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
. . . . . .. . . . . . . . 340 A4 Some linear algebra . 261 4 Models with dependent input . . . . . .. . . . . .. . . . . 336 A3 Matrixexponentials . . . . . . . .. . 279 X Simulation methodology 281 1 Generalities . . . . . . . . . . . . . . . . . 259 3 The renewal model . .CONTENTS vii IX Ruin probabilities in the presence of heavy tails 251 1 Subexponential distributions . . . . . . . . . . . . . . . . 285 3 Importance sampling via Lundberg conjugation . . . 326 Appendix 331 Al Renewal theory . . . . . . . 251 2 The compound Poisson model . . . . .. . . . . . .. . . . . . . 331 A2 WienerHopf factorization . . .. . . . . . . . 297 2 Further applications of martingales . .. . . . . . . . . .. . . . . 316 5 Principles for premium calculation . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . 264 5 Finitehorizon ruin probabilities . . . . . . . . . . . . . . . . .. . . . . . . . . . . .. . .. . . . .. . . . . . . . . . 304 3 Large deviations . . . . 292 6 Sensitivity analysis . . . . . . . . . . . .. . 344 AS Complements on phasetype distributions . . . . . . . . . . . 306 4 The distribution of the aggregate claims . 271 6 Reservedependent premiums . . . . . . . 281 2 Simulation via the PollaczeckKhinchine formula . . . . . .. . .. . . . . . . . . . . . . 350 Bibliography Index 363 383 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . 287 4 Importance sampling for the finite horizon case . . . 290 5 Regenerative simulation . . . . . . . . . The twobarrier ruin problem . . . . . .. . . . . . . . . . . .. . . . . . . . . . . . . . 323 6 Reinsurance .. . .. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . 294 XI Miscellaneous topics 297 1 The ruin problem for Bernoulli random walk and Brownian motion. .. . . . . . . . . . . . . . . . . .
This page is intentionally left blank .
the idea was close to expand these to a short book on the subject. Since I was to produce some handouts for the students anyway. As an excuse: many of these projects were related to the book. I have deliberately stayed away from discussing the practical relevance of the theory. and the series editor Ole BarndorffNielsen for their patience. Let me take this opportunity to thank above all my publisher World Scientific Publishing Co. Apart from these remarks. 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. University of Copenhagen. One reason for writing this book is a feeling that the area has in the recent years achieved a considerable mathematical maturity. and other projects absorbed my interest. The course was never realized. I was invited to give a course on ruin probabilities at the Laboratory of Insurance Mathematics. this applies to longrange dependence which is intensely studied in the neighboring ix . that it can only say something about very simple models and questions. and has been an active area of research from the days of Lundberg all the way up to today.Preface The most important to say about the history of this book is: it took too long time to write it! In 1991. But the pace was much slower than expected. Risk theory in general and ruin probablities in particular is traditionally considered as part of insurance mathematics. It has obviously not been possible to cover all subareas. if the formulations occasionally give a different impression. it would not be fair not to say that the practical relevance of the area has been questioned repeatedly. In particular. and the result is now that the book is much more related to my own research than the initial outline. the book is basically mathematical in its flavour. Thus. and my belief was that this could be done rather quickly. which has in particular removed one of the standard criticisms of the area. However. 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!). it is not by intention.
13. some basic discussion can be found in the books by Biihlmann [82] and Gerber [157]. Concerning ruin probabilities. A book like this can be organized in many ways.13. Finally.45. see also Schmidli [325] and the references in Asmussen & Taksar [52]. for which I apologize to the reader and the authors of the many papers who ought to have been on the list.g. The main motivation comes from statistical data for network traffic (e. Chapters IXX then go in more depth with some of the special approaches for analyzing specific models and add a number of results on the models in Chapters IIIVII (also Chapter II is essentially methodological in its flavor). for the effects on tail probabilities. Willinger et al.2. see e. IX. Chapters IIIVII introduce some of the main models and give a first derivation of some of their properties. IV.13 and XI. 111. Resnick & Samorodnitsky [303] and references therein. Asmussen.g.2. More recently. incorporate 11. Hojgaard & Taksar [35] and Paulsen & Gjessing [284]. an area which is becoming increasingly important.15. For a second reading. [381]).3. I intend to keep a list of misprints and remarks posted on my web page. It is obvious that such a system involves a number of inconsistencies and omissions. Here is a suggestion on how to get started with the book.lth.1.maths .se/matstat / staff/asmus and I am therefore grateful to get relevant material sent by email to asmusfmaths . The present book is in between these two possibilities. In addition. Good luck! I have tried to be fairly exhaustive in citing references close to the text. The book does not go into the broader aspects of the interface between insurance mathematics and mathematical finance. VI. the standard stochastic control setting of diffusion models has been considered. e.13 and IX. http:// www. In the classical setting of CramerLundberg models.4a. the first part of 11. IV. VII. see in particular Michna [259].89. IV. X. One is by model. For a brief orientation. another by method. VIII. read Chapter I. some papers not cited in the text but judged to be of interest are included in the Bibliography. 111.5. Another interesting area which is not covered is dynamic control. I regret that due to time constraints.x PREFACE field of queueing theory. it has not been possible to incorporate more numerical examples than the few there are.se Lund February 2000 Soren Asmussen . VII. The rest is up to your specific interests.14.g.lth.6 (to understand the PollaczeckKhinchine formula in 111.2 more properly). Hojgaard & Taksar [206].
supported by Center for Mathematical Physics and Stochastics (MaPhySto). of which there are not many at this stage . Section VII . Parts of II.6. IV. 111 .2 by Rafal Kulik . Fig. Section VIII. More substantial remarks.PREFACE xi The second printing differs from the first only by minor corrections. 5. were produced by Lone Juul Hansen .1 by Bjarne Hojgaard and the table in Example 111. . Parts of X. 5 from Asmussen & Kliippelberg [36] with the permission from Elsevier Science . not least the more complicated ones. A number of other figures were supplied by Christian Geisler Asmussen . Lund September 2001 Soren Asmussen Acknowledgements Many of the figures . Aarhus.5 from Asmussen [21] with permission from CRC Press.1 and X. many of which were pointed out by Hanspeter Schmidli . as well as some additional references continue to be at the web page.4 from Asmussen. 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. Schmidli & Schmidt [47] with the permission from Applied Probability Trust .6 is reprinted from Asmussen & Schmidt [49] and parts of IX.3 are reprinted from Asmussen & Rubinstein [46] and parts of VIII.8 . 3 is reprinted from Asmussen & Nielsen [39] and parts of IX. Fig.
This page is intentionally left blank .
M = (1. results and topics to be studied in the rest of the book. and give a very brief summary of some of the models.1) We also refer to t/) ( u) and 0(u.Chapter I Introduction 1 The risk process In this chapter . (1.Rt.2) (O<t<T Ro=ul.3) sup St. They are the main topics of study of the present book. it is frequently more convenient to work with the claim surplus process {St}t>0 defined by St = u . Letting T(u) = inf {t > 0 : Rt < 0} = inf It > 0 : St > u}.4) O<t<oo O<t<T 1 . We denote throughout the initial reserve by u = Ro. The probability O(u) of ultimate ruin is the probability that the reserve ever drops below zero.i(u. we introduce some general notation and terminology. t/i(u) = P (infRt < 0) = P (infR t < 0 t>0 t>0 The probability of ruin before time T is t. (1. T) as ruin probabilities with infinite horizon and finite horizon .T) = P inf Rt < 0 I . For mathematical purposes. as defined in broad terms . MT = sup St. (1. A risk reserve process { Rt}t>o. respectively. is a model for the time evolution of the reserves of an insurance company.
(1. Putting things together.i(u. • Premiums flow in at rate p. the ruin probabilities can then alternatively be written as .5) i. St = E Uk .T) = F (MT > u) = P(r(u) < T). T3.1. (1.b(u) = P (r(u) < oo) = P(M > u).pt. . and T1 is the time of the first claim. and Nt = min {n > 0 : 0rn+1 > t} = max {n > 0: Un < t}• The size of the nth claim is denoted by Un. That is.. 1.6) Sofar we have not imposed any assumptions on the risk reserve process. INTRODUCTION be the time to ruin and the maxima with infinite and finite horizon. the number Nt of arrivals in [0. respectively. However. Figure 1.1 ..E Uk. t] is finite. (1. we see that Nt Nt Rt = u + pt . the time of arrival of the nth claim is an = T1 + • • • + Tn.7) k=1 k=1 The sample paths of {Rt} and {St} and the connection between the two processes are illustrated in Fig. We denote the interarrival times of claims by T2. the following setup will cover the vast majority of the book: • There are only finitely many claims in finite time intervals. Thus. per unit time. say.2 CHAPTER I.
b(u) = 1 for all u.s. (1. A further basic quantity is the safety loading (or the security loading) n defined as the relative amount by which the premium rate p exceeds p. allowing a countable infinity of jumps on Fig. • Brownian motion or more general diffusions. and hence O(u) < 1 for all sufficiently large u. Some main examples of models not incorporated in the above setup are: • Models with a premium depending on the reserve (i.1 Assume that (1. 1.(. 1. however. say 10% . rl= pP P It is sometimes stated in the theoretical literature that the typical values of the safety loading 77 are relatively small.8) The interpretation of p is as the average amount of claim per unit time. However.) V 0.s. on Fig. for example.. a basic references is Gerber [127]. The models we consider will typically have the property that there exists a constant p such that Nt a E Uk k=1 p. We study this case in Ch. It would appear obvious.1.20%. of course. we shall. not discuss whether this actually corresponds to practice. however. and hence . If 77 < 0. t * oo. one may well argue that Brownian motion in itself could be a reasonable model.1. then M = oo a. since any modeling involves some approximative assumptions. VII. We shall discuss Brownian motion somewhat in Chapter IV. that the insurance company should try to ensure 77 > 0. and in fact: Proposition 1.1. We shall not deal with this case either. immaterial.e. one could well replace Rt by Rtnr(u) or RtA.1 the slope of {Rt} should depend also on the level). and the basic ruin probabilities are derived in XI. then M < oo a. 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. • General Levy processes (defined as continuous time processes with stationary independent increments) where the jump component has infinite Levy measure. . For the purpose of studying ruin probabilities this distinction is.. Thus. If 77 > 0.8) holds. but as an approximation to the risk process rather than as a model of intrinsic merit. though many results are straightforward to generalize from the compound Poisson model.
namely that M = oo a. and independent of {(0(t). this needs to be verified in each separate case.i. in connection with risk processes in a Markovian or periodic environment (Chapter VI).10) is a property which we will typically encounter.i. (1. corresponding to the Pdlya process.s. The simplest example is 3(t) = V where V is a r . . rl > 0. and that . If 77 < 0.2 (Cox PROCESSES) Here {Nt} is a Poisson process with random rate /3(t) (say) at time t.T) for {Rt} is given by V)(u) = t/i (u). are i.11) .d.i(u.Tp).10) hold with p constant.s. and independent of {Nt}. If U1.4 CHAPTER I.s. St In concrete models..6EU (on the average.v. M < oo a.Q (say) and U1. t t p  p' t ^ oo. if {(3(t)} is nonergodic.8) that F N. Then the connection between the ruin probabilities for the given risk process {Rt} and those ^(u).. (1.T) = i. (1.3 Assume p 54 1 and define Rt = Rt1p. where {Nt} is a Poisson process with rate . However. If u oo. Proposition 1. Here it is easy to see that p = . zP(u ..8). This case is referred to as the mixed Poisson process.d. U2. Thus p may well be random for such processes. then this limit is > 0 which implies St a$ oo and hence M = oo a.10) Again. Nt)}.oo t 0 J (provided the limit exists).. not all models considered in the literature have this feature: Example 1. 0(u. namely.8) is given by ^t p = EU • lim it (3(s) ds t. and here (1. it is not too difficult to show that p as defined by (1. with the most notable special case being V having a Gamma distribution. tb(u) = 1 for all u holds also when rl = 0.b(u) < 1 for all u when rl > 0. 0 We shall only encounter a few instances of a Cox process. U2. k=1 (1. .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. . INTRODUCTION Proof It follows from (1. However. then similarly limSt/t < 0.. _ St __ k =1 Uk pt a4. are i. The simplest concrete example (to be studied in Chapter III) is the compound Poisson model.
Sundt [354]. Some early surveys are given in Cramer [91]. Schmidli. Notes and references The study of ruin probabilities. Insurance: Mathematics and Economics. [76]. [330]. Since { Rt } has premium rate 1. many results and methods in random walk theory originate from there and the area was ahead of related ones like queueing theory. Embrechts et al. Rolski. the assumption > 0 is equivalent to p < 1. We roughly classify these into two groups . Mitteilungen der Verein der Schweizerischen Versicherungsmathematiker and the Scandinavian Actuarial Journal. Gerber [159]) has a rather different flavour. Pentikainen & Pesonen [101].. some main texts (typically incorporating some ruin theory but emphasizing the topic to a varying degree) are Bowers et al. The term risk theory is often interpreted in a broader sense than as just to comprise the study of ruin probabilities. and we do not get near to the topic anywhere in this book. Segerdahl [334] and Philipson [289]. 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. [101]. U2. In the even more general area of nonlife insurance mathematics. often referred to as collective risk theory or just risk theory. [134]. Note that when p = 1. De Vylder [110]. see also Chapter XI. Daykin et al. the role of the result is to justify to take p = 1. CLAIM SIZE DISTRIBUTIONS 5 The proof is trivial. Straub [353]. Besides in standard journals in probability and applied probability. Grandell [171]. For mixed Poisson processes and Polya processes. Schmidt & Teugels [307] and Seal [326]. in particular. in a number of models.2. Gerber [157]. which is feasible since in most cases the process { Rt } has a similar structure as {Rt} (for example.. Taylor [364].g. while the first mathematically substantial results were given in Lundberg [251] and Cramer [91]. the research literature is often published in journals like Astin Bulletin . Some main later textbooks are (in alphabetical order) Buhlmann [82]. but in probability and applied probability as a whole. The Swedish school was pioneering not only in risk theory. Daykin. Heilmann [191]. An idea of the additional topics and problems one may incorporate under risk theory can be obtained from the survey paper [273] by Norberg. the claim arrivals are Poisson or renewal at the same time). another important early Swedish work is Tacklind [373]. was largely initiated in Sweden in the first half of the century. Hipp & Michel [198]. Cox processes are treated extensively in Grandell [171]. and in fact p < 1 is the fundamental assumption of queueing theory ensuring steadystate behaviour (existence of a limiting stationary distribution). the recent survey by Grandell [173] and references therein. Buhlmann [82].. see e . lighttailed distributions (sometimes the term . Some of the main general ideas were laid down by Lundberg [250].g.. Note that life insurance (e. we shall be able to identify p with the traffic intensity of an associated queue.
g. On the more heuristical side. and heavytailed distributions. Example 2 . For example in the compound Poisson model. In contrast. P B[s]= (8Is ) . 6 has density r(p)xPleax b(x) P and m. B is heavytailed if b[s] = oo for all s > 0. B[s] is finite for some s > 0. Here lighttailed means that the tail B(x) = 1 .f. In particular.x given X > x is again exponential with rate b (this is essentially equivalent to the failure rate being constant). the m. if 1 °O AB Jbos x B(dx) > 0.6 CHAPTER I. the exponential distribution is by far the simplest to deal with in risk theory as well. INTRODUCTION 'Cramertype conditions' is used).g. but different more restrictive definitions are often used: subexponential. a fact which turns out to contain considerable information. Equivalently.3) . a simple stopping time argument shows that this implies that the conditional distribution of the overshoot ST(u) . s<8.u at the time of ruin given r(u) is again exponential u with rate 8. one could mention also the folklore in actuarial practice to consider B heavytailed if '20% of the claims account for more than 80% of the total claims'.2 (THE GAMMA DISTRIBUTION) The gamma distribution with parameters p. As in a number of other applied probability areas. i.2) = 0. 2a Lighttailed distributions Example 2. then the conditional distribution of X . for the compound Poisson model with exponential claim sizes the ruin probability . and can also be interpreted as the (constant) failure rate b(x)/B(x). where B(bo.1 (THE EXPONENTIAL DISTRIBUTION) Here the density is b(x) = beax (2. (2.2 and /LB is the mean of B. The crucial feature is the lack of memory: if X is exponential with rate 6.f.O(u) can be found in closed form.e. regularly varying (see below) or even regularly varying with infinite variance.1) The parameter 6 is referred to as the rate or the intensity.8.B(x) satisfies B(x) = O(e8x) for some s > 0.
JP 1 B(x) r(p ) XP ie ax In the sense of the theory of infinitely divisible distributions..ate (b2 ): L• i=o In the present text.y i=1 where >i ai = 1. then X v Xl + • • • + X. the squared coefficient of variation (s. In particular. An appealing feature is its simple connection to the Poisson process: B(x) = P(Xi + • • • + XP > x) is the probability of at most p . and exponential with rate d. Ruin probabilities for the general case has been studied. are i. one has r(bx. . .v. The exact form of the tail B(x) is given by the incomplete Gamma function r(x. the Gamma density (2. or just the Erlang(p) distribution. 0 < ai < 1. u Example 2 .) VarX1 (EX )2 p is < 1 for p > 1. p.. B(x) = r(p) Asymptotically. p) °° where r (x.3 (THE HYPEREXPONENTIAL DISTRIBUTION) This is defined as a finite mixture of exponential distributions.c. we develop computationally tractable results mainly for the Erlang case (p = 1. P b(x) = r` aibiea. In particular. An important property of the hyperexponential distribution is that its s.. X2. p). u ..2) can be considered as the pth power of the exponential density (2.c. 2. where X1.. CLAIM SIZE DISTRIBUTIONS 7 The mean EX is p/b and the variance Var X is p/b2.1) (or the 1/pth root if p < 1). p) = J tPletdt. x] so that B(x) = r` e..1 Poisson events in [0.. .d. This special case is referred to as the Erlang distribution with p stages. i = 1. > 1 for p < 1 and = 1 for p = 1 (the exponential case).v.). is > 1.i. if p is integer and X has the gamma distribution p. among others. by Grandell & Segerdahl [175] and Thorin [369].2. 0.
6. T) is called the representation. INTRODUCTION Example 2 . which is slightly smaller but more amenable to probabilistic reasoning. resp. q2 q3 (2.6 (DISTRIBUTIONS WITH BOUNDED SUPPORT) This example (i. (or. the Erlang and the hyperexponential distributions. are b(x) = aeTxt. Equivalent characterizations are that the density b(x) has one of the forms q b(x) j=1 = cjxienbx. B(x) = aeTxe where t = Te and e = (1 . However..8 CHAPTER I. The density and c. the restriction T of the intensity matrix of the Markov process to E and the row vector a = (ai)iEE of initial probabilities. equivalently. Example 2 . of which one is absorbing and the rest transient.1 and defer further details to u Chapter VIII.d.7) are possibly complexvalued but the parameters in (2. The couple (a.(2. Important special cases are the exponential.g.8) are realvalued. This class of distributions is popular in older literature on both risk theory and queues.7) q1 b(x) = cjxieWWx + djxi cos(ajx)ea'x + > ejxi sin(bjx)e`ix .4 (PHASETYPE DISTRIBUTIONS) A phasetype distribution is the distribution of the absorption time in a Markov process with finitely many states. a. B(x) > 0 for x < xo) is of course a trivial instance of a lighttailed distribution. We give a more comprehensive treatment in VIII. We give some theory for matrixu exponential distribution in VIII.8) j=1 j=1 j=1 where the parameters in (2. The parameters of a phasetype distribution is the set E of transient states.f. there exists a xo < oo such that B(x) = 0 for x > xo. 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 . . but the current trend in applied probability is to restrict attention to the class of phasetype distributions. 1)' is the column vector with 1 at all entries.e. T) or sometimes the triple (E.5 (DISTRIBUTIONS WITH RATIONAL TRANSFORMS) A distribution B has a rational m..6. 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. a rational Laplace transform) if B[s] _ p(s)/q(s) with p(s) and q(s) being polynomials of finite degree. it is notable from a practical point of view because of reinsurance: if excessofloss reinsurance has been arranged with retention level xo.xo)+ is covered by the reinsurer). Example 2 .f.
(2.pl = 1 W (logx .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). Example 2 . Here failure rates b(x) = b(x)/B(x) play an important role.N(0.N(p. and then b(x) = 0.1). a)/A)a+1' x > a. All moments are finite. a2). There are various variants of the definition around. p is defined as the distribution of ev where V .11) ex log logx 2r p 1 1 2 ( a ) f 1 (lox_P)2} (2. or equivalently as the distribution of a°U+µ where U .1.p a 1 (2.8 (THE LOGNORMAL DISTRIBUTION) The lognormal distribution with parameters a2. CLAIM SIZE DISTRIBUTIONS 9 2b Heavytailed distributions Example 2. It follows that the density is 't (1ogX .7 (THE WEIBULL DISTRIBUTION) This distribution originates from reliability theory. we obtain the Weibull distribution B(x) = eCx'.13) u . (2. the mean u is eµ+a /2 and the second moment is e2µ+2o2. u Example 2 . Writing c = d/r.u l b(x) = d dx or J ax lor 1 exp Asymptotically.9 (THE PARETO DISTRIBUTION) Here the essence is that the tail B(x) decreases like a power of x. In particular. b(x) _ A(1 + (x a The pth moment is finite if and only if p < a .10) The loinormal distribution has moments of all orders. b(x) = crx''le`xr. one being B(x) (1 + X)b(x) (1 + x)a+1' x > 0. the tail is B (x ) 2 x. the exponential distribution representing the simplest example since here b(x) is constant. However.9) which is heavytailed when 0 < r < I. x < a. (2.12) Sometimes also a location parameter a > 0 and a scale parameter A > 0 is allowed.
(2. For p = 1. { s () 1s+3s29s3log(1+2s I p=3.15) x2 + 16x3 ) a3x/2) 3 (1 .e. another standard example is (log x)').'s of the form YX.x6+lr(p) (2.v. Thus.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) . x 4 oo (any L having a limit in (0. i. oo) is slowly varying . 6 is defined as the distribution of et' where V has the gamma density (2.17) where L (x) is slowly varying.16) 11 Example 2. Choudhury & Whitt [1] as the class of distributions of r. (2.12) (here L (x) * 1) and ( 2. u .13). 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). examples of distributions with regularly varying tails are the Pareto distribution (2. INTRODUCTION Example 2. satisfies L(xt)/L(x) 4 1.10 (THE LOGGAMMA DISTRIBUTION) The loggamma distribution with parameters p.14) The pth moment is finite if p < 5 and infinite if p > 5. In general.(1 + 2x + 2x2)e2x) p = 2 (2.2).11 (PARETO MIXTURES OF EXPONENTIALS) This class was introduced by Abate. in particular. in particular. A = 1 and X is standard exponential.(1 + Zx + $ p = 3. where Y is Pareto distributed with a = (p .L( x ). the density is { 3 (1 .1)/p. B(x) = O(xP). u Example 2 . the loggamma distribution (with exponent 5) and a Pareto mixture of exponentials. the loggamma distribution is a Pareto distribution. x + 00. The simplest examples correspond to p small and integervalued. The density is 8p(log x)pi b(x) .10 CHAPTER I.
.3. though the proof of this is nontrivial. Similar discussion applies to the distribution of the accumulated claims (XI.. Namely. Also. the claim size distribution represents of course only one aspect (though a major one).. THE ARRIVAL PROCESS 11 Example 2. (2.. U2. We return to a closer study in IX. and so is the Weibull distribution with 0 < r < 1. but the model also admits a natural interpretation : a large portfolio of insurance holders . which each have a ( timehomogeneous) small rate of experiencing a . By far the most prominent case is the compound Poisson (CramerLundberg) model where {Nt} is Poisson and independent of the claim sizes U1. From a practical point of view. 3 The arrival process For the purpose of modeling a risk process . Thus. one may argue that this difficulty is not resticted to ruin probability theory alone. 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.13 (THE SUBEXPONENTIAL CLASS OF DISTRIBUTIONS) We say that a distribution B is subexponential if xroo lim B `2^ = 2.1. When studying ruin probabilities. it will be seen that we obtain completely different results depending on whether the claim size distribution is exponentially bounded or heavytailed. the subexponential class of distributions provide a convenient framework for studying large classes of heavyu tailed distributions. this phenomenon represents one of the true controversies of the area. 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.1) that any distribution with a regularly varying tail is subexponential. the knowledge of the claim size distribution will typically be based upon statistical data. The reason is in part mathematical since this model is the easiest to analyze. but can never be sure whether this is also so for atypical levels for which far less detailed statistical information is available. However.18) B(x) It can be proved (see IX. for example the lognormal distribution is subexponential (but not regularly varying).4) or even to completely different applied probability areas like extreme value theory: if we are using a Gaussian process to predict extreme value behaviour. We give some discussion on standard methods to distinguish between light and heavy tails in Section 4f. and based upon such information it seems questionable to extrapolate to tail behaviour.
getting away from the simple Poisson process seems a crucial step in making the model more realistic. its basic feature is to allow more variation (bursty arrivals ) than inherent in the simple Poisson process. are i. has some mathematically appealing random walk features . Historically. Some of them have concentrated on the marginal distribution of NT (say T = one year ). in just the same way as the Poisson process arises in telephone traffic (a large number of subscribers each calling with a small rate). such that 8(t) = . but with a general not necessarily exponential distribution ). see 11. in Chapter VII). In order to prove reasonably substantial and interesting results . and also that the ruin problem may be hard to analyze .. so that . where {/3 (t)}too is an arbitrary stochastic process .3(t) fluctuating over time. when Jt = i.d. in particular to allow for certain inhomogeneities. The point of view we take here is Markov dependent random walks in continuous time (Markov additive processes ).. the negative binomial distribution. we study this case in VI . 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 . the periodic and the Markov modulated models also have attractive features . to be studied in Chapter V. radioactive decay (a huge number of atoms each splitting with a tiny rate ) and many other applications. Nevertheless . with a common term {Nt} is a Markovmodulated Poisson process . To the author 's knowledge . This model can be intuitively understood in some simple cases like { Jt} describing weather conditions in car insurance . epidemics in life insurance etc.6. However . An obvious example is 3(t) depending on the time of the year (the season). Cox processes are. This applies also to the case where the claim size distribution depends on the time of the year or . The one we focus on (Chapter VI) is a Markovian environment : the environmental conditions are described by a finite Markov process {Jt }too. too general and one neeed to specialize to more concrete assumptions . however. e. IV (and. Another one is Cox processes..8 (t) is a periodic function of t. A more appealing way to allow for inhomogeneity is by means of an intensity . I.12 CHAPTER I.i. not many detailed studies of the goodnessoffit of the Poisson model in insurance are available . The compound Poisson model is studied in detail in Chapters III.. INTRODUCTION claim . gives rise to an arrival process which is very close to a Poisson process. Mathematically. which facilitate the analysis. with the extension to premiums depending on the reserve.e. In others. found the Poisson distribution to be inadequate and suggested various other univariate distributions as alternatives . the first extension to be studied in detail was {Nt } to be renewal (the interarrival times T1 .(3. 5. This model . it is more questionable whether it provides a model with a similar intuitive content as the Poisson model. T2.g. 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.
6) . interacting particle systems. 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 . R = p(R) in between jumps. Similarly. A general release rule p(x) means that {Vt} decreases according to the differential equation V = p(V) in between jumps. extreme value theory. A stochastic process {Vt } is said to be in the steady state if it is strictly stationary (in the Markov case. others being branching processes. that quite often the emphasis is on computing expected values like EV. dam/storage processes. and here (4. ruin probabilities for risk processes with an input process which is renewal. it is a recurrent theme of this book to stress this connection which is often neglected in the specialized literature on risk theory. stochastic differential equations.1) holds as well provided the risk process has a premium rule depending on the reserve. Markovmodulated or periodic can be related to queues with similar characteristics. this amounts to Vo having the stationary distribution of {Vt}).T) = P(VT > u). queueing theory. The ones which appear most related to risk theory are queueing theory and dam/storage processes. it is desirable to have a set of formulas like (4. 4 A summary of main results and methods 4a Duality with other applied probability models Risk theory may be viewed as one of many applied probability areas. The M/G/1 workload process { Vt } may also be seen as one of the simplest storage models. methods or modeling ideas developed in one area often has relevance for the other one as well.1) permitting to translate freely between risk theory and the queueing/storage setting. It should be noted. and a lot of information on steadystate r. More generally. and the limit t 4 oo is the steadystate limit. The study of the steady state is by far the most dominant topic of queueing and storage theory. with Poisson arrivals and constant release rule p(x) = 1.'s like V is available.4. 0(u) = P(V > u).1) where V is the limit in distribution of Vt as t + oo. In fact. others are quite different. genetics models. In the setting of (4. and which seems well motivated from a practical point of view as well. A SUMMARY OF MAIN RESULTS AND METHODS 13 the environment (VI. point processes and so on.0 (u. Thus. Mathematically.v. however. Some of these have a certain resemblance in flavour and methodology. reliability. (4. time series and Gaussian processes.1). stochastic geometry. this gives only f0 O°i (u)du which is of limited .
3. • The compound Poisson model with premium rate p(x) depending on the reserve and exponential claim size distribution B. Here Vi(u) is given in terms of a matrixexponential function ( Corollary VIII. see Boxma & Cohen [74] and Abate & Whitt [3].'s like the environmental process {Jt} in a Markovmodulated setting. as is typically the case. A prototype of the duality results in this book is Theorem 11.3.g. Here O(u) = peryu where 3 is the arrival intensity. the two areas. Similarly. Here ?P(u) is explicit provided that . the functions w x f d 1 exdx () . see Corollary III. p = 0/8 and y = 8 .8.1.. Example VIII. have to some extent a different flavour. 4b Exact solutions Of course .p(y) y^ Jo p(x) can be written in closed form.1 . Vi(u. INTRODUCTION intrinsic interest .6. Thus . • The compound Poisson model with some rather special heavytailed claim size distributions.1 is a sample path relation should be stressed : in this way the approach also applies to models having supplementary r. The infinite horizon (steady state ) case is covered by letting T oo. . which gives a sample path version of (4. the ideal is to be able to come up with closed form solutions for the ruin probabilities 0(u). e . though overlapping. B(x) = ebx.3.v. which can be expanded into a sum of exponential terms by diagonalization (see.14 CHAPTER I. 3.1) in the setting of a general premium rule p(x): the events {VT > u} and {r (u) < T} coincide when the risk process and the storage process are coupled in a suitable way (via timereversion ). The qualifier 'with just a few phases ' refers to the fact that the diagonalization has to be carried out numerically in higher dimensions.3.1). • The compound Poisson model with constant premium rate p = 1 and B being phasetype with a just few phases .T). 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. 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. The fact that Theorem H. see Corollary VII.2). • The compound Poisson model with a claim size distribution degenerate at one point.
the formulas ( IV. where Furrer [150] recently computed ii(u) as an infinite series involving the Mittag. say the fast Fourier transform (FFT) as implemented in Grubel [179] for infinite horizon ruin probabilities for the renewal model.S(u) 1S(oo) f °D exp {. f {eXp U LX 2. Grubel & Pitts [132] and Grubel & Hermesmeier [180] (see also the Bibliographical Notes in [307] p. [s. O(u. the second best alternative is a numerical procedure which allows to calculate the exact values of the ruin probabilities.1) is the explicit form of the ruin probability when {Rt} is a diffusion with infinitesimal drift and variance µ(x). T) du dT 0 TO 00 in closed form than the ruin probabilities z/'(u). esuTb( u. Ab(u). Also Brownian models or certain skip free random walks lead to explicit solutions (see XI . Abate & Whitt [2]. A SUMMARY OF MAIN RESULTS AND METHODS 15 • The compound Poisson model with a two step premium rule p(x) and B being phasetype with just a few phases. • An astable Levy process with drift . Embrechts. it is easier to find the Laplace transforms = f e8 . see VIII. the only example of something like an explicit expression is the compound Poisson model with constant premium rate p = 1 and exponential claim size distribution . T) themselves. 191).u(y)/a2(y) dy} 4c Numerical methods Next to a closedform solution. We don't discuss Laplace transform inversion much. For the finite horizon ruin probability 0(u. . a2 (x): Ip (u) = where S(u) = f °O exp {. but are somewhat out of the mainstream of the area . 1).b(u)du .2) is the natural scale.4. However. A notable fact ( see again XI. (u. relevant references are Grubel [179]. T) can then be calculated numerically by some method for transform inversion. T).Lef$er function. Given this can be done.7.1) are so complicated that they should rather be viewed as basis for numerical methods than as closedform solutions.ff 2µ(y)/a2(y) dy} dx . Here are some of the main approaches: Laplace transform inversion Often.f f 2µ(y)/a2(y) dy} dx  (4.
. as the solution of linear differential equations or by some series expansion (not necessarily the straightforward Eo U'u/n! one!).4) 00['Y]1)'Y = 0. (4.or integral equation. most often it is more difficult to come up with reasonably simple equations than one may believe at a first sight.16 CHAPTER L INTRODUCTION Matrixanalytic methods This approach is relevant when the claim size distribution is of phasetype (or matrixexponential). and in quite a few cases (Chapter VIII). However. Differential. whereas for the renewal arrival model and the Markovian environment model U has to be calculated numerically. see VIII.3.g.7. it states that i/i(u) .f. 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.and integral equations The idea is here to express 'O(u) or '(u.Ce"u.) B[s].3) in the compound Poisson model which is an integral equation of Volterra type. For the compound Poisson model with p = 1 and claim size distribution B with moment generating function (m. 0(u) is then given in terms of a matrixexponential function euu (here U is some suitable matrix) which can be computed by diagonalization. In the compound Poisson model with p = 1. and in particular the naive idea of conditioning upon process behaviour in [0. 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. One example where this is feasible is the renewal equation for tl'(u) (Corollary III.1) and y > 0 is the solution of the Lundberg equation (4. U is explicit in terms of the model parameters. and carry out the solution by some standard numerical method. u * oo. dt] most often leads to equations involving both differential and integral terms.p)/(13B'[ry] . T) as the solution to a differential. 4d Approximations The CramdrLundberg approximation This is one of the most celebrated result of risk theory (and probability theory as a whole). which can equivalently be written as f3 [7] = 1 +13 .3) where C = (1 .
