eAt = A'(t). and in fact any rank 1 matrix can be written on this form.DZ(ent .s.I) (A.e. .h.I) . the rows are proportional to it. and the columns to h. h ® it reduces to hit in standard matrix notation. Note that h ® it has rank 1. 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. ()®(6 f 6/ 7f 8^ 7 8 )=! ^)( 6 7 8 )=(6^ 7^ 8^) \ u Example A4.26) follows by integration by parts: t f t /' xeAx dx = [x {xe7r + D(eAx . (A. the formulas involving O(e6t) follow by PerronFrobenius theory.D + D2 + O(ebt).24). (A. B'(t) = e7r + DAeAt = eir + (I . of (A. respectively. in block notation i2h A®B= ( a11B a21 B a12B a22 B Example A4.I)} dx. Interpreting 7r. the r..2 Let it be a row vector with m components and h a column vector with k components.2e7r .I)}.J {xe^r + D(e . Equivalently.
28) In particular. A2 = v2 are row vectors and C1 = h1.k)! ( n0 n=0 t=0 k=0 J _ ® Ak ®Blk r ^.3v'6. each of which is A ® I or I ® B.3f 4v/.5v'8 5vf9 11 A fundamental formula is (A1B1C1) ®(A2B2C2) = (A1 (9 A2)(B1 (9 B2)(C1®C2).30) eA+B = eAeB function generalizes to Kronecker notation (note that in contrast typically only holds when A and B commute): Proposition A4. (A. (A B)' = eA®B e! L 1=0 0 . C2 = h2 are column vectors.31) Indeed. if A ® I occurs k times. then the Kronecker sum is defined by A(1) ®A(2) = A(1) ®Ik2 + k ®A(2).4vf.3V8.(A. it follows that e® ® e B An _ 0o oo oo Bn 7 I F n! = ` k! (I . Using (A.29). if Al = vi.4 eA® B = eA ®eB.29) If A and B are both square (k1 = ml and k2 = m2). such a factor is Ak (&B 1k according to (A.A9. then v1B1h1 and v2B2h2 are real numbers.3vV/72f 20. and v1B1h1 • v2B2h2 = v1B1h1 ® v2B2h2 = ( v1(&v2 )( B1(&B2 )( h1(&h2 ) .31). Proof We shall use the binomial formula A crucial property is the fact that the functional equation for the exponential t / l (A ®B)t = I k Ak 0 B1k k=0 (A. (A. and the number of such factors is precisely given by the relevant binomial coefficient.5v/.APPENDIX 347 Then A®B = 2 f 20. (AED B)1 = (A®I+I(9 B)l is the sum of all products of t factors.50 6 7 6 4f 4.
Then 2 0 ire At h • ve Bt kdt = (^®v)(A®B)1(e A®Ba . P(t) Yt(2) }. From what has been said about matrices of {Yt( 1). where transition matrix of the bivariate Markov chain {X n1).3 < 0 Lemma A4 . represents ces Q( 1). P8 = exp {sQ} = exp {s (Q(1) ®Q(2)) } . Let further it. we have P8 = Pal) ® p(2).I)(h ® k).6 Suppose that A and of B. Thus . and Q = Q(1) ® Q (2) = Q(1) ® I + I ® Q(2) (A. (A.33) .4 can easily be obtained by probabilistic be the sstep transition reasoning along the same lines . k any column vectors. first term on the r . Yt(2 ) }. the same time.32) is the intensity matrix of the bivariate continuous Markov process {Yt(1). Ps 1) = exp {sQ ( 1) } > p(2 ) = exp {sQ(2) } can therefore be rewritten as Taking s = 1 for simplicity . { 1't(1) }.s. resp .5 Many of the concepts and results in Kronecker calculus have p(2) is the intuitive illustrations in probabilistic terms. X ) }. { On the other hand. Q(2). v whenever a is an eigenvalue of A and 0 is an eigenvalue be any row vectors and h. in the definition (A. A special case of Proposition A4. Yt(2) where independent Markov processes with intensity matri{y(2) } are {Y(1) }. {Yt(1). n2 n1 ) {X(2) } are independent Markov chains with transition matrices P(1).32). P(2). P8 = Pal ) ® P82) exp {Q ( 1) ® Q(2)1 = eXp {Q( 1) } ® exp {Q(2) } Also the following formula is basic: B are both square such that a +. and the form of the bivariate intensity matrix reflects the fact that Yt(2) } cannot change state in both components at due to independence . Let P8f P(Sl). the {Yt(2) } transitions in the {Yt(1) } component and the second transitions in the component . h. independent Markov chains. p = P(1) ® {X }.348 APPENDIX Remark A4.
. h can be chosen with strictly positive elements. . so that by asssumption A ® B is u invertible. h = e and v = 7r (the stationary row vector).. see e.. Then: (a) The spectral radius Ao = max{JAI : A E sp(A)} is itself a strictly positive and simple eigenvalue of A.3 whenever a is an eigenvalue of A and 3 is an eigenvalue of B... we have AO = 1. Similarly. Now note that the eigenvalues of A ® B are of the form a +.. and appeal to (A.12). Then the eigenvalue Ao with largest real part is simple and real. [APQ] X. which can be found in a great number of books. . = j and atk_li. in such that io = i. A is called aperiodic if the pattern of zero and nonzero elements is the same as for an aperiodic transition matrix. i. and the corresponding left and right eigenvectors v. h can be chosen with 3By this.7 Let A be a p x pmatrix with nonnegative elements. We call A irreducible if the pattern of zero and nonzero elements is the same as for an irreducible transition matrix. 4c The PerronFrobenius theorem Let A be a p x pmatrix with nonnegative elements. then IN < Ao for all A E sp(A). .g.1 and references there (to which we add Berman & Plemmons [63]): Theorem A4. (b) if in addition A is aperiodic.29). il.APPENDIX 349 Proof According to (A.. and if we normalize v. . That is. we mean that the pattern of nonzero offdiagonal elements is the same as for an irreducible intensity matrix. then An = Aohv+O(µ") = Aoh®v+O(µ") for some u. .. E (0. the integrand can be written as ( 7r (9 v)( eAt ® eBt )(h ®k ) = ( 7r ®v)(eA (DBt)(h (& k). p there should exist io. . . (A.The PerronFrobenius theorem has an analogue for matrices B with properties similar to intensity matrices: Corollary A4.8 Let B be an irreducible3 p x pmatrix with nonnegative offdiagonal elements. ao). > 0 for k = 1. Here is the PerronFrobenius theorem. n.34) Note that for a transition matrix. . h such that vh = 1. f o r each i. j = 1. and the corresponding left and right eigenvectors v.
let {Yti°i } be a Markov process with initial distribution a and intensity . Ao). note that we can write the phase generator T as Q . not only in the tail but in the whole distribution.e. the phasetype distribution B(a) with representation (. To this end. but is an easy consequence of the PerronFrobenius theorem. Furthermore. The content is that B is approximately exponential if the exit rates ti are small compared to the feedback intensities tij (i # j).2). Example A3.8 is most often not stated explicitly in textbooks.35) for some p E (oo.350 APPENDIX strictly positive elements. 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). I. the analogy of this procedure with unformization. we have A0 = 0. let t = (ti)iEE # 0 have nonnegative entries and define T(°) = aQ . then eBt = ea0thv + O(eµt) = eA0th ® v + O(et t) (A. relate the eigenvalues of B to those of B via (A. Note that for an intensity matrix.(ti)diag where Q = T + (ti)diag is a proper intensity matrix (Qe = 0). A5 Complements on phasetype distributions 5a Asymptotic exponentiality In Proposition VIII.1. it was shown that under mild conditions the tail of a phasetype distribution B is asymptotical exponential. Corollary A4.(3. if we normalize v. h = e and v = 7r (the stationary row vector). h such that vh = 1. the condition is that t is small compared to Q. T(°)) is asymptotically exponential with parameter t* _ r EiEE aiti as a 4 oo..(ti)ding. For example. Then for any (3. Bi° (x) + at*x Proof Let { 4 } be the phase process associated with B(a) and (°) its lifelength. Proposition A5.1 Let Q be a proper irreducible intensity matrix with stationary distribution a. 10) and use the formula me at e Bt = e 00 Antn = e .8. The next result gives a condition for asymptotical exponentiality.n t AL n=0 n! (cf.
We can think of ( ( a) as the first event in an inhomogeneous Poisson process ( Cox process ) with intensity process matrix aQ . 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}. dx/ti] or not. 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. a' = a . from which the phase process is terminated . it states that the state. {t Y( a) } v>0 . Then a(a'V)/a (aV) a' 1.g.(a) > x . We can assume that Jta) = Yt(°).)_ = Y(a) = 1'aS(a) = Ya(av)^ it follows that Pi ((. we get dx F (Idx = j) = (1 + qij t )Sij + qij dt.jEE. = YQ(x). By the law of large numbers for Markov processes . and that Yt(a) = Yat for all t.Yj(av) = j f . prove a somewhat more general result which was used in the proof of Proposition VI. Since JJ(.9.aE where 0 < e < 1). a . in fact . fo tY dv/t a$' t*.1. has a limit distribution: Proposition A5. v/ t. and write Yt = Yt(1). Then {Ix} is a Markov process with to = Yo. Hence O ((a) aa. and this easily yields a(x)/x a' 1/t*. a'/a + 1.bij) Hence the intensity matrix of { Ix} is (qij/ti)i.2 Pi (c(a) > x. J(()) _ = i) + at•x t tt' . t < (a).x (1 . Proof Assume first ti > 0 for all i and let I. We shall .YQ(av) = j) Pi ( ci(a'V) > x. In addition to the asymptotic exponentiality.APPENDIX 351 ((1) etc. Let further V be exponential with intensity V and independent of everything else. Conditioning upon whether { Yt} changes state in [0.a' + oo (e. J^O)_ = j) Pi (v(aaV) > x.
5b Discrete phasetype distributions The theory of discrete phasetype distributions is a close parallel of the continuous case. . 2. Example A5. u Notes and references Propositions A5. Then: (a) The point probabilities are bk = aPklp.+ at*x • a't' L ` at t* t* J Reducing the state space of {Ix } to {i E E : t. a = b = (bk)k=1. so is the geometric distribution.4 Any discrete distribution B with finite support. zkbk is za(I .p)k1 p. ' pk 0 k>1 11 Theorem A5. Gnedenko & Kovalenko [164] and Glasserman & Kou [162]). (c) the nth moment k 1 k"bkis 1)"n!aP"p.j) and initial distribution a. > 0}. However.. Keilson [223]..zP)'p. See also Korolyuk. 1 k=1 1 0 otherwise.Pe.. 2.g. and thus the parameter p of the geometric distribution u can be identified with the exit probability vector p. K}.5 Let B be discrete phasetype with representation (P. the simplest discrete phasetype distribution: here E has only one element.2 do not appear to be in the literature. a) if B is the lifelength of a terminating Markov chain (in discrete time) on E which has transition matrix P = (p.1 and A5. (b) the generating function b[z] _ E' . Example A5. is discrete phasetype. so we shall be brief. a). these results are in the spirit of rare events theory for regenerative processes (e.} is said to be discrete phasetype with representation (E. k>1. Et II I a(a^V) > x) at' .. Penev & Turbin [238]..... Indeed. an easy modification of the argument yields finally the result for the case where t. k = 1. P. = 0 for one or more i.x k > K.. . let E and Pkj j=k1. . Then P is substochastic and the vector of exit probabilities is p = e . A distribution B on {1.352 rr Ia(a'V) Ei I ( > x) P APPENDIX L at (Yo (aV) . with point probabilities bk = (1 ..3 As the exponential distribution is the simplest continuous phasetype distribution. say bk = 0..