4. . However. See Chapter IX. For example. Diffusion approximations Here the idea is simply to approximate the risk process by a Brownian motion (or a more general diffusion) by fitting the first and second moment. in such cases the evaluation of C is more cumbersome. Diffusion approximations are easy to calculate. the claim size distribution should have an exponentially decreasing tail B(x). but typically not very precise in their first naive implementation. in some cases the results are even more complete than for light tails.6) are by far the best one can do in terms of finite horizon ruin probabilities '(u. In particular. However. Approximations for O(u) as well as for 1(u. In fact. u > oo. a Markovian environment or periodically varying parameters. J B dx. T) for large u are available in most of the models we discuss. incorporating correction terms may change the picture dramatically. the exact solution is as easy to compute as the CramerLundberg approximation at least in the first two of these three models. It has generalizations to the models with renewal arrivals. In the case of heavytailed distributions. (4. corrected diffusion approximations (see IV.7 and IV.2. This list of approximations does by no means exhaust the topic. The CramerLundberg approximation is renowned not only for its mathematical beauty but also for being very precise. some further possibilities are surveyed in 111 . for the compound Poisson model ^(u) p pu In fact . T). Large claims approximations In order for the CramerLundberg approximation to be valid. often for all u > 0 and not just for large u. other approaches are thus required.6) 4e Bounds and inequalities The outstanding result in the area is Lundberg's inequality (u) < e"lu. when the claim size distribution is of phasetype. and use the fact that first passage probabilities are more readily calculated for diffusions than for the risk process itself. 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.
The standard suggestion is to observe that the mean residual life E[U . 4f Statistical methods Any of the approaches and results above assume that the parameters of the model are completely known. of being somewhat easier to generalize beyond the compound Poisson setting.k (U(`) . T].i. say degenerate at m. . and to plot the empirical mean residual life 1 N . This procedure in itself is fairly straightforward. they have however to be estimated from data. in the compound Poisson model. the difficulty comes in when drawing inference about the ruin probabilities.. one may question whether it is possible to distinguish between claim size distributions which are heavytailed or have an exponentially decaying tail. can we trust the confidence intervals for the large values of u which are of interest? In the present author's opinion.U(k)) i =k+ i . .x U > x] = B(x) f '(yx)B(dx) typically has a finite limit (possibly 0) in the lighttailed case and goes to oo in the heavytailed case. This is proved for the compound Poisson model in 111. For example.18 CHAPTER I. as a general rule. We return to various extensions and sharpenings of Lundberg's inequality (finite horizon versions. In practice. lower bounds etc.3).d. it has the advantage of not involving approximations and also. INTRODUCTION Compared to the CramerLundberg approximation (4. to have smaller ruin probabilities than when B is nondegenerate with the same mean m. more importantly.8. . . which is a standard statistical problem since the claim sizes Ui.) at various places and in various settings. it is a general principle that adding random variation to a model increases the risk. How do we produce a confidence interval? And. . However. though not too many precise mathematical results have been obtained. For example. one expects a model with a deterministic claim size distribution B.. UNT are i. empirical evidence shows that the general principle holds in a broad variety of settings. obtained say by observing the risk process in [0.. However. When comparing different risk models. fitting a parametric model to U1. . 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. e. UNT may be viewed as an interpolation in or smoothing of the histogram). 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. given NT.g.
Simulation may be used just to get some vague insight in the process under study: simulate one or several sample paths. formula (5.4. < U(N) are the order statistics based upon N i. to observe whether one or the other limiting behaviour is apparent in the tail.v's) which can be generated by simulation. 5 Conventions Numbering and reference system The basic principles are just as in the author's earlier book Applied Probability and Queues (Wiley 1987. The infinite horizon case presents a difficulty. UN. good methods exist in a number of models and are based upon representing the ruin probability zb(u) as expected value of a r. and look at them to see whether they exhibit the expected behaviour or some surprises come up. Klnppelberg & Mikosch [134]. A main problem is that ruin is typically a rare event (i.. The chapter number is specified only when it is not the current one..2.e.d. Still.. where U(1) < . CONVENTIONS 19 as function of U(k). claims U1.2.3) and Section VI. because it appears to require an infinitely long simulation.. and in fact methods from that area can often be used in risk theory as well . this is a straightforward way to estimate finite horizon ruin probabilities. . 4g Simulation The development of modern computers have made simulation a popular experimental tool in all branches of applied probability and statistics. formula VI.3 (or just VI. Thus Proposition 4.(5.i. and of course the method is relevant in risk theory as well. For example. (or a functional of the expectation of a set of r. reference [14]. but is not very satisfying.v. in all other chapters than VI where we just write ..5.3) or Section 3 of Chapter VI are referred to as Proposition VI. having small probability) and that therefore naive simulation is expensive or even infeasible in terms of computer time. Truncation to a finite horizon has been used. However. We look at a variety of such methods in Chapter X. respectively.3). in this book referred to as [APQ]). the more typical situation is to perform a Monte Carlo experiment to estimate probabilities (or expectations or distributions) which are not analytically available. The problem is entirely analogous to estimating steadystate characteristics by simulation in queueing/storage theory. . and also discuss how to develop methods which are efficient in terms of producing a small variance for a fixed simulation budget. See further Embrechts.
log E[s] where b[s] is the m. squared coefficient of variation. B[s] the m.d.c. . Abbreviations c. B(x) = P(X < x) = fx.i. if B(x) . and for a defective probability distribution IIGII < 1. h + 0. cumulative distribution function P(X < x) c. b[s] is defined always if Rs < 0 and sometimes in a larger strip (for example.g.d. E. IIGII the total mass (variation ) of a (signed ) measure G .o. w.ceax. mation. random variable s. for a probability distribution IIGII = 1.p. B(x) the tail 1 . oo). i.g.f.r. B is concentrated on [0.f.B(x) = P(X > x) of B.v.20 CHAPTER L INTRODUCTION Proposition 4.f. formula (5. r. (A.g. B(dy). cumulant generating function. References like Proposition A.4. infinitely often l.t.h.2.g. The Laplace transform is b[s].e. with probability Mathematical notation P probability. EX2/(EX)2. as for typical claim size distributions. . A different type of asymptotics: less precise.3) or Section 3.The same symbol B is used for a probability measure B(dx) = P(X E dx) and its c. i. left hand side (of equation) m. n! 27r nn+1/2en. n i oo. In particular.g. If. moment generating function.s.f. E expectation.29) refer to the Appendix.v. independent identically distributed i. right hand side (of equation) r. or a more precise one like eh . (moment generating function) fm e82B(dx) of the distribution B.s.f. with respect to w.h. say a heuristic approxi1 + h + h2/2.f. then for 1s < 5).Used in asymptotic relations to indicate that the ratio between two expressions is 1 in the limit.d. see under b[s] below.
R(s) the real part of a complex number s. a.. N(it. . the ith entry is 1 and all other 0.e. only have finitely many jumps in each finite interval. often the term 'cadlag' (continues a droite avec limites a gauche) is used for the Dproperty. I(A) the indicator function of the event A. (the dimension is usually clear from the context and left unspecified in the notation). Usually. F o r a given set x1. 0 marks the end of a proof. xa. Unless otherwise stated. of numbers. Xt_ the left limit limstt X8f i.5. oo) the space of Rvalued functions which are rightcontionuous and have left limits. In the Frenchinspired literature. the processes we consider are piecewise continuous. a2) the normal distribution with mean p and variance oa2. all stochastic processes considered in this book are assumed to have sample paths in this space. row vectors have lowercase Greek letters like a. (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 have uppercase Roman or Greek letters like T. E[X. 7r. i. the ith unit row vector is e'i.e. though slightly more complicated. Usually. i. 21 D [0. an example or a remark. Notation like f3i and 3(t) in Chapter VI has a similar . intensity interpretation. . Thus.e. Matrices and vectors are denoted by bold letters.. and column vectors have lowercase Roman letters like t. the value just before t. A.A] means E[XI(A)]. CONVENTIONS {6B the mean EX = f xB(dx) of B ABA' the nth moment EXn = f x"B(dx) of B. 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.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). . 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.
cf.g. cf. FL. e. ry The adjustment coefficient. cf. I.5. J the rate parameter of B for the exponential case B(x) = eby.1.22 CHAPTER L INTRODUCTION B the claim size distribution. I. 'q the safety loading . EL the probability measure and its corresponding expectation corresponding to the exponential change of measure given by Lundberg conjugation. though slightly more complicated.1.5. VI. interpretation. p the net amount /3pB of claims per unit time. Notation like BE and B(t) in Chapter VI has a similar. . or quantities with a similar time average interpretation. 111.
used in Chapter VI on risk processes in a Markovian (or periodic) environment. however. The duality results in Section 3 (and. not crucial for the rest of the book. The topic is. More precisely. 5 on random walks and Markov additive processes can be skipped until reading Chapter VI on the Markovian environment model. The reader should therefore observe that it is possible to skip most of the chapter. Sections 4. The general theory is. The likelihood ratio approach in Section 2 is basic for most of the models under study. 5) are. the level of the exposition is. in particular at a first reading of the book. 23 . fundamental ( at least in the author' s opinion) and the probability involved is rather simple and intuitive. in most cases via likelihood ratio arguments. somewhat more advanced than in the rest of the book. All results are proved elsewhere . however. strictly speaking. Due to the generality of the theory. the relevance for the mainstream of exposition is the following: The martingale approach in Section 1 is essentially only used here.Chapter II Some general tools and results The present chapter collects and surveys some topics which repeatedly show up in the study of ruin probabilities. in part. however. Sections 4. When encountered for the first time in connection with the compound Poisson model in Chapter III. a parallel selfcontained treatment is given of the facts needed there.
T(u) < oo] + 0 = eryuE [e7Vu).T(u) < cc] = e7uE {e7f(u) I T(u) < cc] z/.1 is basic for the study of the compound Poisson model in Chapter III.2) takes the form 1 = E [e'ys(). Our first result is a representation formula for O(u) obtained by using the martingale optional stopping theorem . V) (u. the second term converges to 0 by (b) and dominated convergence (e7ST < eryu on {r(u) > T}). (1. We get 1 = Ee7So = E e'Y S(. the time to ruin r(u) is inf It > 0 : St > u}.0. claim size distribution B and p = . Let e(u) = ST(u) . (b) St a$ oo on {T(u) = oo}.QµB < 1.5 can be skipped. using r(u) A T invokes no problems because r(u) A T is bounded by T).. SOME GENERAL TOOLS AND RESULTS The ladder height formula in Theorem 6. Example 1 .2 Consider the compound Poisson model with Poisson arrival rate . StUit.24 CHAPTER II. T(u) > T] . T(u) < T] + E [eryST .2) As T > oo.u denote the overshoot. however.(u). Proposition 1. f1 . and the ruin probabilities are ip(u) = P (T(u) < oo).)AT = E [e7ST(°). The more general Theorem 6. 1 Martingales We consider the claim surplus process {St} of a general risk process. and in the limit (1. T) = P(T(u) < T).1 Assume that (a) for some ry > 0. {e'YS° }t>0 is a martingale. Thus N. 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.
r" where y = S .1 are satisfied. Since {St} has stationary independent increments. and thus Ee'rs° = 1.6a for details) that typically a solution to the Lundberg equation K(y) = 0 exists.ap. From this it is readily seen (see III. of the normal distribution.. By standard formulas for the m.= e"(') where K(a) = .d.1 is satisfied. B(x) _ edx. and (b) follows from p < 1 and the law of large numbers (see Proposition III. the conditional distribution of the overshoot e(u) = U . MARTINGALES 25 where {Nt} is a Poisson process with rate . condition (a) of Proposition 1.Q and the U. O(u ) < e7".f.u + x is again just exponential with rate S.Q(B[a] .g. Since {St} has stationary independent increments. Proof Since c(a) = /3 (B[a] .Q. Eeas° = e"(°) where n(a) = a2a2/2 . Thus. Thus 00 E [e'rt (") I T(u) < oo] = I e5e  dx = f 5edx . it follows that E [e7st+v I J] = e"rstE [e7(st+vSt) I Ft] = e7StEe"rs° = elst where .3/6 < 1.x. the conditions of Proposition 1.i.(„)_ = x is that of a claim size U given U > u .4 (LUNDBERG ' S INEQUALITY ) tion 1 .3 Assume that {Rt} is Brownian motion with variance constant o.1) . are i.1) shows that Eels.1) . with common distribution B (and independent of {Nt}).1.5 For the compound Poisson model with B exponential. Example 1 . Thus. the ruin probability is O(u) = pe.1. 1.a.a it is immediately seen that y = S . From this it is immediately seen that the solution to the Lundberg equation ic(y) = 0 is y = 2p/a2. and thus Ee7s° = 1.1.2.2 and drift p > 0. u Corollary 1. The available information on this jump is that the distribution given r(u) = t and S. and p =. and thus by the memoryless property of the exponential distribution .2(c)). Under the conditions of Proposi Proof Just note that C(u) > 0.Ft = a(S" : v < t). Now at the time r(u) of ruin {St} upcrosses level u by making a jump .a = a . Then {St } is Brownian motion with variance constant o2 and drift p < 0. the martingale property now follows just as in Example 1. A simple calculation (see Proposition III./3. u Corollary 1.
and F. However. A]. u Notes and references The first use of martingales in risk theory is due to Gerber [156]. Proof Just note that ^(u) = 0 by continuity of Brownian motion. The interesting concept is therefore to look for absolute continuity only on finite time intervals (possibly random. F).6 N S = { lim Nt I t +00 t gJ are both in F. oo). F(S) = P(S) = 1. Theorem 2.1) 'though not always: it is not difficult to construct a counterexample say in terms of transient Markov processes.v. Delbaen & Haezendonck [103] and Schmidli [320]. cf. P on (DE. 0 and claim size distributions B. then z/'(u) = e7" where 'y = 21A/a2. Two such processes may be represented by probability measures F.3 below).. and in analogy with the theory of measures on finite dimensional spaces one could study conditions for the RadonNikodym derivative dP/dP to exist. Thus the sets S = I tlim +oot t =. P correspond to the claim surplus process of two compound Poisson risk processes with Poisson rates /3.Ft}too and the Borel afield F. A E Ft. hence so is Nt = limfyo N2`i. Grandell & Schmidli [131]. we look for a process {Lt} (the likelihood ratio process) such that P(A) = E[Lt.F). P are then singular (concentrated on two disjoint measurable sets). A somewhat similar u argument gives singularity when B $ B. the parameters of the two processes can be reconstructed from a single infinite path. But if a $ ^ . as shown by the following example this setup is too restrictive: typically'. which we equip with the natural filtration {. [172].26 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Corollary 1. More recent references are Dassios & Embrechts [98].6 If {Rt} is Brownian motion with variance constant a2 and drift p > 0. Example 2 . and by the law of large numbers for the Poisson process . B. 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. on (DE. The number Nt F) of jumps > e before time t is a (measurable) r. and is further exploited in his book [157]. (2. Embrechts. I.1 Let F.e. . then S and S are disjoint . Grandell [171].
1 J (2. under the assumptions of (ii) we have for A E rg and s < t that A E Ft as well and hence E[L8.2(i).F). Then Ft (A) = E[Lt. A E Ft . (i) If {Lt}t> o is a nonnegative martingale w.r. define P by Pt (A) = E[Lt. If r is a stopping time and G E PT. 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 . This proves (i). that the restriction of P to (DE. ({.Tt) is absolutely continuous w. Ft). Then Lt > 0 and ELt = 1 ensure that Pt is a probability measure on (DE.Pt}adapted process {Lt}t>o (2. F the Borel o•field and P a given probability measure on (DE. A]. then {Lt} is a nonnegative martingale w..F such that (2. Hence E [_ . By the martingale property.F8] = L8 and the martingale property.2 Let {Ft}t>o be the natural filtration on DE. using the martingale property in the fourth step. if for some probability measure P and some {. (ii) Conversely. P be as in Proposition 2. Lets < t. then { 1 P(G) = EG . ({Ft} .t.1) holds. A]. F) such that ELt = 1.Ft}. A E F8.Pt)) The following result gives the connection to martingales.2.r. Proposition 2. The truth of this for all A E Y.t.3 Let {Lt}. . Conversely. implies that E[LtI. the restriction of P to (DE.1) and nonnegativity by letting A = {Lt < 0}. G C {T < oo}. Finally. P) such that LLt = 1. G l ] = E [_I(G)E[LTIFT] ] = E { _I(G)Lr ] = P(G). Hence the family {Pt} is t>o consistent and hence extendable to a probability measure F on (DE.e. we have E [ LTIFT]1 = LT on {T < T}. _. Lt < 0] can only be nonnegative if P(A) = 0. A] = E[Lt. A E F. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 27 (i. then there exists a unique probability measure Pon .r. . G ] = E [LT .Y) such that P(A) = Pt(A).1) holds.t. Proof Under the assumptions of (i). Then P(A) = E[Lt.2) Proof Assume first G C {T < T} for some fixed deterministic T < oo. ELt = 1 follows by taking A = DE in (2.A] = EE[LtI(A)IF8] = EI(A)E[LtIFB] = EI(A)L8 = PS(A).
1) in Proposition 1. r(u) < oo] occuring there than with the (conditional) expectation E[e'r{(u ) Jr(u) < oo] in (1. first in the Markov case and next (Sections 4. each F. The problem is thus to investigate which characteristics of {Xt} and {Lt} ensure a given set of properties of the changed probability measure. u From Theorem 2. the natural filtration {.4) compared to (1.t. Rt) or Xt = (Jt.Ft} .. applying (2. (2.s. say. we assume for simplicity that {Xt} has Dpaths. Now just rewrite the r.28 CHAPTER II. St). For the definition.2) follows by monotone convergence. and this problem will now be studied.3) to G of{r < T} we get 1111 F(Gn {r <T}) = E[ 1 . Xt = (Jt. 1 Since everything is non negative. First we ask when the Markov property is preserved. 5) for processes with some randomwalklike structure.Rt} the claim surplus process and {Jt} a process of supplementary variables possibly needed to make the process Markovian. both sides are increasing in T. The crucial step is to obtain information on the process evolving according to F.4) Proof Letting G = {r(u) < oo}. Lr(u) 11 The advantage of (2.1). SOME GENERAL TOOLS AND RESULTS In the general case .r.1: Corollary 2.Gn {r <T} .3 we obtain a likelihood ratio representation of the ruin probability V) (u) parallel to the martingale representation (1.1.h. is nonnegative and has Ey Lt = 1 for all x. we need the concept of a multiplicative functional. is Markov w.O(u) = eryuE[e'YC(u). in continuous time (the discrete time case is parallel but slightly simpler). {St} = {u . of (2.4 Under condition (a) of Proposition 1.t. A change of measure is performed by finding a process {Lt} which is a martingale w..1) is that it seems in general easier to deal with the (unconditional) expectation E[eryVu). To this end. we have F(G) = V )(u).r.2) by noting that 1 = ersr(„) = e1'ue7Ou). Consider a (timehomogeneous) Markov process {Xt} with state space E. and letting T * oo. where {Rt} is the risk reserve process. one would typically have Xt = Rt. T(u) < oo]. t. Xt = St. . (2. In the context of ruin probabilities.
the Markov property can be written E. t. or. Similarly. Zt.5 Let {Xt} be Markov w.Ft }. 0 .v. By definition of Px.F8measurable r.Ft]. Then the family {Px}xEE defines a timehomogeneous Markov process if and only if {Lt} is a multiplicative functional.(Xtitl) with all t(i) < t + s.[Y. the natural filtration {Ft} on DE.v. where Ot is the shift operator. (2.Ft] = Ex.'s of the form fl' f. Ex[Lt+8Zt(Y8 o et)] = Ex[LtZt(Y8 o 0t)(L8 o Bt)]. t and let Px be the probability measure given by t.5) Pxa. Indeed.s.r. since Ext [L8Y8] = E[(Y8 o et)(L8 o 8t)I.v..Pt+8measurable r.t. (2. (2. o 9t = V.7).2.. The converse follows since the class of r. Y8. o 9tI.'s of the form Zt • (Y8 o 0t) comprises all r. s. Proof Since both sides of (2.YB] for any Ftmeasurable r.6) implies (2. let {Lt} be a nonnegative martingale with Ex Lt = 1 for all x.(A) = Ex [Lt.6) for any .7). which in turn is the same as Ex[Lt+8Zt • (V8 o Bt)] = Ex[Lt • (L8 o 91)Z1 • (Y8 o et)] (2.7) for any Ftmeasurable Zt and any .Ft+8measurable.5) are Tt+e measurable.Ftmeasurable. since Zt • (Y8 o Ot ) is . LIKELIHOOD RATIOS AND CHANGE OF MEASURE 29 on DE and define {Lt} to be a multiplicative functional if {Lt} is adapted to {.5) is equivalent to Ex[Lt+8Vt+8] = E8[Lt • (L8 o 91)Vt+8] for any .Y8f t < s. A]. Vt+e.8) which is the same as (2. and then L. which is the same as Ex[Zt(Y8 o Bt)] = E8[ZtEx. Lt has the form Lt = 'Pt ({x }0<u<t) for some mapping cot : DE[O. oo).v. nonnegative and Lt+8 = Lt•(Lso9t) (2. t] * [0. for all x. (2.T9measurable Y8. The precise meaning of this is the following: being . ({Xt+u} 0<u<8) Theorem 2.v. this in turn means Ex[Lt+8Zt(V8 oet)] = Ex[LtZtExt[L8Y8]].
. Thus R = Ro + f p(R8) ds . A more elementary version along the lines of Theorem 2. t] = LtExt L8 = Lt.Ft] = LtE[L8 o 9t I. R = p(R)).1) The initial condition is arbitrary.. (using the Markov property in the second step) so that the martingale property is automatic. The formulation has applications to virtually all the risk models studied in this book. The storage process {Vt }o<t<T is essentially defined by time reversion. .5 can be found in Kuchler & Sorensen [240]. UN.6 For {u . (3.. Indeed. then Remark 2. {Vt} . 3 Duality with other applied probability models In this section.. 0 < vl < . < aN < T. see Dynkin [128] and Kunita [239]. . In between jumps. . CN. We work on a finite time interval [0. In between jumps. reflection at zero and initiar condition Vo = 0. we shall establish a general connection between ruin probabilities and certain stochastic processes which occurs for example as models for queueing and storage.At where At = k: vk <t U. t. u Notes and references The results of the present section are essentially known in a very general Markov process formulation. SOME GENERAL TOOLS AND RESULTS to define a timehomogeneous Markov process. with a proof somewhat different from the present one. . The corresponding claim sizes are Ul. the arrival epochs are Qi... aN where or* = T UN_k+l.. . and the time to ruin is 7(u) = inf {t > 0: Rt < 0}.. The result is a sample path relation. } E[Lt+B I.e.. Ro = u (say).. More precisely . it xEE suffices to assume that {Lt} is a multiplicative functional with Ex Lt = 1 for all x..30 CHAPTER H. the premium rate is p(r) when the reserve is r (i. 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. and thus for the moment no parametric assumptions (on say the structure of the arrival process) are needed.
_. Note that these definitions make {Rt} rightcontinuous (as standard) and {Vt} leftcontinuous.3.. That is. {Vt} remains at 0 until the next arrival).. instead of (3...3) (u) Proof Let rt' denote the solution of R = p(R) subject to r0 = u.1..AT_t.__..11 4.T) = inf Rt < 0 P (O<t<T P(r(u) < T) be the ruin probability.1 Define r(u) = inf It > 0: Rt < 0} (r(u) = oo if Rt > 0 for all t < T) and let ii(u. V = p(V)). V)(u. In particular. (3. .__.1 The events {T(u) < T} and {VT > u} coincide..___ . 3.1) we have Vt = At  f P(Vs)ds where A= U= AT .x.. (3.____•_.2) k: ok <t and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0._: 1} 0 011 =T01N ^N3 To 0 011 014 01N Figure 3. Then rt°) > rt°) for all t when u > v.T) = P(VT > u)..e. :. Theorem 3. The sample path relation between these two processes is illustrated in Fig. DUALITY WITH OTHER APPLIED PROBABILITY MODELS 31 decreases at rate p(r) when Vt = r (i.
d. Hence RQ„ > 0 for all n < N. Hence if n satisfies VVN_n+1 = 0 (such a n exists. and in between rainfalls water is released at rate p(r) when Vt (the content) is r. Historically.2 Consider the compound Poisson risk model with a general premium rule p(r).i.1 with Ro = u = u2). see Siegmund [344].3). if and only if O(u) < 1 for all u. 3. Some further relevant more general references are Asmussen [21] and Asmussen & Sigman [51].32 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Suppose first VT > u (this situation corresponds to the solid path of {Rt} in Fig. we have RQ„ < 0 so that indeed r(u) < T. Theorem 3. Then similarly VVN = r0. if nothing else n = N). the connection between risk theory and other applied probability areas appears first to have been noted by Prabhu [293] in a queueing context.Ul < roil  Ul = RQ„ Va1V_1 < RQ2. Then the storage process {Vt} has a proper limit in distribution. u A basic example is when {Rt} is the risk reserve process corresponding to claims arriving at Poisson rate . Proof Let T ^ oo in (3.U1 = Rol.1 and its proof is from Asmussen & Schock Petersen [50]. 3. 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. and a general premium rule p(r) when the reserve is r.1 with Ro = u = ul).3 and being i. we can repeat the argument and get VoN_1 > Ra2 and so on. the distinction between right.U1 > roil . and since ruin can only occur at the times of claims. Then Vo. We get: Corollary 3. Suppose next VT < u (this situation corresponds to the broken path of {Rt} in Fig.2 from Harrison & Resnick [188]. = r(VT) . we have r(u) > T. The arrival epochs correspond to rainfalls.and left continuity is immaterial because the probability of a Poisson arrival at any fixed time t is zero). Nevertheless.T l . Resnick & Tweedie [79]. with distribution B. If VaN > 0. and so on. Historically. say V. one may feel that the interaction between the different areas has been surprisingly limited even up to today. . u Notes and references Some main reference on storage processes are Harrison & Resnick [187] and Brockwell. Thus we may think of {Vt} as having compound Poisson input and being defined for all t < oo. Corollary 3. and then '0 (u) = P(V > u). Then the time reversibility of the Poisson process ensures that {At } and {At } have the same distribution (for finitedimensional distributions. this represents a model for storage.
. N min (Y1 + • • • + YNn) n=0..d. has a proper limit W in distribution as n + oo.Y1 according to Wo = 0.. can be viewed as the reflected version of the random walk with increments Z1.. Xo = 0. there is an analogue of Theorem 3. the Lindley process Wo. Theorem 4. Then the events {r(u) < N} and {WN > u} coincide...1. Here F is a general probability distribution on R (the special case of F being concentrated on {1.N From this the result immediately follows. WN = YN .. W2.1. n=0. WN be the Lindley process generated by Z1 = YN.s...min n=0.2 The following assertions are equivalent: (a) 0(u) = P(r(u) < oo) < 1 for all u > 0. W1.. Z2...1) Thus {Wn}n=o.. Most often. ...YNn+1) n=0..w. evolves as a random walk with increments Z1i Z2.. 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.e. {Wn}n=0.. (c) The Lindley process {WN} generated by Zl = Y1. ..1 Let r(u) = inf In: u + Y1 + • • • + Yn < 0}.. = Xo + Y1 + • • • + Y. I. where the Yi are i ...i.1..1.h. 1} is often referred to as simple random walk or Bernoulli random walk).. For a given i....1. For discrete time random walks . as long as the random walk only takes nonnegative values. 0 Corollary 4. with common distribution F (say).. . . just verify that the r .2) satisfies the same recursion as in (4.1)). . is defined by assigning Wo some arbitrary value > 0 and letting Wn+1 = (Wn + Zn+1)+• (4. .1.. Let further N be fixed and let Wo. generated by Z1..... . Z2 = YN_1 i .Yl min (YN .. Z2. of (4..... N min (Y1 + + Yn). Z2 . ..N (4.. . 0). Z2 = Y2.. i.2).d. .. (b) 1/i(u) = P(•r(u) < oo) > 0 as u * oo. if Wo = 0 then (Z1+•••+Zn) WN = Zl+•••+ZN.. Proof By (4. hits (oo. R valued sequence Z1.4.2) (for a rigorous proof. ZN = .. In particular.. W1......1 in terms of Lindley processes . and is reset to 0 once the r.
e. the condition 00 F(YI+•••+ Yn<0)<00 n=1 is known to be necessary and sufficient ((APQ] p. Y1) has the same distribution as (Y1. (e)Yi+•••+Yn 74 . Combining these facts gives easily the equivalence of (a)(d)... a sufficient condition for (e) is that EY is welldefined and > 0. (Z1 + • • • + Zn) = m and P(W > u) = P(M > u) = i (u ). . (e). Remark 4 .s. . .... W v m and P(W > u) = P (m > u) = 0(u). Next consider change of measure via likelihood ratios... YN in Theorem 4.. or M < oo a. equivalently.i. . .1 is actually not necessary .1 is equivalent to WN D MN = (Z1 + .5 does not necessarily lead to a random walk: if.s. By Kolmogorov's 01 law.l. doubly u infinite (n'= 0... (d) #. .l. a Markovian change of measure as in Theorem 2.. either M = oo a. Similarly. 176) but appears to be rather intractable.2 and Theorem 4.. . For a random walk.1.g. YN).1 have the same distribution for n = 0.the result is a sample path relation as is Theorem 3.g.2.. 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. The converse follows from general random walk theory since it is standard that lim sup (Y1 + • • + Yn) = oo when Y1 + • • • + Yn 74 oo. N. In general.. The following result gives the necessary and sufficient condition for {Ln} to define a new random walk: ..ooa. .. SOME GENERAL TOOLS AND RESULTS (d) m = inf. the Lindley processes in Corollary 4.3 The i. ±2.) and defines Zn = Yn.. then the restrictions of Fx.1. . .1... ±1. (Yi + • • • + Yn) > oo a.d. . Proof Since (YN.N so that WN _P4 M = supra=0.s .. In that case .) sup n=0.. 0 By the law of large numbers. Px to Fn are equivalent (have the same null sets) so that the likelihood ratio Ln exists.1..34 CHAPTER II.. assumption on the Z1. ZN or.s.=o... + Z.o. the Y1. Thus the assertion of Theorem 4.. One then assumes Yn to be a stationary sequence. Clearly. there is a more general version of Corollary 4. w.
3) 1Pxa. RANDOM WALKS INDISCRETE OR CONTINUOUS TIME 35 Proposition 4. In that case. implying g (x. .Y2) = h(1'i)h(I'a).g. Y) f (Y)] for all f and x. Y ) f (Y)] = E[g(O.nrc(a )} (4.4. Y) = h(Y ) a.g. 100 )..4). where h (y) = g(0.. we get L2 = L1 (L1 o91 ) = h(Y1)g(X1.5 Consider a random walk and an a such that c(a) = log F[a] = log Ee° ' is finite. h(Yn) (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.. this means E(g(x. Proof If (4. Then the change of measure in Theorem 2.. the changed increment distribution is F(x ) = E[h(Y). (a) = log F(a] is the c. Conversely. The corresponding likelihood ratio is Ln = exp {a (Y1 + • • • + Yn) . Then the change of measure in Theorem 2.3) holds for n = 1. Y < x]. y ). cf. (4.4..4 Let {Ln} be a multiplicative functional of a random walk with E_L.5 corresponds to a new random walk with changed increment distribution F(x) = e'(a) Jr e"'F(dy) .4) ({Ln} is the familiar Wald martingale . Y1).3) holds.. In particular. of F)..f.5 corresponds to a new random walk if and only if Ln = h(Y1) . We get: Corollary 4. = 1 for all n and x.s.s. For n = 2. Breiman [78] p. Since L1 has the form g (Xo. u A particular important example is exponential change of measure (h(y) = e°y'(") where r. and so onforn =3. for some function h with Eh(Y) = 1. the random walk property implies Ex f (Y1) = Eo f (Y1 ). and define Ln by (4. e.
Roughly.. given by a the increment distribution F(x) = P(Xn+l . they arise as models for the reserve or claim surplus at a discrete sequence of instants. {Xt} can be written as the independent sum of a pure drift {pt}. the claim surplus process for the compound Poisson risk model . i.Xn < x). The simplest case is 3 = JhvMM < oo.). say the beginning of each month or year . we are . say by recording the reserve or claim surplus just before or just after claims (see Chapter V for some fundamental examples). or imbedded into continuous time processes . However. a Brownian component {Bt} (scaled by a variance constant) and a pure jump process {Mt}.. The appropriate generalization of random walks to continuous time is processes with stationary independent increments (Levy processes). {Xt} is a random walk. 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.5) Note that the structure of such a process admits a complete description.6) More precisely. Xt (n)t(n1) being independent whenever t(O) < t(1 ) < . corresponds to a process with stationary independent increments and u = p. .7) for all e > 0. (4. < t(n) and with Xt( i)_t(i_l) having distribution depending only on t(i) .36 CHAPTER II. which corresponds to the compound Poisson case: here jumps of {Mt} occur at rate 0 and have distribution B = v/0 (in particular .Xt)I.. the pure jump process is given by its Levy measure v(dx).. but we omit the details ). a positive measure on R with the properties e J x2v(dx) < oo. . (4. this description needs some amendments. with premium rate p. e f x:IxJ>e} v(dx) < oo (4. Xt =Xo+pt+oBt+Mt. In discrete time.1). An equivalent characterisation is {Xt} being Markov with state space R and E [f (Xt+e . Xt(2)_t(l). SOME GENERAL TOOLS AND RESULTS Discrete time random walks have classical applications in queueing theory via the Lindley process representation of the waiting time . In continuous time.t(i . the tradition in the area is to use continuous time models. The traditional formal definition is that {Xt} is Rvalued with the increments Xt(1)_t(o). v2 = 0 and v = 3B).Ft] = Eof (X. see Chapter V.e. In risk theory. the interpretation is that the rate of a jump of size x is v(dx) (if f of Ixlv (dx) = oo. however.
Then the storage process {Vt} has constant release rate 1.4. 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). i. then Ee'(xtxo) = Eoeaxt = etx(a). 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.s. cf.Q and distribution B of the service times of the arriving customers. jxJ v(dx) < oo.v. A different interpretation is as the workload or virtual waiting time process in an M/G/1 queue. VT + V for some r.10) . Proposition 4.] Processes with a more complicated path structure like Brownian motion or jump processes with unbounded Levy measure are not covered by Section 3. and b(u) = P(V > u). and the reflected version is then defined by means of the abstract reflection operator as in (4. is easily seen to be f3pB < 1. where VT is the virtual waiting time at time T in an initially empty M/G/1 queue with the same arrival rate /3 and the service times having the same distribution B as the claims in the risk process. having Poisson arrivals with rate . where c(a) = ap + a2a2/2 + f 00 provided the Levy measure of the jump part {Mt} satisfies f". oo].6 In the compound Poisson risk model with constant premium rate p(r) . T) = P(VT > u). defined as a system with a single server working at a unit rate.7 If {Xt} has stationary independent increments as in (4.2).1)v(dx) (4. O(u. First assume in the setting of Section 3 that {Rt} is the risk reserve process for the compound Poisson risk model with constant premium rate p(r) = 1.1.e. Corollary 4.min Xt (4.6). Furthermore. Now consider reflected versions of processes with stationary independent increments.8) O<t<T (assuming Wo = Xo = 0 for simplicity). 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. Chapter III. [The condition for V < oo a. (ex .3 and decreases linearly at rate 1 in between jumps. has upwards jumps governed by B at the epochs of a Poisson process with rate . V E [0. WT = XT .
5) and get E [f(Xt+B . t.10) is the LevyKhinchine representation of the c. Eea(µt + QBt) = et{aµ +a2OZ/2}. Then the Markov process given by Theorem 2.f. Q2 = v2.g. By explicit calculation . v(dx) = e9xv (dx). u Note that (4. Xt +B . In particular.1. we use the characterization (4. we show in the compound Poisson case ( IlvIl < oo) in Proposition III. Proof For the first statement . Theorem 4 . if Lt = e9(xt .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 . let e" (a ) = Eoeaxl. Xt Xo) with E2Lt = 1 for all x. 5 has stationary independent increments as well. (4.Xt)L8 o 0tIFt] = E [f (Xt+s . Chung [86]). use the representation as limit of compound Poisson processes.xo)tk ( e).1 ) v(dx) . e. Then l e" (a) = Eo [ Li ea "] = eK (9)Eo {e ( a+9)x1 J I = er(a+o )K(B) R(a) = K(a + 0) .38 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Proof By standard formulas for the normal distribution.Xo is necessarily infinitely divisible when {Xt} has stationary independent increments.1 that E eaMt = exp fmoo In the general case . then the changed parameters in the representation (4.Xt)I'Ftl = E [f(Xt+B .g. X8) = Eof (X8)L8 = Eof (X8)• For the second. 8 Assume that {Xt} has stationary independent increments and that {Lt} is a nonnegative multiplicative functional of the form Lt = g(t.6) are µ = µ + Oo2 .1)eexv(dx). of an infinitely divisible distribution (see. This is of course no coincidence since the distribution of Xl .11) (eax .Xt)g(s. .Xt)I Ftl = Eof (X8)g(s..
Example 4 . U2.3 and claim size distribution B.3B[B]. B(dx) = B[9] B(dx).l3 and claim u size distribution B.o[f (S8)g(J8)]. <t whenever the RadonNikodym derivative dB/dB exists (e. the structure of MAP's is completely understood when E is finite: 2and only there .Ft] = Ejt. 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... dB/dB = b/b when B.3 =. Thus (since µ = p = 1.11 For an example of a likelihood ratio not covered by Theorem 4. 0. 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 .0.9 If X0 = 0.4). corresponding to p = 1. Then we can write v(dx) _ /3eOxB(dx) = / (dx). b with b(x) > 0 for all x such that b(x) > 0). (5. v(dx) _ .(3 = . Ei instead of P2. Recalling that U1.3 and claim size distribution B # B. . one reason is that in parts of the applied probability literature. MARKOV ADDITIVE PROCESSES 39 Remark 4.0 in the following. are the arrival times and U1.g. Ei..2.. u 5 Markov additive processes A Markov additive processes.(3B(dx). it is then easily seen that Lt = H dB(Ui) i:o. is defined as a bivariate Markov process {Xt} = {(Jt. abbreviated as MAP in this section2.St)g(Jt+s)I. Example 4 .5. and let the Px refer to the claim surplus process of another compound Poisson risk process with Poisson rate. As for processes with stationary independent increments . then the martingale {eex(t)tk(e)} is the continuous u time analogue of the Wald martingale (4. the corresponding claim sizes .8. . where . b = a = 0) the changed process is the claim surplus process of another compound Poisson risk process with Poisson rate . we write Pi. MAP stands for the Markovian arrival process discussed below.1) For shorthand . a = 0. B have densities b..10 Let Xt be the claim surplus process of a compound Poisson risk process with Poisson rate .
In continuous time (assuming Dpaths).J1=j)= Fij (dx) Pij In simulation language. by generating Yn according to Hij when J„_1 = i. If all Fij are concentrated on (0. v. {St} evolves like a process with stationary independent increments and the parameters pi. let {Jt} be standard Brownian motion on the line.1 For a MAP in discrete time and with E finite.40 CHAPTER H.) If E is infinite a MAP may be much more complicated. which we omit and refer to Neveu [272] or cinlar [87]..[a) = (Ei[easl.g.i. {Jt} is specified by its intensity matrix A = (Aij)i. with the Y„ being interpreted as interarrival times.jEE• On an interval [t.9 EE = (iii&ij[a])i j EE . the converse requires a proof.it = A. consider the matrix Ft [a] with ijth element least Ei . SOME GENERAL TOOLS AND RESULTS In discrete time. An alternative description is in terms of the transition matrix P = (piA. . vi(dx) in (4. a MAP is the same as a semiMarkov or Markov renewal process. In addition.Sr_1. t+s) where Jt . Y2. Fn[a] = F[a]n where P[a] = P . 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. Y1 E dx) where Y„ = S„ .f.6) depending on i. As a generalization of the m. Jn = j. oo). the distribution of which has some distribution Bij. 1 J1 ='^])iJEE = (Fij[a])i . this means that the MAP can be simulated by first simulating the Markov chain {J„} and next the Y1.. Proposition 5. (That a process with this description is a MAP is obvious. As an example.o(Ji = j.... a jump of {Jt} from i to j # i has probability qij of giving rise to a jump of {St} at the same time.jEE (here pij = Pi(J1 = j)) and the probability measures Hij(dx)=P(Y1 EdxlJo=i. a MAP is specified by the measurevalued matrix (kernel) F(dx) whose ijth element is the defective probability distribution Fij(dx) = Pi.
2 Let E be finite and consider a continuous time Markov additive process with parameters A.1 )v(dx). Then. qij. By PerronFrobenius theory (see A.1)) . assume that the Markov chain/process {Jt} is ergodic. \ diag Ft[a] = Ft[a]K. 013 . Bij (i. we infer that in the discrete time case the . Then the matrix Pt[a] with ijth element Ei [east. where K[a] = A+ (r.ijgij(Bij[a] . Jt = k] { xk kEE j la] . aSt h = (1 + Ajjh) Ei [east .qkj + k?^j qkj Bkj [a] } = Ei [east. j E E) and So = 0. this means that F't+h [a] = Ft[a] II+h(rc(i)(a)) +hA+h(Aijgij(Bij[a]1)) I.4c).1) } (recall that qjj = 0). a= . In matrix formulation . Jt = k] { 1 . J1 = A which in matrix formulation is the same as Fn+1 [a] = Fn[a]F[a]. kEE Jn = k]Ek[e"Y" . Proof Let {Stt) } be a process with stationary independent increments and pa rameters pi . pi. vi(dx) (i E E). u Proposition 5. up to o( h) terms.5. vi(dx). Jt = j] is given by etK[a]. MARKOV ADDITIVE PROCESSES Proof Conditioning upon (Jn.(')(a)) diag + (). which in conjunction with Fo[a] = I implies Ft[a] = etK[a) according to the standard solution formula for systems of linear differential equations. 00 r(i) (a) = api + a2ot /2 + f (e° . Jt = j] (1 + htc (j) (a)) j + Ak j qk j (Bk +h E Ei [east . Jn+1 = A] = 41 Ei[ e 5„. u In the following. Jt = j] Ejesh'^ + E Ak j hEi [ease . Sn ) yields Ei[easn+ '.
c(a) (and h(")). Eie"sth^a) = e'Pt[a]h( a) = e. (5. and write k = k(°).r. as will be seen from the following results. its derivatives are 'asymptotic cumulants'. Proposition 5. just note that [a]h(a) = eietK (a)h(a) = etK(a)h(a). 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.42 CHAPTER II.4c).etx It then follows that E feast+^(t+v)K(a)h(a) I ^tl l . Proof By PerronFrobenius theory (see A. u Let k(a) denote the derivative of h() w. h(") are only given up to a constants.h(a)vva)etw(a). Jeast. and we shall take V(a)h(a) = 1. of a random walk.3 Ei [east. Yrh(a ) = 1.5 EiSt = tK'(0) + ki .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).e=e°tk. we are free to impose two normalizations. The corresponding left and right eigenvectors v("). cf. a. Since v(").Eikjt = ttc'(0) + ki . Then h(°) = e.2) where 7r = v(°) is the stationary distribution.f. . Corollary 5.7. We also get an analogue of the Wald martingale for random walks: Proposition 5. and appropriate generalizations of the Wald martingale (and the associated change of measure) can be defined in terms of .tK(a)h(a) J jj it L o is a martingale. Proof For the first assertion. Corollary 5. cf.4. In particular.4 Eie"sth(a) = h=a)et?(").t. Corollary 5. Jt = j] . Furthermore. h(") may be chosen with strictly positive components. The function ic(a) plays in many respects the same role as the cumulant g.
St]2 = t2/c'(0 ) 2 + 2ttc (0)vk .7 No matter the initial distribution v of Jo. E=ST = tc'(0)E7. (5. Corollary 5. 43 Ei [Steast h(a) + east k^a)1 = et"(a) (kia) + tic (a)hia)) . For the second . u The argument is slightly heuristic (e.4. summing and letting a = 0 yields E„ [St + 2Stkj. we differentiate (5.. for a random walk: Corollary 5.6 For any stopping time T with finite mean.4) . [E. .a) + ttc (a)2hia ) Multiplying by v=.3) Let a = 0 and recall that h(°) = e so that 0=°) = h(o) = 1. MARKOV ADDITIVE PROCESSES Proof By differentiation in Proposition 5.5.5 yields + W (a)k. (5.g. one obtains a generalization of Wald's identity EST = Er • ES. 8 Also for E being infinite (possibly uncountable ). tam E tSt a (0). t a oo. Ee"st typically grows asymptotically exponential with a rate ic(a) independent of the initial condition (i.2ttc (0)Evkjt + 0(i)..") }) . the existence of exponential moments is assumed ) but can be made rigorous by passing to characteristic functions.3) to get Ej [St a " st h i(a ) + 2Ste"st k(a) + e"st k^a) J etI(a) (kia )' + ttc (a)ki") + t {ic"(a)h.5. Squaring in Corollary 5.+ k. ] = t2tc (0)2 + 2tK'(0)vk + ttc"(0) + O(1) . the distribution of Jo). In the same way. Remark 5 .Eikjr . there is typically a function h = h(") on E and a ^c(a) such that Ey a"st t" (") * h(x). Since it is easily seen by an asymptotic independence argument that E„ [Stkjt] u = trc'(0) E„kjt + O(1). t im v^"St = '(0) Proof The first assertion is immediate by dividing by tin Corollary 5. subtraction yields Vary St = tic"(0) + O(1).e. More precisely.
gha(i. V. Then {Lt } is a multiplicative functional. 0 Proposition 5..6. 1) (i. Remark 5. h(Jo) Lo is a Px martingale for each x E E.10 Let {(Jt.3b and Remark VI. see.for the present purposes.6) We shall not exploit this approach systematically. 0) = n(a) h(i). inconvenient due to the unboundedness of ea8 so we shall not aim for complete rigour but interpret C in a broader sense.5) is a martingale can be expressed via the infinitesimal generator G of {Xt } = { (Jt. let ha(i.6. we take the martingale property as our basic condition below (though this is automatic in the finite case).4 that { h(Jt) easttK(a) L o (5.f (x) tyo t provided the limit exists. u forsEE). this is. St)} be a MAP and let 0 be such that h(Jt) OStt. St) } as follows. G is defined as Gf (x) = lim Exf (Xt) . s) = ea8h(i).e. First.. however. Usually. In view of this discussion . where {Jt} is deterministic period motion on E = [0.9 The condition that (5. this leads to h(i) + tcha( i.s.5 defines a new MAP. xEE .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. For t small . however. some extra conditions are imposed. Jt = (s+t) mod 1 P8a.5. From (5.5) is a martingale . 0) = h(i )( 1 + ttc(a)).(9) {Lt}t>o = . Given a function h on E. An example beyond the finite case occurs for periodic risk processes in VI. and the family {f LEE given by Theorem 2. We then want to determine h and x(a) such that Ejeasth (Jt) = etK(a)h(i). (5.44 CHAPTER IL SOME GENERAL TOOLS AND RESULTS for all x E E. in particular that f is bounded.
0<b<oo. Here Oh(e) is the diagonal matrix with the h=e) on the diagonal.c(0)e = tc(0)e .11 Consider the irreducible case with E finite.1) .Qi < oo and Bi a probability measure. this gives a direct verification that A is an intensity matrix: the offdiagonal elements are nonnegative because Aij > 0.. qij = r.1) holds for the P.ic(0)e = ic(0)Oh e) h(e) .7) In particular.St)sl(e) h(Jt) 45 The proof that we have a MAP is contained in the proof of Theorem 5. Bi. . 0 < qij < 1 and Bij [0] > 0.10 is given by P = eK(e) Oh e) F[e]Oh('). 0<q<1. (5.. if vi(dx) is compound Poisson. 1 + q(b . MARKOV ADDITIVE PROCESSES Proof That { Lt} is a multiplicative functional follows from L8 ogt = h(Jt+s) es(St+ . vi (dx) = f3 Bi(dx) with . u Theorem 5. in the discrete time case. Ai = µi + 0Q.(0)j. then also vi (dx) is compound Poisson with e Ox ^i = /3iBi[0]. Bi [0] Remark 5. ^i = of qij Bij [0] 1 + qij ( Bij [0] . Bi(dx) = Bi(dx). and by A = Oh(°) K [0]Oh(e ) vi(dx) = e"xvi (dx).7(dx) Bij [0] Bij(dx) in the continuous time case . In particular. one can directly verify that (5.1) eft ea' f ij (dx) = Hij (dx) Hij [0] . Then the MAP in Proposition 5. We omit the details. That the rows sum to 1 follows from Ae = Oh(e) K[O]h(B) .tc(0)e = 0 .12 The expression for A means h(e) Aij = hie) Aij [1 + gij(Bij[0] i 0 j.11 below in the finite case.5. In the infinite case . That 0 < qij < 1 follows from the inequality qb <1.
..t.tc(0)' )Ah() = Oh(o) K[a + 0]Oh() . In matrix notation . Jt = j] = hie) .tc(0)I. Jl = j] :(Yi E dx. . (dx). Similarly.e) Consider first the discrete time case . v. Now we can write K[a] =A+A ) ( K[a + 0] . Hence the same is true for H=j and H. Letting a = 0 yields the stated expression for A. are probability measures .8) h(.tc (') (0) corresponds to the stated parameters µ.13) for matrixexponentials . .r. it follows that indeed the normalizing constant is H1 [0].8). Jt = A. is absolutely continuous w. First note that the ijth element of Ft[a] is etK(e)Ej [e(a+B)st E:[east Jt = j] = Ej[Lteas' .K [O])Oh(e) (0) l + ( A + (tc(') (a + 0) .. a = 0 in (5. Further Fib (dx ) = P=(YI E dx. v= . Jl = j) = Ei[Lt.11.tc(') (8)/ d)ag h 7 Aiiii (Bii[a + 0] . this means that Ft[a] = etw ( e)Ohc) Ft[a + 9]oh (e) (5. H1. F:j with a density proportional to eei .Bay [0]) That k(') (a + 0) . in continuous time ( 5. Yi E dx. (dx) of a process with stationary independent increments follows from Theorem 4. since Hij.46 CHAPTER II.8) yields et'[a] = Ohie )et (K[a +e]K(e)I)Oh(°) By a general formula (A. Here the stated formula for P follows immediately by letting t = 1. SOME GENERAL TOOLS AND RESULTS Proof of Theorem 5. this implies k[a] = A 1 ) (K[a + 0] .8. This shows that F. Ji = j) h(e) eeyK(B)p h(8) h(e) eexK ( h=e) e)Fi.
however. 7+ < oo). the literature on the continuous time case tends more to deal with special cases.1) = Aij4ij(Bij[a] .(u) = inf {t > 0 : St > u} to ruin in the particular case u = 0 . which. and is typically defective. h. Notes and references The earliest paper on treatment of MAP's in the present spirit we know of is Nagaev [265].+ < x. however. [226] and Miller [260]. < x) = 11 (S. 0]. [262] in discrete time. Conditions for analogues of Corollary 5.e. see also Fuh & Lai [149] and Moustakides [264].6. For the Wald identity in Corollary 5.Bij[0]) = hjel)ijgijBij[0](Bij[a] . Though the literature on MAP's is extensive. 6 The ladder height distribution We consider the claim surplus process {St } of a general risk process and the time 7. THE LADDER HEIGHT DISTRIBUTION 47 Finally note that by (5. is slightly less general than the present setting.6. has no mass on (oo.3 for an infinite E are given by Ney & Nummelin [266].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]. an extensive bibliography on aspects of the theory can be found in Asmussen [16]. [261]. 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).. Much of the pioneering was done in the sixties in papers like Keilson & Wishart [224].. [225]. hardly a single comprehensive treatment. there is. oo). . Note that G+ is concentrated on (0.7). i. Write r+ = T(0) and define the associated ladder height ST+ and ladder height distribution by G+(x) = 11 (S.)Ajjgij(Bij[a+0] .
Also. = ST+(1) Figure 6. 6. the sum of all the ladder steps (if rl > 0.d.1.i.. Recall that B(x) = 1 . has no mass on ( 0. Theorem 6 . a fact which turns out to be extremely useful. The first ladder step is precisely ST+. In other cases like the Markovian environment model.1. the ladder heights are i.1 The term ladder height is motivated from the shape of the process {Mt} of relative maxima.1) The interpretation of R+(A ) is as the expected time {St} spends in the set A before T+. it follows that for g > 0 measurable.48 CHAPTER K. see Fig. 1 For the compound Poisson model with p = 01LB < 1. i. SOME GENERAL TOOLS AND RESULTS M ST+(2) Sr. g(y)R+(dy) = E f g(St)dt. 6. To illustrate the ideas. by approximation with step functions . Thus. define the prer+occupation measure R+ by R+(A) = E f o "o I(St E A. 0 f T+ (6. 0].1. o 00 (6. In simple cases like the compound Poisson model.ST+(1) and so on.00 ). which gives an explicit expression for G+ in a very general setting.. i. the second ladder height (step) is ST+(2) . The main result of this section is Theorem 6. On Fig. Here bo(x) _ B(x)/µB. oo). G+ is given by the defective density g + (x) =.e.e. the dependence structure seems too complicated to be useful). there are only finitely many).5 below. and the maximum M is the total height of the ladder.B(x) denotes the tail of B.(3B(x ) = pbo(x) on (0. they have a semiMarkov structure (but in complete generality.2) . the second ladder point is ST+(2) where r+(2) is the time of the next relative maximum after r+(1) = r+. where basically only stationarity is assumed. In any case. at present we concentrate on the first ladder height.2. R+ is concentrated on (oo. we shall first consider the compound Poisson model in the notation of Example 1.T+ > t)dt = E f 0T+I(St E A) dt. For the proof of Theorem 6.
2 R+ is the restriction of Lebesgue measure to (00. St S* t a Figure 6. {St }o<t<T is constructed from {St}o<t<T by timereversion and hence.6. has the same distribution as {St}o<t<T.2. P(STEA.2(a): T+ > t Figure 6. 0 < t < T) P(STEA. 49 Proof Let T be fixed and define St = ST . That is.St<0. ST < ST_t.ST_t.2(b): r+ < t Thus.ST<St. since the distribution of the Poisson process is invariant under time reversion. see Fig. THE LADDER HEIGHT DISTRIBUTION Lemma 6 . 6. 0 < t < T.T+>T) = P(STEA.O<t<T). 0].O<t<T) = P(STEA.ST<St.0<t<T) = F(ST E A. .
U + St_ E A. T+ > t] 0 _ /3 f E[B( A . Fig.3 where the bold lines correspond to minimal values. E A} precisely when r+ > t.y) (here we used the fact that the probability of a jump at u t is zero in the second step. G+(A) = Q f 0 B(A . SOME GENERAL TOOLS AND RESULTS Integrating w. But since St 4 oo a.T+ > t] dt 0 T+ _ /3E f g( St) dt = 0 f g(y) R+(dy) 0 00 where g(y) = B(A .3 Lemma 6 ..3 G+ is the restriction of /3R+*B to (0.St).50 CHAPTER II. it follows that R+ (A) is the expected time when ST is in A and at a minimum at the same time . and since the jump rate is /3. and (6.t dT. 6. Figure 6.2) in the last). cf. for A C (0. oo).St _). oo). .y)R+(dy) 00 Proof A jump of {St} at time t and of size U contributes to the event IS.r..St _)I(r+ > t). s. this is just the Lebesgue measure of A. we get G+ (A) = f 00 /3 dt E[B(A . That is. The probability of this given { Su}u<t is B(A .
6. Uk) for those k for which ak .S8 )t> o = {St }t>o for all s > 0. h]} /h (by stationarity.M o 08 shifted by s is defined the obvious way.z)B(dz) _ f I(x < z)B(dz) _ f (x). cf. obviously. The points in the plane (marked by x on Fig.. Fig. THE LADDER HEIGHT DISTRIBUTION 51 Proof of Theorem 6. U k)} k=1 a is as a marked point process M *. assuming basically stationarity in time and space. as a point process on [0. The traditional representation of the input sequence {(TT. this is equivalent to the risk process {St*} being stationary in the sense of (6.3 yields g+ (x) = . 0 Generalizing the setup.s.. oo). {St+8 .1.. we define the arrival rate as E# { k : ak E [0 . 2. We call M * stationary if M* o B8 has the same distribution as M* for all s > 0. we consider the claim surplus process {St }t>o of a risk reserve process in a very general setup.:T1 +•••+Tk <t}. . 4 (the points in the plane are (ak .e.1 With r+(y) = I(y < 0) the density of R+. 6 . The first ladder epoch r is defined as inf It > 0 : St > 0} and the corresponding ladder height distribution is * G+ (A) = P(S** E A) = P(ST+ E A. The 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..Q f r+(x . i.T+ < oo). Uk) (k = 1.4) are (ak. . Nt St =>Uk k=1 t where Nt = max{k = O. In the stationary case. the first component representing time (the arrival time o. i. 6 .) where ak = Ti + • • • + Tk .4). The marked point process .e. oo) x (0..* ) and the second the mark (the claim size Uk ). this does not depend on h). Lemma 6.s > 0).
. h. h] Eco(M*) = 1 E f co(M o Bt)dt.5) represents the conditional distribution of M* given vi = 0. Oh becomes the approximate probability F(ri < h) of an arrival in [0.e..2. Note also that (again by stationarity) the Palm distribution also represents the conditional distribution of M* o Ot given an arrival at time t. where T is the first arrival time > 0 of M and h > 0 an arbitrary constant (in the literature. i. k: vk E [0.4 Given a stationary marked point process M*.s.4 Consider a finite Markov additive process (cf. and let T = T2 denote the first proper interarrival time . o. We represent M by the sequence (Tk. = 0 . Section 5) which has pure jump structure corresponding to pi = a = 0. i 1 U2 Us 1_ 0 or Q2 $ U3 *1 L 0 7 X I 11 1 Figure 6.QiBi(dx). h] and the sum approximately ^o(M*)I(ul < h).g.. Uk) k=1. The two fundamental formulas connecting M* and M are Eco(M) = aE E. See. As above . letting h J. 0. vi(dx) = . SOME GENERAL TOOLS AND RESULTS M* U. . the r. Assume {Jt} irreducible so that a stationary distribution 7r = (1i)iGE exists. most often one takes h = 1).. we define its Palm version M as a marked point process having the conditional distribution of M* given an arrival at time 0 . where TI = 0. e. V(M* o eak ).5) does not depend on h.52 CHAPTER II. Example 6 . Sigman [348] for these and further aspects of Palm theory. of (6. This more or less gives a proof that indeed (6.
6 Under the assumptions of Theorem 6. qij when {Jt} jumps from i to j and have mark distribution Bij. j) and let the initial mark Ul have distribution Bi when i = j and Bij otherwise.5.OF(x). Before giving the proof. THE LADDER HEIGHT DISTRIBUTION 53 Interpreting jump times as arrival times and jump sizes as marks.O for i = j and iriAijgij/.s. Jt = j is iri(3i /. we note: Corollary 6. the ruin probability .O fo "o F(x)dx = .6iBi + Aijgij Bij j#i iEE iEE 0 Theorem 6 . If Jt_ = i. Then the ladder height distribution G+ is given by the (defective) density g+(x) = . v. Note in particular that the Palm distribution of the mark size (i. 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 . let U0 be a r. an arrival for M* occurs before time t + dt w. and by some additional arrivals which occur w. and that p = 0EU0 < 1.p. This follows by noting that iP*(0) = IIG+JI = J0 "o g+(x)dx = . After that.p. Assume that St * ./.6. It follows that we can describe the Palm version M as follows . let the arrivals and their marks be generated by {Jt} starting from Jo = j. Jo) w. having the Palm distribution of the claim size and F (x) = F(Uo < x) its distribution .e.p.oo a. we get a marked point process generated by Poisson arrivals at rate /3i and mark distribution Bi when Jt = i.= i. . the probability aij of Jt . A stationary marked point process M* is obtained by assigning Jo distribution Tr. aij for (i.O for i # j.. 5 Consider a general stationary claim surplus process {St }t>o.*(0) with initial reserve u = 0 is p = /3EU0.OEU0. First choose (Jo_. the distribution of Ul) is the mixture B = E aii Bi + aij Bij J = j#i !i J.
. in (oo.St <.Su_ <0. We then represent M by the mark (claim size ) Uo of the arrival at time 0.. 0). Now conditionally upon At . and the kth preceding claim arrives at time t .1] here the r .0<u<tIAt) = P(St EA.. 6.).$St_ u. in (0. moves down linearly at a unit rate in between jumps and starts from S0 = U. that M* and M have doubly infinite time (i. 105) shows that one can assume w.0<u<tIAt) = P(St EA.0<u<t) = P(StEA.o.s. 0<u<t) = P(St EA. T+ = t given the event At that an arrival at t occurs .5.o. Proof of Theorem 6. (k = St}t>o 1.Su< 0. the mark at time Qk is denoted by Uk.l.5. { Su}0<u<t is distributed as a process {Su} . which makes an upwards jump at time . 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.. A standard argument for stationary processes ([78] p.54 By (6. . are point processes on (oo .(left limit) when 0 < it < t and is illustrated on Fig . oo) p(t) = P(St EA. oo) x (0 . The sample path relation between { Su } and { Su } amounts to S„ = St .. The last property is referred to as insensitivity in the applied probability literature.Q_k and has size U_ k..o<u<t where a claim arrives at time t and has size Uo. 2. has a very simple interpretation as the average amount of claims received per unit time ..St<Su.g. Then clearly * G+ (A) = P(ST+ E A) = Consider a process { f p(t)f3dt. the arrival times 0 < 0'1 < Q2 < .Su<0.Mt)..0<u<t) = P(St EA. oo)). h.e. It follows that for A C (0.A.St*_ u. SOME GENERAL TOOLS AND RESULTS V` (0) = E E Uk k: ak E [0. CHAPTER H.. oo ) and the arrival times 0 > 0_1 > a_2 > . Let p(t) be the conditional probability that ST+ E A.5). .
NIt)dt . Since So = U0. 2 therefore immediately shows that L(dy) is Lebesgue measure on (oo. the support of L has right endpoint U0. cf. Uo].. A sample path inspection just as in the proof of Lemma 6 . 0 < u < t } is the event that { Su } has a relative minimum at t . the left endpoint of the support is oo. t a oo. Mt)dt = i3EL(A) o"o . time instants corresponding to such minimal values have been marked with bold lines in the path of { St}. and we let L(dy) be the random measure L(A) = fo°° I(St E A. Fig.6. 6.5 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.5 where it = { St < Su. In Fig.s. THE LADDER HEIGHT DISTRIBUTION 55 { A Su}0<u<t U0 U0 \t tt u>0 N U_1 Figure 6. Thus.5. G' (A) = 3 f P(St E A. 6.
.6 is Bjork & Grandell [67]. A further relevant reference related to Corollary 6. [147]. SOME GENERAL TOOLS AND RESULTS = OE f 0 I(Uo>y)I (yEA)dy = Q f IP (Uo>y)dy A 0o a fA P(y) dy• 0 Notes and references Theorem 6.5 is due to Schmidt & coworkers [48].2. [263] (a special case of the result appear in Proposition VI.1).56 CHAPTER II.
are i. i=1 i=1 An important omission of the discussion in this chapter is the numerical evaluation of the ruin probability. see Chapter IV.e.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. being of the form Rt = Rt+Bt + Jt where {Rt } is a compound 57 . 3). say. and independent of {Nt}. St = uRt = EUi t. • the claim sizes U1. exact matrixexponential solutions under the assumption that B is phasetype (see further VIII. • the premium rate is p = 1. Thus . A common view of the literature is to consider such processes as perturbed compound Poisson risk processes . Some possibilities are numerical Laplace transform inversion via Corollary 3. i. i. 4. Panjer's recursion ( Corollary XI. {Rt} and the associated claims surplus process {St} are given by Nt Nt Rt = u+t EUi. and assume that • { Nt}t>o is a Poisson process with rate j3. U2. with common distribution B.4 below . 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 .6) and simulation methods ( Chapter X). For finite horizon ruin probabilities .. .d..
1) = t(p . {Bt} a Brownian motion and {Rt} a pure jump process. Write pB^) = EUn' YB = Pali = EU.s.g. Proof It was noted in Chapter I that p . Furrer [150]. See e. [324]. m.1 is the expected claim surplus per unit time.f. For (c). and Schlegel [316]. 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)..)3t (fit' k t} = etk(8) exp {st '3t + B[s]f Finally.+Uk)P(Nt = k) k=O e8t k=O B[s]k . P = PAB = 1/(1 + rl) Proposition 1. and that B(k)[0] = Pak). Dufresne & Gerber [126].t = t(p .1 (a) ESt = t(13µ$ .t = E E [ U k k=1 k=1 Nt . we shall start by giving the basic formulas for moments. 1 Introduction For later reference.6pBa). we get Ee8st = 00 e8t c` Ee8 (U1+. e . say stable Levy motion. where K(k) (0) is the kth derivative of is at 0..'s etc.t = fltpB .Rt.1) .g. 0 . A more formal proof goes as follows: Nt r Nt ESt = E > U k .u . for (d) just note that the kth cumulant of St is tic(k) (0). (d) The kth cumulant of St is tf3p(k) for k > 2. and this immediately yields (a).58 CHAPTER III. cumulants .1). THE COMPOUND POISSON MODEL Poisson risk process. Schmidli [319]. of the claim surplus St . (c) Ee8St = et" (8) where c(s) = f3(B[s] .1).t = E[Ntµs] . (b) Var St = t. We do not spell out in detail such generalizations.
then St> SnhV>Snhh. and the value is then precisely v..3EU01 = 1µs where rt is the safety loading. INTRODUCTION 59 The linear way the index t enters in the formulas in Proposition 1. Proof We first note that for u. we have Sok .Sok_l = Uk . 2. Obviously. then St 4 co. Sn+0 .h < St < S(n+1)h + h.d. we get a discrete time random walk imbedded in the claim surplus process {St}.2 (DRIFT AND OSCILLATION) St/ta3'p1 ast >oo. we need the following lemma: Lemma 1. and there are at least two ways to exploit this: Recalling that ok is the time of the kth claim. so that {Sok } is a random walk with mean EUET = EU. where Tk is the time between the kth and the (k . St = oo.3 If nh < t < (n + 1)h. if t = nh + v with 0 < v < h. (d) If 17 = 0.4. S„+V > S„ . rather to view {St} directly as a random walk in continuous time. In particular. then lien inft. (a) No matter the value of 77..Tk are i. The right hand inequality in (1.1 is the same as if {St} was a random walk indexed by t = 0. however.3) is proved similarly.V.. cf.1 = . We return to this approach in Chapter V. 1. v > 0.. lim supt. In this way.1)th claim. then St 00. For example. II.i. Here is one immediate application: Proposition 1. St = oo. The point of view in the present chapter is. u + v]. which is often used in the literature for obtaining information about {St} and the ruin probabilities. The connections to random walks are in fact fundamental. Indeed.Tk. then Snh . meaning that the increments are stationary and independent. the Uk . (b) If 77 < 0. (c) If 77 > 0.1.S„ attains its minimal value when there are no arrivals in (u. obviously 0(u) = F(maxk Sok > u). . For the proof.