A. a. The discrete counterpart is the negative binomial distribution with point probabilities bk k1) (1 k = r. as is seen by minor modifications of Example A5. 11 Example A5. Jt t > U1 + U2.2 The form of these results is easily recognized if one considers two independent phase processes { Jt 1) }.. a' . T= ( 0 T(2) ) (A. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2). B2 be phasetype with representations (E(1). initial distribution a and phase generator T.1 This corresponds to a convolution of r geometric distributions with the same parameter p. _ i E E(1) T(1) t(1)a(2) i E E(2) .6 (CONVOLUTIONS) Let B1. { Jt 2) } with lifetimes U1 . r .a(2).T(1)).a(1)... . Then {Jt} has lifetime U1 + U2 .r + 1. U2.. (E(2).{ 0.6. Then the convolution B = B1 * B2 is phasetype with representation (E. and hence the negative binomial distribution is discrete phaseu type. and a=1).36) in blockpartitioned notation (where we could also write a as (a (1) 0)). resp.APPENDIX 353 5c Closure properties Example A5. 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. A reduced phase diagram (omitting transitions within the two blocks) is am E(1) t(1) a(2) (2) t(2) Figure A.6 is the Erlang distribution Er which is the convolution of r exponential distributions. resp.T(2)).
In risk theory.O)B2 (0 < 0 < 1) is phasetype with representation (E. we need to restart the phase process for B w.0)a(2) E(2) Figure A.E) where a(°) = fAa(a)v(da). a mixture of more than two phasetype distributions is seen to be phasetype. P(N = n) = (1 . Equivalently. T) and C = EO°_1(1 .37) (1) (1 . Then the mixture B = 9B1 + (1 .T. Then it is trivial to see that B(") is u phasetype with representation (a(").354 APPENDIX Example A5. a.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. and o'i Oa.0)ai2). i E E(1) T 0 I (A. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2).. one obvious interpretation of the claim u size distribution B to be a mixture is several types of claims. p at each termination. To obtain a phase process for C.8 (FINITE MIXTURES) Let B1. i E E(2) 0 T(2) =IT (in blockpartitioned notation.T(2)). resp.i.'). and consider B(") = fA B(a) v(da) where v is a probability measure on A.0)a(2))). Let B(") be the corresponding phasetype distribution. are i.p)pn1.4 .3 In exactly the same way. with common distribution and N is independent of the Uk and geometrically distributed with parameter p. U2. Example A5. a.. then C is the distribution of Ul + • • • + UN. Example A5. this means that a = (Oa(1) (1 .T(1)). A reduced phase diagram is 0a(1) E(1) A . if U1. B2 be phasetype with representations (E(1). Thus.p.d.a(1). (E(2).p)pn1B*n..10 (GEOMETRIC COMPOUNDS) Let B be phasetype with representation (E.a(2). a reduced phase diagram is f a E t Figure A.
T). cf. a(2). { 4 } as exit of {Jt}. then U1 +• is phasetype with representation (E. f2. Equivalently.d. Minor modifications of the argument show that 1. j E F}. then C is the distribution of U1 + • • • + UN. To obtain a phase representation for C .. of F. To see this. E). resp.2.g. then U1 + • • + UN is zeromodified phasetype with representation (a. v. Example A5 . T(1) ® T(2)). Proposition VIII. a.x)+. i. a.. if {Jt} is a phase process for U. 13 (MINIMA AND MAXIMA ) Let U1. . U2 be random variables with distributions B1.1.11 (OVERSHOOTS) The overshoot of U over x is defined as the distribution of (U . T + pta).a(1). P)..f. It is zeromodified phasetype with representation (E.T) if U is phasetype with representation (E. { Jt2) } be independent with lifetimes U1. say with distribution F. If U1 has a different initial vector. it follows by mixing (Example A5.. say v.. T) and C = F. we then let the governing phase process be {Jt} _ {(411 Jt2))} 2) interpreting exit of either of {4 M }. +UN 2. if U1.2. Example A5. a. T(2) ). (E(2). Note that this was exactly the structure of the lifetime of a terminating renewal u process. with common distribution B and N is independent of the Uk with P(N = n) = f.. then Jy has distribution aeTx. v. X independent of U. let the initial vector be a ® v and u let the phase generator be I ® T + P ® (ta).T) where F[T] = J0 "o eTx F(dx) u is the matrix m.APPENDIX 355 and C is phasetype with representation (E.v. a(1) ® a(2 ). U2. B2 of phasetype with representations (E('). T + ta.9) that (U . but the same T. For U1 A U2.aeTx. If we replace x by a r.°. Example A5 .T + pta). ..X)+ is zeromodified phasetype with representation (E.TWWW). Corollary VIII.aF[T]. 12 (PHASETYPE COMPOUNDS ) Let fl. be the point probabilities of a discrete phasetype distribution with representation (E. U2. let {Jtl)}. cf. let the phase space be E x F = {i j : i E E. Thus the representation is (E(1) x E(2). let B be a continuous phasetype distribution with representation (F. if B is defective and N + 1 is the first n with U„ = oo.°_1 f„ B*?l..7. Indeed. Then the minimum U1 A U2 and the maximum U1 V U2 are again phasetype. are i. resp.
oo). Now we can find first a sequence {Dm} of distributions with finite support such that D. and vice versa.(Sn) with Sn = n/b.8. Then from above. r # oo.B(bk) I < 1/n for n > k. The mean of B„ is n/Sn = b and the variance is n/Sn = b2/n. the initial vector is (a(1) (& a (2) 0 0). The general case now follows easily from this. oo) can be approximated 'arbitrarily close' by a phasetype distribution B: Theorem A5.(bk) + B(bk) for all k.(n) = D.(bk) + B(bk) for all k as n * oo.. That is. Thus the state space is E(1 ) x E(2) U E(1) U E( 2). and let Bn be the Erlang distribution E..356 APPENDIX For U1 V U2. 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 . we need to allow { Jt. we can assume that ID. Here are the details at two somewhat different levels of abstraction: (diagonal argument .. elementary) Let {bk} be any dense sequence of continuity points for B(x). any distribution B on (0.. and the closedness of the class of phasetype distributions under the formation of finite mixtures.14 To a given distribution B on (0. 5d Phasetype approximation A fundamental property of phasetype distributions is denseness . i= 1 C.. Proof Assume first that B is a onepoint distribution.xq(n)(n)}. Hence it is immediate that Bn 4 B.} of phasetype distributions such that Bn 3 B as n + oo. Let the support of Dn be {xl(n)..(bk)'. there is a sequence {B.n = I:pi(n)Er v ( __ ) n) ) a= 1 . see Neuts [269] (where the proof. By the diagonal argument (subsequent thinnings). say degenerate at b. Then we must find phasetype distributions Bn with B. with weight pi(n) for xi(n). cf..2) } to go on (on E(2)) when { i 1) } exits. the fact that any distribution B can be approximated arbitrarily close by a distribution with finite support. however. Example A5. q(n) q(n) pi(n)a .. relies more on matrix algebra than the probabilistic interpretation exploited here).
and that cp is known to be continuous.. that this procedure should be used with care if ^p(B) is the ruin probability O(u) and u is large. k < n.r. Since PET is closed under the continuous operation of formation of finite mixtures. oo) approximation Assume that we can compute a functional W(B) when B is phasetype. the topology for weak convergence) PET of the class PET of phasetype distributions contains all onepoint distributions. if information on Bo is given in terms of observations (i. E E.. Corollary A5. But To is the class G of all distributions on [0.d. one would use the B given by some statistical fitting procedure (see below). f2.e. If Cpl (B) and ^02(B) are weakly continuous.n( b k ) . Let E be the class of functions f : [0. replications). the class CO of all discrete distributions.B(bk )I < . say on the claim size distribution B in risk theory.n (bk) .t. In particular. . for some a < oo. oo) * [0. i. then it is immediate that WI(B) = p2(B) for all distributions B on [0. u Theorem A5.n. 2. It should be noted... PIT contains all finite mixtures of onepoint distributions. x 4 oo. oo) and any fl.. however. and we can take Bn = Cr(n).D(bk)I < n. 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.. k < n.i.(x)Bf.. For a general Bo. oo).. u 2 (abstract topological ) The essence of the argument above is that the closure (w.APPENDIX 357 Hence we can choose r(n) in such a way that ICr( n). compute W(B) and use this quantity as an approximation to cp(B0). i = 1.14 is fundamental and can motivate phasetype assumptions. Hence G C PET and L = PIT. Then ICr( n ). we can then approximate Bo by a phasetype B.( dx) * f r f{(x)B(dx). oo) such that f (x) = O(e«x).15 To a given distribution B on (0 . there is a sequence {Bn} of phase type distributions such that Bn Di B as n 4 oo and f ' f.
f (x)B(dx). for each i.. Bn=En z f f (x)Bn(dx) fof (x)B(dx) = ° (A. liminf B. \\ 0 Corollary A5. i = 1. ...14 Dn has been chosen such that 00 1 °° f fi(x)D n(dx ) < 1++ '  o \ n o f fi(x)B(dx). n. i=1. By (A.. we may assume that in the proof of Theorem A5. i = 1. .39). oo). then cc f (x)Bn ( dx) = (?!c ) e'= ..16 To a given distribution B on (0 .f ' f (x)B(dx). f° xtBn(dx ) * f °° x`B( dx). . and the case of a general f then follows from the definition of the class E and a uniform integrability argument. and hence it is sufficient to show that we can obtain limsup n4oo fi(x)Bn(dx) < Jo 0 f fi( x)B(dx ). 2. there is a sequence {Bn} of phase type distributions such that Bn Di B as n + oo and all moments converge.. f00 fi(x)Cr. Now returning to the proof of (A. n B=az. TO (A..n(dx) + f 0 fi(x)Dn(dx).  APPENDIX B implies that 00 o o 00 n. i = 1.oo J fi(x)B..38) We first show that for each f E E.f (z) = f = 1 1 1 1n/ o .2 .358 Proof By Fatou' s lemma..39) Indeed. ..n(dx) < 1+. if f (x ) = e°x. n.38 ).(dx) > J fi(x)B(dx).... and hence we may choose r(n) such that L 9l) f (x)Cr(n). .