{Snh}n=o.1). Thus using Lemma 1. There is also a central limit version of Proposition 1. is a discretetime random walk for any h > 0..4 The ruin probability 0(u) is 1 for all u when 77 < 0. If rl > 0. Corollary 1.1.. and < 1 for all u when 77 > 0..p. at least once. and hence it folz lows from standard central limit theory and the expression Var(St) = tf3pB (Proposition 1.1.1. This contradicts u St400. Assuming that each risk generates claims at Poisson intensity /3 and pays premium 1 per unit time. Snh u = 00 (the lemma is not needed for (d)). where the size of the portfolio at time t is M(t).o. lim supn_. {Snh}n=o.. . then {St} upcrosses level 0 a.2.. 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.5 The limiting distribution of St . (c) are immediate consequences of (a).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). However. 1 since St 4 oo) and repeating the argument. it suffices to prove 4'(0) = F(M > 0) < 1. and (b).1(b)) that the assertion holds as t 4 oo through values of the form t = 0.._.s.3. Considering the next downcrossing (which occurs w. 0 Snh = 00. it is seen that upcrossing occurs at least twice. h.. u 307). h A similar argument for lim sup proves (a). Notes and references All material of the present section is standard.3.60 CHAPTER III. THE COMPOUND POISSON MODEL Proof of Proposition 1. Snh/n a4' ESh = h(p .2. 2h.. hence by induction i.. Part (d) follows by a (slightly more intricate) general random walk result ([APQ]. 169) stating that lim infra. or by a general result on discrete skeletons ([APQ] p.2: Proposition 1.6 Often it is of interest to consider size fluctuations. is a discrete time random walk. The general case now follows either by another easy application of Lemma 1. p. 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 .t . Proof The case of 17 < 0 is immediate since then M = oo by Proposition 1. if P(M > 0) = 1. Remark 1 . For any fixed h..
Note that the distribution B0 with density bo is familiar from renewal theory as the limiting stationary distribution of the overshoot (forwards recurrence time ). we may view the ladder heights as a terminating renewal process and M becomes then the lifetime. 1 The distribution of M is (1. Summing over n. and we further get information about the joint conditional distribution of the surplus and the deficit. cf.e.1) representing the distribution of M as a geometric compound.1 provides a representation formula for 0(u).just before ruin is again B0.1) is not entirely satisfying because of the infinite sum of convolution powers. This follows simply by noting that the process repeats itself after reaching a relative maximum. It is crucial to note that for the compound Poisson model. equivalently. The decomposition of M as a sum of ladder heights now yields: 00 Theorem 2 . and we shall here exploit the decomposition of the maximum M as sum of ladder heights.1. Fig. The following results generalizes the fact that the conditional distribution of the deficit ST(o) just after ruin given that ruin occurs (i. As a vehicle for computing tIi(u).IIG+II) (the parenthesis gives the probability that there are no further ladder steps after the nth ). 1e. n=0 (2. where G+ is given n=0 by the defective density g+ (x) = 3B (x) = pbo(x) on (0..1.. (2. The expression for g+ was proved in Theorem 11.34 or A. Note that this . but we shall be able to extract substantial information from the formula.2. i. 0 Alternatively. 11. Theorem 2. the ladder heights are i. nevertheless.P) E PnBon(u) . Here bo(x) _ Proof The probability that M is attained in precisely n ladder steps and does not exceed x is G+ (x)(1 . which we henceforth refer to as the PollaczeckKhinchine formula.IIG +II)EG+ . d. Combined with i/i(u) = P ( M > u). [APQ] Ch. IV. oo ). B(x)/aB.6. we can rewrite the PollaczeckKhinchine formula as 00 (u) = P (M > u) = (1 . the formula for the distribution of M follows . We assume throughout rl > 0 or.6. THE POLLACZECKKHINCHINE FORMULA 61 2 The PollaczeckKhinchine formula The time to ruin r(u) is defined as in Chapter I as inf It > 0: St > u}. p < 1. that r(0) < oo) is Bo: taking y = 0 shows that the conditional distribution of (minus) the surplus ST(o). Thus .
cf. see for example [APQ]. For the study of the joint distribution of the surplus ST(u)_ just before ruin and the deficit ST(„). 7r(0 ) < oo) = Q 3 Special cases of the PollaczeckKhinchine formula The model and notation is the same as in the preceding sections. (c) the marginal distribution of ST(o)_ is Bo . W are independent.V)W) where V. the PollaczeckKhinchine formula is often referred to as Beekman 's convolution formula. [62]. ST(o) > y. ladder heights so that the results do not appear not too useful for estimating 0(u) for u>0. V is uniform on (0.just after ruin.2 and it gives an alternative derivation of the distribution of the deficit ST(o) Notes and references The PollaczeckKhinchine formula is standard in queueing theory. Theorem 2. In the risk theory literature. Theorem A1. . cf. (a) 11 (ST(o)_ > x. 2 The joint distribution of (ST(o )_. 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.i.2(a) is from Dufresne & Gerber [125]. Theorem 2 . ST(o )) given r (0) < oo is the same as the distribution of (VW. f +b (b) the joint distribution of (ST( o).1 is traditionally carried out for the imbedded discrete time random walk. Feller [143] or Wolff [384]. 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).6. (d) the marginal distribution of ST(o)_ is B0. However. The proof of Theorem 11.d. Asmussen & Schmidt [49]. where it requires slightly more calculation. cf.5. the form of G+ is surprisingly insensitive to the form of {St} and holds in a certain general marked point process setup.5. (1 . As shown in Theorem 11. We assume rt > 0 throughout. in this setting there is no decomposition of M as a sum of i. and the conditional distribution of ST(o)_ given ST(o)_ = z is Bo z) The proof is given in IV.6.62 CHAPTER III. 1) and W has distribution Fw given by dFyy/ dB(x) = x/µB. see Schmidli [323] and references there. Again. Beekman [61]. ST(o)) is given by the following four equivalent statements: B(z) dz. there is a general marked point process version.
3.3 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. use Laplace transforms.e. B0 is exponential with rate S and the result can now be proved from the Pollaczeck Khinchine formula by elementary calculations .2 If B is exponential with rate S. SPECIAL CASES OF POLLACZECKKHINCHINE 3a The ruin probability when the initial reserve is zero 63 The case u = 0 is remarkable by giving a formula for V)(u) which depends only on the claim size distribution through its mean: Corollary 3. the result follows . I. 3b Exponential claims Corollary 3.p. and hence this overshoot has the same distribution as the claims themselves .p. The result can. For a failure at x. 1 .p) E pn S n x n. also be seen probabilistically without summing infinite series . 0 .p)pSe a ( l v)x = p( S . Let r ( x) be the failure rate of M at x > 0. Alternatively. then. Thus r(x) = S(1 . however .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.O)e(b0)x.1 e ax = n1 (n . the formula for P(O) holds in a more general setting..1)1 00 ( 1 . Integrating from u to oo. Thus .p) = S . But claims are exponential . Bon is the Erlang distribution with n phases and thus the density of M at x > 0 is (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. a further relevant reference is Bjork & Grandell [67]. the current ladder step must terminate which occurs at rate S and there must be no further ones which occurs w.6. hence without memory. As shown in 11.
y)f3 (y) dy. II. Corollary 3.p + G+ * Z(u) = 1 . (3.3.p + f u Z(u .y)/3B (y) dy. (3. the survival probability Z(u) = 1 .4) zu P(M > u . 3c Some classical analytical results Recall the notation G+(u) = f^°° G+(dx).3) Equivalently. we show that expression for /'(u) which are explicit (up to matrix exponentials) come out in a similar way also when B is phasetype.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 . and weights 1/2 for each.+ = y yields P(M>u.+ >u.s.4) can be derived by elementary algebra from (3.S. we use the PollaczeckKhinchine formula in Chapter IX to show that b(u) . Then the first term on the r. 0 Proof Write o (u) as P(M>u) = P(S. (b) use stopped martingales .1) For a heavytailed B.i(u) satisfies the defective renewal equation Z(u) = 1 . (3.3).1. (Example VIII.y)G+(dy ) = f U V(u .h. then 24 1 V.+ <u.3.T+ <oo). We mention in particular the following: (a) check that ip (u) = pe (60)u is solution of the renewal equation (3. E. cf.y)G+(dy) For the last identity in (3. THE COMPOUND POISSON MODEL In VIII. just insert the explicit form of G+.3) below. T+ <00) (3.64 CHAPTER III. The case of (3.S. u + oo. (3.+ <U. A variety of proofs are available . is ?7+ ( u).2) Notes and references Corollary 3. if 3 = 3 and B is a mixture of two exponential distributions with rates 3 and 7. and conditioning upon S.4) is similar (equivalently.g.1 p pBo(u). (u) 35eu + 35e6u. u .3)).2). 2 is one of the main classical early results in the area.T+ <oo)+P(M> u.
3.pBo[s] no (1 . Griibel [179] and Thorin & Wikstad [370] (see also the Bibliographical Notes in [307] p.7) s +.P)PB 2(1 . either of these sets of formulas are what many authors call the PollaczeckKhinchine formula.p)s s /3 . SPECIAL CASES OF POLLACZECKKHINCHINE Corollary 3.6) 00 = (I . g. Bo of B0 as m e8u B(du) = B[s] . We omit the details (see. which yields the survival probability as 00 f u }t Z(u) = f f3eRtdt 0 from which (3.Ps s(.PPB2) EM2 = PPB) + QZPBl 2(1 ./3B[s] which is the same as (3. Griibel & Pitts [132].7). eau B(u) du = f PB 3PB SPB 0 o (3. The approach there is to condition upon the first claim occuring at time t and having size x .5). it is not surprising that such arguments are more cumbersome since the ladder height representation is not used. 191). [APQ] pp.Ee8M) f ao e8' ( u)du = a8uP (M > u)du = 0 o 1 ( 1+ (1 . 0 Notes and references Corollary 3. In view of (3.4 The Laplace transform of the ruin probability is 65 fo Hence Ee8M 00 e8uiP(u)du .3 .3 is standard .5 The first two moments of M are 2 EM .p = (1 .3 . Also (3. numerical inversion of the Laplace transform is one of the classical approaches for computing ruin probabilities.g.p) E p"Bo[s]" = 1 . Corollary 3.7) and Corollary 3.s . . [APQ] pp.. Embrechts.p)s .s .1 Bo[s] = f oc..(3B[s] 1 .g. Some relevant references are Abate & Whitt [2]. e.5). (3. Of course. for example.f.8) Proof This can be shown.4) can be derived by elementary but tedious manipulations. by analytical manipulations (L'Hospital's rule) from (3.P)pB' (3.)3B[s]) (3. see e.5) Proof We first find the m./3B[s] . 111112 or Feller [143].p)2 3(1 . 206207).(3 . In fact.5 can be found in virtually any queueing book.
9).u + 1 )]k = QZ(u) .y<1)dy 0<u<1 1 < u < oo uu ulhu 1a+/3 J0 uZ(y)dy U Z(y) dy 113+0 For 0 < u < 1./3Z(u .9) shown for n .u)]k1 ku+1) [/3( k .3+ 18+ J0 Z(uy). THE COMPOUND POISSON MODEL 3d Deterministic claims Corollary 3.)3(1 .1)! k=1 u1 .Q) 3e.Q (k 1 k= n  [O(k .4) for Z( u) means f lhu Z(u) = 1.u) [N(k .1).1).u)]k d 1 u) _ a) n ( du ( k! (1  .u) [p(k .1 < u < n and let Z(u ) denote the r./32(u . .6 If B is degenerate at p.Q) k=0 k! E e0( = /32(u) . then p) 1: ep(k u/.66 CHAPTER III. of (3.s.u)]k k! k0 The renewal equation (3.3I( 0<y<1)dy Z(y)/3I(0<u.9) follows for 0 < u < 1.h.z/'(u) takes the form Z(u) L^J L. Z^ =eR(k. we may assume p = 1 so that the stated formula in terms of the survival probability Z(u) = 1 .u) a)Qea" + (1 . eO('u) [)3(k . Assume (3.u)]k k! (1 L3) 1: e_O(ku) NIN (k (k . For n < u < n + 1.u/p)]k ko k! Proof By replacing {St} by {Stu/p} if necessary. 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. differentiation yields Z(u) _ /3Z(u) .
and define rce by (4. (4. but will now be repeated for the sake of selfcontainedness.1) or equivalently. co(a) = rc(a + 9) .2).g. 0 Notes and references Corollary 3.1) . we just have to multiply (4.rc(9) = .3B[9].3) by t. See also Iversen & Staalhagen [208] for a discussion of computational aspects and further references.g.r.f. 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) . or equivalently BB[a] = B[^+ Repeating for t 54 1.a.(9).) The adaptation of this construction to stochastic processes with stationary independent increment as {St} has been carried out in 11. Formalizing this for the purpose of studying the whole process {St}. it follows that Z(u) = 2(u) for n<u<n+1. 4 Change of measure via exponential families If X is a random variable with c.4) . The answer is yes: inserting in (4.4. (4.1) .6 is identical to the formula for the M/D/1 waiting time distribution derived by Erlang [139].f. of F9.1 that c(a) = /3(B[a] . The question then naturally arises whether ie is the c.4) works as well.4.(9) is welldefined. We could first tentatively consider the claim surplus X = St for a single t.3B = . in terms of the c. say t = 1: recall from Proposition 1. we set up .a.f. F and c.2) shows that the solution is Ox [O]0]. K(a) = logEe'X = 109f 00 eaxF(dx) = logF[a].d.g. 00 the standard definition of the exponential family {F9} generated by F is FB(dx) = e°xK(e)F(dx).2) (Here 9 is any such number such that r. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 67 Since Z(n) = 2(n) by the induction hypothesis.f.Qe(Bo[a] . B9(dx) = B[9] B(dx). and thus (4. (4.
Then P(G) = Fo(G) = EB [exp {BST + TK(O)} . EeeBSt + tk(B) = 1.8) By standard measure theory. (4. t < T. the corresponding expectation operator is E9. = exp {BST . it suffices to consider the case where Z is measurable w. for G E FT. and thus (4. Z is measurable w.5) for the density. the PBT) are mutually equivalent on. oo) governing a given compound Poisson risk process with arrival intensity. Then FB denotes the probability measure governing the compound Poisson risk process with arrival intensity. But let Xk = SkT/n . Then the Xk are i. G C {T < oo}.f.t.5) for the density of n i.2. .2 For any fixed T. (4. and define 09. The following result (Proposition 4...7) now follows by taking Z = eBST+TK(e)I(G) u Theorem 4 . replications from Fe (replace x by xi in (4. G].S(k_1)Tln.0e and claim size distribution Be.10) .9) Proof We first note that for any fixed t. .3 Let T be any stopping time and let G E FT. and dP(T) dP^T) That is.4).FT. with common c. . THE COMPOUND POISSON MODEL Definition 4.t.r. n) for a given n.i..FTn) = Q(SkTIn : k = 0. and PBT) the restriction of PB to FT.1) and multiply from 1 to n). v(Xi. Let FT = o(St : t < T) denote the o•algebra spanned by the St. .g. G]. The identity (4.1 Let P be the probability measure on D[0. BB by (4. Ti(a)/n.1.Tic (0)} . . Xn). .nr.d. (4.r.d.6) F(G) = Po (G) = EB [exp {BST + Ttc(0)} .7) Proof We must prove that if Z is FTmeasurable.3 and claim size distribution B. in particular the expression (4.68 CHAPTER III.(9)} (4.i.8) follows by discrete exponential family theory. (4. then EBZ = E [Ze9ST _T"(9)I . (4. with T taking the role of n) is the analogue of the expression exp{8(x1 + • • • + xn) . Proposition 4.
5 Lundberg conjugation Being a c.1) _ 1 + a.g. according to what has just been proved. GT C_ Jr < T}. Then G E FT.r is deterministic. t = T . .ST) + (T .9) holds for G as well. (0) + rc'(0)a = 0 + ES1 a = a (p .10).1) . Then GT = G n Jr < T} satisfies GT E FT.7 1 Some discussion further supporting this statement is given in the next section.7) holds.5.9) holds with G replaced by GT.. subject to the basic assumption ij > 0 of a positive safety loading. 5.r)rc(9)}I . the typical shape of rc is as in Fig.1 It is seen that typically) a ry > 0 satisfying 0 = r.1(a).f. and hence (4.FT]] = EB [exp { BST + Trc(9)} I(G)] . c(a) is a convex function of a. Given FT. 77 Thus. so that PG = EeE0 [exp { 9ST+Trc(9)}I(G)I FT)] = Ee [exp { BST + rrrc(O)} I(G)EB [ exp {9 (ST . The behaviour at zero is given by the first order Taylor expansion c(a) r.(Y) = 13(B['Y] . Now consider a general G. (a) rc (a) (b) KL(a) 'Y 'Y Figure 5. Thus by (4. Ee [exp { BST +Trc(9)} I(G) FT)] = 1. Thus. Letting T t oo and using monotone convergence then shows u that (4. (4. LUNDBERG CONJUGATION 69 Now assume first that G C Jr < T} for some deterministic T.
1 Consider the case of exponential claims. b[s] = 5/(b . (5. an equivalent version illustrated in Fig.1) (or (5. 5. 5. u It is a crucial fact that when governed by FL. (5.3) cf. Equation (5.1) 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.4) yields /3L = b and that BL is again exponential with rate bL =. THE COMPOUND POISSON MODEL exists .1) is known as the Lundberg equation and plays a fundamental role in risk theory . Thus B[7] = 6/.3. the claim surplus process has positive drift > 0. and (4.1) . Thus.2)) is 7 = 5/3. e.g.2) 7 Figure 5.a = i(a + 7). .2 s As support for memory.s). It is then readily seen that the nonzero solution of (5.3. G = {T(u) < oo} in Theorem 4.3.1(b). Fig.1(b). Taking T = r(u). 5. (5.4) ELS1 = #L(0) cf.2 is B(7) = 1 + ^. we further note that ( 5. we write FL instead of F7. Example 5 .70 CHAPTER III. Note that KL (a) = /L (BL [a] . Fig.QL instead of /37 and so on in the following . Lundberg conjugation corresponds to interchanging the rates of the interarrival times and the claim sizes. the Lundberg exponent.
LUNDBERG CONJUGATION 71 to vanish so that Theorem 4. which shows that G(L) (dx) = e7xG +(dx) = e7x /3 (x) dx.u be the overshoot and noting that PL(T(u) < oo) = 1 by (5. T(u) < oo] . take first 0 = ry. V)(u) < e7u.3 takes a particular simple form.r. e(u) has a limit i.5) Theorem 5 .1e.G+ L) (x) G+L) (x) IL(+) µ+L) L) where G+L) is the FL. Proof Just note that e(u) > 0 in (5.P Y j o' xeryxOB (x) dx /3k [Y] .(u)} .+ E A} in Theorem 4.3 (THE CRAMERLUNDBERG APPROXIMATION) i'(u) .1.8) .7) is the same as (5. 0 Theorem 5 . Then P(ST+ E A) = EL [exp { 7S?+} .2 (LUNDBERG'S INEQUALITY) For all u > 0.4).ascending ladder height distribution and µ+ its mean.1p . (5. (5.e7x)G+(dx).3. we therefore have ELe7t(u) + C where C ELe7 (00) = µ+) f e7(1 . (5. Letting e(u) = ST(u) . where C .1 (5. see A .G+L)(x)) dx ry^+L) J 00 f 0 (1 . T = T+.(oo) (in the sense of weak convergence w. PL ) with density 1 .6) Proof By renewal theory. Since a7' is continuous and bounded. V) (u) = P(T(u) < oo) = EL [exp {ryS. G = {S. we can rewrite this as 0(u) = e"ELe7^(u).6 ).Ce7u as u 4 oo.t. ST+ E A] . To this end.5.5).7) 0 and all that is needed to check is that ( 5.
11) so that I 7B ['Y](B[7]1) BI [7]Q VP (7) 72 7 (using (5.72 CHAPTER III.4 Consider first the exponential case b(x) = Seax.8) yields +L) J0 xel'B ( x) dx (5.1) (5.3 (this was found already in Example 5. Using (5.1 = ^7 The accuracy of Lundberg's inequality in the exponential case thus depends on how close p is to one.10) VW = JI c* e° (x) dx = a (B[a] . that 7 = S . . THE COMPOUND POISSON MODEL In principle. From this it follows.")G + (dx ) = 1  J0 00 3B(x) dx = 1p. of course.1 above) and that C = p. we get L where 00 (1 . 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. u .12) Example 5 . this solves the problem of evaluating (5. Noting that SIG(L)II = 1 because of (5.4).1)) and 7µ+L) = 'y/3 [7] 7 1/0 = /3B ['y] .7).e. or equivalently of how close the safety loading 77 is to zero. (5. Then 0(u) = pe(a_Q)u where p = /3/S. but some tedious (though elementary) calculations remain to bring the expressions on a final form.1 .
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. Though of course it is based upon purely empirical grounds.8µBo  Ss' u where 6 = µB/µBo = 2µa/µB 2) . Mathematically.2) was suggested by De Vylder [109]. Notes and references The approximation (7.1.)3 )PBo µB  . Grandell [171] pp. and hence the ruin probability approximation is b(u) e(bAln)u. or equivalently that /3 is only slightly smaller than /3max = 1/µ8. 1924. (/3max .g. 7c The heavy traffic approximation The term heavy traffic comes from queueing theory.(bA*)". the premiums exceed only slightly the expected claims.p = (/3max 0)µB. That is.s(/3max . THE COMPOUND POISSON MODEL These three equations have solutions 9/3µB2)3 30µa2)2 3µa2) (3) P+ (3) ' 0 . Proposition 1.3 )1 } _ 1p 1 . but has an obvious interpretation also in risk theory: on the average.Ps(/3max . cf. Proposition 7. we shall represent this situation with a limit where /3 T fl but B is fixed.p . p* _ . [174]) shows that it may produce surprisingly good results./3)M converges in distribution to the 2a exponential distribution with rate S = B' Proof Note first that 1 .P .b(u) = p*e./3)] 1 ./3)PBo PB . heavy traffic conditions mean that the safety loading q is positive but small.7) that Ee$(Amex /j)M _ 1p _ 1p Eo [s (0max 1 .1.3* /S. we have according to the PollaczeckKhinchine formula in the form (3. the approximating risk process has ruin probability z.80 CHAPTER III.p + p { 1 1p ti 1 .1 As /3 f Nmax. Letting Bo be the stationary excess life distribution.(3)2 P PB 2µB 2µB Letting /3* = /3/P.PBo [s (/3max .3 and Corollary 3.
VIII).1 1 .2. obviously Corollary 7. These results suggest the approximation Vi(u) e6(0_.p _ 2rl11B PB p.4) suggested by the Cramer Lundberg approximation and Proposition 6. the term light traffic comes from queueing theory.B AB ) 6()3max _'3) = However . 2 provides the better mathematical foundation. It is worth noting that this is essentially the same as the approximation (2) z/i(u) Ce.2 If .g. This follows since rl = 1/p . Numerical evidence shows that the fit of (7.0)u. and hence 2µ2B 1 . or equivalently that 0 is small compared to µB . Notes and references Heavy traffic limit theory for queues goes back to Kingman [230].ze a2unµB laB (7. u * oo in such a way that (3max . the premiums are much larger than the expected claims . then P(u) 4 e6„ Proof Write z'(u) as P((/3max . 7d The light traffic approximation As for heavy traffic . as well as it is needed for the interpolation approximation to be studied in the next subsection. The present situation of Poisson arrivals is somewhat more elementary to deal with than the renewal case (see e .3) is reasonable for g being say 1020% and u being small or moderate./3)u * v.l3)M > (/3max .7./3)u). the first results of heavy traffic type seem to be due to Hadwiger [184]. [APQ] Ch. while the approximation may be far off for large u. However . . but has an obvious interpretation also in risk theory: on the average .ryu . we shall represent this situation with a limit where 3 10 but B is fixed.p. in risk theory heavy traffic is most often argued to be the typical case rather than light traffic . That is . Of course. Mathematically. VARIOUS APPROXIMATIONS FOR THE RUIN PROBABILITY 81 Corollary 7. light traffic is of some interest as a complement to heavy traffic . light traffic conditions mean that the safety loading rl is positive and large . We return to heavy traffic from a different point of view (diffusion approximations) in Chapter IV and give further references there . In the setting of risk theory.Q T /3max.
En'=2 • • • = O(/32) so that only the first terms matters. i. Again. the Poisson case is much easier than the renewal case.= 1 aJ 1 a 0+ 1 = = p. Light traffic does not appear to have been studied in risk theory. U > u] = /3iE(U . ao n=1 00 n=1 (u) P) anllBBon(U) onPaBon(u) • Asymptotically. 0(u) /3 J B(x)dx = /3E[U .3 As .e.5) u Proof According to the PollaczeckKhinchine formula. cf. Asmussen [19] and references there.u)+. u Note that heuristically the light traffic approximation in Proposition 7. [97].u. 10 ( u The alternative expressions in (7.T > u). by monotone time T of the first claim .3 is the same which comes out by saying that basically ruin can only occur at the F(U . .Q limIP ( u) + Q lim z/'(u) Amax &0 amax ATAm. and hence 00 (U) /3pBBo (u) = 0 / B(x)dx. 0 u Notes and references Light traffic limit theory for queues was initiated by Bloomfield & Cox [69]. The crude idea of interpolating between light and heavy traffic leads to 0 (u) C1 . 7e Interpolating between light and heavy traffic We shall now outline an idea of how the heavy and light traffic approximations can be combined. For a more comprehensive treatment. Indeed. (7.5) follow by integration by parts. ( 3 J O B dx. Omax max m.(3 10. Sigman [347].82 CHAPTER III.T > u) = J o" B(x + u)/3eax dx . 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 . see Daley & Rolski [96]. THE COMPOUND POISSON MODEL Proposition 7. z/' (u) convergence P(U .
Instead./3)) . B(2).3 and use similar notation for %(B) (u) = (u). where further references can be found .6) is . ^IE) exist: 1 (B) HT QmsxQ hm J e e6" 2µE/µE2)'" = e(1 6)" =  Q1Qm. 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. (U). we see that the following limits HT) (u').3).3n. Substituting v = u(. . The adaptation to risk theory is new. ^ LT Q maxQ m"^ Qlo V LT) ( CHT(v) (say). that is. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS 83 which is clearly useless .x .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. 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. Notes and references In the queueing setting .Wmax f(x ) dx + pee6mQ.VHT) ( ax QmQ ) h (B) ( . Al . "/Qmex Cu) CLT(u ( /3max 0) + O16 CHT( U(Qmaz . . 8 Comparing the risks of different claim size distributions Given two claim size distributions B(1). the idea of interpolating between light and heavy traffic is due to Burman & Smith [83 ]. to get nondegenerate limits . with rate 1/µB = /3max. (7. even if the safety loading is not very small. no empirical study of the fit of (7. available. ) M. z/i(E) (u) = pe(QmaxQ)u.O0 M. [84]. Another main queueing paper is Whitt [380]. f / Qmax B(x)dx 00 eQmaxxdx 4/ Qmax 00 QmaxQ amaze" and the approximation we suggest is J B(x) dx = cLT(v) (say). however.O(E)(u) 1 (1 . Thus . Let OLT) (u) denote the light traffic approximation given by Proposition 7. one may hope that some correction of the heavy traffic approximation has been obtained.8. _(E) (u).
most often the term stoploss ordering is used instead of increasing convex ordering because for a given distribution B. B(2)) if f fdB(1) < f fdB(2) for any convex function f. or the existence of random variables U(l). U(2) such that U(l) has distribution B('). 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.1 is quite weak. whereas (consider x2) B(2) has the larger variance. .84 CHAPTER III. B(2) and PB(1) = µB(2). we can assume that 1) < St 2l for all t. In terms of the time to ruin. A weaker concept is increasing convex ordering: B(1) is said to be smaller than B(2) (in symbols. Proposition 8. B(' <. an equivalent characterization is f f dB(') < f f dB (2) for any nondecreasing convex function f. XI. this ordering measures difference in variability.2 If B(') <j. and a particular deficit is that we cannot compare the risks of claim size distributions with the same mean: if BM <d B(2) and µB«) = /IB(2). we have the convex ordering. cf.' 1)(u) < V)(2) (U) for all u. u Of course. Rather than measuring difference in size.6. B(') <d B(2)) if B(1)(x) < B(2)(x) for all x.ill(u) < V)(2) (U) for all u. B(') <i. for more detail and background on which we refer to Stoyan [352] or Shaked & Shantikumar [337].1 If B(') <d B(2). Bill is said to be convexly smaller than B(2) (in symbols. B(2)) in the increasing convex order if f BM (y) dy < f 00 Bi2i (y) dy x x for all x. Taking probabilities. In the literature on risk theory. equivalent characterizations are f f dB(') < f f dB (2) for any nondecreasing function f. Proof According to the above characterization of stochastical ordering. U(2) distribution B(2) and U(1) < U(2) a. THE COMPOUND POISSON MODEL To this end. 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 . then Bill = B(2). this implies St T(l)(u) > r(2)(u) for all u so that 17(I) (U) < oo} C_ {T(2)(u) < oo}. Recall that B(') is said to be stochastically smaller than B(2) (in symbols.s. Here convex ordering is useful: Proposition 8. then i. the proof is complete. Finally. Proposition 8.
3 If B(1) <. Hence by the PollaczeckKhinchine formula . A general picture that emerges from these results and numerical studies like in Example 8. Proof Consider the light traffic approximation in Proposition 7. and consider the following claim size distributions: B1: the standard exponential distribution with density ay. u We finally give a numerical example illustrating how differences in the claim size distribution B may lead to very different ruin probabilities even if we fix the mean p = PB. B(2). Then V. then /'(')(u) < 0(2)(u) for all u. then B(1) <.5 If '0(1)(u) < p(2) (U) for all u and a. larger variance is paramount to larger second moment.p ) E /3"µ"Bo2)* n(u) _ V(2) (u) n=1 = Corollary 8. say to p.e. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS Proof Since the means are equal. The heavy traffic approximation (7. Proof If f is convex. with fixed mean.4) certainly supports this view: noting that. The problem is to specify what 'variation' means. and here is one more result of the same flavor: Corollary 8. (D) (u) < O(B) (U ) for all u.1 and µB at 1 so that the safety loading 11 is 10%.(1) (. Example 8. A partial converse to Proposition 8. from which the result immediately follows.. This u implies that D <.3 provides another instance of this.2 is the following: Proposition 8.. Corollary 8. B.u) = (1 _ P) E /3npnBo( 1):n(u) n=1 00 < (1. we have Bol) (x) f ' B(1) (y) dy < ' f' B(2) (y) dy = Bo2) (x)• µ 85 I. . A first attempt would of course be to identify 'variation' with variance. B(2).6 Fix /3 at 1/1.4 Let D refer to the distribution degenerate at 'LB . it is seen that asymptotically in heavy traffic larger claim size variance leads to larger ruin probabilities.8. we have by Jensen 's inequality that E f (U) > f ( EU).6 below is that (in a rough formulation) increased variation in B increases the risk (assuming that we fix the mean).1. Bo1) <_d Bo2) which implies the same order relation for all convolution powers.
One then obtains the following table: U005 U0. = 0. Note to make the figures comparable. Pellerey [287] and (for the convex ordering) Makowski [ 252].4142. We return to ordering of ruin probabilities in a special problem in VI. van Heerwarden [189]. A2 = 3. B3: the Erlang distribution with density 4xe2x.lA. However. the behaviour of which is governed by a parameter 9. A standard example from queueing theory is . B. B3 the comparison is as expected from the intutition concerning the variability of these distributions. and consider a = 5%.e.0' U0.01%. in comparison to B2 the effect on the ua does not show before a = 0..9A2e'2r where A. THE COMPOUND POISSON MODEL B2: the hyperexponential distribution with density 0. Kluppelberg [234]. B4: the Pareto distribution with density 3/(1 + 2x)5/2. 0.) = a. 9 Sensitivity estimates In a broad setting.1358. 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).86 CHAPTER III. Let ua denote the a fractile of the ruin function. 1/)(u. In terms of variances o2. all distributions have mean 1. 1%.1%.4. we have 0r3 = 2 < or2 = 1 < 02 = 10 < 04 = 00 so that in this sense B4 is the most variable. i.001 u0. For B1i B2.e'\1x + 0. which appears to be smaller than the range of interest in insurance risk (certainly not in queueing applications!). 11 Notes and references Further relevant references are Goovaerts et al. [166]. with the hyperexponential distribution being more variable than the exponential distribution and the Erlang distribution less.01%. and this is presumably a consequence of a heavier tail rather than larger variance.000. sensitivity analysis (or pertubation analysis) deals with the calculation of the derivative (the gradient in higher dimensions) of a performance measure s(O) of a stochastic or deterministic system. 0.
and s(9) the expected sojourn time of a customer in the network. if = a e(6A)u.1 Consider the case of claims which are exponential with rate 8 (the premium rate is one). Then a p ao = 00 Qa/. it follows that ' is approximatively normal N(0. say estimated from data. In particular . t].9. and hence a _ e(60)u + u e(60)u = ( i + which is of the order of magnitude uV. where Q2 = fl ( l2 1113 / _ Ou2v)2. a/3 0 . Then the arrival rate /3(P) for { R(P) } is )31p.3. Then ib = Pe(613)u. In the present setting. u Proposition 9. while /3 = j3 is an estimate. Proof This is an easy time transformation argument in a similar way as in Proposition 1. a0 as ao 80 19P . the standard deviation on the normalized estimate ^/1' (the relative error ) is approximatively . Example 9.. or we may be interested in aV)/0/3 as a measure of the uncertainty on '0 if 0 is only approximatively known. i.(u) for large u. Thus. obtained say in the natural way as the empirical arrival rate Nt/t in [0. For example. a2/t). with 0 the vector of service rates at different nodes and routing probabilities. we may be interested in a'/ap for assesing the effects of a small change in the premium. Let R(P) = Rtli. SENSITIVITY ESTIMATES 87 a queueing network. s(9) is of course the ruin probability t' = Vi(u) (with u fixed) and 0 a set of parameters determining the arrival rate 0. the distribution of %3 0 is approximatively normal N(0„ Q/t). Similar conclusions will be found below. Assume for example that 8 is known.19P a/ .1.01/2u.e. Then if t is large .2 Consider a risk process { Rt} with a general premium rate p.Ap). Thus at p = 1. the premium rate p and the claim size distribution B. where the partial derivatives are evaluated at p = 1. increasing in u. and hence the effect of changing p from 1 to 1 + Ap corresponds to changing /3 to /3/(1 + Op) /3(1 .
4) (9 . (9. it suffices to fix the premium at p = 1 and consider only the effects of changing . THE COMPOUND POISSON MODEL As a consequence.()YC = 1 +y/ /3 \ Q2 From this (9.1 or Proposition 9. ^) .5) are similar. (9. In the case of the claim size distribution B. mathematically a proof is needed basically to show that two limits (u * oo and the differentiation as limit of finite differences) are interchangeable. but we shall concentrate on a special structure covering a number of important cases. (3+'y)PC (0+7. Differentiating w.3 70 = 'Ye = = 7 /3(1 we(e +'y. Consider first the adjustment coefficient y as function of 3. for the ruin probabilities .^)] 1(/3+y)we(9+'y.10) below. However . so that heuristically we obtain '00 50ryu = Coe"u . Of course. and the proofs of (9.w(O.3. 9. 3) ( 9 . () Proof According to (9.6 below for some discussion of this assumption).88 CHAPTER III.()(0 +'0) ' (9 . u Now consider the ruin probability 0 = 0 (u) itself.((dx ) = exp {Ox + (t(x) . but must look for approximations for the sensitivities 0. we can rewrite the Lundberg equation as w(9+ y. e.3 or/and B.w(6. /3 yields w e(e + Y. various parametric families of claim size distributions could be considered.6) As will be seen below. ()} p(dx) . (.4).2) (see Remark 9. we cannot expect in general to find explicit expressions like in Example 9. x > 0 (9.uypCe7u urypO. and write yp = 8y/8/3 and so on .r. Similar notation for partial derivatived are used below. 5) (Q+'Y)[we(0+7.3) follows by straightforward algebra.g. Consider first the case of 8/8/3: . Viei '0(.()^ 1 .()wC(e.Owe (9. () = log(1 + y//3). Proposition 9.(/3 + y)we(9 + 7.t. namely that of a twoparameter exponential family of the form Bo. The most intuitive approach is to rely on the accuracy of the CramerLundberg approximation . this intuition is indeed correct.0 = t/'(u) and the CramerLundberg constant C.3.
BarndorffNielsen [58]).w(9. u 0 Then Z = z + F * Z and F is a proper probability distribution .w(O.x)B(x) dx + J U W(u .4 As u oo. ()} . z2(u) _ 1 ^ e'ri`i7i( u .2 of the Appendix ). w((9 + a.g.10) (9.we(9 . PF = (1 .x) F(dx ) f u J C F(dx) = C as u 4 oo. Z= zl + z2 where zl (u) = e7u J m B(x)dx.11) Ee. () . Further write de = [we (9 +'y.8) Letting cp = e0/e/3 and differentiating (9. it holds that 89 a ue ryu a/3 Q(1 P) 7C2 Proof We shall use the renewal equation (3. we note the formulas Ee. z2(U) = e7" J u b(u . Hence by a variant of the key renewal theorem (Proposition A1.x).3) for z/'(u). () . and alsoo zl(u) + 0 because of B['y ] < oo.9.x). But from the proof of Theorem 5. () . O} (9. () exp {w(9 + a. (9. Combining these estimates .p)/C'y.w(9.8) (Section 5). 0(u) = /3 Ju"O B(x) dx + f 0 0(u .St (U) Ee.QB(x) dx.4t (U)e°`U = which are wellknown and easy to show (see e.8). F(dx) = e'yy/3B(x)dx.(e"U = = wS(O. the proof is complete.C). we get p(u) = J "O B(x) dx + J U O(u . ()} .x)B(x)dx. 11 For the following.12)). ()] exp {w (O + y. By dominated convergence.3(x) dx. u 0 Proceeding in a similar way as in the proof of the CramerLundberg approximation based upon (9. Z(u)/u a C//3PF where PF is the mean of F.9) (9. () .3 (see in particular (5.([a] = exp {w(9 + a. we multiply by e7" and let Z(u) = elt" cp(u). Be. SENSITIVITY ESTIMATES Proposition 9.
w(e.2) holds.1) B(dy) 'f '[t(y) .wc (O. ()](e7v .12) f exp {O y + (t(y) . ()]B(dy) dx.we (0. 8^ ue7u. this implies Z = z + F * Z.6C do 89 1p 8( 1p Proof By straightforward differentiation. oo z2 (u) f C . C) .w( (0. z = zl + z2. ())B(dy) dx.lB(x) dx = e7uzl(u) + e7°zz(u) + V(u T where zl (u) = . 2 z 07P N ue7u (3C de . THE COMPOUND POISSON MODEL [we(e+7. ()]B(dy) dx x 0 0C T ON O .5 Assume that (9.90 CHAPTER III.w(0. 8 8() 8( (9.8) that cp(u) . 01 (i+) do = +'Y. )}B(dy)• Letting cp it thus follows from (9.11). Then as u > oo. ()]e7vB(dy) 'fCd 7 c .9)(9.wc(O.x)f3 f ^[t(y) .e7x/3 f 00 [t(y) .x). By dominated convergence and (9. ^)} [wc (0 + 7.w (9. F(dx) = e7x.wc(9. u Z2(U) = e7° f u ^/i(u .QB(x)dx.6e7u f "o f[t(y) . ^) .w( (0. C)] (1 + 7 ) Proposition 9. ()} 1z(dy) = f [t(y) . 0 x Multiplying by e7" and letting Z(u) = e"uV(u).
6 Consider the gamma density b (x ) = Sa xa.. by inserting in the above formulas.) log(9) = %F(a) logs where %1 = F'/]F is the Digamma function..w((9.1 . Z(u) /3C 91 o c'o e11(t (y) .12) follows. ue_Yu 'C2d( 8a 8( 1 p .1edz = 1 exp {Sx + a log x . and also zj (u) 4 0 because of f Hence. and the proof of the first one u is similar.yu/3C2do u86 89 1p' az/) = 8z/. We get w( (0. that C = a. SENSITIVITY ESTIMATES as u 3 oo. () = log r(a) . ( 9. a /(S .(log r(a) a log S)} • r(a) 1. Here (9.rye) S 5rya. w(e. () ='I'(t.Y)a+1 ' (9./35' a/i'y + aryl 625ry.13) (9. . U 7µF from which the second assertion of (9.a/35a&y' ' (9.3ary tog('Finally. < = a. ())B(dy) < oo. t(x) = logx.pa+1 .12) takes the form y) alp a .16) (9. () = C/9 = a/S.14) de = d( 7!3 76 = 7e = log ( \ ( \5a_ / \SSry ) 72 . 9 = S.2) holds with p(dx) = xldx.15) (9.9.a log S = log r(c) .ry) 5a1 cry (5 .17) (9.QS 1 .C log(9).. Example 9. It follows after some elementary calculus that p = a)3/5 and. we (9.18) (05 + 57 _'3_y .Sry a/32 + a/37 + /37 .
THE COMPOUND POISSON MODEL Example 9. C) = B = Yc = de = do = .log c = 2 In particular.w(9.2) with µ(dx) = 2x3zrdx.1 16 +ry c C22ry 2( = + 70 We (e. Straightforward but tedious calculations . Be."62 . which we omit in part . further yield .3Ee. () = Cc .CZ try)} 1 C C2 try .7 Consider the inverse Gaussian density ( b(x) Zx37 exp This has the form (9. for a < a* = z (.([Y] . 9 = .21og 2. w(e.3. C = . ()} = exp {c (C .l3 of Section 6a needed for the existence of ry becomes e^Q > 1+62 / 2.92 CHAPTER III.2 .S[a] = exp {w (9 + a.9) (() . C) .2a) } Thus the condition B[a*] > 1 + a* /. t(x) _ .2 log (0.1 = eXP {c(C .22.
thus.. In general. 10 Estimation of the adjustment coefficient We consider a nonparametric setup where /3. to our knowledge.10. .+UNT) > 1.8 The specific form of (9. kT (a) = /T (BT [a] . in which case we can just let t(x) = 0. in u which case the extension just described applies. ESTIMATION OF THE ADJUSTMENT COEFFICIENT Finally. 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). by the LLN both F (NT = 0) and F (PT > 1) converge to 0 as T . Van Wouve et al. Note that if NT = 0.3C2de 1p' z a = c .a.. and we estimate y by means of the empirical solution ryT to the Lundberg equation.2 of the parameters. BT [a]= NT ^` e"U. To this end.12) takes the form a = a 93 ar.1) . for which we refer to X. and hence explicit or asymptotic estimates are in general not possible. Comparatively less work seems to have been done in risk theory. Thus. sj=1 and let YT be defined by IKT('ryT) = 0.. (9. the exponent of the density in an exponential family has the form 01 tl (x) + • • • + 9ktk (x). That it is no restriction to assume k < 2 follows since if k > 2. we can just fix k . if 1 PT = /3TNT(U1+. ae t 1lEY u S _ . However . Finally if k = 1.cue_7u)3C P Remark 9. [379] consider a special problem related to reinsurance. queueing networks) are typically much more complicated than the one considered here. then BT and hence ryT is undefined. the main tool is simulation. the models there (e. Notes and references The general area of sensitivity analysis (gradient estimation) is currently receiving considerable interest in queueing theory. the results presented here are new. let NT 16T = ^T . then ryT < 0. B are assumed to be completely unknown.2) is motivated as follows. Also.g. or Ct(x).oo. However.7 and references there. we have assumed k = 2 and ti (x) = x. Thus. the exponent is either Ox.
3T . Lemma 10 .(27)/K'(7)2.: N ()3.i3)(B[7] 1) + (3(BT[7]  .B[7]) + B [7] .b[Yp'V21 T { (E[7] .B[7]) 0+ Iv/o(b[y].2) rT(7) N N (0. we need a lemma.1) .7 + (.2 As T * oo.B[7]2 n Hence ( 10. since NT /T .)vl+ N CO.3) Proof Since Var(eryU) = we have B[7]. .Q and Anscombe 's theorem. vfoVFB[2y]./^ B[27] . B[27] . then (10.a BT[7] I B[7] I + .94 CHAPTER III.'s.T y .. (10.1)2 + E[27] . More generally. THE COMPOUND POISSON MODEL Theorem 10 . Hence KT(7) = (F' + (OT a(B[7l 0))((BT [7] .1 As T 4 oo.v. If furthermore B[27] < oo. it is easy to see that we can write \ V1 1 l _ . N ( n[7]. 1) r.B[7]2 V2 . For the proof.If . B [7]2 (10.B[7]2 }) ( T 0 .'Y . a2 where a2 = /3r.2) follows from NT/T a4' .: N 0. B[2'Y]  /3T ) .1) 'YT .1) .3/ T). 16T where V1. V2 are independent N (0. 7T a4' 7.
10.E) < 4T(7T) < 4T(7 + E).e) < 0 < r.1 By the law of large numbers. y + E) eventually. it follows that 7T7 KT(7T) . 6"Y (10.E ) < 4T(7T) < (7 +0' which implies 'T(ry4) a$' r. OT a 95 u 4 /3. NT BT [a] Hence r. ESTIMATION OF THE ADJUSTMENT COEFFICIENT which is the same as (10. and the truth of this for all e > 0 implies ryT at 'y.c'(7) N (0' T (2(7) / N (0.(ry . where ryT is some point between ryT and ry. °7IT) . Let 0 < E < ry. Combining ( 10. Now write KT(7T)  kT(7) = 4T(7T)( 7T 7). If ryT E (7  we have KT(7 . lcT(a ) 4 /c(a)..e. 0 are estimated from data . 7T E (y .3).Q. Then r. Proof of Theorem 10.4) and Lemma 10.2. BT[a] 3 B[a].'T(a) = 1 E Uie°U' a$' EUe "u = B'[a].e. NT i =1 n'(a) for all a so that for all sufficiently large T K7 .'(y). To this end .4) + E). first note that e7TU N (e7U u2e27Uo'2/T) 7 .1 can be used to obtain error bounds on the ruin probabilities when the parameters . I.(ry + e) and hence KT(7 .KT(7) kT(7) K'(7) . Theorem 10.E) < 0 < kT(7 + E) for all sufficiently large T . By the law of large numbers.
3 or equivalently p > 1/2 or 11 < 100%. Mammitzsch [253] and Pitts. various alternatives have been developed.T = 3TKT ( 21T)IKT (^T)2 is the empirical estimate of vy and fc.. i. = 1. 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. with a tail of the form P(Y > y) ..e.T VIT where r7ry. Deheuvels & Steinebach [102]. Letting Wo = 0. Hipp [196].g.g.5%). [197].C1e"a ( see e.. Notes and references Theorem 10.1 is from Grandell [170]. Embrechts & Mikosch [133].96 CHAPTER III. U2. Asmussen [23]) can then be used to produce an estimate of ry. satisfies b(.. . ft. wn = inf{t > W. For this reason .f.. the nth busy cycle is then [Wn1. Vt = St . Further work on estimation of y with different methods can be found in Csorgo & Steinebach [94].Q. This approach in fact applies also for many models more general than the compound Poisson one. i . Herkenrath [192]. 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. t]}. Csorgo & Teugels [95]. 6 < 2.d. V. Griibel & Embrechts [292].. Wn).. For example .. > 0 for some t E [Wn_ 1.ueryuU ".1 : Vt = 0. Frees [146]. it means 2 (8 . if B is exponential with rate 8 so that ry = 8 .Wn) are i . A major restriction of the approach is the condition B[2ry] < oo which may be quite restrictive.info< „< t S.i. and the known fact that the Y„ = max Vt tE[W„1.) = a (e.0) < 5.96 if a = 2.e.
Chapter IV The probability of ruin within finite time This chapter is concerned with the finite time ruin probabilities 0(u. See Fig. generalizations to other models are either discussed in the Notes and References or in relevant chapters. The safety loading is q = 1/p . the Poisson intensity is 0 and the claim size distribution is B with m. Unless otherwise stated. B[•] and mean AB. 0. The notation is essentially as in Chapter III.s.1 (the role of ryy will be explained in Section 4b). defined as solution of c(ry) = 0 where ic(s) _ /3(B[s] .g. the premium rate is 1. 'y) where c(a) attains it minimum value. exists. 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. In particular.1) .f.1 where p = 13µB. 97 . it is assumed that i > 0 and that the adjustment coefficient (Lundberg exponent) y. Further let 'Yo be the unique point in (0.
2 that E [T(u)k.. 1 FL. (u) is exponential with rate 0 w. PL = 6/0 = 1/p > 1).u is the overshoot. EL refer to the exponentially tilted process 3 with arrival intensity S and exponential claims with rate / (thus .1 In the compound Poisson model with exponential claims with rate S and safety loading 77 > 0.(.r. In particular. using that the overshoot l. E[T(u) I T(u) < 00 ] = ELT (U). By the likelihood identity III.) = e7u ELe'Y^(u) ELT(U)k = e'Yu b ELT(u)k = O(u)ELT(u)k.(4. 7. FL and independent of T(u).9).2) Proof Let as in Example 111. we have for k = 1. the time of ruin is T(u) and ^(u) = ST(t&) .t. 1 Exponential claims Proposition 1.(U) < 00] = ELT(u)ke'YS. PROBABILITY OF RUIN IN FINITE TIME Figure 0.1) (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 (1.98 CHAPTER IV. Var[T(u) I T(u) < 00] = VarL T( U) . .1 The claims surplus is {St}.5 .
the Laplace transform of the time to ruin is given by Eea7( u) = E [eaT (u).V/ is as asserted. Let 0 > yo be determined by ^c(0 ) = a.1) .6 + a)0 .1.2 In the compound Poisson model with exponential claims with rate 6 and safety loading rl > 0. u + ELe(u) _ PL . 0 Proposition 1.1)2VarLT(u) + 2 Ca 1I VarLT(u). This means that /3(6/(6 .3) B = 0(a) = + (6/3a)2+4a6 2 and hence that the value of ic(yo) Proof It is readily checked that yo = 6 . Since Sr (u) and (PL .h.B = a.s.0) . EXPONENTIAL CLAIMS For (1 . 1).(PL .2). .2) is aLELT( u) .12 Thus the l."(ry) = 26//32.1)T(u))2 = UL where = s. Wald's second moment identity yields 2 EL (Sr(u) . of (1. is V1rLSr( u) +VarL ((PL .s.1)ELT(u). which leads to the quadratic 02 + (/3 ./3 .1 /3u + 1 u + 1 //3 = 6/3 6/01 For (1./3) . where = eBu I 1 .1//32 (6/)3 1)2 26(/3u + 1)/(6 .h.1)) ELST(u) ELT(u) (PL . the 1.s.1)T(u) are independent with QL the same mean . we have by Wald's identity that (note that ELSt = t(pL .h.1 (6)3)2 which is the same as the r. T(u) < oo] fora > r.1)T(u)) = VarLe(u) + (PL .6a = 0 with solution 0 (the .(yo) = 2V .I (1.6.
1 . St Ti F..4. M(u) T(u) = T + E Tk k=1 where T = T(0) is the length of the first ladder segment . ...v.T+ Ti a t U T I 1 a i F. Note that it follows from Proposition 1. Fig. Using 5 = 6 . But by the fundamental likelihood ratio identity ( Theorem 111.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. .4) The interpretation of this that T(u) can be written as the independent sum of T(0) plus a r. are the ladder heights which form a terminating sequence of exponential r. T(u) < oo] = EB [exp {aT(u ) . T2 ..3. Cf. 1.. and M(u)+1 is the index of the ladder segment corresponding to T(u). the result follows.100 CHAPTER IV.1 where Y1. Ti. T(u) < oo] = e... Y2. are the lengths of of the ladder segments 2. Y(u) belonging to a convolution semigroup . (1.Y1 Y2 Figure 1. PROBABILITY OF RUIN IN FINITE TIME sign of the square root is + because 0 > 0).3 that we can write EeaT( u) = eeuEe 017(o).9ST(u) +T(u)!c(0)} . More precisely.0.3) we have E [e«T(u ).v.'s with rate 5..
UN. and exponential with rate S = 1. let (cf.T.T are conditionally i.T is the residual service time of the customer being currently served and U2 .3 sin0 + 29) f3(0) = 1+/32/cos9.i (u. Note that the case 6 # 1 is easily reduced to the case S = 1 via the formula V. .3 Assume that claims are exponential with rate b = 1. i.6.T + • • • + UN. If QT = N > 0.1(u.. T) to be evaluated by numerical integration: Proposition 1.4. density xN lex/(N .cos (u/. Since U1 .I ex cos B cos j O dB fo " ... cf. .1. including the customer being currently served).1.ST).v. For j = 0. Then V(u. EXPONENTIAL CLAIMS 101 For numerical purposes .e.6(u) = Vfl/j l(Su.T.1 )!.T) = P(VT > u) where {Vt } is the workload process in an initially empty M/M/1 queue with arrival rate 0 and service rate S = 1.T the service times of the customers awaiting service .k + 1).. the following formula is convenient by allowing t.T) 1 I fl(O)h(0) fdO where (1.6) fl(9) f2(0) = = fexp {2iTcos9(1+/3)T+u(/cos91) cos (uisin9) . the conditional distribution of VT given QT = N is that of EN where the r. EN has an Erlang distribution with parameters (N. then VT = U1. . [4]) 00 (x/2)2n+3 Ij (x) OnI(n+j)! . 2. T. U2. 1).. Let {Qt} be the queue length process of the queue (number in system. Corollary 11. .i. .T.d.. UN.. where U1. 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 .0. Proof We use the formula .
PROBABILITY OF RUIN IN FINITE TIME denote the modified Bessel function of order j.8 ) yields F(QT > k + 1) .3(k +1)/ 2ei(k + l)6 (.1)] L _112 /(k+1)/2 [. f3(0) . 912.cos((k + 2)9)] d9. in particular equations (1.cos (( k + 1)0)] f3(9) Hence the integral expression in (1.(31/2eie . k k2 + $k+1 E bj 00 t j .(31/2 cos (( k + 2)9) .1 R [. 00 E '3j/2 cos(je) j=k+1 00 _ j=k+1 ^j/ zeij = . similar formulas are in [APQ] pp.38).13(k +l)/2ei(k +1)9 R E .31 /2eie L 1)] 1 I/31/2eie .44). Then (see Prabhu [294] pp. (1.)3k+1 = e(1+0)T e201/2Tcos 7r 0 e )3(k +l)/2 [31/ 2 cos ( kO) .102 CHAPTER IV. and define tj = e(1+R)Taj/2Ij(2vT T).112 l 1( k +1)/2 [ 31/ 2 cos(kO) .ie .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 . let I _ j (x) = Ij (x).1 00 ok+lR 00 j=k1 +1)/2e .)3k +1 tj g'(QT >.k + 1) = 1 k +1 + bj j=00 j=00 00 j=kk+1 j=k1 By Euler 's formulas.i(k +1)e R [/3( klal/2e:0 (01 /2 e . 8789) 00 E aj j= 00 = 1.
or. Ui < x I / (note that P(St < x ) = F(x + t. however. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Since P(QOO > k + 1) = flk+1. T) in terms of F(. t )). there are several misprints in the formula there.0(u.e = e' COS a cos(uf31/2 sin 0).. it follows as in (1.7) that _ [^ au ak+l (30 k L. We allow a general claim size distribution B and recall that we have the explicit formula z/i(0) _ P(7(0) Goo) = p. k! k=O k0 i/z Co Uk ate" o'/z e . Seal [327] gives a different numerical integration fomula for 1 . The first formula. the numerical examples in [12] are correct). 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). t) = P .T) which. equivalently. E Fk. expresses V)(0. We first prove two classical formulas which are remarkable by showing that the ruin probabilities can be reconstructed from the distributions of the St. u Notes and references Proposition 1. F(x. . T). 2 The ruin probability with no initial reserve In this section .2. oo (u)31/2e^e)k = )3k z cos(k9) = R k. from the accumulated claim distribution N. Related formulas are in Takacs [359]. going back to Cramer.3 was given in Asmussen [12] (as pointed out by BarndorffNielsen & Schmidli [59]. The rest of the proof is easy algebra. we are concerned with describing the distribution of the ruin time T(0) in the case where the initial reserve is u = 0. is numerically unstable for large T. and the next one (often called Seal's formula but originating from Prabhu [293]) shows how to reduce the case u 54 0 to this. however.
1 1 . meaning that we interchange the two segments of the arrival process of {St}o<t<_T corresponding to the intervals [0. T T o where the second equality follows from II.T].104 CHAPTER IV. .T))dv E^T I(M(v. Proof For any v E [0.S„ 0 <t<Tv STS„+St_T+v Tv<t<T as the event that IS. ") } is at a minimum at time t.T)) 1 fT P(M(v.1 In formulas. 2. Then 1 .3) with A = (0.t)= {Stv) < SM.T))dv. co ).i.1. [v. 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.(. resp. T) = P(Tr(0) > T) = P(M(0.T) T F(x.T)) does not {Stv)} depend on v. f T lStv)} 0<t<T by a 'cyclic translation'.(6. Stv^ _ Define M(v.T)dx. See Fig.b (0. T].0<w<t} St+v . we define a new claim surplus process St StM NJ Figure 2.(0. PROBABILITY OF RUIN IN FINITE TIME Theorem 2 .
Sv. v<t<T}n{ST<STSv+St. T Theorem 2 . It follows that if M(0 .T)) dv f T I(M(0.v<t<T} = {ST<StSv. 0<t<v} = {ST < St ..2. T T o i =1 Let f (•. THE RUIN PROBABILITY WITH NO INITIAL RESERVE 105 Now consider the evaluation of fo I(M(v. or it occurs. We claim that if M(0. T)). T) = M(0. then M(v. v). letting w = inf It > 0 : St_ = mino<w<T Sw}. Fig 2. Hence T TE f I( M(v. in which case there is a last time o where St downcrosses level u. v < t < T} n M(0.Tt))f(u+t. If ST < 0. we can take v E (w E.T)f(I z /)(0.. then i fT I(M(v. T.T) and Sv < 0 on M(0. this integral is 0 if STv) . Obviously.T) = F(u+T. v)) dv = ST T T o (note that the Lebesgue measure of the v for which {St} is at a minimum at v is exactly . T) occurs. cf.2. w) for some small E. T) as {ST<St+ vS. v). T].xdx.T) occurs.ST on M (0. v) = M(0. v).T) occurs or not as long as ST < 0. Proof The event {ST < u} = { Ei T Ui < u + T j can occur in two ways: either ruin does not occur in [0. Indeed. T) occurs.t)dt.2 10(u. T)) dv. we can write M(v. there exist v such that M(v. t) denote the density of F(•. . where the last equality follows from ST < St on M(0. T)) dv = TEST = T fP(ST < x) dx T T NT 1 f P(ST < x) dx = 1 f P Ui T . It is then clear from the cyclical nature of the problem that this holds irrespective of whether M(0. For example. t). ST > 0. 0<t <Tv}n{ST<ST Sv+St T+v.
2 .t). Then P(T(0) E • I T(0) < oo) = P(T_ (Z) E •). Hence P(ST<u) = 1 . The following representation of T(0) will be used in the next section.Tt))P(StE[u.(0.T) = {St < 0. z > 0.T) = .z.p. ST_ _ z}. 2.3 Define r_ (z) = inf It > 0 : St = z}. which is independent of St and has the stationary excess distribution B0. The proof is combined with the proof of Theorem 111. PROBABILITY OF RUIN IN FINITE TIME u Q II T Figure 2. u + dt] and there is no upcrossing of level u after time t. E [t. O(T .u+dt]). ST_ _ z} . ST_ _ z}.T)+ J0 T (1V. {St > . Proof of Theorem 111. C*(z. 0 < t < T. define St = ST . For a fixed T > 0. {S t > z. 0 < t <T . Let Z be a r. 0 < t <T. which occurs w.106 CHAPTER IV.2.b(u.v. Proposition 2.T) = C(z. u which is the same as the assertion of the theorem.2 Here o. t + dt] occurs if and only if St E [u.ST_ t_ and let A(z.2.
3.ST(o) >y. Figure 2.3). Thus P(Sr(o)_>x. u which is the assertion of Theorem 111. A(z.2.T)) = P(Cx. T(0) < oo) B(y B(z) + z) f3B(z) dz = 3 f °^ B(y + z) dz = f3 + x v f B(z) dz.T + dT]. Hence integrating (2. {St }o<t<T have the same distribution . z + dz]. we therefore have P(A(z.1) yields P(ST(o)_ E [z. . ST(o)_ E [z. 7( 0) < oo) = P (C(z)) dT.2. 2.T).T))f3B(z) dz dT. r(0) < oo) = 3R(z) dz JP(C(z. and since {St}o<t<T. (2.3 But by sample path inspection (cf.T)).T(0)<oc) = f x F(U > y + z U > z) P(Sr(o)_ E [z.T) = C*(z.1) z T . It follows by division by P(ST(o)_ E [z. z + dz].2.T))dT = Off(z) dz P(T_ (z ) < oo) = 3B(z) dz. z + dz].1) that P(T(0) E [T. T(0) < oo) = OR(z) dz in (2. z + dz]. z + dz]) = P(A(z. Proof of Proposition 2. T + dT] I S7(o)_ E [z. Fig. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Then 107 P(r(0) E [T.
because of77>0. r(a) denotes the solution < 'Yo of the equation a = ic(r (a)) = .1) where a > r. Let T_ (y) be defined as Proposition 2. [329].1 Eear( y) = eyr(a). a martingale proof is in Delbaen & Haezendonck [103]. ^(0) E dx] (recall that ^(0) = Sr(o)) and write ga[b] = f OD ebxga(x) dx.3 was noted by Asmussen & Kl(ippelberg [36]. I L Let ga(x) be the density of the measure E[ear(°). cf.108 Hence CHAPTER IV.(yo). T(0) < oo) = dTP(T_(Z) E [T. 2. Tak'ecs [359].1) . T(0) < oo) 0 = dT f 0 P(C(z))P(Z E [z.T+dT]). see in addition to Prabhu [293] also Seal [326]. z + dz]. who instead of the present direct proof gave two arguments.1. r(0) < oo. In the setting of general Levy processes. Proof Optional stopping of the martingale I er (a) 9 t.(3(B[r( a)] .2. PROBABILITY OF RUIN IN FINITE TIME ]P(7(0) E [T. Theorem 2. one based upon a result of Asmussen & Schmidt [49] generalizing Theorem 11.3.5a). Proposition 2. (3. Lemma 3. Lemma 3 . Note that T_ (y) < oo a.r(a).5 and one upon excursion theory for Markov processes (see IX. Notes and references For Theorems 2.s. z + dz].T + dT] T(0) < oo) dT f ' P(C(z))P(Sr( o)_ E [z.1 and the present proof is in the spirit of Ballot theorems. some relevant references are Shtatland [338] and Gusak & Korolyuk [181].2 ga(x) = Qexr(a) f "o eyr(a)B(dy) x . 3 Laplace transforms Throughout in this section.6.c(r(a)) l = l er( a)se+at } u yields 1 = eyr(a)Eear(y).
u . y + dy].r(a) b . Corollary 3. time T(u): u u Here is a classical result : the double m.x)(a) B(dy)• Lemma 3 .3. Hence eb"du E[eaT(").ga [b] 0 TO Using Lemma 3. b .g.2 P(Z E [y. (u . LAPLACE TRANSFORMS 109 Proof Let Z be the surplus .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)) .f. £(0) E dx) = /3B(x + dy) dx and hence ga(x) = f e r)/3B(x + dy) _ /3 f x e(v.(v) = ev''(a).ic(b)/b x(b) + a eb"E[eaT(" ).4 E[eaT (o).x)ga (x) dx where za(u) = f.3 ga[b] = c(b) Proof + b + a . T(u) < oo] du = Proof Define Za(u) = E [eaT(" ).1] evr(a)B(dy)[ b . Then by Proposition 2. It is then easily seen that Za(u) is the solution of the renewal equation Za (u) = za (u) + fo Z.5 f 00 o a/r(a) . E[ear (o) I T(0) < oo . Further by Theorem 111.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(0) < oo] = 20[b] = za[b] (9a[b] 9a[0])/b 1 .3.ST(o)_ just before ruin .2. rr(0) < oo) = 1_ r(a) Proof Let b = 0. Z = y] = EeaT.r(a) = a [B[b] B[r(a)]] . (Laplace transform) of the ruin Corollary 3. r(u) < oo).ga [b] 1 .°° ga(x)dx.3.
6.(u ) = o(u) a.mL > E T(u) < 00 ) 40. Then given r(u) < 00. and take basically the form of approximations and inequalities. mu ) ( 0 m < ML '(u) 1 m > rL..h(u. note that by Wald's identity u + EC(u) = ES. Proposition A1. where _ 1 _ 1 1 C ML w(ry) 6B'[7J 1 . cf.UProof The assumption 11 < 0 ensures that P(T(u) < oo) = 1 and r(u) a4' oo. (4. P = /3µB > 1.3). uoo u using e. Later results then deal with more precise and refined versions of this statement. Theorem 4 . T(u)/u mL as u + oo. i.r(u) = Er(u) • ES.mu D 2 4 N(0. St/t 1 1/m. Then as u * oo.s. t T(u) T(u) T(u) t m = lim = lim = lim Utioo u + Sr(u) u+oo S. and hence a. for any c > 0 P( Further. By Proposition 111. we need the following auxiliary result: Proposition 4.3LELU 1 1p' is in some appropriate sense critical as the most 'likely' time of ruin (here C is the CramerLundberg constant).110 CHAPTER IV.s.1)Er(u) .. That is.e. The first main result of the present section is that the value umL. = (p .1 Assume 77 > 0. (u) t. This proves the first assertion of (4.2 Assume ri < 0. For the second . T(u) a. PROBABILITY OF RUIN IN FINITE TIME 4 When does ruin occur? For the general compound Poisson model. u 1 ET(u) 1 p1 u where Pw2 = 311B)m3• 7(u) .1.w ) v/. the known results are even less explicit than for the exponential claims case.00 St = lim . For the proof.1) i. for any m T(u) u .2.
N(0. apB ) . proving (4.t/m D (2) 111 . this can be rewritten as u + 1(u) .2 of [86]) and (4. 1'r(U) . T) for T which are close to the critical value umL). If Z . Theorem 7. though it is not easy to attribute priority to any particular author. According to Anscombe' s theorem (e.7 6  11 Proof of Theorem 4. For (4 .mL >E By Proposition 4. PL (•)+ 0. of (4. Notes and references Theorem 4.3. 4a Segerdahl's normal approximation We shall now prove a classical result due to Segerdahl. Tu) T( u) .r(u)/m T(u) ti µB2) Z. note first that ( Proposition 111.1).2) follows immediately from u (4. WHEN DOES RUIN OCCUR? and that Ee(u)/u a 0. T(u) < oo f / 00) e7uE L [e_7 (t1).1 The l.1).mu (2) '• m3/2 µB 7 .1 is standard.h.2.3). and as a timedependent version of the CramerLundberg approximation.1 (by considering 0(u.mu m .6µB2) Z v m (3µB2) Z. .mL U > E.5) St .1. the result comes out not only by the present direct proof but also from any of the results in the following subsections. Thus.4.g. cf.6. 4). implying T(u) .1).1) is T (u)  U mL P( T (u) < I > E. which may be viewed both as a refinement of Theorem 4.^ N (o. and (4. T (u) < 00 J 0(u) e7'PL U \ I T u) .s. again Proposition A1. the same conclusion holds with t replaced by r(u).