However.l3µb < 1.} of phasetype distributions such that Bfz + B as n * oo and Yn 4 ry where ryn = y(Bn. and therefore the following result is highly relevant as support for phasetype assumptions in risk theory: Corollary A5. from a more conceptual . 0 as i * oo. Notes and references Theorem A5. We shall formulate the problem in the slightly broader setting of fitting a phasetype distribution B to a given set of data (1i . The present section is a survey of some of the available approaches and software for inplementing this.14 is classical.17 To a given /3 > 0 and a given distribution B on (0... then Bn['Y + ei] * B[y + ei] > 1 + 7 Q implies that 'yn < ry + ei for all sufficiently large n . (N. the remaining results may be slightly stronger than those given in the literature.16. (N or a given distribution Bo. I.e.> y for some sequence {ei} with ei E (0. there is a sequence {B. there is substantial advantage in assuming the claim sizes to be phasetype when one wants to compute ruin probabilities. This is motivated in part from the fact that a number of nonphasetype distributions like the lognormal. For practical purposes. 5e Phasetype fitting As has been mentioned a number of times already./3). . . . but are certainly not unexpected.. . The adjustment coefficient is a fundamental quantity. and in part from the fact that many of the algorithms that we describe below have been formulated within the setup of fitting distributions./3) is defined as the unique solution > 0 of B[y] = l+y/j3. . oo) with B[y +e] < oo for some e > y = 7(B. /3) = ry for all n. the problem thus arises of how to fit a phasetype distribution B to a given set of data (1. O We state without proof the following result: Corollary A5. one can obtain 7(Bn. the loggamma or the Weibull have been argued to provide adequate descriptions of claim size distributions.3).APPENDIX 359 In compound Poisson risk processes with arrival intensity /3 and claim size distribution B satisfying . If ei > 0.18 In the setting of Corollary A5. . Proof Let fi(x) = el'r+E. lim inf > is proved similarly. lim sup ryn < 7. . the adjustment coefficient 'y = 7(B. e ) and ei J.
d.f.. we have constructed a sequence { B. The likelihood function is maximized by a local linearization method allowing to use linear programming techniques.g.} of phasetype distribution such that Bo. In a series of papers (e. and this is what matters when using phasetype distributions as computational vehicle in say renewal theory. d.. we do not not want to perform matrix calculus in hundreds or thousands dimensions). defined by the absence of loops in the phase diagram . It seems therefore a key issue to develop methods allowing for a more general phase diagram. [202]. The characteristics of all of these methods is that even the number of parameters may be low (e..g . three for a mixture of two Erlangs ). Asmussen & Nerman [38] implemented maximum likelihood in the full class of phasetype distributions via the EM algorithm .. and we next describe two such approaches which also have the feature of being based upon the traditional statistical tool of like maximum likelihood.'s). [70]) restrict attention to acyclic phase type distributions . [216] ). Schmickler (the MEDA package. 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. (N is the empirical distribution Be. giving mass 1 /N to each S=. reliability or queueing theory.g.360 APPENDIX point of view the two sets of problems are hardly different : an equivalent representation of a set of data (1 . 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. g. one could argue that the results of the preceding section concerning phasetype approximation contains a solution to our problem : given Bo (or Be). The observation is that the statistical problem would be straightforward if the whole ( EAvalued) phase process { Jtk)} o<t<( k associated with each observa .g. Johnson & Taaffe considered a mixture of two Erlangs (with different rates ) and matched (when possible ) the first three moments . the number of phases required for a good fit will typically be much larger. A number of approaches restrict the phase type distribution to a suitable class of mixtures of Erlang distributions . cf. 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. Of course. and in practice this sets a limitation to the usefulness (the curse of dimensionality . e . and used a nonlinear programming approach . A method developed by Bobbio and coworkers (see e. and as fitted distribution we may take B. . at a a number of selected points . [317] ) has considered an extension of this setup.f. the L1 distance between the c . for some suitable large n. risk theory. . B„ The problem is that the constructions of {B„} are not economical : the number of phases grows rapidly.
..T (n)(TiI(1. it is easy to see that N (k Ea(n). (N) tJk Ea ( n). EN where ai = N 1 I ((k) = i) tii=i iEE. e.T(n) (Nik IC1. eieT(n)((k.. = j) f k=1 k =1 tE[0. it seems open whether the restriction to the acyclic case is a severe loss of generality. (n+1) _ Ea (n). The general idea of the EM algorithm ([106]) is to replace such unobserved quantities by the conditional expectation given the observations..(N) = E Ea(n). the methods of [70] and [38] appear to produce almost identical results. .(k] (Ti is the total time spent in state i and Nii is the total number of jumps from i to j).T(n) k=1 I (Jti) dt o \f a(n)eT(n )(kt(n) N f:i a(n)eT(n)xei . then the estimators would be of simple occurenceexposure type..g.g. jEEA.. one is lead to an iterative scheme. In fact. . .. N Ti = I(J= i) dt.. E. . Thus.. (N ) (^ 54 k )+ and similarly for the cn+1) The crux is the computation of the conditional expectations.APPENDIX 361 tion Sk was available.x)t(n) 1 and this and similar expressions are then computed by numerical solution of a set of differential equations..T(n) (Ti ^^ 1. In practice. since this is parameterdependent. Nii = = .
This page is intentionally left blank .
Stegun ( 1972 ) Handbook of Mathematical Functions (10th ed. New York. Queueing Systems 5 . New York. [14] S. 14. [3] J. 587. Chichester New York. Asmussen ( 1982 ) Conditioned limit theorems relating a random walk to its associate. Adv.). Probab. 20. [7] E. G.Bibliography [1] J. Whitt (1998) Explicit M/G/1 waitingtime distributions for a class of longtail service time distributions .). Arfwedson ( 1955) Research in collective risk theory. Arfwedson ( 1954) Research in collective risk theory. Preprint. [11) S. 191223. 37. 345368. Skand. [2] J. Abramowitz & I. AT&T. 38. 143170. Appl. Abate. J. 53100. 311338. Choudhury & W. Asmussen (1985) Conjugate processes and the simulation of ruin problems. 1985. Sparre Andersen (1957) On the collective theory of risk in the case of contagion between the claims . Dover. [9] G.A. Aktuar Tidskr. Optimizations Software. Queueing Systems 16. Anantharam (1988 ) How large delays build up in a GI/GI11 queue. pp. [15] S. 363 . [5] G. Transactions XVth International Congress of Actuaries. ed. Queueing Systems 10. Asmussen (1984) Approximations for the probability of ruin within finite time . II. New York. 101116. 253267. Arndt (1984) On the distribution of the supremum of a random walk on a Markov chain . Whitt (1994) Waitingtime tail probabilities in queues with longtail service time distributions . ibid. [4] M. Aktuar Tidskr. 213229. John Wiley & Sons. Scand. Borokov. [8] G. B. The Mathematical Scientist 14. Appl.L. [12] S. In: Limit Theorems and Related Problems (A. Alsmeyer (1991) Erneuerungstheorie. Stoch. Abate & W. Abate & W. Stuttgart. [13] S. with applications to risk reserve processes and the GI /G/1 queue. Teubner. 219229. Asmussen (1989a) Aspects of matrix WienerHopf factorisation in applied probability. [10] K. Asmussen (1987) Applied Probability and Queues. 1984 .G. Skand. [6] V. Proc. 57. Whitt ( 1992) The Fourierseries method for inverting transforms of probability distributions . The case of equal risk sums. 3157 . Act.
Chakravarty. Frey. A. Stoch. Hojgaard (2000) Rare events simulation for heavytailed distributions. 1936. Asmussen (1992a) Phasetype representations in random walk and queueing problems. Rolski & V. 189201. Probab. Asmussen & K. MatrixAnalytic Methods in Stochastic Models (A. Ann. Asmussen (1998c) A probabilistic look at the WienerHopf equation. Riga Aviation University. Act. Asmussen (1991) Ladder heights and the Markovmodulated M/G/1 queue. Asmussen & B. [30] S. 2149. Ann. Asmussen (1989b) Risk theory in a Markovian environment. [33] S. J. 96107.S. 4966. 1999 . T. Bernoulli 6. stationary distributions and first passage probabilities. J. 303322. Asmussen (1999) On the ruin problem for some adapted premium rules. Asmussen & B.M. Methods V Problems (J. Hojgaard (1999) Approximations for finite horizon ruin prob abilities in the renewal model.K. Dshalalow ed. SIAM Review 40. Asmussen (1992b) Light traffic equivalence in single server queues. 193226. Asmussen. Kliippelberg (1994) Large claims approximations for risk processes in a Markovian environment. Appl. [23] S. Binswanger & B. Stoch. [34] S. 315. CRC Press. 37. Probabilistic Analysis of Rare Events (V. [20] S. [18] S. Scand. 297318. [25] S. Random Processes 2. Asmussen. eds. [26] S. Alfa & S. Hojgaard (1996) Finite horizon ruin probabilities for Markovmodulated risk processes with heavy tails. Andronov. 1996. [28] S. Scand. 25. Kalashnikov & A.).364 BIBLIOGRAPHY [16] S. 2. Asmussen (1995) Stationary distributions via first passage times. 313326. Schmidt (1995) Does Markovmodulation increase the risk? ASTIN Bull. 313341. [19] S. Probab. [17] S. [31] S. [21] S. Th. Scand. Proc. Statist. eds. Extremes 1 . 69100.). Act. Binswanger (1997) Simulation of ruin probabilities for subexponential claims. Asmussen (1998b) Extreme value theory for queues via cycle maxima. Appl. 2943. J. Marcel Dekker. [29] S. Probab. Bladt (1996) Renewal theory and queueing algorithms for matrixexponential distributions. Ann. J. Florida. [32] S. K. 772789. 20. Asmussen (1998a) Subexponential asymptotics for stochastic processes: extremal behaviour. Act. [27] S. New York. Floe Henriksen & C. [24] S. Appl. 137168. . 555574. 1989 . Asmussen & M. Appl. L. 106119. Asmussen (2000) Matrixanalytic models and their analysis. Boca Raton. Bladt (1996) Phasetype distributions and risk processes with premiums dependent on the current reserve.). Asmussen. 354374. Astin Bulletin 27. [22] S. Asmussen & M. Advances in Queueing: Models. Proc. 8. Scand. 27. 79102. 54. Stochastic Models 11. Asmussen (1992c) Stationary distributions for fluid flow models and Markovmodulated reflected Brownian motion.