4). (oo. Then the distribution of T(u) . with w2 as in (4.v. oo).VU T. one has 9 (r(u)_rnu) Ef (^(u)) * E.6) whenever f. then e(u) and r(u) are asymptotically independent in the sense that.T ( u')] = E[ T ( ul /4 . g are continuous and bounded on [0. Hence Ef (Vu )) 9 (T(u.5) For the proof. PROBABILITY OF RUIN IN FINITE TIME Corollary 4. Then h(u) 4 h(oo) = E f (6(oo)).e(u') oo w .f ( (oo)) . S( u ) < ul/4] < ET(ul / 4) = O(ul/4). we get E[ T (u) . resp . e'°'/b (u. we can replace T(u) by r(u'). letting Z be a N(0.))I h(ul /4  ^(u)) I(6 (u') C ) f < ul /4 + f(e(u') . E9(Z) (4.(u. Then for any y.6).um.) is readily seen to be degenerate at zero if ST(u•) > u and otherwise that of T(v) with v = u . Using ( 4. and thus in (4.112 CHAPTER IV. P because of ^(u') .)mu \ h(oo)Eg (r(ul) .T(u') given F.ST( u') = u1/4 . oo ). using that ul/4 .t.ul/4.a C4'(y )• ( 4.3). we need the following auxiliary result: Proposition 4.^(T(u')).w2) r. O .4 (SIAM'S LEMMA) If 71 < 0. Let h(u) = E f (^(u)).u1/4)I(S(u') > u1 /4) h(oo) + 0.l:(oo) (recall that rt < 0). Proof Define u' = u .mul h(oo)Eg(Z). and similarly as above we get E[f(^(u)) I Fr(u.r.L+YWLV'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).
see also von Bahr [55 ] and Gut [182]. PL(T(u ) < umL + ywL) 113 4 C4(y). 4b Gerber's time. 0.7) to be valid is that T varies with u in such a way that y(T) has a limit in (. y u) < e 7v" . WHEN DOES RUIN OCCUR? Proof of Corollary 4. see Asmussen [12] and Malinovskii [254]. 3 is due to Segerdahl [333]. The present proof is basically that of Siegmund [342]. y u) < .T) Ce7"4 (T . Cf. where we used Stain's lemma in the third step and (4. yy by 1 K.3 in terms of Edgeworth expansions .umL wI V"U u (4. u needs to be very large). y > k'(7) . For practical purposes . ELe7E (") .4.7) To arrive at this . CL Fig. in practice one would trust (4. Notes and references Corollary 4 .5) and solve for y = y(T). define ay.1.5 '(u . Segerdahl 's result suggests the approximation b(u. oo ) as u * oo.3 ery"z/i(u . Thus . 10) '5(u) . umL + ywL f) = e"P(T (u) < umL + ywL) = EL [e7V ").4) in the last. e7v" y < ^'(7) (4 .9) ( 4 .oo.8) Note that ay > 7o and that 7y > •y (unless for the critical value y = 1/ML). just substitute T = umL + ywL in (4.dependent version of Lundberg's inequality For y > 0. The precise condition for (4. Theorem 4. T(u) < umL + ywL f.7) whenever u is large and ly(T)l moderate or small (numerical evidence presented in [12 ] indicates . that for the fit of (4. For refinements of Corollary 4.7) to be good.yK(ay)• (4.z/)(u . . also Hoglund [204]. however .(ay) = 17 7y = ay .
v"U.2. However. we have rc(ay) < 0 and get (u) . Hoglund [203] treats the renewal case. u Differentiating w. .114 CHAPTER IV.yu ) = eayuEav [e .8 below . see MartinLM [257] . yu 11 < T(u) < oo j < eayu +Y UK(ay) Remark 4. if y > 1/ic'(y).5. yy is sometimes called the timedependent Lundberg exponent. we arrive at the expression in (4. yu < T (u) < oo 1 l e ayuEav [eT ( u)K(ay). PROBABILITY OF RUIN IN FINITE TIME Proof Consider first the case y < 1/K'(y). yu) < C+(ay)e7a„ where l C+(ay) = sup f 00 eayR(xy)B( .3 yields easily the following sharpening of (4.9): Proposition 4. From the proof it is seen that this amounts to that a should maximize ayic(a). the point is that we want to select an a which produces the largest possible exponent in the inequalities.8). and generalizations to more general models are given in Chapter VI.r. 5 is due to Gerber [156 ]. which shows that the correct rate of decay of tp(u. f Some urther discussion is given in XI.1). T(u) < yu] < eayu + yUr(ay) Y < eayuEav [ eT(u)K(av )L T(u) < yu} Similarly. a.7 i. An easy combination with the proof of Theorem 111. yu ) = < eayuEay [eay^ ( u)+T(U)K ( ay). the bound a7y° turns out to be rather crude . Numerical comparisons are in Grandell [172 ].Y' (u.6 It may appear that the proof uses considerably less information on ay than is inherent in the definition (4. For a different proof. dy) Notes and references Theorem 4 . which may be understood from Theorem 4. In view of Theorem 4. Then ic(ay) > 0 (see Fig . yu) is e 'Yyu/ . 0.h(u. and hence t.8).ay4(u)+ T(u)K(ay ).6.b (u.t. who used a martingale argument.
T(u) < yu] . T(u) suggests heuristically that l t/. For any a > yo. and in case of ruin probabilities the approach leads to the following result: Theorem 4 . (4. then ay > 0.. not inequalities. u 4 oo.(u.e. WHEN DOES RUIN OCCUR? 115 4c Arfwedson's saddlepoint approximation Our next objective is to strengthen the timedependent Lundberg inequalities to approximations.yu) c ay . As a motivation. Using Lemma 111. We thereby obtain that T is 'in the center' of the Padistribution of T(u).ayuEay f eay^ ( u)+T(u)K(ay). it is instructive to reinspect the choice of the change of measure in the proof.ay and get Ea e ayf (00) y _ 'Ya( ayKal lay C 1 .4.'(y ). yu ) ayay e ryyu ayay 27ry/3B"[ay] u Proof In view of Stam 's lemma.8 If y < 1/ic'(ry). the choice of ay. [eT(u )K( ay). yu) = e. and ii(u) .yyu y l ay I 21ry/3B" [ay] V fU_ u + 00. Proposition 4. This idea is precisely what characterizes the saddlepoint method. Ea .13) ..e. (0) r1 (a) ' I. (4.ayC() . if we want EaT(u) .ay a.z. then the relevant choice is precisely a = ay where y = T/u. The traditional application of the saddlepoint method is to derive approximations. we have ryas = ay . and b(u.11) ' If y > 1/ r .6 with P replaced by Pay and FL by Pay.c(&) = ic(ay) is < 0.5. i.. then the solution &y < ay of .ay y 'Yay  ay .2 yields EaT(u) u u r.^3 ]1/ Bay [lay . yu ) eaauEaye . the formula 0(u.: T.i(u.12) < yu] Here the first expectation can be estimated similarly as in the proof of the CramerLundberg ' s approximation in Chapter III. (4.
1)3 = (jB"[ay]l (Pay .ay) ay +.a. 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).1.1B[ay]1 ) y(ay .ay)K(ay) ay ayI&YI For the second term in (4.c'(a) _ /3a/(8 . and the equation ic'(a) = 1/y is easily seen to have .(j (1 .13). PROBABILITY OF RUIN IN FINITE TIME ry I i .a)2 .7ruw2 Inserting these estimates in (4.B[ay] /ay &y y(ay .4). V < 01 Ir 00 er(ay)"1'2"'x eyur.1) under Pay mation (4.11) follows.c(ay)ul/2W p 2ir = eyu(ay) dz 1 rc(ay ) 2.116 CHAPTER IV.1) .3(5/(S .ay + ayl /BLay] .13) rigorously.a) . .I ay &y a ^c'(ay) a (1 +. The difficulties in making the proof precise is in part to show (4.ay ) r. where V is normal(0. i B[7ay . a nr=.(ay) _ y(ay .1)3 = y3/3B"[ay]. T(U) < yu] = eyuk (ay)E''ay (ek(ay )"1/2WV. Then ic(a) = . (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 . it seems tempting to apply the normal approxiyu + ul/2wV. we get heuristically that Eay Ler (u)r(ay).l'B)y /(Pay .12) is 0 entirely similar.9 Assume that B(x) = eay.13). The proof of (4. Writing r(u) and W2 = I3ay{. (4. Example 4.
g. A related result appears in BarndorffNielsen & Schmidli [59]. (5..ay)3 0 3/2 and (4. yu) when y < 1/ic'('y) = p/1 . 2 = Var(Si ) the variance. c a 00. DIFFUSION APPROXIMATIONS solution ay=5 117 V 1 (the sign of the square root is .i )( v s vc ('3 + s _2 / . 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. in discrete time: if p = ES.f.5.p.8 is from Arfwedson [9]. and next to note that such an approximation in particular implies that the first passage probabilities are close./4 ^y for 1/i (u.3+52 1+/351/y' sy 7 B ii[ay] 25 _ 251/2(1 + y)3/2 (5 . then { __ .tcp) Lo {Wo ( t)}t>0 . . is the drift and o.1..1) .because the c. y) a''y" L '3 _ fl ) 51 /4(1 +1IY)3/4 \.11) gives the expression '31/4 ( . The mathematical result behind is Donsker's theorem for a simple random walk {Sn}n=o.. It follows that 5^y =5ay = /«y =f3+ay=l3+d 1+1/y' V 1+^1/y /35 1+1/y /3' ay ay =Qay say =. is undefined for a > 5). 0 Notes and references Theorem 4.= (s.
p. this is an easy consequence of (5.3) whenever c = cp f oo as p 1 p. Indeed .tcpp) y = { WC (Sct) pct) } {Wo( t)}t>o (5.z } {W_1(t )}t>o (5. PROBABILITY OF RUIN IN FINITE TIME where {W( (t)} is Brownian motion with drift S and variance (diffusion constant) 1 (here 2 refers to weak convergence in D = D[0. (5.1) with S.3..7c). where p is the critical premium rate APBTheorem 5 .tp). oo)). 0 .e. n/c < t < (n + 1)/c. for the purpose of approximating ruin probabilities the centering around the mean (the tcp term in (5. cf.a = Snp) and the inequalities Sn )C .118 CHAPTER IV. St = EN` U= .2) t>o where p = pp = p . a2 =/3µB2) Proof The first step is to note that { WC (St P) . We want an approximation of the claim surplus process itself. such that the claim size distribution B and the Poisson rate a are the same for all p (i. Letting c = a2/pp.1)) is inconvenient..t} _ {W_1(t)} . of which a particular case is the claim surplus process (see the proof of Theorem 5. and consider the limit p j p.p/c < St(p) < S((n+l)/ c + Pp/c. p. This is the regime of the diffusion approximation (note that this is just the same as for the heavy traffic approximation for infinite horizon ruin probabilities studied in III.1 As p J. Lemma 111. However. and this can be obtained under the assumption that the safety loading rt is small and positive.3) takes the form LI S(P) { a2 to2/µ2 + t LI S (P) { a2 ta2/µ2 {W0(t)}.1 below). we shall represent this assumption on 77 by a family {StP) L of claim surplus processes indexed by the premium rate p.1. Mathematically. we have o {i!t s: . + {Wo(t ) . It is fairly straightforward to translate Donsker's theorem into a parallel statement for continuous time random walks (Levy processes).
DIFFUSION APPROXIMATIONS Now let Tp(u) = inf{t>0: S?)>u}.5) Note that letting T * oo in ( 5.(u) ti IG(oo.t. .h.( ^ I + e2( \ I .6) This is the same as the heavy traffic approximation derived in III.Ta2 /p2).1. For practical purposes .7c.. we obtain formally the approximation V.u). is 1/ip (ua2 /IpI.h. w.s.1.. ulpI/a2).T) IG(Tp2/ a2).8 or [APQ] p.1 I 7= . 196. u) of r( (u) (often referred to as the inverse Gaussian distribution) is given by IG(x. has a continuous distribution..s. ^ p2 Proof Since f 4 SUP0<t<T f (t) is continuous on D a. (ua2 To2 op \ IPI > IG ( T .1 . Because of the direct argument in Chapter III. and in fact some additional arguments are needed to justify (5. 199. 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. C.s.f I \\\ J \ (5. 119 It is wellknown (Corollary XI. 263) that the distribution IG(•. u) =PIT( (u) < x) = 1 . and the r. u). (5.5.e. 1. since ti(u) has infinite horizon . TS(u)=inf{t>0: WW(t)>u}. ('. ulpl /a2) = e2"1µl / or2. u) is defective when < 0. However.2 As p j p. see Grandell [ 168].r.4) Note that IG(. we omit the details . (5. Corollary 5 . is IG(T.5). any probability measure concentrated on the continuous functions. the continuity argument above does not generalize immediately.2 suggests the approximation u 0(u. the continuous mapping theorem yields sup W Sz2 to lP 4 sup Wi(t)• O<t<T O<t<T a2 Since the r.h. [169] or [APQ] pp. Corollary 5.6) from Theorem 5. (.
5) for the compound Poisson model which does not require much more computation. in Asmussen [12]. 0) { 2 StQ2 /µ2 D { W_ i(t)}t>o t>o D 2 where p = pe = pe .6 of [APQ]. B0 * Boo. in the next subsection we shall derive a refinement of (5. the B9. the simplicity of (5. (5. Then as 0 _+ 90.120 CHAPTER IV. We conclude this section by giving a more general triangular array version of Theorem 5. The first application in risk theory is Iglehart [207]. pt? 4 peo. In contrast.1 and Section VIII. . See for example Billingsley [64]. Assume further that 039µB6 < pe. the claim size distribution B9 and the premium rate p9 depends on 0. pe .t. The picture which emerges is that the approximations are not terribly precise.6) therefore does not appear to of much practical relevance for the compound Poisson model. e. Theorem 5. However. In view of the excellent fit of the CramerLundberg approximation.r.. we have ^A.1. as an example of such a generalization we mention the paper [129] by Emanuel et al. For claims with infinite variance.g. PROBABILITY OF RUIN IN FINITE TIME Checks of the numerical fits of (5.5) 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. All material of this section can be found in these references. However. and which is much more precise. Further relevant references in this direction are Furrer [151] and Boxma & Cohen [75].6) are presented. Furrer. a2 = ae = 00µa6 Notes and references Diffusion approximations of random walks via Donsker's theorem is a classical topic of probability theory. on the premium rule involving interest.3 Consider a family {Ste) } oc claim surplus processes indexed by a parameter 9. that 00 4090. in particular for large u. The proof is a straightforward combination of the proof of Theorem 5.00µB6 + 0. [169].5) and (5.Pe. and two further standard references in the area are Grandell [168].Po = 09µB6 . Michna & Weron [152] suggested an approximation by a stable Levy process rather than a Brownian motion. as 0 * 00 and that the U2 are uniformly integrable w. such that the Poisson rate Oe. for more general models it may be easier to generalize the diffusion approximation than the CramerLundberg approximation.
ao (0) _ /c(s + 9 . this is because in the regime of the diffusion approximation .'(yo) = 0 and let 90 = 'Yo. risk process with safety loading 77 > 0 correspond to 9 = 0 . PB('r(u ) < oo) < 1 for 9 < 0. In terms of the given risk process with Poisson intensity . Then r.T) = Peo(r(u) < T) for 90 < 0. and we want to consider the limit 77 10 corresponding to Oo f 0.6. Since Brownian motion is skip free. whereas there we let the given 3B. 77 is close to zero. claim size distribution B . . this idea ignores (among other things) the presence of the overshoot e(u). 9o T 0. and we are studying b(u. let P9 refer to the risk process with parameters Q9 = QoB0[9] = QB[9 9o]. Then EOU' = Boki[0] = Biki[eo]/E[9o] and "(s) = k(sBo)k(9o).1 > 0.c(s) = . However . 0(0) = 0. 77 = 1/p . Determine yo > 0 by r.Q (B[s] . Bo(dx) = B[eo]B(dx). 3.4.6. For each 9. Let PO refer to the risk process with parameters e9oz Qo = QB[90].s and p = /3µB < 1.90) . P9(r (u) < oo) = 1 for 9 > 0. In this setup. which we have seen to play an important role for example for the CramerLundberg approximation ./c(9 .90) and the given risk process corresponds to Poo where 90 = 'yo. 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. 2. B9(dx) =Bale] Bo(dx) e9z keo)z = B[9 .90] B(dx). The setup is the exponential family of compound risk processes with parameters ( B9 constructed in III. The objective of the corrected diffusion approximation is to take this and other deficits into consideration.1) . .9(s) = Ico ( s + 9) . this means the following: 1. it is more convenient here to use some value 9o < 0 and let 9 = 0 correspond to n = 0 (zero drift).
C) = 2A + (2 . 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 > 0.3) this implies (take u = 1) Ego exp { . tu2 ) i IG (t. bl IG(t81. S2 = 3E0U2 Bier [Yo] 3B"[Yo] Write the initial reserve u for the given risk process as u = C/Oo ( note that C < 0) and.u. Theorem 5. for brevity.2' where as ususal ry > 0 is the adjustment coefficient for the given risk process. (.(y) = 0. i. .T) 1+u2 (6.e. write r = T(u).1) . () where h (A.() The idea of the proof is to improve upon this by an O (u1) term (in the following.2) .1) IG(x.. The first step in the derivation is to note that µ = k (0) = r0 (00) . The corrected diffusion approximation to be derived is (u. PROBABILITY OF RUIN IN FINITE TIME Recall that IG(x. (6. means up to o(u1) terms): .3 applies and yields 1061 U61 Stdlu2/CZdi {W_1(t)}t>0 t>0 which easily leads to 1 StU2 {W( J(t)1t>0 { u S1 t>o Y'(u. u) = euh(a . (01. the solution of r. u) = IG(x/u2.7(u)/u2} eh(A. .S.122 CHAPTER IV. One has (6.Varo S1 = f30Eo U2 = S1.. 9otc0" (0) = 0061 = ul. Vargo S. C . 1) • Since L eatIG (dt. (. 0o to. and Si = QoEoU2 = Q B"'['Yo Eo U3 ].. IGu+u2. (U. C.C. _ ^(u) = ST ..
u is Eeazead2/++ Eeaz[1 + ab2/u] where the last expression coincides with the r. we have p =.h. A numerical illustration is given in Fig. But the Laplace transform of such a r.2) is indeed o(u1). is the c.5) according to (6.1 + 629. 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.exp { h(A.1 As u + oo.5) Once this is established . of (6. of a (defective) r. To arrive at (6.2 ). The solid line represents the exact value .3. .d. .s.3 = 0. in (3) and (4).z .ry2 .v.7. 1% in (2) and (4). it holds for any fixed A > 0 that Ego exp { Ab1rr(u)/u2} . that the saddlepoint approximation of BarndorffNielsen & Schmidli [59] is a serious competitor and is in fact preferable if 77 is large] . and the dotted line the corrected diffusion approximation (6. (6. distributed as Z . the formal Laplace transform inversion is heuristic: an additional argument would be required to infer that the remainder term in (6.f. The initial reserve u has been selected such that the infinite horizon ruin probability b(u) is 10% in (1) and (3). In ( 1) and (2).6. we get by formal Laplace transform inversion that C 2 u.3). 1.4.1 + u2 I Indeed. Note. calculated using numerical integration and Proposition 1. 6 .2).v. 9o T 0 in such as way that C = Sou is fixed.1 below is exact. bl I IG I t +2 . . however . which is based upon exponential claims with mean µB = 1. CORRECTED DIFFUSION APPROXIMATIONS 123 Proposition 6.52/u where Z has distribution IG (•. that whereas the proof of Proposition 6.'yu /2)(1 + b2/u)} + Aug 1I J .s. just replace t by Tb1/u2. .h. the r. however. p = 0.
T) 111 0.1 proceeds in several steps. (Inc 0s 0. the fit at p = 0.T) 0. see Asmussen [12].7.4 may not be outstanding but nevertheless.00 0. Note that the ordinary diffusion approximation requires p to be close to 1 and '0 (u) to be not too small.TI CHAPTER IV.011 L1 60 T IM 11.07 0. OM 0.W21 0.01 0.(061 0. For further numerical illustrations.2 e.1 W IU.7 or at values of Vi(u) like 1% is unsatisfying. A51 7(SAT 3 3 h(X.() Lemma 6.05{ 0.OOIi O.08 a.19)2 11 20 20 i0 T 1n0 Figure 6. Similarly.T1 00.u2 2u3 (e . BarndorffNielsen & Schmidli [59] and Asmussen & Hojgaard [34].1 It is seen that the numerical fit is extraordinary for p = 0.02 I 90 120 160 2W A0 Z WT 40 80 120 160 100 240 280 T 111 WI.08 0.aa1 . and all of the numerical studies the author knows of indicate that its fit at p = 0.0 0.^) .111 W(U.199 0. it gives the right order of magnitude and the ordinary diffusion approximation hopelessly fails for this value of p. .124 0.. The proof of Proposition 6..114 0.EB 0 p ex p ( 7 S h ^)u . PROBABILITY OF RUIN IN FINITE TIME 0.
2u (B3 .+ h (A.co (e) . in Lemma 6..4) that the r.r0 (00)) } Replacing B by 8/u and Bo by C/u yields e(B() = E eo exp { (e .61a2T (B3 . the formulas Po(C(0) > x) Po(C(co) > x) imply 1 °° Po(ST(o) > x) = EIU fIP0 (U>y)dy .6) u U3 Lemma 6 . CORRECTED DIFFUSION APPROXIMATIONS Proof For a>0..(3) J t _ aa1T l + eh(A.7) 2 2 . C) 1 1 + u2/ 111 + 2u CZ Z  (2A + ()1/2 J 1 Proof It follows by a suitable variant of Stam's lemma (Proposition 4.00)(u +C)  'r (. (6.T (co (8/u) .6.C2 = 2).2 behaves like C l Eeo eXp r _ ^81T 1 Sl u2 1 u 2u3 [1+h(AC) S .h. () 62 Eeo exp u u2 J .1) h(A.co ((/u)) } Let 8 = (2a + (2)1/2 = h(). 3 lim Eof (u) = EoC(oo) = a2 Ep = 3EoU2 uroo Proof By partial integration . + a1b2 + .. the result follows. exp ue } al 1J 3 exP I [2). (6..3 EoU2 + 103OoEoU3 + " 2 6 Using d2 .(3)Eea LauT exp i 3J .s.C)C/u .4 Ea. () + C and note that 2 KO (0) = 102. 1 / Po(C(0) > y) dy EoC(0) x k EDUk + 1 k Eo[(0)k+1 EoC(0) _ (k + 1)EoU' EoC(^) _ (k + 1) Eo£(0) Lemma 6 . 1 = PB(T < oo) = Eo0 exp 125 {(B .
yu/2) h(A.4. There are two reasons for this : in this way.(2A + ()1/21 exp S h(A.x. () .2 (^/2 + 3y9o + 390) + O(u3).S) d e 62 . Thus a2 y = 290 + O (u2). 5 exp { _h(A) (1 + / y u J)) exp 1. PROBABILITY OF RUIN IN FINITE TIME The last term is approximately (e 3 (3) 27.\+ (2 (3 e 2u [ (2. yu/2).6  d h(A. we get the correct asymptotic exponential decay parameter ^/ in the approximation ( 6.e h(aS)h (^^ 262 exp {_h(.2.\ + () 1 2 / .2) for O(u) (indeed . Thus by Taylor expansion around ( = 90u. yields +90 62 0 + O(u 3) 2u2 +O(u 3). letting formally T * oo yields 7/)(u) C'e7u where C' = e7a2).() I 1 + u2 ) y . yu/2) 11+ 62 I} S 1 \\\ u/11 l 62 (3 2u 2A Proof Use first (6. l Lemma 6 .() .6) and 7co (Oo) = ico('y + Bo) to get 0 = 21 (^/2 + 2y90) + 1112 (_Y3 + 3_Y200 + 3y9o) + O(u4). we get h(A. and inserting this and 9o 2 = S/u on the r. 0 The last step is to replace h(A.2u [2A+ (2 3 .h (A. 2 and (6. and the correction terms which need to be added cancels conveniently with some of the more complicated expressions in Lemma 6. C) ( 1+ u2 The result follows by combining Lemma 6 . 2 + 00 = .126 CHAPTER IV.1 (y/2 + Oo)u .s.7) and using eh(a.h. [2+ (2 .. () by h(\.
with the translation to risk processes being carried out by the author [12]. 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. Fuh [148] considers the closely related case of discrete time Markov additive processes. The adaptation to risk theory has not been carried out. 0 1 Proof of Proposition 6. We shall now generalize this question by asking what a sample path of the risk process looks like given it leads to ruin. ()} . the 'typical' value (say in sense of the conditional mean) was umL. .5 in Lemma 6. that is. HOW DOES RUIN OCCUR? exp { h (x.1: Just insert Lemma 6. () I 1 + u 2 ) } S 1 . the approach to the finite horizon case is in part different and uses local central limit theorems. The answer is similar: the process behaved as if it changed its whole distribution to FL. 7 How does ruin occur? We saw in Section 4 that given that ruin occurs.1 (y/2 + Oo)u )} 1 (i + U ) [2+ C2 2u 62 S Pt^ exP { J 62(2 exp { h(A. u Notes and references Corrected diffusion approximations were introduced by Siegmund [345] in a discrete random walk setting.7. i. Hogan [200] considered a variant of the corrected diffusion approximation which does not require exponential moments.(i+ 62 exP{ h(A.4. () (i+a ) 2A + (2 .e.T) has not been carried out and seems nontrivial. The corrected diffusion approximation was extended to the renewal model in Asmussen & Hojgaard [34]. ()} 3 h (A. and to the Markovmodulated model of Chapter VI in Asmussen [16]. 'yu/2) 127 ( i+ M pz^ exP { h (A. In Siegmund's book [346]. the analogous analysis of finite horizon ruin probabilities O(u. () I 1 + u2 )I 2u L 2A+C2_(2 exp { _h. His ideas were adapted by Asmussen & Binswanger [27] to derive approximations for the infinite horizon ruin probability 'i(u) when claims are heavytailed. the same as for the unconditional Lundberg process.
the Poisson rate changes from .TT(u) _measurable.r. .1.(u)_ is that i.1 Let {F(u)}u>0 be any family of events with F(u) E F.(u). Recall that . (u) and satifying PL(F(u)) * 1.T. {ST(u)+t . then P(u) and FL coincide on . Then also P(u)(F(u)) + 1.2 If B is exponential. ^(u) is exponential with rate 8 w. {St}0< t<T(u)) Proof Write e'rsr(u ) = e'rue'r£(u).3L and the claim size distribution from B to BL. F(u)c] ti e' ru]PL (F(u)`) > 0.EL[e7S. .128 CHAPTER IV.3 to . Note that basically the difference between FT(u) and . Theorem 7 . . so in the in the proof. FL As example.. P(u) and rate = aL w.F. In the exponential case.r. the numerator becomes e'ruELe7^ (u)PL(F( u)t) = e7uCFL (F(u)°) when F(u) E . PROBABILITY OF RUIN IN FINITE TIME changed its arrival rate from 0 to /3L and its claim size distribution from B to BL.FT(u) = o' (T(u ). stating roughly that under F(u).F.vi(u) Ce'Yu Corollary 7. and let M(u) be the index of the claim leading to ruin (thus T(u) = Ti + T2 + .(u)_ and ^(u) are independent . 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.t.t. u * oo. Recall that 13L = (3B[ry] and BL(dx) = e'rxB(dx)/B[7]. r(u) < oo) . In fact. Proof P(u) (F(u)c) = F(flu)c. + TMOO ).ST(u)}t> o is just an independent copy of {St}t>o).(u) is not ..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(u)c] P(r(u) < oo) ?P(U) < EL[e7u.(u)_ and similarly the denominator is exactly equal to Ce7u. we give a typical application of Theorem 7.
A somewhat similar study was carried out in the queueing setting by Anantharam [6]. who also treated the heavytailed case.3 M(u) pcu) 1 . . HOW DOES RUIN OCCUR? Corollary 7. Proof For the first assertion. This is currently a very active area of research. 129 M(u) >2 I(Tk < x) M(tu) p(u) M(u) >2 I(Uk < x) BL(x).(1 . see further XI. Notes and references The results of the present section are part of a more general study carried out by the author [11]. however.7. From a mathematical point of view.3.eALx) M(u) k=1 u The proof of the second is similar. take I(Tk < x) . the subject treated in this section leads into the area of large deviations theory.eaLx. the queueing results are of a somewhat different type because of the presence of reflection at 0.
This page is intentionally left blank .
d.. . We use much of the same notation as in Chapter I. with the same distribution A (say) for T2.. Then the arrival process is stationary which could be a reasonable assumption in many cases (for these and further basic facts from renewal theory.. and the one corresponding to T1 = s by 1/i8 (u). the one corresponding to the stationary case by 00)(u). {St} is the claim surplus process given by I. are i. In the socalled zerodelayed case.i. D'2.1. with Nt = # {n: Un <t} the number of arrivals before t. the Tn are independent.7).Chapter V Renewal arrivals 1 Introduction The basic assumption of this chapter states that the arrival epochs O'1. AA t*oo lim St = lim ESt t tioo t = p . . .(1..1 (T1 = a1). Thus the premium rate is 1.1 Define p = !µ. and M is the maximum of {St}. Proposition 1. of the risk process form a renewal process: letting Tn = Qn . The ruin probability corresponding to the zerodelayed case is denoted by 1/'(u).. ..Q.. A different important possibility is Al to be the stationary delay distribution A° with density A(x)/µA. U2. T3.Then no matter the distribution Al of T1i B.1). the claim sizes U1. with common distribution B. see A. r(u) the time to ruin. the distribution Al of T1 is A as well. Var(St) = 11Ba2A + I4AaB lim t goo t PA 131 .
For (1 .0 > 0. OFF. and (1 .St] = a(p .a is really the accumulated claims over a period u of length a. If the environment is Markovian with transition rate A from on to off and u from OFF to ON. The simplest case is of course the Poisson case where A and Al are both exponential with rate 0. This has a direct physical interpretation (a large portfolio with claims arising with small rates and independently). Nt ESt = E E UI Nt t = ENt•pB .2 (DETERMINISTIC ARRIVALS) If A is degenerate. stating that E[Nt+a . such that no arrivals occur in the off state. the . Proposition 1. However . RENEWAL ARRIVALS lim E [St+a . 3) follows similarly by Blackwell 's renewal theorem.t. s + t µA PA 0 Of course. Nt = Var(PBNt) + E(4Nt) Q2 2 0`2 A tpB B + o(t). say at a. by the elementary renewal theorem (cf. we get similarly by using known facts about ENt and Var Nt that Nt Var(St) = Var E Nt U.1 gives the desired interpretation of the constant p as the expected claims per unit time. t 4oo Proof Obviously.3 (SWITCHED POISSON ARRIVALS) Assume that the process has a random environment with two states ON. but the arrival rate in the ON state is . one could imagine that the claims are recorded only at discrete epochs (say each week or month) and thus each U.132 Furthermore for any a > 0. Here are two special cases of the renewal model with a similar direct interpretation: Example 1. Nt + EVar U. 2). From this ( 1.Nt] * a/PA.1). after E. The renewal model is often referedd to as the Sparre Andersen process. the definition 77 = 1/p . Sparre Andersen whose 1959 paper [7] was the first to treat renewal assumptions in risk theory in more depth. CHAPTER V.1 of the safety loading appears reasonable here as well. Thus.1) ENt/t + 1/µA. Example 1 .1) follows . A.
integrate (1..} with {S(d)} a discrete time random walk with increments distributed as the independent difference U . and then the whole process repeats itself). initial vector (1 0) and phase generator 11 However. (an arrival occurs necessarily in the ON state. S o<t<oo n=0. the first term represents the probability F(U1 .. For the stationary case. the fundamental connections to the theory of queues and random walks. and for historical reasons.i.2. 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 .1. Ao.y)B(dy).oFF}.t. the relevance of the model has been questioned repeatedly.r.4) w.4) with phase space {oN. (1. . we feel it reasonable to present at least some basic features of the model. Therefore. U1 . However.s > u) of ruin at the time s of the first claim whereas the second is P(r(u) < oo. A is phasetype (Example 1.1.4 The ruin probabilities for the zerodelayed case can be represented as 0(u) = P(M(d) > u) where M(d) = Max {Snd) : n = 0. 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^. if for nothing else then for the mathematical elegance of the subject.1. INTRODUCTION 133 interarrival times become i.4) fo Indeed. arrival times. in general the mechanism generating a renewal arrival process appears much harder to understand.s < u). More precisely. we have From this the result immediately follows.s.. and the present author agrees to a large extent to this criticism.d... The following representation of the ruin probability will be a basic vehicle for studying the ruin probabilities: Proposition 1. u For later use.. Proof The essence of the argument is that ruin can only occur at claim times. 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 .T between a claim U and an interarrival time T.