Eng. Insurance: Mathematics and Economics 20. Rubinstein (1999) Sensitivity analysis of insurance risk models. Insurance: Mathematics and Economics 10. Adv. Asmussen & S. Dshalalow ed. J. Appl. Sc. Adv. Asmussen & R. Asmussen & D. Asmussen & C. . Asmussen & V. (to appear). [36] S. Probab. Wiley. [47] S. [48] S. Asmussen & R. J. Opns. 11251141. 31. 365372. 913916. Boca Raton. 421458. Proc. [38] S. Statist.A. Koole (1993). [51] S. 19. first passage problems and extreme value theory for queues. 105119. Oper. Probab. Math. Taksar (2000) Optimal risk control and dividend distribution policies . Perry (1992) On cycle maxima. Schock Petersen (1989) Ruin probabilities expressed in terms of storage processes. Advances in Queueing: Models. 419441. Statistica Neerlandica 47. with applications to insurance risk Stoch. 20. Appl. H. Supplementary Volume (Kotz. Nielsen (1995) Ruin probabilities via local adjustment coefficients . Proc. Stochastic Models 8. Asmussen & K. Asmussen & V. [45] S. Olsson (1996). 259274. [37] S. [40] S. 429466. Nerman & M. B. Res.Y. 120. Schmidt (1993) The ascending ladder height distribution for a class of dependent random walks. Asmussen & G.BIBLIOGRAPHY 365 [35] S. Taksar (1997) Controlled diffusion models for optimal dividend payout. Marked point processes as limits of Markovian arrival streams. Asmussen & C.M.). Read eds. 410433. [52] S. 58. Appl. Rolski (1991) Computational methods in risk theory: a matrixalgorithmic approach. Asmussen & T. Probab. Asmussen & M. [42] S. [41] S. Management Science 45. Th. [39] S. 30. Methods 8 Problems (J. 23. 10.). Rolski (1994) Risk theory in a periodic environment: Lundberg's inequality and the CramerLundberg approximation.Y. Johnson. Stoch. Sigman (1996) Monotone stochastic recursions and their duals. Finance and Stochastics 4. 115. Probab. O. Res. Appl. 64. Example of excessoffloss reinsurance for an insurance corporation. O'Cinneide (2000/2001) Matrixexponential distributions [Distributions with a rational Laplace transform] Encyclopedia of Statistical Sciences. 736755. [43] S. O'Cinneide (2000/01) On the tail of the waiting time in a Markovmodulated M/G/1 queue. Asmussen & T. Schmidt (1999) Tail approximations for nonstandard risk and queueing processes with subexponential tails. Asmussen.A. 103125. Schmidli & V. CRC Press. Scand. 299324. Asmussen & H. Rubinstein (1995) Steadystate rare events simulation in queueing models and its complexity properties. Appl. Asmussen . Fitting phasetype distributions via the EM algorithm. 422447. Kliippelberg (1996) Large deviations results for subexponential tails. Asmussen & C. J. [49] S. [46] S. Schmidt (1995) Ladder height distributions with marks. Florida. Appl. [50] S. Hojgaard & M. Asmussen. Prob 32 . [44] S. Inf.
Cambridge University Press. New York. Beekman (1985) A series for infinite time ruin probabilities. [67] T. Schmidli (1995) Saddlepoint approximations for the probability of ruin in finite time. C. Teugels (1987) Regular Variation. [69] P. Act. Asmussen & J.T. Springer . 661667.M. Goldie & J. J. Scand. [71] P. J. 221232. Boogaert & A. 148156. [58] O. [66] N. Act. 1985 . Baker (1977) The Numerical Solution of Integral Equations. Berlin.L. 610. Athreya & P. Bingham (1975) Fluctuation theory in continuous time. Oxford. J. [62] J. [72] P. Grandell (1985) An insensitivity property of the ruin probability. Bobbio & M. Bingham. Probab. 4148. Beekman (1974) Two Stochastic Processes. Bjork & J. Grandell (1988) Exponential inequalities for ruin probabilities in the Cox case. Wiley. Insurance: Mathematics and Economics 4. SpringerVerlag. New York. Trans. Soc.H. Teugels ( 1996) Convergence rates for M /G/1 queues and ruin problems with heavy tails . [61] J. Crijns (1987) Upper bounds on ruin probabilities in case of negative loadings and positive interest rates. 275279.H. Scand. [64] P. 169186. Billingsley (1968) Convergence of Probability Measures.H. Appl. Stochastic Models 10. Appl. Probab.A. [63] A. Scand. Boogaert & V. Borovkov (1976) Asymptotic Methods in Queueing Theory. Scand.366 BIBLIOGRAPHY [53] S. Telek (1994) A bencmark for PH estimation algorithms: results for acyclic PH. 832840. J. [59] O. 1995. 9599. Wiley. New York.R. Ney ( 1972) Branching Processes . [54] K.B . Appl.A. 11811190.J. [60] J. 9. 190204. . [55] B. [56] B. [68] T. BarndorffNielsen (1978) Information and Exponential Families in Statistical Theory. Scand. [73] A. 1988 . Berman & R. 77111. Insurance: Mathematics and Economics 9. 705766. [65] N. Plemmons (1994) Nonnegative Matrices in the Mathematical Sciences. J. [57] C. von Bahr (1975) Asymptotic ruin probabilities when exponential moments do not exist. Bloomfield & D. Bjork & J. Halsted Press. Clarendon Press. Actuaries 21. Probab. Adv. J. 1974. 33.L. Insurance: Mathematics and Economics 6. 129134. Cambridge. de Waegenaere (1990) Simulation of ruin probabilities. J. SIAM. [70] A. Act. 1975. Act. Cox (1972) A low traffic approximation for queues. 7. von Bahr (1974) Ruin probabilities expressed in terms of ladder height distributions. Beekman (1969) A ruin function approximation. Act. BarndorffNielsen & H.
907920. Wiley. Cohen (1998) The M/G/1 queue with heavytailed service time distribution. [76] N.I. [88] J. Berlin. Reading. [78] L. 3539.A.BIBLIOGRAPHY 367 [74] O. [80] F. Proc. Area Commun. [79] P. Jones & C.C. Geb. Nesbitt (1986) Actuarial Mathematics . 392433. Malgouyres (1983) Large deviations and rare events in the study of stochastic algorithms . Boxma & J.M. 127130.). Soc. Fort & G. 62. 51. 14331453. D. Schrage (1987) A Guide to Simulation. Gerber.W. Herzog (1977) The phase concept : approximations of measured data and performance analysis. Burman & D. [86] K. NorthHolland. 105119. IEEE Trans. Croux & N. Amsterdam.R. Fox & L. J. 93121. Probab. Illinois.Y. Tech. Reiser eds. Biihlmann ( 1970) Mathematical Methods in Risk Theory.A. Bucklew ( 1990) Large Deviation Techniques in Decision. [77] P. Jr. Amsterdam. Stockholm. Adv. H.L. J.. Computer Performance (K. Goovaerts & F. Bratley. Bell. Cramer (1955) Collective risk theory. Academic Press. Chandy & M. Insurance : Mathematics and Economics 5. Sel. . Z. SpringerVerlag. Veraverbeke (1990). 34. Aut. [84] D. 14. M. [83] D.U.Y . Queueing Systems 33. Syst. Insurance : Mathematics and Economics 9. Bux & U. Wahrscheinlichkeitsth. [92] H. [91] H. Springer. 24. Resnick & R. Oper. Philos.R. IEEE J.L. Appl. New York.J. Cox (1955) Use of complex probabilities in the theory of stochastic processes. Control AC28.R. New York. Smith ( 1986) An asymptotic analysis of a queueing system with Markovmodulated arrivals . AddisonWesley. [81] J. Burman & D . Skandia Jubilee Volume.L. verve. [82] H. Itasca. 2338.) NorthHolland. Broeck .L.). Cohen (1982) The Single Server Queue (2nd ed. Hickman. Brockwell . [90] D. Res. Simulation and Estimation. [87] E. 177204. 749763.J.C. II. [93] K. [75] O. Cambr. Stockholm. Breiman ( 1968) Probability. 313319.W. Cramer (1930) On the Mathematical Theory of Risk.W. S. 16. Nonparametric estimators for the probability of ruin. J. Tweedie (1982) Storage processes with general release rule and additive inputs . Cohen (1999) Heavytraffic analysis for the GI/G/1 queue with heavytailed service time distributions . B. Smith ( 1983) Asymptotic analysis of a queueing system with bursty traffic. Cottrell. The Society of Actuaries . Boxma & J. [89] M. New York San Francisco London. De Vylder (1986) Ordering of risks and ruin probabilities. Chung (1974) A Course in Probability Theory (2nd ed. [85] W. Bowers . The Jubilee volume of Forsakringsbolaget Skandia. Qinlar (1972) Markov additive processes.
Editions de 1'Universite de Bruxelles. Devroye (1986) Nonuniform Random Variate Generation. Forsiikringsmatematiska Studier Tillagnade Filip Lundberg. Dassios & P. Steinebach ( 1991 ) On the estimation of the adjustment coefficient in risk theory via intermediate order statistics . Blotter der deutschen Gesellschaft fur Versicherungsmathematik XVII. [112] F. J. Oper.L. T. [109] F. Res. 17. Oper. Zeitouni (1992) Large Deviations Techniques and Applications. Scand. [100] A. Act. Daley & T.P. Teugels ( 1990) Empirical Laplace transform and approximation of compound distributions . Daley & T. Comp. Pentikainen & E. 227229. 435436. 121131. Embrechts (1989). Laird & D. De Vylder & M. Rolski (1984) A light traffic approximation for a single server queue. Haezendonck (1987) Classical risk theory in an economic environment. De Vylder & M. Dembo & O. [102] P. Jones and Bartlett. De Vylder (1978) A practical solution to the problem of ultimate ruin probability. Res. Springer. Csorgo & J. Goovaerts (1984) Bounds for classical ruin probabilities. New York. Boston. 1990 . Delbaen & J. Aktuar. [105] A. 3. 135159. [96] D. Scand. Insurance: Mathematics and Economics 4. [108] F. 9 . Rolski (1991) Light traffic approximations in queues. 7083 (1969). Davidson (1946) On the ruin problem of collective risk theory under the assumption of a variable safety loading (in Swedish). Suppl.D. [95] M. [104] F. Statist. 88101. van Dawen (1986) Ein einfacher Beweis fur Ohlins Lemma. Rubin (1977) Maximum likelihood from incomplete data via the EM algorithm. Daykin. Delbaen & J. [103] F. Roy. Insurance: Mathematics and Economics 7. 16. N. . English version published in Skand. Math. Insurance: Mathematics and Economics 6. Dempster. Insurance: Mathematics and Economics 3. Pesonen (1994) Practical Risk Theory for Actuaries. Goovaerts (1988) Recursive calculation of finitetime ruin probabilities. Appl. Stockholm. 3750. 85116. Haezendonck (1985) Inversed martingales in risk theory. Martingales and insurance risk. [107] L. 27. J. 181217. De Vylder (1996) Advanced Risk Theory. [101] C. Tidskr. Deheuvels & J. Probab. J. Math. [97] D. Appl.. Math. Stochastic Models 5. Csorgo & J. Chapman & Hall. Steinebach (1990) On some alternative estimators of the adjustment coefficient in risk theory. J. 5771. Soc. [99] R. De Vylder (1977) A new proof of a known result in risk theory. [106] A. Actuarial J..B. [110] F. 583602. [111] F. [98] A. 22. 624628. 114119.M. Insurance: Mathematics and Economics 10.368 BIBLIOGRAPHY [94] M. 201206.
1984 . Ruin estimates for large claims .J. 1993 . Dickson & J. [114] D. Astin Bulletin 21. J. Dickson & J.C. Scand. Insurance : Mathematics and Economics 22. Dickson (1994) An upper bound for the probability of ultimate ruin. Scand. Gray (1984) Approximations to the ruin probability in the presence of an upper absorbing barrier. Dickson (1992) On the distribution of the surplus prior to ruin . J. Dickson & H.W. 107124. Insurance : Mathematics and Economics 11. Camb. 118. Embrechts .M. Dickson & C. Hipp (1998) Ruin probabilities for Erlang (2) risk processes.C.R.K.M. Gerber & E. . J. 193199. 1741. 177192. Emanuel . 1994 .C. Dynkin (1965) Markov Processes I. O'Connell (1995) Large deviations and overflow probabilities for the general singleserver queue. J. [117] D. 147167. 105115. Scand.C. Harrison & A. Insurance : Mathematics and Economics 10. 229232. J. 6180. Dickson (1995) A review of Panjer's recursion formula and its applications. Act. British Actuarial J. J. [115] D. Dickson & H. 174186. 1975 . [129] D . 3745. Duffield & N. [126] F.M. Act. J. Act. Insurance : Mathematics and Economics 7. Soc.C. Gerber (1988) The surpluses immediately before and at ruin. Berlin Gottingen Heidelberg. Hipp (1999) Ruin problems for phasetype(2) risk processes.S. [119] D. 2000 .R.M. [120] D.U. 1993 .G.B. Insurance: Mathematics and Economics 25. Schmidli (1993) Finitetime Lundberg inequalities in the Cox case . 4962. [127] F.M. J.U.U.C.J. [121] D. [130] P. Insurance: Mathematics and Economics 7. Math. Djehiche (1993) A large deviation estimate for ruin probabilities. Dufresne & H. with applications.M. Taylor ( 1975) A diffusion approximation for the ruin probability with compounding assets.C. Scand.C.R. J.M. Philos. Dickson & C. van Doorn & G. [122] B. Insurance: Mathematics and Economics 19. [125] F. Act. Act. Dufresne & H. Embrechts ( 1988 ).M. Springer . [118] D. Shiu (1991) Risk theory with the Gamma process. Res. Scand.C. 191207. Scand. [116] D. H. Proc. 1984 .M. Act. Waters (1999) Ruin probabilities wih compounding. 5159. [123] E. and the amount of claim causing ruin. 363374.BIBLIOGRAPHY 369 [113] D. Oper. Letters 7. [124] N. Dufresne. Gray (1984) Exact solutions for ruin probability in the presence of an upper absorbing barrier.R. Grandell & H. 251262. Gerber (1991) Risk theory for the compound Poisson process that is perturbed by a diffusion.M. 131138. Act.C. 1. [131] P. [128] E. 269274. Waters (1996 ) Reinsurance and ruin . Regterschot (1988) Conditional PASTA. Scand. 4259.
[144] P. Reprinted 1948 as 'The theory of probabilities and telephone conversations' in The Life and work of A. 181190. On the excursions of Markov processes in classical duality. Skand. Insurance : Mathematics and Economics 1. Griibel & S. J. [148] C. Th. J. Konig. 5572. Probab. first passage times and moments of ladder variables in Markov random walks. Probab. Nyt Tidsskrift for Matematik B20. [135] P. 14661474. [146] E. Schmidli (1994) Ruin estimation for a general insurance risk model. Probab. . Probab. Insurance: Mathematics and Economics 10. Appl. Villasenor (1988) Ruin estimates for large claims. Lai (1998) Wald's equations. Franken.D. Tidsskr. Sci. Proschan & D. [149] C. Wiley. Appl. New York. Appl. 29. with applications. [137] P.K. U. Embrechts.). M.370 BIBLIOGRAPHY [132] P. Wiley. Insurance: Mathematics and Economics 7. Embrechts. Mikosch (1991). 38. 404422. Danish Academy Tech. 695712. [142] W. Wiley. Pitts (1993) Some applications of the fast Fourier transform in insurance mathematics . [133] P. Esscher (1932) On the probability function in the collective theory of risk.W. 623637. Mikosch (1997) Modelling Extremal Events for Finance and Insurance. [138] P. 175195.J. Astin Bulletin 16. Statist. 5975. Embrechts & J. Frenz & V. [134] P. Trans.W. Akt.A. F. 26. Probab. Scand. [145] P. 29. Springer. Schmidt (1982) Queues and Point Processes. 269274. 17.). D. J. 159178. [141] F. 1998 . Fuh (1997) Corrected ruin probabilities for ruin probabilities in Markov random walks. Veraverbeke (1982) Estimates for the probability of ruin with special emphasis on the possibility of large claims . 5974.D.M. Feller (1966) An Introduction to Probability Theory and its Applications I (3nd ed. 35. Fields 75. Erlang 2. [136] P. Approximations for compound Poisson and Polya processes. Embrechts & T. 8190. [139] A. Adv. Appl. Frees (1986) Nonparametric estimation of the probability of ruin. Fitzsimmons (1987). Rel. Embrechts & N. Adv. Act. 3339. Heidelberg. Probab. J. Schmidt (1992) An insensitivity property of ladder height distributions. Maejima & J. Embrechts. Appl. Statistica Neerlandica 47. New York. R.K. A bootstrap procedure for estimating the adjustment coefficient. Adv. Walkup (1967) Association of random variables. Fuh & T. Esary. [147] M. Math. C. Ann.L. [150] H. Embrechts & H. Erlang (1909) Sandsynlighedsregning og telefonsamtaler . Teugels (1985). 566580. 131137. Kliippelberg & T. [143] W. 616624. Arndt & V.L. Feller (1971) An Introduction to Probability Theory and its Applications II (2nd ed. Jensen. Furrer (1998) Risk processes perturbed by astable Levy motion.D. [140] J.
Life Insurance Mathematics. [162] P. Furrer & H. de Vylder & J. Mitt.E. Gnedenko & I. [168] J. Math.J. Basel. A. Haezendonck (1984) Insurance Premiums. Appl.U. Schweiz. . [163] P. Insurance : Mathematics and Economics 7. 5. [158] H.. Gerber (1988) Mathematical fun with ruin theory. NorthHolland.). Kou (1995) Limits of first passage times to rare sets in regenerative processes. Scand. Glasserman & S. 31A. Probab.U. [153] H. Amsterdam. Gani. London. 6370. Gerber (1986). Probab. [169] J.U. [152] H. Appl. Schweiz. Schmidli (1994) Exponential inequalities for ruin probabilities of risk processes perturbed by diffusion. [155] H. Kovalenko (1989) Introduction to Queueing Theory (2nd ed. 1977. Aktuarietidskrift 1970 . 105115. University of Pennsylvania. Kluwer. F. Gerber (1973) Martingales in risk theory. 73. Probab. Kaas. 463480. Glasserman (1991) Gradient Estimation via Perturbation Analysis. Grandell (1978) A remark on 'A class of approximations of ruin probabilities'. Act. Springer. Suppl. Weron (1997) Stable Levy motion approximation in collective risk theory. 205216. Glynn & W.. R. 28. Amsterdam. Gerber (1979) An Introduction to Mathematical Risk Theory. 1523. Michna & A.J. In Studies in Applied Probability (J. S. van Heerwarden & T. 2336. Skand.S. Gerber (1981) On the probability of ruin in the presence of a linear dividend barrier. Boston. J.G. Scand. Insurance: Mathematics and Economics 20.V. Scand. Goovaerts. Insurance: Mathematics and Economics 15. Gerber (1971) Der Einfluss von Zins auf die Ruinwahrscheinlichkeit. 97114. Bauwelinckx (1990) Effective Actuarial Methods. [154] H. ETH Zurich.). Huebner Foundation Monographs. Gerber (1970) An extension of the renewal equation and its application in the collective theory of risk. Act. Vers. Furrer. J. 3752. contained in the authors PhD thesis. [161] P. J. Appl. Mitt. Adv.71. Goovaerts.W.N. Unpublished. 1978 . [156] H. [159] H.U. Galambos & J. Z.U. eds. 1981 . Furrer (1996) A note on the convergence of the infinitetime ruin probabilities when the weak approximation is astable Levy motion. 7778. Vers. Whitt (1994) Logarithmic asymptotics for steadystate tail probabilities in a singleserver queue. 424445. Griibel (1996) Perpetuities with thin tails. [157] H. [165] M. NorthHolland. Grandell (1977) A class of approximations of ruin probabilities. Goldie & R. [164] B.U.M. Dordrecht. J. 131156.BIBLIOGRAPHY 371 [151] H. 205210. Ver. [160] H. Math. Ver.U. Act. [167] C. Birkhauser . Ann. [166] M.
1. AkademieVerlag. Winter School on Stochastic Analysis and Appl . [179] R. . 143158. Grigelionis (1996) Lundbergtype stochastic processes. Opns . Tinbergen Institute Research Series 20. 7590.243255. Proc. 400409. [178] B.J. [181] D. Grandell (1997) Mixed Poisson Processes. Griibel & R. New York. [177] B. Appl. [173] J. 3041. Gyllenberg & D. 197214. M. mathematische Wirtschaft . Proc . Berlin. Insurance: Mathematics and Economics 26. Rink. Harrison (1977) Ruin problems with compounding assets . [189] A. Ibragimov. Appl. 8. KTH.S. [182] A. Korolyuk (1969) On the joint distribution of a process with stationary increments and its maximum . Appl. 6779. 33. 3 . 332. [172] J. Harrison & S.O. SpringerVerlag. 14. [171] J. [186] J. Grigelionis ( 1992) On Lundberg inequalities in a Markovian environment. 54. [175] J. Probab. Grandell ( 1992) Finite time ruin probabilities and martingales . 347358. Grigelionis ( 1993) Twosided Lundberg inequalities in a Markovian environment. Grandell & C. Stoch. [183] M. 5 . Hadwiger (1940) Uber die Wahrscheinlichkeit des Ruins bei einer grossen Zahl von Geschiiften . [185] J. Matem. Grandell (1979) Empirical bounds for ruin probabilities. [174] J. Statistical Algorithm 265. Lemoine (1977) Limit theorems for periodic queues. M.. Harrison & A. [188] J. Astin Bulletin 29. Appl.A. Grandell (1990) Aspects of Risk Theory. Grandell (1999) Simple approximations of ruin functions. Chapman & Hall.I. [176] B. Res.V. Proc.). Opns . Resnick (1976 ) The stationary distribution and first exit probabilities of a storage process with general release rule . Stoch. Aktuar Tidskr. Silvestrov (2000) Cram€rLundberg approximations for nonlinearly perturbed risk processes . Res. Amsterdam. Harrison & S. Informatica 2. M. Segerdahl (1971) A comparison of some approximations of ruin probabilities. Math.and Sozialforschung 6. Math. 14. Applied Statistics 40. Preprint. Gusak & V. M. 355365. ed. SpringerVerlag. [180] R. Probab. Hermesmeier (1999) Computation of compound distributions I: aliasing errors and exponential tilting. [184] H . [187] J. van Heerwarden (1991) Ordering of Risks: Theory and Actuarial Applications. Skand. Liet. Gut (1988 ) Stopped Random Walks . Probability Theory and Mathematical Statistics (I.131135. 167176. Resnick (1977) The recurrence classification of risk and storage processes . Th. J. Theory and Applications. Archiv v. Grnbel (1991) G/G/1 via FFT. 5766.372 BIBLIOGRAPHY [170] J. 566576. E.I.