(2. U Figure 2. b=1 !=1 where {Nt } is a Poisson process with rate . A typical sample path of {Rt } is illustrated in Fig. the claims and the premium rate are negative so that the risk reserve process .1) +ry. If . St = t . are independent of {Nt} and i. At the time of death . 2.1. Theorem 2 . the claim surplus process are given by Nt Nt Rt = u+^U.0* (u) = P (rr* (u) < oo) where rr* (u) = inf It > 0: Rt < 0} ) .a*PB• > 1. each of which receive a payment at constant rate during the lifetime . i.1 r* (u) One situation where this model is argued to be relevant is life annuities.d.134 CHAPTER V. Using Lundberg conjugation . t. we shall be able to compute the ruin probabilities i(i* (u) for this model very quickly (. The initial reserve is obtained by prepayments from the policy holders.3* pB.3* (say ) and the U. then 0*(u) = e 'r" where ry > 0 is the unique solution of 0 = k*(ry) = *(B*[ry] . the remaining part of the prepayment (if any ) is made available to the company. RENEWAL ARRIVALS 2 Exponential claims. with common distribution B* (say) concentrated on (0. resp .1 If.Ut. then 0 * (u) = 1 for all u > 0. That is . < 1.1) . The compound Poisson model with negative claims We first consider a variant of the compound Poisson model obtained essentially by signreversion . A simple sample path comparison will then provide us with the ruin probabilities for the renewal model with exponential claim size distribution. 00).
g. Fig. B. the safety loading of { St} is > 0. Define T_ (u) = inf It > 0 : St = u} . 2.2.2 Assume now .3*.* (a) = log Ee'st I.2 sup St = inf St = 00 t>o t>o and hence 0* (u) = 1 follows.f. 0 Now return to the renewal model. Let B(dx) = ^e7x B*(dx). of {St} is c(a) = is*(a7). (a) is*(a) (b) . B*.s. Proof Define 135 St =u . T_ ( u) < 001 e7"P(T_ (u) < oo) = e"V)* (u).Rt. and the Lundberg conjugate of {St} is { St } and vice versa. then by Proposition 111. Then the c. If I3*pB* < 1. Since ic'(0) < 0.0. 0) and has typically the shape on Fig.1.(a) 7 Figure 2.(u ) < oo) = E {e7sr_ (u).2(b). EXPONENTIAL OR NEGATIVE CLAIMS [Note that r. 2. Hence y exists and is unique. St=Rtu=St.UB. T_ (u) = inf { t > 0 : St = u 'r* (u). Hence T_(u) < oo a. cf. > 1 . B* [7] and let {St} be a compound Poisson risk process with parameters .2(a). Then { St } is the claim surplus process of a standard compound Poisson risk process with parameters 0 *.. and thus 1 = P(T. . Then the function k* is defined on the whole of (oo.
.136 CHAPTER V.1 means that M* is exponentially distributed with rate ry.. 3* = 6.4 goes as follows: define 7r+ = P(M(d) > 0) and consider {St*} only when the process is at a maximum value.e...)(u) _ 1r+e7" where ry > 0 is the unique solution of 1 = Ee'Y(uT ) = S 8 A[. we get Ee'M(d) = Ee°M* _ Y/(..+Tn U1 Un. and (2 . RENEWAL ARRIVALS Theorem 2 . and hence the failure rate . with the probability that a particular jump time is not followed by any later maximum values being 1 . However. Hence M* max {To + Ti + • • • + Tn .u+ and lr+. the failure rate of this process is y. alternatively termination occurs at a jump time (having rate 8). and 5PA > 1.Ui .f. Now the value of {St*} just before the nth claim is To +T1* +.1.2). with rate S (say). Taking m.. 2. A variant of the last part of the proof.2 If B is exponential.Y] (2.1) + ry = 0 which is easily seen to be the same as (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 ..1.a) = 1 . then .7r+ 7r EeTo b/(Sa) + +.'s and noting that V)*(u) = P(M* > u) so that Theorem 2. According to Theorem 2. and from Fig .. u Hence P(M(d) > u) _ 1r+e'r".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 .4. respectively.• • • .. To + max {Ul+•••+UnTI. which has the advantage of avoiding transforms and leading up to the basic ideas of the study of the phasetype case in VIII..g.. 1) means that 8(A[ry] . T2 = U2..•. Then B* = A. To + M(d) in the notation of Proposition 1..1.1 it is seen that ruin is equivalent to one of these values being > u.Un } = max St = t>0 n=0.Ti = U1.Tn} n=0..Tr+.Y a I.
B^d) where Aad> (dt) = ^[ a] A(dt).. 3a The imbedded random walk The key steps have already been carried out in Corollary 11.7r+) and hence r+ = 1. : S(d) > u} .. which states that for a given a.2. the imbedded discrete time random walk and Markov additive processes. 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^.3. Hence the failure rate of M(') is 6(1 . resp. Thus a ladder step terminates at rate b and is followed by one more with probability 7r+.4.5.3 A[a] B[a] F( d) [a +)3] F(d) [a] = Fad) [^] Letting M(u) = inf in = 1. letting P(d) refer to the renewal risk model with these changed parameters .B(dx). CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 137 is b(1. Bads (dx) = .y/b.6. 0 3 Change of measure via exponential families We shall discuss two points of view. hence exponential with rate b.. Furthermore.7r+). we have ] A[a )3] E«d'efl' = Bad> [a] A ad> [Q] = B[a +.7r+) = ry and hence P(M(d) > u) = P(M(d) > 0)e7u = 7r+e'r". the relevant exponential change of measure corresponds to changing the distribution F(d) of Y = U . we see that ry = 6(1. This follow since. 111.2. consider instead the failure rate of M(d) and decompose M(d) into ladder steps as in II. a ladder step is the overshoot of a claim size. The probability that the first ladder step is finite is 7r+. However.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] .. Putting this equal to y.
.T is nonlattice. the evaluation of C is at the same level of difficulty as the evaluation of i/i(u) in matrixexponential form. 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 .C(°)eryu where C(O) = C0[7] ..(u) . 7µA . VIII.1 implies Cu) = e«uE ( 7d)e«^(u) . For the stationary case. ik.u the overshoot . let 7 > 0 be the solution of r. RENEWAL ARRIVALS be the number of claims leading to ruin and ^(u) u = SM(u) . 187) and thereby for ^(u) to be nonlattice w. we get: Proposition 3.2 p. We have the following versions of Lundberg' s inequality and the CramerLundberg approximation: Theorem 3 . 00)(u) . O(u) = eauE (d)ea{ (u)+M(u)K (d)(a) .r. i . (d) (7) _ 0.138 CHAPTER V.1 For any a such that k(d)' (a) > 0..1) is explicit given 7. in the easiest nonexponential case where B is phasetype.2 In the zerodelayed case.1). Consider now the Lundberg case.t. (a) '(u) < eryu. to converge in distribution since p(yd) (r(0) < oo) = 1 because of r (d)' (y) > 0.4.C8e7u where Cs = Ce78B[7]. For claim (b). and claim (a) follows immediately from this and e (u) > 0.p)/($B'[7] . In fact. Corollary 3. just note that F7(d) is nonlattice when F is so . cf.Ce"u where C = limu. Proof Proposition 3. This is known to be sufficient for ^(O) ]p (d) ([APQ] Proposition 3. (b) V)(u) . E(d)e 1' (u).e. provided the distribution F of U .3 For the delayed case Tl = s.
5. 0) = ah (s) . 3b Markov additive processes We take the Markov additive point of view of II. B(x) = o(e7x) and dominated convergence.(°) ( Ce8B[7] Ao(ds) similar manner. The underlying Markov process {Jt} for the Markov additive process {Xt} = {(Jt..dt(ah ( s) + h'(s)) so that Gha ( s.4)./c. we look for a function h(s) and a k (both depending on a) such that Gh. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES Proof Using (1. According to Remark 11. h(s) = e(a +x( a))8 (3. where G is the infinitesimal generator of {Xt} = {(Jt.5.(s.dt ) eadt = h ( s) .y) B(dy) 0 For the stationary case..1) (normalizing by h(0) = 1). To determine boundary 0. E8h0 (Jdt.9.0 ) = Eo[ha ( Jdt. we invoke the behavior at the 1 = h«(0.h'(s). The expressions are slightly more complicated and we omit the details. 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] . y) = e°yh(s). Equating this to tch(s) and dividing by h(s) yields h'(s)/h(s) _ . (s.1) = C(O).3. Let P8f E8 refer to the case Jo = s. delayed version of Lundberg's inequality can be obtained in a e7u. Sdt) = h(s . (u + s . 0 0 . 0) = tc(a)h(s). we get r u +8 e"8(u) 139 e7uB(u + s) + 4 0 + L 00 J e7(v8)e7(u+8v). For s > 0. IPA 0 Of course. St)} and h. Here K.Sdt] = Ee'uh(T) means 1 = f ' e°^B(dy) f ' h( s)A(ds).a .St)} can be defined by taking Jt as the residual time until the next arrival.
. RENEWAL ARRIVALS B[a]A[a . A[a . Proposition 3.rc(a)] = B = Ba[13]Aa[5]. T2.13]A[b . Remark 3 . St)}too by letting the likelihood ratio Lt restricted to Yt = a((J. [e1U1 + 6T2ea ( U1s)stc ( a)e(a+K(a ))( T2s)I B[a +.e(«+k(a))t esy A(dt). .2) As in 11. (3..140 CHAPTER V. Further.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. we can now for each a define a new probability measure Pa. however. . resp. . An easy extension of the argument shows that U1. since JT. rc(a) = 0 (B[a] .. .a . (3. ] = E. [a + /3] A[b . U. Aa as asserted .2).rc(a)] B[a] A[a . An important exception is. . i. Note that the changed distributions of A and B are in general not the same for Pa.4 The probability measure Pa.( a+r' (a))(Jt s) h(s) where c(a) is the solution of (3. J n+1 u are independent with distribution Aa for the Tk and Ba for the Uk. Ba(dx) = B(dx). = J8 = T2. T2 are independent with distributions Ba. Ease AU1+6T2 [ AU1+6T2 = Ea a LT. .1)a in agreement u with Chapter III.8.tK( a)e. Ba where Aa (dt) .e..rc(a)] = 1.s and P(d). the determination of the adjustment coefficient ry where the defining equations rc(d) (ry) = 0 and rc(ry) = 0 are the same.s governing {(Jt.2) means 1 = B[a]/3/(/3+a+rc (a)).c(a)] which shows that U1.a .. 5 For the compound Poisson case where A is exponential with rate .s(Jo = s) = 1 follows trivially from Lo = 1. resp..c(a)] B[a] Proof Pa.5.S„):0<v< )be Lt = eaSt tK(a) h(Jt) = east .
t. .y yuAa y [ay + K(ay) .(u. and assume that T„ is the time between the arrivals of customers n .. see in particular Dassios & Embrechts [98] and Asmussen & Rubinstein [45].. u The approach via the embedded random walk is standard.1 and n. Let M(u) be the number of claims leading to ruin . XII. The virtual waiting time Vt at time t is the residual amount of work at time t. yu ) < e7yu. that is. which is the same as the asserted inequality for 0.rc( ay)] = e(aa+(r ))sb[a ]e7yu L y1 In particular. yu) = F'ay. or FCFS = first come first served) queueing discipline and renewal interarrival times.5. . [APQ] Ch. defined as the single server queue with first in first out (FIFO.yx(ay). Then "^ e(ay+w(aY))8 Ys(u. The claim for the zerodelayed case follows by integration w.4.6 Let y < let ay > 0 be the solution of ic'(ay) = 1/y. T(u) < yu h(JT(u)) < eayu+yuk(ay ) ( Eia y Le(a(+k(ay))s v.yu ) e7vu A[ay . Using the Markov additive approach yields for example the following analogue of Theorem IV. it is easily seen that ic(ay ) > 0.4. not after time T. The actual waiting time Wn of customer n is defined as his time spent in queue (excluding the service time). Then J(rr(u)) = TM(u)+1 and hence Ws(u. the amount of . the time from he arrives to the queue till he starts service. Notes and references 4 The duality with queueing theory We first review some basic facts about the GI/G/1 queue. see e. A(ds). 2. say finite horizon ruin probabilities where the approach via the imbedded random walk yields results on the probability of ruin after N claims. . yu).g.s eaysr(")+r(u ) K(ay) h (s) .4.5 Proposition 3.)+1 e J j e(ay+w(ay))8 e .ay+ray))TM(. and U„ the service time of customer n. 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. that is. Label the customers 1. For the approach via Markov additive processes. Proof As in the proof of Theorem IV. for the zerodelayed case zp8(u.r. and define yy = ay .
Thus. (4. and we have P(W > u) = V.. (4. Thus Vos}1 _ = (Wn + Un .2) (b) as t * oo. but interchanging the set . then Wn v M.142 CHAPTER V. equivalently.. the proposition follows. equivalently.1. It then jumps to VQ„ . the waiting time a customer would have if he arrived at time t. Then: (a) as n + oo.3 Assume rl > 0 or. Also {Zt}o<t<T evolves like the leftcontinuous version of the virtual waiting time process up to just before the Nth arrival.1 and Corollary 11. in which case {V} remains at zero until time on+1. Then P(r(u) < T) is the probability z/iiNi (u) of ruin after at most N claims. and we have P(V > u) = ?/iiol(u).. p < 1. since customer n arrives at time on.Tn)+."^ Vi(N) (u).4. Let the T there be the random time UN. 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. (u).• • • Tk ).Tn)+ Proof The amount of residual work just before customer n arrives is VQ„ . and combining with (4. The traffic intensity of the queue is p = EU/ET. we get: Corollary 4. an+1) = [on. 0 Applying Theorem 11. whereas in [On . If W1 = 0. Vt converges in distribution to a random variable V. we have Wn = Van(left limit). RENEWAL ARRIVALS time the server will have to work until the system is empty provided no new customers arrive (for this reason often the term workload process is used) or.3. . (4.1).1 Wn+1 = (Wn + U. and obviously z/'(u) = limN.1) The following result shows that {Wn} is a Lindley process in the sense of II. on + Tn) the residual work decreases linearly until possibly zero is hit.4.n1 (U1 +• • •+Uk Tl . but we shall present a slightly different proof via the duality result given in Theorem II.2 Let Mnd) = maxk=o.4: Proposition 4.+ Un.4.3) Proof Part (a) is contained in Theorem 11. Wn converges i n distribution to a random variable W..1.4): Proposition 4.2.
. Then the arrivals of {Rt} in [0. Then K(x) = J x00K(x .. resp .T* = y yields K(x) = P ((W + U* . u Now return to the Poisson case . hence (since the residual lifetime at 0 and the age at T have the same distribution .oo in Proposition 4. and hence in particular ZT is distributed as the virtual waiting time just before the Nth arrival. i.. Letting n oo in Corollary 4. Hence for x > 0... Then the corresponding queue is M/G/1. by an obvious reversibility argument this does not affect the distribution . K(x) = P(W < x). Ti) and similarly for the U.2. x > 0. where U*. In fact . we obtain: Corollary 4. However.. namely W1 = 0 in (a) and Vo = 0. and we get: . (4. as WN.5 (LINDLEY'S INTEGRAL EQUATION) Let F(x) = P(U1T1 < x). It follows that P(WN > u) =. { 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 . A.T* are independent and distributed as U1.2). (4.Ao in (b).T*)+ < x) = P(W + U* . but this requires some additional arguments (involving regeneration at 0 but not difficult) that we omit.(N)(u) has the limit tp(u) for all u.1.T* < x) fK(x_y)F(dy) (x > 0 is crucial for the second equality!).y)F(dy). Corollary 4. cf.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.le) the same is true for the timereversed point process which is the interarrival process for { Zt}o<t < T• Thus as before .T*)+. conditioning upon U* . we get W = (W + U* . THE DUALITY WITH QUEUEING THEORY 143 (T1..4. T1 .5) Proof Letting n . TN) with (TN.e.. For part (b). T] form a stationary renewal process with interarrival distribution A.4 The steadystate actual waiting time W has the same distribution as M(d). we let T be deterministic .. T1.. which implies the convergence in distribution and (4.
Proof For the Poisson case.6 For the M/G/1 queue with p < 1.5) is in fact a homogeneous WienerHopf equation.144 CHAPTER V. Asmussen [24] and references there. VIII). 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. Some early classical papers are Smith [350] and Lindley [246].5) looks like the convolution equation K = F * K but is not the same (one would need (4. The equation (4. see e.5) to hold for all x E R and not just x > 0). W v V. RENEWAL ARRIVALS Corollary 4. That is. the actual and the virtual waiting time have the same distribution in the steady state. Cohen [88] or [APQ] Ch. Note that (4. the zerodelayed and the stationary renewal processes are identical. 0 Notes and references The GI/G/1 queue is a favourite of almost any queueing book (see e .g. implying P(W > u) = P(V > u) for all u. Hence '(u) = Ali(°)(u). .g.
i=1 0 and r(u) = inf It > 0: St > u}. here it exists whenever A is irreducible which is assumed throughout. Thus.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 . The intensity matrix governing {Jt} is denoted by A = (A. {St} denotes the claim surplus process. 145 Oj( u.(3i when Jt = i.)iJEE and its stationary limiting distribution by lr. and can be computed as the positive solution of WA = 0.. M = supt>o St.f pi. Ire = 1. dv. • The premium rate when Jt = i is pi. .T) = Pi (T(u) < T). {Jt} describes the environmental conditions for the risk process. As in Chapter I. The ruin probabilities with initial environment i are '+ki(u) = pi(T(u ) < oo) = Pi (M > u). t St = E Ui . • Claims arriving when Jt = i have distribution Bi. N.
P = E 7riPi. we can approximate A(i) with a phasetype distribution (cf.1) iEE Then pi is the average amount of claims received per unit time when the environment is in state i.11 below. and we have f3. the operational time argument given in Example 1.T(n)). Example 1. 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. t(i) = T(')e are the exit rates. cf. say.1 implicitly assumes that the sojourn times of the environment in the normal and the icy states are exponential. Proposition 1. meaning that accidents occuring during icy road conditions lead to claim amounts which are different from the normal ones. Example 1 . Thus. An example of how such a mechanism could be relevant in risk theory follows. and p is the overall average amount of claims per unit time.2 (ALTERNATING RENEWAL ENVIRONMENT) The model of Example 1. leading to E having two states n. = iii when j E E(i). a(i). which is clearly unrealistic. u . We let p Pi = /ji/AB. and assume that weather conditions play a major role for the occurence of accidents. According to Theorem A5. we shall assume that pi = 1. f3i and claim size distributions Bn.a('). i and corresponding arrival intensities Qn. we could distinguish between normal and icy road conditions.146 CHAPTER VI. this is no restriction when studying infinite horizon ruin probabilities. MARKOVIAN ENVIRONMENT where as usual Pi refers to the case Jo = i.4) with representation (E(i). Bi. Assume similarly that the sojourn time in the normal state has distribution A(n) which we approximate with a phasetype distribution with representation (E('). r^ = P (1. one expects that 3i > on and presumably also that Bn # Bi. T(=)). in blockpartitioned form.2. respectively. say. /3 = Nn when j E E(n). the intensity matrix is A OW) T(i) T(n) t(n)a(i) where t(n) = T(n)e. Unless otherwise stated.5 below.1 Consider car insurance. assume that the sojourn time in the icy state has a more general distribution A(i).14. For example. with rates Ani and Ain. cf. Then the state space for the environment is the disjoint union of E(n) and E(i)..
St = SB=(t).j = . it = Jel(t). 4 (SEMIMARKOVIAN ENVIRONMENT) Dependence between the length of an icy period and the following normal one (and vice versa) can be modelled by semiMarkov structure. t(n) = T("i)e. 1 . A('^). u Example 1 . the parameters are ^ij = aid/pi. In the car insurance example.tEH is a transition matrix. This amounts to a family (A(")) ?CH Of sojourn time distributions. and similarly for the normal period. Approximating each A('?) by a phasetype distribution with representation (E('l). i E E(n) }. we assume again pi = 1 so that the claim surplus is Nt St = ?Ui_t. i8f n1.. such that a sojourn time of type rt is followed by one of type c w. where W = (w.p.5 (MARKOVMODULATED PREMIUMS) Returning for a short while to the case of general premium rates pi depending on the environment i.. such that the icy period is of two types (long and short) each with their sojourn time distribution A('L). one could for example have H = {i1.. resp.... say. (9) where q = CHI.. i ) : n E H. One way to model this would be to take A(') to be Coxian (cf. depending only on 77.3i/pi. but assume now that the arrival intensity changes during the icy period..3i.1. and 1/ii(u) = t/ii(u). .. . The simplest model for the arrival intensity amounts to . n8}... Indeed.n.3 Consider again the alternating renewal model for car insurance in Example 1. u Example 1.3.2. iq (visited in that order) and letfOil >. is the probability that a long icy period is followed by a short normal one. w. T(9) +wggt(9)0. the state space E for the environment is { ('q. MODEL AND EXAMPLES 147 Example 1 .4) with states i1. Then for example wi.Q. let T 9(T) = f pi. T(1) +w11t(1)a(1) w12t (1)a(2) w21t(2)a(1) w1gt(1)a(9) w2gt ( 2)a(q) T(2) +w22t( 2)a(2) A = wg1t(9)a(1) wg2t(9)a(2) . say it is larger initially.1. 0 Then (by standard operational time arguments) {St } is a risk process in a Markovian environment with unit premium rate. and . u From now on.T(n)). dt.J017.a(n). Example VIII. Qi = ..>.
Pi (Ti E dx. )3*.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.e(A(Oi)d'sg)xe. The key property for much of the analysis presented below is the following immediate observation: Proposition 1. t l=1 Note that the last statement of the proposition just means that in the limit. we put )3* = E 7fi/3i. . MARKOVIAN ENVIRONMENT We now turn to some more mathematically oriented basic discussion. qij = 0 in the notation of Chapter 11.7 The Pidistribution of T1 is phasetype with representation (ei. B* = 1 /^* Bi. iEE iEE )3 These parameters are the ones which the statistician would estimate if he ignored the presence of Markovmodulation: Proposition 1.148 CHAPTER VI. one can associate in a natural way a standard Poisson one by averaging over the environment. the Markov additive structure will be used for exponential change of measure and thereby versions of Lundberg's inequality and the CramerLundberg approximation. dx. the empirical distribution of the claims is B*. N > 1(Ul < x) a4 B*(x). More precisely. . Next we note a semiMarkov structure of the arrival process: Proposition 1. 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. Nt Nt a . 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.5. JT1 = j) = Qj • e.(3iBi(dx). In particular. 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).A .(Qi)diag)• More precisely. o = 0.8 As t oo. vi(dx) = .
the limiting distribution of the first claim size U1 is B*. i.x) = Nt E > I (Uk) X) Nt Bi(x) 1=1 iEE k=1 iEE 1: t5 Bi( x) = B*(x). 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$. cf. y Ni) i Nti) t a.9 Consider a Markovmodulated risk process {St} with param eters Ni. given {Jt}0<t<0. Bi(x). In particular.* (u) for all u.. 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. Nt a' t t iEE Also. A. B*.4. ^j 7riNi. oo) as a 4 oo.aA .. and furthermore in the limit JT. Bi. this converges to the exponential distribution with rate 0* as a * oo. 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 . Hence Nt'> a .7. and let {St °i} refer to the one with parameters Pi. Then {St)} + {St*} in D[0. has distribution (7ri)3i //3*)iEE and is independent of Ti. e.2.. iEE 13 A different interpretation of B* is as the Palm distribution of the claim size. Then it is standard that ti lt '4' iri as t > oo. Proof According to Proposition 1. By Proposition A5. zli( (u) .(/3i)aiag).6.. In particular. the Fidistribution of T1 in {St(a ) } is phase type with representation (E. N + oo Hence 1 Nt 1 N`+) Nits Nt E I ( Ut <. aA. Bi. {St} to the compound Poisson model with parameters 0 *. However . Conditioning .1. Example 11. Proposition 1. we may view Nt`i as the number of events in a Poisson process where the accumulated intensity at time t is Niti..
and in fact P1 = 31AB1 = 9 3 1 2 (5 3 3 1 1 2 1 5 7 1 81 70 ' _ 19 4 5 P2 = ..=1.. marked by thin.s = 1o in state 2. 0 Example 1. shows similarly that in the limit (T2. lines in the path of {St}.2}..g. 1. those of type E7 with intensity z s = 5 in state 1 and with intensity z . with T2 being exponential with rate . from Theorem 3.2.1 of [145]. we may imagine that we have two types of claims such that the claim size distributions are E3 and E7. U. which also yield O(a) (u..).s = o in state 1 and with intensity 1 . the paths of the surplus process will exhibit the type of behaviour in Fig. B2=1E3+4E7.. oo). since E3 is a more dangerous claim size distribution than E7 (the mean is larger and the tail is heavier).. B1=3E3+2E7.10 Let E_{1. resp. T) + ?P* (u. Claims of type E3 arrive with intensity 2 . 5 5 where E5 denotes the exponential distribution with intensity parameter 5 and a > 0 is arbitrary. and (at least when a is small such that state changes of the environment are infrequent). From this the convergence in distribution follows by general facts on weak convergence in D[0.2 +2 2 = 3. is as in {St }. On Fig. state 1 appears as more dangerous than state 2.l3* and U2 having distribution B*. e.1. T) for all u and T.. Continuing in this manner shows that the limiting distribution of (T. MARKOVIAN ENVIRONMENT upon FT. 9 . thick.FT. > 1.150 CHAPTER VI. U2) are independent of . The fact that indeed 0(a) (U) 3 0* (u) follows..31µB 2 = 2 5 3 7 70 Thus in state 1 where p. 1. Computing the parameters of the averaged compound Poisson model. s 5 in state 2. Thus. That is. the overall drift is negative since it = (2 2) so that p = 71P1 + 112P2 = 7.. the company even suffers an average loss. we first get that 3 (3* = 2.1 with periods with positive drift alternating with periods with negative drift. there are p = 2 background states of {Jt}.2. 132=2. A= (  a a ) \ a a 5 5 J 9 3 2 a1=2.
01 /.s. note first that EN Uk')/N a4' µgi. 0 The definition (1.1 a..1) of the safety loading is (as for the renewal model in Chapter V) based upon an asymptotic consideration given by the following result: Proposition 1. For (b). Proof In the notation of Proposition 1.11 (a ) ESt/t * p . a fraction r.1 Thus.3* = 3/4 of the claims occur in state 1 and the remaining fraction 1/4 in state 2. (b) St/t * p .1). Hence B* = 415E3+5E7/ +4 ( 51E3 +5 E7) = 1E 3 +2E7.1. 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). iEE . t * oo. = P.8.(3. Hence (i) Nti) 1 U(i) k' N(i)k=1 E t 4 St + t = iEE Nt t 1: 7ri Qi µs. the averaged compound Poisson model is the same as in III. MODEL AND EXAMPLES 151 Figure 1. That is.1. t + oo.
and so on.a form a renewal process . X3. The case 77 > 0 is similarly easy. If 77 > 0. and a more comprehensive treatment in Asmussen [16]. Theorem II.4. and hence M = 00. and hence wn /n a4.1 of St / t is > 0. [315]. having the Pidistribution of X. n n Thus {SWn l is a discrete time random walk with mean zero. then M = 00 a. also + . [212]. See Meier [258] and Ryden [314]. [APQ]. Now let r) = 0. with some important improvements being obtained in Asmussen [17] in the queueing setting and being implemented numerically in Asmussen & Rolski [43].Jt=i}. with X2. The proof of Proposition 1. Eiw. .1 and the Corollary are standard.Eiw o'o Eiw • E ^ifjµs.. The mainstream of the present chapter follows [16]. MARKOVIAN ENVIRONMENT Corollary 1.. X2 =SW2 So. some early studies are in Janssen & Reinhard [211]. There seems still to be more to be done in this area. 38) Eiw1 = 1/ir. X 1 =Sty. In risk theory. s.2(a) p. and involves a version of the .0i(u) < 1 for all i and u. Proof The case 77 < 0 is trivial since then the a. + Xn SWn ](1 a .s... then M < oo a.152 CHAPTER VI..1)Eiw = 0.. and hence 1/ii(u ) = 1 for all i and u. 0 Notes and references The Markovmodulated Poisson process has become very popular in queueing theory during the last decade.\ i and EiX1 Ei f 13 J. limit p . EiX = 0.1 jEE = (p .. PB.Jt=i}. [302]. Then by standard Markov process formulas (e. w2=inf {t>w1:Jt_#i. see [APQ] p. dt . Since the X„ are independent . let some state i be fixed and define w=wl=inf{t >0:Jt_#i. 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 . .. Proposition 1.ld.1(b) is essentially the same as the proof of the strong law of large numbers for cumulative processes. Now obviously the w.s. Statistical aspects are not treated here.. and hence oscillates between 0o and oo so that also here M = oo.12 If 77 < 0.g.. 136 or A. see the Notes to Section 7.
6. e. the definition of .j E E. •) * G +(k.(u) = Pi(M < u) = e' E G+ (u)(I . we get the same ladder height distribution as for the averaged compound Poisson model. let G+(i. •). n=0 (2. 00 (2.Jt=j)dt. j.2(a) below ) where the ladder height distribution is evaluated by a time reversion argument.2 (a) The distribution of M is given by 00 1 .j. but is substantially more involved than for the compound Poisson case .5.IIG +II)e. . j. see also Example II. B* in Section 1. oo)). we define the convolution operation by the same rule as for multiplication of realvalued matrices. j.dx). j.x. G+ is the matrix whose ijth element is E G +(i. cf.EA. 6. T+ < oo) and let G+ be the measurevalued matrix with ijth element G+(i.x).i. oo) = J ao 0 G+(i. Thus. k.3*B *(y)dy. For measurevalued matrices. oo)) = f R(i. Let further R denote the preT+ occupation kernel.g. The form of G+ turns out to be explicit (or at least computable). dx)/jBj(y . by specializing results for general stationary risk processes (Theorem II . for i. (y. •)• kEE Also. IIG+ II denotes the matrix with ijth element IIG+(i. That is.2) R(dx)S((y . •) II = JG+(i. Define the ladder epoch T+ by T+ = inf It : St > 0} = r(0). Proposition 2.Jr+ =j.1 irG+(dy)e =. and S (dx) the measure valued diagonal matrix with /3 Bj(dx) as ith diagonal element. However . T R(i. THE LADDER HEIGHT DISTRIBUTION 153 PollaczeckKhinchine formula (see Proposition 2.a/i.j..A) =ZI(St E.1) 0 (b) G+ (y.6*.2. only with the product of real numbers replaced by convolution of measures. which represents a nice simplified form of the ladder height distribution G+ when taking certain averages : starting {Jt} stationary. j. Proposition 2.4) we obtain the following result .A) = Pt(ST+ E A.
1 for an illustration in the case of p = 2 environmental states of {Jt}. 0  x Figure 2. JJ = j.IIG+II)e. From this (2.6. The u proof of (2. we need as in Chapters II. . MARKOVIAN ENVIRONMENT Proof The probability that there are n proper ladder steps not exceeding x and (x)ej. resp. lines in the path of {St}. the intensity matrix A* has ijth element * 7r ^i3 7ri and we have Pi(JT = j) = 7rj P2(JT = i)7ri (2. III to bring R and G+ on a more explicit form . thick. {mx} is a non terminating Markov process on E. marked by thin. and let further {my} be the Evalued process obtained by observing {Jt } only when {St*} is at a minimum value.2 useful .2) is just the same as the proof of Lemma 11.3 When q > 0. G+ the probability that there are no further ladder steps starting from environment j is e^ ( I . St < S* for u < t.154 CHAPTER VI.1 The following observation is immediate: Proposition 2. and that the environment is j at the nth when we start from i is e . To this end . see Figure 2. hence uniquely specified by its intensity matrix Q (say). we need to invoke the timereversed version {Jt } of {Jt} .3.1) follows by summing over n and j.3) We let {St*} be defined as {St}. only with {Jt} replaced by {Jt } (the /3i and Bi are the same ). To make Proposition 2. That is. mx = j when for some (necessarily unique) t we have St = x.
0 mms1   ^O \ T. the sequence {Q(n)} A* defined by Q(O) = .2 .. = x}. Q( n+l) _ ^. The definitions are illustrated on Fig.(/3i) diag. In general. Note that the integral in the definition of W(Q) is the matrix whose ith row is the ith row of _ 3 f e2Bi(dx).0.4 Q satisfies the nonlinear matrix equation Q = W(Q) where 0 co(Q) = n* .(/3i)diag + T S(dx) eQx.and a jump (claim arrival) occurs at time t. {S. For example the excursion of depth 2 has one subexcursion which is of depth 1. and the excursion is said to have depth 1 if each of these subexcursions have depth 0. } is a minimum value at v = t. and the excursion ends at time s = inf {v > t : S. Furthermore.*. If there are no jumps in (t. s]. Proof The argument relies on an interpretation in terms of excursions.2. we say that the excursion has depth 0.2 where there are three excursions of depth 1. An excursion of {St*} above level x starts at time t if St = x.2. Figure 2. and S(dx) is the diagonal matrix with the f3iBi(dx) on the diagonal. we recursively define the depth of an excursion as 1 plus the maximal depth of a subexcursion. corresponding to two subexcursions of depth 0. 2. ( Q( n)) converges monotonically to Q. Otherwise each jump at a minimum level during the excursion starts a subexcursion. THE LADDER HEIGHT DISTRIBUTION 155 Proposition 2.
Theorem 2 . It is clear that { mini } is a terminating Markov process and that { mio) } has subintensity matrix A* . Writing out in matrix notation . of the definition to make U be concentrated on (co. By considering minimum values within the excursion. The proof of Q = W(Q) then immediately carries over to show that the subintensity matrix of {mil) } is cp (Q(o)) = Q(l).(01)diag = Q. A) = L' U(j.156 CHAPTER VI. either due to a jump of {Jt } which occurs with intensity A= j. A). StEA .. Now let {m ( n) } be {mx } killed at the first time i7n (say) a subexcursion of depth at least n occurs . we first compute qij for i $ j. 7rE Proof We shall show that Fi(Jt=j. Then a jump to j (i. i.j +/3ipij. or through an arrival starting an excursion terminating with J. = j. h.5 R(i.Qi + )%pij) Now just note that t pij and insert (2. 0)). it becomes clear that pij = r [eQh] 0 ij Bi (dy) • (2.4) To show Q = cp(Q). (2.St EA. mx+dx = j) occurs in two ways . p1^) Define a further kernel U by f U(i. A) = f Pi(mx = j) dx eie4xej dx A u (2. Q = W(Q) follows. the subintensity matrix of {min+i ) } is cp (Q(n)) = Q(n +l) which implies that qgj +1) = \!. Similarly.s. It follows that qij = A.4).6) .. e. j.5) A (note that we use A = {x : x E Al on the r.St <S*.T+>t) _ ^iF 7ri (JJ =i. Suppose mx = i.u< t). 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. Similarly by induction .j. Fi(mh =i ) = 1 + =hflh+Qihpii+o(h) implies qii = 'iii /i +)3ipii.