Probab. Act. Probab. Herkenrath (1986) On the estimation of the adjustment coefficient in risk theory by means of stochastic approximation procedures. Karlsruhe. Heilmann (1987) Grundbegriffe der Risikotheorie. Z. Iversen & L. Hoglund (1974) Central limit theorems and statistical inference for Markov chains. Foruth Prague Symp. Hoglund (1990) An asymptotic expression for the probability of ruin within finite time.B.BIBLIOGRAPHY 373 [190] P. Taksar (2000) Optimal proportional reinsurance policies for diffusion models. J.. Hoglund (1991). 1998 . Scand. [201] L. 8996. 11631174. [207] D. on Asymptotic Statistics (P. Klugman (1984) Loss Distributions. [203] T.). Hesselager (1990) Some results on optimal reinsurance in terms of the adjustment coefficient. Hipp (1989a) Efficient estimators for ruin probabilities. New York. Verlag Versicherungswirtschaft e. Ann. Asmussen & 0. 305313.M. Probab. [198] C. J. [206] B. Mathematical Statistics. 5770.A. Chalmers University of Technology and the University of Goteborg. Appl. 19. Scand. Michel (1990) Risikotheorie: Stochastische Modelle and Statistische Methoden. Hipp (1989b) Estimators and bootstrap confidence intervals for ruin probabilities. [193] U.a program for fitting phasetype distributions. Statist. Ann. [192] U. Versicherungswirtschaft.V. 12981310. Hogg & S. J. Nachrichten 30. Mandl & M. 6.L. [200] M. 25. [199] R. [204] T. Proc. Horvath & E. Hermann (1965) Bin Approximationssatz fiir Verteilungen stationarer zufalliger Punktfolgen. [196] C. [195] B. Heidelberger (1995) Fast simulation of rare events in queueing and reliability models. Act. Hipp & R. The ruin problem for finite Markov chains. ACM TOMACS 6. verve. Hogan (1986) Comment on corrected diffusion approximations in certain random walk problems. [205] T. Prob. Insurance: Mathematics and Economics 5. Hill (1975) A simple general approach to inference about the tail of a distribution.R. Haggstrom.L. 259268. Wiley. Wahrscheinlichkeitsth. 1990 . 8095. 285293. 377381. 23. Electronic Letters 35. Huskova. [202] 0. . 18. 29. 123151. [194] 0. Geb.V. Iglehart (1969) Diffusion approximations in collective risk theory. Astin Bulletin 19. Ann. [208] V. Karlsruhe. S. J. No. Math. 378389. 166180. 4385. eds. Studies in Statistical Quality Control and Reliability 1992 :4. Insurance: Mathematics and Economics 5. 285292. Hojgaard & M. Appl. 3. [197] C. Nerman (1992) EMPHT . [191] W. Staalhagen (1999) Waiting time distributions in M/D/1 queueing systems. Willekens (1986) Estimates for the probability of ruin starting with a large initial reserve.
Stochastic Models 1. [218] V. Stochastic Models 5. Amer. ibid. Statist. 173190. [217] R. J. [226] J. 31. Proc.J.G. 866870. Kao (1988) Computing the phasetype renewal and related functions.M. Lazar (1998) Subexponential asymptotics of a Markovmodulated random walk with a queueing application. Probab.P. Janssen & J. Addenda to for processes defined on a finite Markov chain. 547567. 239256.374 BIBLIOGRAPHY [209] D. Proc. [227] J. Kluwer. 61. 325247. Math. Taaffe (1989/90) Matching moments to phase distributions. Johnson & M. Soc. Goovaerts (1986) General bound on ruin probabilities. Jelenkovic & A. [228] J. [221] E. Wishart (1964) A central limit theorem for processes defined on a finite Markov chain.G. NEC Research Institute. Kalashnikov (1997) Geometric Sums: Bounds for Rare Event with Applications. Keilson & D. 35. 37. New York.M. 118. Manuscript.M. 187193. [215] J. Technometrics 30. 6.G. Act. Jagerman (1985) Certain Volterra integral equations arising in queueing. Ann. 60. Insurance: Mathematics and Economics 5. [225] J. Act. Astin Bull. Proc. Cambridge Philos. 561563.H. [213] P. [222] S.B. [219] V. 283306. [210] D. Karlin & H. Scand. Soc. Astin Bull. Kalashnikov (1996) Twosided bounds of ruin probabilities. Kemperman (1961) The Passage Problem for a Markov Chain. Kennedy (1994) Understanding the WienerHopf factorization for the simple random walk.J. 259281. Probab. 3. Reinhard (1985) Probabilities de ruine pour une classe de modeles de risque semiMarkoviens. Wishart (1964).A. Keilson (1966) A limit theorem for passage times in ergodic regenerative processes. 6.L. [223] J.C. Taylor (1981) A Second Course in Stochastic Processes. 123133.M. Appl. 63. 116129. 8793. 711743. . 11. University of Chicago Press. Cambridge Philos. Chicago. Janssen (1980) Some transient results on the M/SM/1 special semiMarkov model in risk and queueing theories. ibid. Cambridge Philos. Jensen (1953) A Distribution Model Applicable to Economics. Wishart (1964) Boundary problems for additive processes defined on a finite Markov chain. [211] J. [212] J.M. [220] V. 165167. Jagerman (1991) Analytical and numerical solution of Volterra integral equations with applications to queues. Oxford. Keilson & D. 4151. Kaas & M. [216] M. Clarendon Press. Kalashnikov (1999) Bounds for ruin probabilities in the presence of large claims and their comparison. [224] J. Munksgaard.R. Appl. N. J. J. 1996. Jensen (1995) Saddle Point Approximations. J. N. Princeton. [214] A. Academic Press. Copenhagen. Soc. Keilson & D. 15.
J. Kuchler & M. Klnppelberg (1989) Estimation of ruin probabilities by means of hazard rates. 6075. Proc. Ramaswami (1999) Introduction to MatrixAnalytic Methods in Stochastic Modelling. 48. Kingman (1964) A martingale inequality in the theory of queues. B24. [234] C. [239] H. [244] T.F.F. Scand. Stadtmiiller (1998) Ruin probabilities in the presence of heavytails and interest rates . Act. Philos. SpringerVerlag. 1995 . 1998 . Roy. Quart. 359361. 60 . Prob. 259264. Soc. [230] J. . [248] D. [246] D. Sorensen (1997) Exponential Families of Stochastic Processes.F. [233] C.J. Kingman (1962) On queues in heavy traffic. Lindley (1952) The theory of queues with a single server. Insurance: Mathematics and Economics 8. 383392. Kunita (1976) Absolute continuity of Markov processes.. Lemoine (1989) Waiting time and workload in queues with periodic Poisson input.a Linear Algebraic Approach. Turbin (1973) An asymptotic expansion for the absorption time of a Markov chain distribution. J. Lehtonen & H. Nyrhinen (1992a) Simulating levelcrossing probabilities by importance sampling. 132141. Klnppelberg & T. Lecture Notes in Mathematics 511. (235] C. Soc.J. Lipsky (1992) Queueing Theory . 858874. Insurance : Mathematics and Economics 12. Act. 25. [238] V. Adv. Nyrhinen (1992b) On asymptotically efficient simulation of ruin probabilities in a Markovian environment. Lehtonen & H. Appl. Math. Actuarial J. Proc. [241] G. Klnppelberg (1993) Asymptotic ordering of risks and ruin probabilities. Lemoine (1981) On queues with periodic Poisson input. Kliippelberg (1988) Subexponential distributions and integrated tails. SpringerVerlag. 4958.C. J. 26. [231] J. Quart. 4477. 283284. Soc. Lucantoni (1991) New results on the single server queue with a batch Markovian arrival process. Probab. Penev & A. Korolyuk. [247] L.S. [242] A. Cambr. I. Appl. Klnppelberg & T.P. [243] A. Stochastic Models 7. New York. Seminaire de Probabilties X . 146. J. 133135 (in Russian). 277289.F. Appl. J. Latouche & V. [236] C. Appl. Oxford 12. J. Kingman (1961) A convexity property of positive matrices . Philos. 889900. Macmillan. Bernoulli 1 . [232] C. SIAM. Probab.BIBLIOGRAPHY 375 [229] J. 279285.C. J. Scand. 24. 154168. Statist. 18. 125147. Mikosch (1995) Modelling delay in claim settlement. [237] C. Mikosch (1995) Explosive Poisson shot noise with applications to risk retention. Scand. Camb. Probab. Kliippelberg & U.C. 390397. [245] T. [240] U. Cybernetica 4.
J. MeierHellstern & M . . 11 . Proc.V. [255] V. Probab. J.S.K. J. Probab. Res. 1996 . Ann. Appl. F. with an application . Adv. Ney & E. Nagaev (1957) Some limit theorems for stationary Markov chains. Scand. 161174. Eigenvalue properties and limit theorems . Malinovskii ( 1996) Approximation and upper bounds on probabilities of large deviations of ruin within finite time. Englunds Boktryckeri AB. 124147. Probab.376 BIBLIOGRAPHY [249] D. Math. Probab.D. Appl. Soc. Act. 429448. Opns. Almqvist & Wiksell. 2. J. Malinovskii ( 1994) Corrected normal approximation for the probability of ruin within finite time. Lundberg (1926) Forsdkringsteknisk Riskutjdmning.D.K. 561592. 1986 . [258] K . Scand. 4859. J. [256] A. Probab.D. [262] H . 2939. 58. Ann. . [259] Z . 1994. Lucantoni . A fitting algorithm for Markovmodulated Poisson processes having two arrival rates . Michna ( 1998) Selfsimilar processes in collective risk theory.M. 378406. 147149. Math. 676705. Act. [257] A. [254] V . 763776. [264] G. [252] A. Stoch. Europ. [250] F. Probab . Cambridge Philos. Th. Neuts ( 1990) A single server queue with server vacations and a class of nonrenewal arrival processes . 834841. J. Ann. 22. Statist. Gut & J. [260] H. 29 .V. K. MartinL3f (1986) Entropy. Proc. Stockholm. Miller (1962 ) A matrix factorization problem in the theory of random variables defined on a finite Markov chain . Act. 15. MartinL6f (1983) Entropy estimates for ruin probabilities . [266] P. II Aterforsdkring av Kollektivrisker . 286298. Appl. Scand. Schmidt ( 1993) On ladder height distributions of general risk processes . 58 . Miller ( 1962 ) Absorption probabilities for sums of random variables defined on a finite Markov chain . Uppsala. Miller ( 1961 ) A convexity property in the theory of random variables defined on a finite Markov chain. Appl. Cambridge Philos. a useful concept in risk theory. 223235. 3. Anal. Mammitzsch ( 1986 ) A note on the adjustment coefficient in ruin theory. Appl. Nummelin (1987) Markov additive processes I. 268285. Miyazawa & V. Makowski ( 1994) On an elementary characterization of the increasing convex order . Hoist eds. Meier ( 1984). Soc. Lundberg (1903) I Approximerad Framstdllning av Sannolikhetsfunktionen. Probability and Mathematical Statistics (A. 31. 36. Insurance : Mathematics and Economics 5. 1261 (0?)1270. Appl.S. [261] H. Moustakides ( 1999) Extension of Wald 's first lemma to Markov processes. [251] F. 32. 370377.). [265] S. [263] M . [253] V. J.F.
Paulsen (1993) Risk theory in a stochastic economic environment. Willekens (1987) Second order behaviour of distributions subordinate to a distribution with finite mean . Appl. 10081026. Omey & E. Insurance: Mathematics and Economics 22.BIBLIOGRAPHY 377 [267] M . 71. Probab. [276] C. Neuts (1977) A versatile Markovian point process. Norberg ( 1990) Risk theory and its statistics environment . Appl. J. New York. Stoch. Neuts ( 1981 ) MatrixGeometric Solutions in Stochastic Models. J. Pakes (1975) On the tail of waiting time distributions. Gjessing (1997b) Present value distributions with applications to ruin theory and stochastic equations. O'Cinneide (1990) Characterization of phasetype distributions.F. Appl. [268] M . [284] J. [279] E. Abh. Stoch. [285] J. J. [272] J . Appl. E75B . Neuts ( 1978) Renewal processes of phase type. 75. [277] J. 445454. Math. 30. Marcel Dekker. . Nyrhinen (1999) Large deviations for the time of ruin. 157. 215223. [274] H . Ohlin (1969) On a class of measures for dispersion with application to optimal insurance. Sem. [271] M. 273299. 46.F. Proc. Proc. IEICE Trans. [270] M . [269] M . 21. Baltimore . Probab. Commun. Stoch. Nyrhinen ( 1998 ) Rough descriptions of ruin for a general class of surplus processes.a survey. Probab. Models 6. Neveu ( 1961 ) Une generalisation des processus a accroisances independantes. Proc.G. Neuts ( 1989) Structured Stochastic Matrices of the M/G/1 Type and their Applications. Insurance: Mathematics and Economics 20. 123144. 555564. 16. 764779. Adv. 12551265. Hamburg. 733746. [280] A. Appl. London. [273] R.F. Probab.F. Johns Hopkins University Press. Paulsen (1998) Ruin theory with compounding assets . Paulsen ( 1998) Sharp conditions for certain ruin in a risk process with stochastic return on investments.A. Naval Research Logistics Quarterly 25. 249266. Willekens (1986) Second order behaviour of the tail of a subordinated probability distributions. Stoch. 327361. [281] J. Stoch. [283] J. 36. Astin Bulletin 5 .F. [278] E. Stochastic Models 3 . Neuts (1992) Models based on the Markovian arrival process. 311342. Appl. Paulsen & H. [275] H.K. [282] J . Proc. Paulsen & H. Statistics 21. Gjessing (1997a) Optimal choice of dividend barriers for a risk process with stochastic return on investments. 12. Appl. 316. 135148.K. 339353. Appl. Omey & E.
[291] S. Prabhu (1965) Queues and Inventories.U. 32. Gjessing (1997c) Ruin theory with stochastic return on investments. Ann. Reinhard (1984) On a class of semiMarkov risk models obtained as classical risk models in a markovian environment. Embrechts (1996) Confidence bounds for the adjustment coefficient. On the ruin problem of collective risk theory. Math. Appl. [297] R.378 BIBLIOGRAPHY [286] J. [292] S. 139143. [289] C. 1989 . Appl.U.L. 222261. Soc. Griibel & P. Wiley.K. de Smit (1986) The queue M/G/1 with Markovmodulated arrivals and services. [288] S. 46. [302] J. Probab. Pitts. Paulsen & H. Prabhu (1961). Adv. Appl. 147159.A. Inst. New York. 29. Act. Res. 757764. [298] V. [290] E. Heidelberg. Pellerey (1995) On the preservation of some orderings of risks under convolution. 215246.R. Insurance: Mathematics and Economics 16. [293] N. 4568. Res. 2330. Wiley. Probab. Math. Rogers (1994) Fluid models in queueing theory and WienerHopf factorisation of Markov chains.M. Scand. Pitts (1994) Nonparametric estimation of compound distributions with applications in insurance.K. [300] C. Math. [299] C. Math. Prabhu & Zhu (1989) Markovmodulated queueing systems. Oper. [304] B.U. 568576.M. Pitman (1980) Subexponential distribution functions. Rao (1965 ) Linear Statistical Inference and Its Applications. Ann. 537555. Tidskr. Opns. 30. Probab. Resnick & G. 820827. Statist.G.G. [305] C. Insurance Risk. Probab. 2343. 4. 337347. Queues. [295] N. Philipson (1968) A review of the collective theory of risk. Skand. Appl. 691707. Ramaswami (1980) The N/G/1 queue and its detailed analysis. 390413. Rolski (1987) Approximation of periodic queues. Berlin. Appl. Springer. Aktuar. Queueing Systems 5. Prabhu (1980) Stochastic Storage Processes.M. Ramsay (1984) The asymptotic ruin problem when the healthy and sick periods form an alternating renewal process. Austr.U. [306] T. Pyke (1959) The supremum and infimum of the Poisson process. New York. Wiley.H. [296] N. 28. 61. Adv. [294] N. 465483. [287] F. Adv. 117133. J. 12.J. Ann. [303] S. [301] G. Ann. Samorodnitsky (1997) Performance degradation in a single server exponential queueing model with long range dependence. 19. Math. J. Insurance: Mathematics and Economics 3. 11. Statist. . and Dams. 29A. J. Ripley (1987) Stochastic Simulation . 235243.J. Schock Petersen (1989) Calculation of ruin probabilities when the premium depends on the current reserve. 965985. Probab . Regterschot & J. Stat. R. Astin Bulletin XIV..
Lecture Notes of the 8th Summer School on Probability and Mathematical Statistics ( Varna). Schmidli ( 1999b) Perturbed risk processes: a review . [320] H. Schmidli. 4068. Schmidli ( 1994) Diffusion approximations for a risk process with the possibility of borrowing and interest. Ann. 2001 . Wiley. Wiley. Scand. 687699. Seal ( 1969) The Stochastic Theory of a Risk Business. Ryden (1994) Parameter estimation for Markov modulated Poisson processes. 7. Schmickler ( 1992 ) MEDA : mixed Erlang distributions as phasetype representations of empirical distribution functions . Schlegel ( 1998) Ruin probabilities in perturbed risk models. [319] H. Schmidli ( 1995 ) CramerLundberg approximations for ruin probabilities of risk processes perturbed by a diffusion . [324] H . Act. Probab . Melamed (1998) Modern Simulation and Modelling. [325] H . Schmidt & J. Statist. 5. Appi. 1997. 93104. Data Anal. [311] R. Insurance: Mathematics and Economics 22. [317] L. [318] H.L. J. 262286. Appl. Rubinstein & A. Appl. Statist. [322] H. 431447 [316] S. [312] R. 417421.J. Science Culture Technology Publishing . 365388.Y. Singapore. Act. [308] S. Schmidli ( 1999a) On the distribution of the surplus prior and at ruin. Stochastic Models 6. Schmidli ( 1997b) An extension to the renewal theorem and an application in risk theory. [309] H. [315] T. [314] T. Scand. Schmidli ( 1996) Martingales and insurance risk. . Math. Schmidli ( 2001 ) Optimal proportional reinsurance policies in a dynamic setting . J. Operationsforsch. Stochastic Models 10 . Stoch. Rudemo (1973) Point processes generated by transitions of a Markov chain.Y. 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. Adv. Rolski. 11. [313] M. Stochastic Models 10. Schmidli ( 1997a) Estimation of the Lundberg coefficient for a Markov modulated risk model .M. Astin Bulletin 29. Proc. 227244. Wiley. Wiley. 5. 5. Wiley. 131156. J. Probab. 21. Shapiro (1993) Discrete Event Systems: Sensitivity Analysis and Stochastic Optimization via the Score Function Method. Comp. Insurance: Mathematics and Economics 16. 795829. Probab.Y. [326] H. Teugels (1999) Stochastic Processes for Insurance and Finance. [321] H. H. Rubinstein & B. Th. 4857. V. [323] H.BIBLIOGRAPHY 379 [307] T. 145165. 121133. [310] R. Rubinstein (1981) Simulation and the Monte Carlo Method. 135149. 155188. Ryden (1996) An EM algorithm for estimation in Markovmodulated Poisson processes. Ross (1974) Bounds on the delay distribution in GI/G/1 queues.
11. Weiss (1995) Large Deviations for Performance Analysis. [331] G. [334] C: O. Sengupta (1990) The semiMarkov queue: theory and applications. Ann. Siegmund (1975) The time until ruin in collective risk theory. [337] M. Shin (1987) Convolution of uniform distributions and ruin probability. Wiley. J. [346) D . 279299. 75. Sigman ( 1994 ) Stationary Marked Point Processes: An Intuitive Approach. Seal (1978) Survival Probabilities. Act. Seber (1984) Multivariate Observations. Shantikumar (1993) Stochastic Orders and Their Applications. Siegmund (1976) The equivalence of absorbing and reflecting barrier problems for stochastically monotone Markov processes . 673684. Scand. [332] C : O. 159180. Mitt. Math.W. Adv. Shaked & J. Seal ( 1972). 4383. 701719. Queueing Systems 11. Aktuar Tidsskr. [336] B.G. 77100. Stockholm. Chapman and Hall. Aktuar Tidsskr. [335] B . 11. Versich. Grenander). Wiley. Scand. Chapman & Hall. Appl. Probab . Siegmund (1976) Importance sampling in the Monte Carlo study of sequential tests. Statist.L. Seal (1974) The numerical calculation of U(w. 171178.F. Math. 72. 2236.A. t). Almqvist & Wiksell. 121139. Risk teory and the single server queue .W. 4. Mitteil. Probab. [345] D.S. Mitt. Math. Academic Press. Versich.the Harald Cramer volume (ed. Segerdahl (1955) When does ruin occur in the collective theory of risk? Skand. [341] E. [342] D. Probab. Segerdahl ( 1942) Uber einige Risikotheoretische Fagestellungen . [330] H. Adv. Appl. SpringerVerlag. Appl. [340] E. pp. the probability of nonruin in an interval (0. Verein Schweiz. Th. 429442. 61. Shwartz & A.S. [339] A. [333] C: O. L. Ann. Sengupta ( 1989) Markov processes whose steadystate distribution is matrixgeometric with an application to the GI/PH/1 queue. Skand. New York. Shin (1989) Ruin probability by operational calculus. Shtatland ( 1966) On the distribution of the maximum of a process with independent increments.380 BIBLIOGRAPHY [327] H. . Insurance: Mathematics and Economics 8. [328] H. 191197.S. Sigman (1992) Light traffic for workload in queues. In Probability and statistics . Verein Schweiz Versich. J. [348] K . 483487. [343] D. Seal (1972) Numerical calculcation of the probability of ruin in the Poisson/Exponential case. . Probab. [338] E.L. Siegmund (1979) Corrected diffusion approximations in certain random walk problems. Verein Schweiz. [344] D. Segerdahl (1959) A survey of results in the collective theory of risk. 157166. [329] H. 4. Siegmund ( 1985) Sequential Analysis . 72. 1987. 243249. t).L. Act. [347] K. 1974. Stochastic Models 6. 1955. 914924. 383413. 21.
Insurance: Mathematics and Economics 16. [350] W. Act. Smith ( 1953) Distribution of queueing times. . Stoyan (1983) Comparison Methods for Queues and Other Stochastic Models (D. Sundt & W. New York. [357] B. Scand. 8199.C. J. Teugels (1997) The adjustment coefficient in ruin estimates under interest force . 235254. Teugels ( 1995 ) Ruin estimates under interest force . Thorin (1986) Ruin probabilities when the claim amounts are gamma distributed. J. Suri ( 1989) Perturbation analysis : the state of the art and research issues explained via the GI/G/1 queue. [366] O.A. Cambridge Philos.L. Stanford & K. Taylor (1986) Claims reserving in NonLife Insurance. Unpublished manuscript. Taylor (1978) Representation and explicit calculation of finitetime ruin probabilities. Takhcs (1967) Combinatorial Methods in the Theory of Stochastic Processes. 725741. Thorin (1977). [363] G. [351] D . Taylor (1979) Probability of ruin with variable premium rate. Karlsruhe. Insurance: Mathematics and Economics 1. [354] B. Appl.L. Sundt & J. 8594. [359] L. Proc. Act. J. Proceeedings of the IEEE 77. Verlag Versicherungswirtschaft e. Thorin (1974) On the asymptotic behaviour of the ruin probability when the epochs of claims form a renewal process. 118. Act.L. Soc 49. Straub (1988) NonLife Insurance Mathematics. Sundt ( 1993) An Introduction to NonLife Insurance Mathematics (3rd ed.C. 1976. New York. J. 65102. 1980 .. [358] R.J. 114137. Astin Bulletin 16 . Scand. [362] G. Teugels (1982) Estimation of ruin probabilities . Act. Scand. 5776.. [355] B . 722. Wiley. J. 149162. Taylor (1979) Probability of ruin under inflationary conditions or under experience ratings .J. Astin Bulletin 24. [369] O. 2739. [365] J. 1978.and largedeviations probabilities in actuarial risk theory.BIBLIOGRAPHY 381 [349] E.S. Springer. NorthHolland. Scand.V. Scand. John Wiley & Sons. 449461. Stroinski ( 1994) Recursive method for computing finitetime ruin probabilities for phasedistributed claim sizes. [352] D.C. [361] G. Hoesman ( 1989) Moderate. Sundt & J. 163175.C. Thorin (1982) Probabilities of ruin. 65102.). 21.L. Slud & C.C. Daley ed. [353] E. [368] O. Insurance: Mathematics and Economics 19. 1974. 1982 .). Probab. [356] B . Adv. Act. [360] G. Actuarial J. Taylor (1976 ) Use of differential and integral inequalities to bound ruin and queueing probabilities . [364] G. Jewell ( 1981 ) Further results on recursive evaluation of compound distributions . 197208. Astin Bulletin 12. Scand. [367] O. Ruin probabilities prepared for numerical calculations.
37.E. [384] R. Reidel. Willmot (1994) Refinements and distributional generalizations of Lundberg's inequality. De Vylder. Waters (1983) Probability of ruin for a risk process with claims cost inflation. PrenticeHall. 4963. [374] F. Wallace (1969) The solution of quasi birth and death processes arising from multiple access computer systems. J. Sherman & D. In: Premium Calculation in Insurance (F. Van Wouve. Thorisson (2000) Coupling. Wald (1947) Sequential Analysis. [383] G. F.E.W. thesis. 142. R. Aktuar.R. 743756. . 137153. D. Astin Bulletin IX. SpringerVerlag. 6785. Goovaerts. Appl. De Vylder & M. [375] N. Res. Taqqu. Stationarity and Regeneration. Wiley. Thorin & N. [371] 0. [380] W. [379] M. Willmot & X. [372] H. Wikstad (1977) Calculation of ruin probabilities when the claim distribution is lognormal. Wolff (1990) Stochastic Modeling and the Theory of Queues. Tidskr. 936952. Whitt (1989) An interpolation approximation for the mean workload in a GI/G/1 queue. Astin Bulletin VII. [381] W.382 BIBLIOGRAPHY [370] 0. Act. Wikstad (1977) Numerical evaluation of ruin probabilities for a finite period.). Insurance : Mathematics and Economics 13. 148164. [376] A. Haezendonck eds. Thorin & N. Lin (1994) Lundberg bounds on the tails of compound distributions. Dordrecht Boston Lancaster. University of Michigan. [382] G. 31. 231246. Goovaerts (1983) The influence of reinsurance limits on infinite time ruin probabilities. Tacklind (1942) Sur le risque dans les jeux inequitables. 1942. Wilson (1995) Selfsimilarity in highspeed traffic: analysis and modeling of ethernet traffic measurements. M. Skand. J. M. [377] V. Statistical Science 10. Willinger. Submitted. Probab. Veraverbeke (1993) Asymptotic estimates for the probability of ruin in a Poisson model with diffusion. 5762. [373] S. [378] H. Insurance: Mathematics and Economics 15. Scand. 1983 . J. Unpublished Ph. VazquezAbad (1999) RPA pathwise derivatives of ruin probabilities. Oper.
5.269. 89. 1415.249250 integral equation 16 Lindley 143 renewal 64. 218 Cox process 4.Index adjustment coefficient 17.299. 9396.203. 196201 inverse Gaussian distribution 76. 180182. 7079. 318319 Erlang distribution 7.203.178184. 308. 97.293294.314316. 111117. 316323 Bessel function 102.328330.318320 change of measure 2630.285292. 2425. 119. 138139. 37. 110113. 3436. 248 WienerHopf 144 interest rate 190. 217. 301 Kronecker product.281. 207 heavytailed distribution 6.228229.200201. 205.135. 117127 corrected 121127 duality 1314.4447. 360 excursion 155156. 80 81. 239. 341.242. 332333 Volterra 192194. 278 gamma distribution 67. 15.and sum 221.121129.287292.251280 heavy traffic 76.137141. 7179.150.301 central limit theorem 60 . 271274. 97129.9899.160167. 79. 201 Brownian motion 3 .185187. 245248.182. 162164. 122. 227229. 302303 diffusion approximation 17. 1819.359 aggregate claims 103106.86.100. 91. 3839. 2526.259261. 135. 3032. 201214. 9496. 4851. 39.249. 226.226.217. 7475.346349 383 . 141144. 189. 1112. 17. 40.308 CramerLundberg model: see compound Poisson model cumulative process 334 dams: see storage process differential equation 16. 283. 17.6779. 170173. 3334.292293 Edgeworth expansion 113. 12 CramerLundberg approximation 1617.307312 compound Poisson model 4. 361 diffusion 3. 323 Coxian distribution 147. 5796. 9293.272. 8283 hyperexponential distribution 7. 117128. 86. 7879. 14.
185187 GI/G/1 141144 M/D/1 6667 equation 16. 141144.340350 multiplicative functional 2830.174.180. 306316 Levy process 3.350361 Poisson process Markovmodulated 12 periodic 12.302. see also sensitivity analysis phasetype distribution 8. nonlinear 155. 44. 162. 108 life insurance 5. 86 periodicity 12.315 inequality 1718.288290. 35. 203204.339 large deviations 129. 261264. 80.285287 queue 14 . 38. 179 NP approximation 318320 Palm distribution 5253. 3947.215250. 39. 171. 25. 145187.178182.146148.269271.227230. 133. 132133. 3947.218221.304 process 2830. 149. 260 Lundberg conjugation 6979 .240244. 6162.201. 6970. 176185. 295. 230. 304305 random walk 3336.161. 175 light traffic 8183 Lindley integral equation 143 process 3334. 257.287. 5758. 7179.234. 113114. 96. 229 M/M/1 101 Markovmodulated 185187 periodic 187 martingale 2426. 37. 144. 138. 203 Markov additive process 12.275278.234240. 108109.139141.298299.261264. 134. 44.108.336339 . 16. 9899. 271274. 41.128129.161164.134135. 38.161. 99. 245 M/G/1 13.123.297299.148. 35. 178 modulation 12. 59. 16. 134135. 176185 nonhomogeneous 60 PollaczeckKhinchine formula 6167.160161.238. 14.348 terminating 215216.384 ladder heights 4756. 227228. 71. 25. 32. 154. 2730. 65. 133.336339 Laplace transform 15. 3639. 267269 Panjer's recursion 320323 Pareto distribution 910. 137139. 251. 213214. 4446. 269 PerronFrobenius theory 4142. 100. 157. 15.287291 INDEX matrix equation . 142 likelihood ratio : see change of measure lognormal distribution 9.152160. 112113.259261. 106108. 234 matrixexponential distribution 240244 matrixexponentials 14.349 350 perturbation 172173. 52 53. 42. 7576.
186. 7475.359361 stochastic control x stochastic ordering 18. 168172 storage process 13. 333334 regular variation 10. 317318 semiMarkov 147. see also matrixexponential distribution regenerative process 264 268. 162. 131144. 280. 107. 9693. 327 . 186187 virtual: see workload rational Laplace transform 8. 1819. 244.273274. 229234. 261264 reservedependent premiums 14. 146. 233. 123. 240. 3032.INDEX 385 waiting time 141. 31. 147. 292294.336339 workload 13. 172173. 189214. 87. 191192. 326330 Weibull distribution 9. 233234. 251. 223226. 338 utility 324. 260 WienerHopf theory 144. 4950.262263. 257. 60. 222. 120 statistics x.279280 Rouche roots 158. 37. 281296 stable process 15. 11. 160.154157. 12. 238 saddlepoint method 115117.244250. 251. 186187 renewal process 131. 141144. 279280 subexponential distribution 11. 8386. 256258. 260 reinsurance 8. 177 timereversion 14. 5455. 251280 time change 4. 332333 model 12. 174. 335336 sensitivity analysis 8693. 89. 152.314. 253. 213. 294296 shotnoise process 314 simulation 19. 331336 equation 64. 307308.
the ^W A l \ i l ' ''' CramerLundberg approximation.g. Markovmodulation or periodicity.T [Ail i The book is a comprehensive treatment of  I i I \ classical and modern ruin probability theory.Vol. worldscientific." Short Book Reviews ISBN 9810222939 mi u inn i nun I I I I I I i in u www. P'i yfliother approximations (e. I 1! Ruin Probabilities . extensions of the classical compound Poisson model to allow f o r reservedependent premiums. Special features of the book are the emphasis on change of measure techniques. Some i (l I JL I J r of the topics are Lundberg's inequality. phasetype distributions as a computational vehicle and the connection to other applied probability areas like queueing theory.Advanced Series on Statistical Science & Applied Probability . y finite horizon ruin probabilities. for heavytailed claim size distributions). 2 A I 11 JjVb l' i  i Yj . "This book is a must for anybody working in applied probability..com 2779 he 9 "789810ll22293211 .. exact solutions. It is a comprehensive treatment of the known results on ruin probabilities.
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.