4].t.7 It is instructive to see how Proposition 2. St E A. and get irPi(Jt =j.6 is hardly all that explicit in general. G+((z.Qi)diag.0<u<t. Jt = i. and to obtain a simple solution in the .Jo=i. K( n (d) the sequence converges monotonically to K. and we let k be the corresponding right eigenvector normalized by Irk = 1.g. St < St U. 0 < u < t) = 7rjPj(Jt =i. the CramerLundberg approximation (Section 3).z+>t) = P. St EA.0<u<t) = P..6(b): from 7rK = 0 we get 7rG+(dy)e = J W 7reKx(fiiBi(dy + x))diag dx • e 0 w(fiiB1(dy + x))col dx f 0 EirifiiBi(y)dy = fi*B*(y)dy• iEE 0 Though maybe Corollary 2.St EA.. We may then assume Ju=Jtu. THE LADDER HEIGHT DISTRIBUTION 157 from which the result immediately follows by integrating from 0 to oo w. we shall see that nevertheless we have enough information to derive. oo))dx. (b) for z > 0.. {Jt }. `` {K(n)} [the W(•) here is of course not the same as in Proposition 2. 0 +1) = cp (K( n)) defined by K(o) = A .. dt.r. u It is convenient at this stage to rewrite the above results in terms of the matrix K = 0'Q'A. 0<u<t). where A is the diagonal matrix with 7r on the diagonal: Corollary 2. (c) the matrix K satisfies the nonlinear matrix equation K = W(K) where W( K) = A ( i) diag + fi J "O eKx S(dx).S„<0. S.St <Su.6). oo)) = f o' eIXS((x + z. x < 0. it is readily checked that 7r is a left eigenvector of K corresponding to the eigenvalue 0 (when p < 1).1 can be rederived using the more detailed form of G+ in Corollary 2. and this immediately yields (2.(.. consider stationary versions of {Jt}.=StSt.StEA. (Jo = j.(Jt=j.6 (a) R(dx) = eKxdx. Remark 2. To this end.2. From Qe = 0. e.
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
r(u) yu o)(y)eavuEav. if y > 1lk'(ry). we let G+ * W(u) be the vector with ith component E(G+(i. oo)) and. However.7. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 165 hiav)e _avuE. (3. r(u) < yu] hiay)C+ h=av)C+ o) (y)eayu+yuw(av).j) * coj)(u) _ f u ^Pj(u . exp {e() + r(u))} .i [e*(u)K(av).i [eT(u)K(av ). we shall need the matrices G+ and R of Section 2..8 ) Then for all i E E and all u > 0. yu) f h(av)e v avuE«v.00 su e7( ( 3. hj P . 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 (3.11 Let Bj (x) C_ = min 1 • inf jEE hj x>o f2° e'r( vx)Bj(dy) ' C+ _ mE 1 Bj(x) J Y x)Bj (dy). Our next objective is to improve upon the constant in front of a7u in Lundberg's inequality as well as to supplement with a lower bound: Theorem 3.5). 1 Similarly.V)i(u. Chie ryu < Vi(u ) < C+hie 7u.5) will produce the maximal ryy for which the argument works.3. we have ic(ay) < 0 and get 'i(u) . (u.(ay)}. for a vector <p(u) = (cpi (u))iEE of functions . as in the classical case (3. We further write G(u) for the vector with ith component Gi(u) = EiEE G+(i. av 'i [h.9) For the proof.y)G+(z.j. yu < r(u) < 00 h 4(u) < h(av)C+o)(y)eavuEav .i I (a) exp {aye(u) + r(u)r. dy)• o iEE jEE .
u IMP:°) (u) I < oo. dy) = aj f Bj(dy . 0 G+(i. 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.12 Assume sup1.j.13 For all i and u.x ) R(i.(0) ] (u) < sup Jt t.x) jEE 00 u 0 //^^ C+E. dx) 100 C .ery(&u+x)Bj (dy) Bj(u Bj (u .166 CHAPTER VI.u Iv 2°)(u)I Pi(rr+(N + 1) < oo) + 0.dy).x) x) jEE 0 E Qj f jEE R(i. Then cpin)(u) sit (u) as n + oo.j. we have G *(N +1) * ^. j. jEE u 0 j. n > oo. Proof Write UN = EN G+ .7. dx).3jhj // f 00 R(i. U = U". j. MARKOVIAN ENVIRONMENT Lemma 3 . °O . Then iterating the defining equation ip(n+1) = G + G+ * V(n) we get W(N+1) = UN * G + G+N+1) * ^(o) However. dy) 00 C+ ijhj f R(i. = Eo G+ G. Hence lim cp(n) exists and equals U * G. _ To see that the ith component of U * G(u) equals ?Pi (u). dy) : 1(u) < C+ > hj u e(1tL)G+(i. if r+ (n) is the nth ladder epoch. 00 f C_ hj f e(Y)G+(i. j.& (u). 00 Thus C+ > hj f"o e7(Yu)G +(i. and define W(n+1) (u) = G(u) + (G+ * tp(n))(u). dx ) Bj (u . j.x ) = Gi(u). dx) f e7( vu)Bj (dy . Lemma 3 .
3. and let 8 = e'(70).13.8) with y replaced by yo and hi by h=7o ). 167 u Proof of Theorem 3.MT<u.11). and using Lemma 3 . (3. j. we get Wo n +1) (u) = ? 7 i ( U ) + E J u gyp. +i . 14 Let yo > 0 be the solution of 'c'(yo ) = 0. jEE estimating the first term in (3.n) ( u .u)G+(i. dy) (3. ST < u] < C+(yo)e7ouEi [h^7o)e70ST1 l T J = C h(7o)e7ou8T .ST). this is obvious if n = 0. and the proof of the lower one is similar. 9 for the last equality in (3.11) C_e7u 57 O+[i. Then 0< Vi (u )  0i(u. T) < C+(')' o)hi7u)e7ou8T .(u) < T ) to 0i (u) which is different from Theorem 3.Pi(MT > u) = Pi(MT < u. dy) jEE u U +C_ hje7( yu)G jEE"" +(i. j. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES proving the upper inequality. Here is an estimate of the rate of convergence of the finite horizon ruin probabilities 'i (u. and assuming it shown for n. it follows that Vi(u) < C_(yo) h=70)e7ou.13) Hence. we have Vii (u) . u The proof of the upper inequality is similar .M>u) = Ei [VGJT (u .11. letting MT = maxo<t<T St.10) C_ 1 f hje7(y. j. (3.12) Proof We first note that just as in the proof of Theorem 3. MT < u.y)G+(i. 13 and the second by the induction hypothesis . We claim by induction that then cpin) (u) > C_ hie7u for all n. dy) jEE o (3.10: Theorem 3 . let C+(yo) be as in (3.T) = Pi(M > u) . taking cps°) (u) = 0. from which the lower inequality follows by letting n * oo.10 ) by Lemma 3 .13 Let first cp=°)(u) = C_ hie"u in Lemma 3.M > u) = Pi(ST<u. j. y]hj = C_ e7uhi. T) = Pi (7.tpi(u. Indeed.
3p.3) to B = Bi does not depend on i.33 or Bi 0 Bj.o. in part from the folklore principle that any added stochastic variation increases the risk.. is the one for the Markovmodulated one in the stationary case (the distribution of J0 is 7r). we refer to . 4 Comparisons with the compound Poisson model 4a Ordering of the ruin functions For two risk functions 0'. but that in general the picture is more diverse... The motivation that such a result should be true came in part from numerical studies. u > 0.1) Obviously. (4.3). B2 <_s. MARKOVIAN ENVIRONMENT Notes and references The results and proofs are from Asmussen and Rolski [44].31:5)32 . M" of the corresponding two claim surplus proceses (note that 0'(u) _ P(M' > u).168 CHAPTER VI.2) alone just amounts to an ordering of the states. where o*(u) is the ruin probability for the averaged compound Poisson model defined in Section 1 and . Bp.5) Note that whereas (4. <s..2) (4.o. this correponds to the usual stochastic ordering of the maxima M'. ". < . It was long conjectured that 0* Vi.. Occasionally we strengthen (4.4) To avoid trivialities. The Markov process {Jt} is stochastically monotone (4. where it has been observed repeatedly that Markovmodulation increases waiting times and in fact some partial results had been obtained. (4.. V)" if z/i'(u) <'c"" (u). [177].0.o. The results to be presented show that quite often this is so. 0"(u) = P(M" > u)) Now consider the risk process in a Markovian environment and define i' (u) _ >iEE irioi(u). The conditions which play a role in the following are: . and finally in part from queueing theory.o.. we also assume that there exist i # j such that either /3i <..3) Bl <_s. this is not the case for (4. we define the stochastic ordering by 0' < s. Further related discussion is given in Grigelionis [176]. For the notion of monotone Markov processes. (4.
< a..1..r (Sr(o) E dx Jr(o) = i.6. b1 < . The first is a standard result going back to Chebycheff and appearing in a more general form in Esary. note that (4...9) follows by considering the increasing functions 3iBi (x) and Oi (u . where 7r2+) = QiµBilri/p.4) hold. . Lemma 4 .. 3 (a) P. 7(0) < oo) = pirf+). Proposition 2. Then V. we need two lemmas.x) dx u o i =1 i=1 (4. 1:7riaibi > E 7riai i=1 i=1 j=1 The equality holds if and only if a1 = .r (JT(o) = i..9) (4. Conditions (4. Proschan & Walkup [140]. Theorem 4 .r(u x)dx. Lemma 4 .2)(4... = b..1 for the first term in (4. (4.10) and (4.3 for the second) *(u) _ /3 *B* (u) +.13* J0 u 0*(u .. Conditioning upon the first ladder epoch.2)(4..6).1) which with basically the same proof can be found in Asmussen & Schmidt [49]. < bp and 7ri > 0 (i = 1.. the second follows from an extension of Theorem I1..* For the proof. = aP or b1 = . also Proposition 2.1 Assume that conditions (4. . Section 4. ^i 7ri = 1. (b) P.6) 7r= fl*B*(u) + p> s=1 +) fu 0 b (u  x)Bt (x) /pB. T(0) < oo) = Bi(x) dx/tcai .7) 7ri. dx (4. Comparing (4. E 7r i Wi(u . p).7) and Lemma 4.8) ^j Tri/iBd(x) .2.3* f uB(x) z/^.5 (cf. and it is in fact easy to show that Vii(u ) < t/j(u) (this is used in the derivation of (4. it follows by a standard . then P P P 7rjbj. 2 If al < . COMPARISONS WITH THE COMPOUND POISSON MODEL 169 Stoyan [352].x) of i and using Lemma 4.. Proof of Theorem 4.4) is automatic in some simple examples like birthdeath processes or p = 2 . 0 Here (4.4) say basically that if i < j .2.9 ) below). we obtain (cf.x)dx _ /3*B*(u) + f u / ^ t=1 > 3 * B* ( ) + f (4. then j is the more risky state .4.3iBi(x)YPi(u .x)B*(x) dx.10) Q*B*(u)+.
of (4. (u) is not in general true: Proposition 4.3µi < 1 for all i.6). Then i/i*(u) < . dominates the solution 0* to the renewal equation (4.3i.h.4 Assume that . Frey. u To see that Proposition 4.4 is the understanding of whether the stochastic monotonicity condition (4. that P P /^ 1r1NiµBi /^2 /^ ^i/ji pBi < 1il3i i=1 i=1 (4. (u) may fail for some u..r (u ) fails for all sufficiently small e > 0. let = ( 1/2 1/2 ) . Bi as e J. Rolski & Schmidt [32]. 0.(0) = V.0*• i=1 But it is intuitively clear (see Theorem 3. of order 101. Notes and references The results are from Asmussen.* (0).8) we get P P '*' (0) = 3* + /3*1* (0) _ > lri'3qqi • E 7i/ipBi .s.170 CHAPTER VI. and from this the claim follows. µB2 = 104. As is seen. they are at present not quite complete.1 of [145] for a formal proof) that z/ii(u) converges to the ruin probability for the compound Poisson model with parameters .. For u = 0.s.11) is of order 104 and the r. What is missing in relation to Theorem 4.. = 102. u Here is a counterexample showing that the inequality tp* (u) < V).h. Proof Since 0. 4b Ordering of adjustment coefficients Despite the fact that V)* (u) < *. Then the l.6). it will hold for all sufficiently large u. Using (4. except possibly for a very special situation . it is sufficient to show that 0'.4 is not vacuous.. 01 = 103.4) is essential (the present author conjectures it is).11) i=1` and that A has the form eAo for some fixed intensity matrix A0.. this ruin probability is /3iPBi.0. (4. µB./3*. Recall that .1 and Proposition 4.2.(0) < b *'(0) for e small enough. i=1 i=1 7'r(0) _ EFioiwi(0) . MARKOVIAN ENVIRONMENT argument from renewal theory that tk. Q2 = 1.
a = E irirci(a).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. Xt)} is a Markov additive process (a socalled Markovian fluid model. (4.1) . (4. This implies that A is strictly convex.7) )i is convex with A'(0) = lim EXt tioo t = iEE 70i = 0..13) (4.g. cf. Further (see Corollary 11. 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] . 0 . Jt = i])' EE = vA+n(6. Asmussen [20]) as discussed in 11.14) 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.12) iEE Theorem 4. Then {(Jt.ld) with generic cycle w = inf{t>0: Jt_54 k.)a. Lemma 4.5. It is clear that the distribution of X. and by Proposition II. it follows by Proposition A1.1) . Proof Define X= f &ids.5. with strict inequality unless a = 0 or bi = 0 for all i E E. is nondegenerate unless bi does not depend on i E E. The adjustment coefficient y* for the averaged compound Poisson model is the solution > 0 of rc*(ry*) = 0 where rc*(a) _ 13*(B*[a] .13) implies A(a) > 0 for all a. (4.5 y < ry*. e.Jt=kI A (the return time of k) where k E E is some arbitrary but fixed state. Now we can view {Xt} as a cumulative process (see A. in particular .g.5.(a) > 0 for all a 0 0. Hence if 5i 54 0 for some i E E.a.. with strict inequality unless rci (y*) does not depend on iEE.4(b) that the limit in (4.14) is nonzero so that A"(0) > 0.2 we have (Ei[e"X'.4.
If rci(y* ) is not a constant function of i E E.15) yields 0 = (Ii(y*)) diage + Aoh'(0) = (rci('Y*)) diage + (Ao . MARKOVIAN ENVIRONMENT Proof of Theorem 4. Rolski & Schmidt [32].) and rc (•).15) once more and letting e = 0 we get .. this implies that the solution y > 0 of K(y) = 0 must satisfy y < y*..12) and rc*(y*) = 0.e7r)1 (Ici(Y*))diage.. Further a(1) = rc(y*) by definition of A(. (4. we get rc (y*) > 0 which in a similar manner implies that u y < y*.16) Differentiating (4. whereas the .5. h'(0) = (Ao . The corresponding adjustment coefficient is denoted by ry(e). y.172 CHAPTER VI. (4. 4c Sensitivity estimates for the adjustment coefficient Now assume that the intensity matrix for the environment is Ae = Ao/ e. the basic equation is (A + (rci(y))diag)h = 0. and our aim is to compute the sensitivity ay e a E=O A dual result deals with the limit a 4 oo.5 is from Asmussen & O'Cinneide [40].Qi and Bi are fixed . Here we put a = 1/e.p yi and compute 8y 8a a=0 In both cases.. h depend on the parameter (e or a). multiply the basic equation by a to obtain 0 = (A0 + e(r£i(y))diag)h. Since ic is convex with rc'(0) < 0 . a = 1 in Lemma 4. we have 7rh' = 0. improving upon more incomplete results from Asmussen.6. In the case of e. Frey. Hence rc (y*) > 0. Thus y(e) * y* as e 10. Notes and references Theorem 4.eir)h'(0). h(0) = e.15) Normalizing h by 7rh = 0. where A. 0 = ((ri(Y))diag + ery (4{('Y))diag)h + (A0 + e(?i'Y))diag)h'. Then > risi = 0 because of (4. Let bi = rci(y*). Hence letting e = 0 in (4. note that y(a) + mins=1.
Inserting (4. Rolski & Schmidt [32]. (4. . Frey.20) and multiplying by el to the left we get 0 = All + 7'(0)rci (0) + 0 (here we used icl (ry(0)) = 0 to infer that the first component of K[7(0)]h'( 0) is 0).18) 0 = 27'(0)p+27r(rs.19) holds.5.8 when ryi < 0 for some i is open. We get 0 = (aAo + ( lc&Y))diag)h. We assume that 0 < y < 7i.. and we have proved: Proposition 4. then 8a a=o All rci (0) Notes and references The results are from Asmussen.17) (4.18). i = 2. 0 = (Ao + ry'(ii(Y)) diag )h + (aAo + (Ki(7'))diag)h'.20) Letting a = 0 in (4. The additional feature of the model is the following: • Certain transitions of {Jt} from state i to state j are accompanied by a claim with distribution Bid. The analogue of Proposition 4. THE MARKOVIAN ARRIVAL PROCESS 173 0 = 27'(0)(ri(`Y *)) diage + 2(ci('Y* )) diag h' (0) + Aoh" (0) . (4.8 If (4. .i(7' *))diagh'(0). the intensity for such a transition (referred to as marked in the following) is denoted by Aii l and the remaining intensity . 5 The Markovian arrival process We shall here briefly survey an extension of the model. p.17) by 7r to the left to get (4. which has recently received much attention in the queueing literature.19) Then 'y ^ ryl as a ^ 0 and we may take h(0) = el (the first unit vector). multiplying (4. (4. and may have some relevance in risk theory as well (though this still remains to be implemented).16) yields 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.
We then let (see the Appendix for the Kronecker E = E(1) x E(2). . A(l) = T. A ( 2) = A (2`1 ) ® A. Bij = B. T). the definition of Bij is redundant for i i4 j.2 for details). let { Jt 1) }. B. and that are determined by A = A(l ) +A(2) where A is the intensity matrix the governing {Jt}. Here are some main examples: Example 5 . Note that the case that 0 < qij < 1. Indeed. Bii = Bi . and the marked transitions are then the ones corresponding to arrivals. is neither 0 or 1 is covered by letting Bij have an atom of size qij at 0.2) A(1) = A(' 1) ® A(1.1 (PHASETYPE RENEWAL ARRIVALS) Consider a risk process where the claim sizes are i.2 (SUPERPOSITIONS) A nice feature of the setup is that it is closed under superposition of independent arrival streams . the Markovmodulated compound Poisson model considered sofar corresponds to A(l) = (. the definition of Bi is redundant because of f3i = 0.4). A(l) = tv. we may let {Jt} represent the phase processes of the individual interarrival times glued together (see further VIII.2).(13i )diag. then {Nt} is a Markov additive process if and only if it corresponds to an arrival mechanism of the type just considered. 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.i. A(1'k) A(2 k1). the claim surplus is a Markov additive process (cf. II. where qij is the probability that a transition i * j is accompanied by a claim with distribution.174 CHAPTER VI. The extension of the model can also be motivated via Markov additive processes: if {Nt} is the counting process of a point process. with common distribution B. Jt2)) (2. that Bii = Bi . u Example 5 . refer to notation) { Jt k) }. and thus 1i = 0. A(1) = A .^) etc. For i = j.6i ) diag. Jt = (Jtl). In the above setting. j(2) } be two independent environmental processes and let E(k). Again . but the point process of arrivals is not Poisson but renewal with interclaim times having common distribution A of phasetype with representation (v. Thus . This is the only way in which arrivals can occur.d.
. more recently.kl is redundant). iN = all BOi2.kj = Bik) B13 4k = Bak) 175  (the definition of the remaining Bij. The versatility of the setup is even greater than for the Markovmodulated model. DIVORCED... with rate ai. the kth policy enters a recovering state.. e.5. where ik = 0 means that the kth policy has not yet expired and ik = 1 that it has expired.... RETIRED.. This means that the environmental states are of the form i1i2 • • • iN with il.... iN..iN. WIDOWED. E 10.. Bilo..iN. E = { WORKING.}. as the Markovian arrival process ( MAP). superpositions of renewal processes. INVALIDIZED. 11.1i2 . Hermann [193 ] and Asmussen & Koole [37] showed that in some appropriate . and that the policy then expires. 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 .iil.iil. the idea of arrivals at transition epochs can be found in Hermann [193] and Rudemo [313]. claims occur only at state transitions for the environment so that AN2.iN are zero and all Bi are redundant. Easy modifications apply to allow for • the time until expiration of the kth policy is general phasetype rather than exponential. 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.1i2. • upon a claim.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). 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.. However .. possibly having a general phasetype sojourn time. DEAD etc.. In fact .iN = a2. Similarly.. say.iN C17 AilO. u Example 5 . i2i .. assume that there is a finite number N of policies. MARRIED. or. Assume further that the ith policy leads to a claim having distribution Ci after a time which is exponential. THE MARKOVIAN ARRIVAL PROCESS Bij. In this way we can model.g.. iN. all Al i2.. Example 5 . Thus.iN = C27 All other offdiagonal elements of A are zero so that all other Bii are redundant.
[248].3*µs • p = f /3(v) dv 0 0 (6. The basic assumptions are as follows: • The arrival intensity at time t of the year is 3(t) for a certain function /3(t). • The premium rate at time t of the year is p(t). Let 1 1 /3* _ f /3(t) dt. continuity would hold in presumably all reasonable examples. MARKOVIAN ENVIRONMENT sense any arrival stream to a risk process can be approximated by a model of the type studied in this section : any marked point process is the weak limit of a sequence of such models . p(t) and B(t) are defined also for t t [0.176 CHAPTER VI. Neuts [271] and Asmussen & Perry [42]. Some main queueing references using the MAP are Ramaswami [298]. p * = 0 p(t) dt.p)/p. we may assume that the functions /3(t). )3 t 1 J (6. one limitation for approximation purposes is the inequality Var Nt > ENt which needs not hold for all arrival streams. for s E E = [0. Obviously. Lucantoni et at. For the Markovmodulated model. but now exhibiting (deterministic) periodic fluctuations rather than (random ) Markovian ones. one needs to assume also (as a minimum) that they are measurable in t. Sengupta [336]. 0 < t < 1. • Claims arriving at time t of the year have distribution B(t). Without loss of generality. Lucantoni [248]. let the period be 1. We denote throughout the initial season by s and by P(8) the corresponding governing probability measure for the risk process. 1). B* = J f B(t) ((*) dt.1) Then the average arrival rate is /3* and the safety loading rt is 77 = (p* . . we talk of s as the 'time of the year'. 1). from an application point of view. Thus at time t the premium rate is p(s + t). a claim arrives with rate /3(s + t) and is distributed according to B(8+0 . By periodic extension. 6 Risk theory in a periodic environment 6a The model We assume as in the previous part of the chapter that the arrival mechanism has a certain timeinhomogeneity. where i f00 xB(°) (dx) _ .2) Note that p is the average net claim amount per unit time and µ* = p//3* the average mean claim size.
St = SeI(t). The behaviour of the periodic model needs not to be seen as a violation of this principle. u Remark 6 .3* = 3A.w(t). the average compound Poisson model is the same as in III. it turns out that they have the same adjustment coefficient. 0 Then (by standard operational time arguments ) {St} is a periodic risk process with unit premium rate and the same infinite horizon ruin probabilities.1.2 Define T 6(T) = p(t ) dt. B*. The claim surplus process {St } two is defined in the obvious way as St = ^N° Ui .8. Thus. 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. we shall see that for the periodic model increasing A increases the effect of the periodic fluctuations. It is easily seen that .10. the discussion in 111.9). In contrast.3(t) = 3A(1 + sin 27rt). equivalently. since the added variation is deterministic.w(t)) dt). and thus the averaged standard compound Poisson models have the same risk for all A. Section 4b).3) Note that A enters just as a scaling factor of the time axis.t. let . p* as an averaged version of the periodic model.1) and Example 1. in agreement with the general principle of added variation increasing the risk (cf. We u assume in the rest of this section that p(t) .1 As an example to be used for numerical illustration throughout this section. and we recall from there that the ruin probability is 24 1 *(u) _ 3 5eu + 35e6u. (6. RISK THEORY IN A PERIODIC ENVIRONMENT 177 In a similar manner as in Proposition 1. p(t) = A and let B(t) be a mixture of two exponential distributions with intensities 3 and 7 and weights w(t) _ (1 +cos27rt)/2 and 1 . respectively. the conditional distribution . Thus . Many of the results given below indicate that the averaged and the periodic model share a number of main features. In contrast. or. not random.(3. Example 6 . In particular. one may think of the standard compound Poisson model with parameters 3*. for Markovmodulated model typically the adjustment coefficient is larger than for the averaged model (cf. The arrival process {Nt}t>0 is a timeinhomogeneous Poisson process with intensity function {/3(s + t)}t>0 .6. of the periodic model as arising from the compound Poisson model by adding some extra variability.
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. [44] (the literature in the mathematical equivalent setting of queueing theory is somewhat more extensive.1) a = J8 .. The exposition of the present chapter is basically an extract from [44].(3(s + t)dt)e«St adt + /3(s + t)dt . and the ruin probabilities are 0(8) (U) = P(s )(r(u) < 00). a) is periodic on R. we obtain E.g.^8 [.3(s + t)dt[B(8+t)[a] .1) dv .a.a be the c. with some variants in the proofs.(1 .T) = P(8)(r(u) <T).tc* (a)] dv then h (. 3 E(8)eaSt = h(s.f.. 0 (5)(u. J Theorem 6 .3(v)(B(vl [a] .8).g. (6. of the claim surplus process. e. [101] . . As usual.a) Proof Conditioning upon whether a claim occurs in [t.e.s .178 CHAPTER VL MARKOVIAN ENVIRONMENT of U.a . given that the ith claim occurs at time t is B(8+t).Q(v) (B(„) [a] . with the underlying Markov process {Jt} being deterministic period motion on E = [0. Notes and references The model has been studied in risk theory by..al.east B(8+t) [a] east . Dassios & Embrechts [98] and Asmussen & Rolski [43].1]) . and define h(s. we start by deriving formulas giving the m. Jt = (s + t) mod 1 P(8) .5 (see in particular Remark 11.1) . The claim surplus process {St} may be seen as a Markov additive process.(8) [eaSt+dt I7t] = = (1 .4) At a first sight this point of view may appear quite artificial.5. a) etw*(a) h(s+t. i. t + dt] or not. 1).g. 6b Lundberg conjugation Motivated by the discussion in Chapter II. of the averaged compound Poisson model (the last expression is independent of s by periodicity).adt +.a) = exp { . To this end. r(u) _ inf It > 0 : St > u} is the time to ruin . let f 8+1 tc *(a) _ (B* [a] . Daykin et. see the Notes to Section 7).f.
* (a) h(s.(e) Let = h( h(Jo.c* (e) {Le.0(s + t)dt[B(8+t)[a] . St)} .(8)east 179 = = = = = E(8)east (1 . a) as well as the fact that rc = k` (a) is the correct exponential growth rate of Eeast can be derived via Remark 11.6. 0) P(8)a.9 as follows. we can write Lo Jt. dt log E(8)east a + f3(s + t) [B(8+t) [a] .1]. a) Thus E(8)east = h(s + t.t. a).3(s + t)[D(8 +t)[a] .1]) .t}t>o = h(s. With g the infinitesimal generator of {Xt} = {(Jt. a) Corollary 6. h(s + t. According to Remark 11.1)dv  o h(t.log h(s. 9) is a P ( 8)martingale with mean one. a) = exp I f t3(v)(kv)[a) . a) et.adt +. St)} and ..6 .1)dv l og E(8) et where atetk•(a) h(t. a) h(s + t. Proof In the Markov additive sense of (6. so that obviously {Lo.9) eastt. u Remark 6. a) . it then suffices to note that E(8)Le.s.3. E (8)east (a +. at + f log h(s + t. RISK THEORY IN A PERIODIC ENVIRONMENT E(8)east+ dt d Et.2.5 The formula for h(s) = h(s. a) = h(s. + v)(B([a] .t} is a multiplicative functional for the Markov process { (Jt.4 For each 0 such that the integrals in the definition of h(t . B) eoSt t. 0) exist and are finite.1]) .t = 1 by Theorem 6.4).5.
Now define 'y as the positive solution of the Lundberg equation for the averaged model. B(s). cf. .180 CHAPTER VI. Sdt) = h(s + dt) eadt (1 . correspond to a new periodic risk model with parameters ex . However.1.(3(s)dt) +.3(s)dt • B(s)[a]h(s) = gha(s. A further important constant is the value yo (located in (0. When a = y.60(t) = a(t)B(t)[0].0) = Kh(s).6 The P(s).2.7 When a > yo. That rc = is*(a) then follows by noting that h(1) _ u h(0) by periodicity.3(s)h(s) + h'(s) +.'y). the requirement is cha(i.a . of St is as for the asserted periodic risk model. (iv) finally.g. J s [.3(s)B(s) [a]h(s). y solves n* (y) = 0. Bet)(dx) = ^ B(t ) (dx).y) = eayh(s). For each 0 satisfying the conditions of Corollary 6. ry)) at which n* (a) attains its minimum. 0 < s < 1. Proposition 6.4.T.3(v)( Bi"i [a] . Equating this to rch (s) and dividing by h(s) yields h(s ) = h(s) = a + . the restrictions of Plsi and Pest to Ft are equivalent with likelihood ratio Le. such that for any s and T < oo. That is. say. see [44] for 11 a formal proof. P(s) (T(u) < oo) = 1 for all u > 0. (ii) use Markovmodulated approximations (Section 6c). St)} with governing probability measures Fes). Lemma 6 . ( iii) use approximations with piecewiese constant /3(s). Proposition 6.tc] dv} (normalizing by h(0) = 1).f. 0) = h(s) + dt {ah(s) .3. MARKOVIAN ENVIRONMENT ha(s. That is.1) . [70] . as above E (s) ha(Jdt. yo is determined by 0 = k* (70) = QB*.6 ( s ) exp { 0( s )&s) [a] + tc . it follows by Theorem II. Proof (i) Check that m.3(s)ks)[a]h(s)} ah(s) 13(s)h(s) + h'(s) +.5 that we can define a new Markov process {(Jt. we put for short h(s) = h(s.
1. has a unique stationary distribution. we need the following auxiliary result . 1). T) = h(s. ^(u) = ST(u) . The proof involves machinery from the ergodic theory of Markov chains on a general state space.8 The ruin probabilities can be computed as (u)+T(u)k'(a) ^/i(8) (u.9) and noting that weak convergence entails convergence of E f (^(u).9) 0(')(u) = h(s. T(u) < (6. the Markov process {(^(u). Wu). RISK THEORY IN A PERIODIC ENVIRONMENT Proof According to (6. the mean number of claims per unit time is p« 181 = Jo 1. Corollary 6. e(cc)) Letting u > oo in (6. The relevant likelihood ratio representation of the ruin probabilities now follows immediately from Corollary 11. a)e«uE (a iP(s) (u) = h( s)e7uE(` ) h(O(u)) To obtain the CramerLundberg approximation from Corollary 3. q) = eryx/h(q)). and we refer to [44]. Here and in the following.6(v) dv Jo ' xe«xB (°) (dx) r^ xe«xB'(dx) = Q'B' [ a] = ^' J 0 = ^c"'(a) + 1. xEJ 0 (s)b(8)(x) > 0. J C R+ such that the B(8).6. a) a > ry0 (6. f (x. 9(u)) for any bounded continuous function (e. which is not used elsewhere in the book.9 Assume that there exist open intervals I C [0. s E I. u which is > 1 by convexity. Lemma 6 . a) e«uE(8 ) e «^ . considered with governing probability measures { E(8) }E[ . have components with densities b(8)(x) satisfying inf sEI.1) the distribution of (l: (oo). 0(u)) * (b(oo).g. (6. and no matter what is the initial season s. say s0.4. a) TI h(9(u).8) (6.2).2. we get: . B(oo)).u is the overshoot and 9(u) = (T(u) + s) mod 1 the season at the time of ruin.10) Then for each a.9(u))} u>0.7) h(B(u).
Ch(s)ery". Theorem 6 .6 for the Markovmodulated model: Theorem 6 .1.9). which may provide one among many motivations for the Markovmodulated approximation procedure to be considered in Section 6c. Vi(8) (u) . we obtain immediately the following version of Lundberg ' s inequality which is a direct parallel of the result given in Corollary 3. For our basic Example 6 . Noting that ^(u) > 0 in ( 6.11) gives an interpretation of h(s ) as a measure of how the risks of different initial seasons s vary. 6. MARKOVIAN ENVIRONMENT Theorem 6.) C = E1 h(B(oo)) u + oo. Among other things.11) Note that ( 6. 10 shows that certainly ry is the correct Lundberg exponent. A=1/4 A=1 A=4 0 Figure 6. At this stage . where e. where C(o) = 1 + info < t<i h(t) .W. elementary calculus yields h(s) = exp { A C 2^ cos 2irs  4^ sin 21rs + 11 cos 41rs . illustrating that the effect of seasonality increases with A.182 CHAPTER VI.10) of Lemma 3. 1. it does not seem within the range of our methods to compute C explicitly.1 In contrast to h.1.16.10 Under the condition (6. this provides an algorithm for computing C as a limit. (6.ir) } Plots of h for different values of A are given in Fig. 11 7/'O (u) < C+°)h(s) ery".
we substitute T = yu in 0(u. We state the results below. yu) 000 (u) .w)e4u dx 9w + 7(1 .42 • exp {J_ cos 27rs .w)e4u .12) As for the Markovian environment model.7e . .3x + (1 .167r I Cu.4.. ay) • (6. Just as in IV. Theorem 6. (ay).(ay) > 0 when y < 1/ic' (7).16 In order to apply Theorem 6. RISK THEORY IN A PERIODIC ENVIRONMENT Thus.13 to our basic example.6.g. 1 ) and all u > 0. Lundberg's inequality can be con siderably sharpened and extended.47r sin 27rs + 167r cos 47rs . 1 (6.7x j dx _7x } _ 6w + 6(1 . we obtain Co) = 1. whereas ay < y. r. where ay is the unique solution of W(ay) =y• (6.17) (6.yr. Consider first the timedependent version of Lundberg's inequality.14) < C+)(y)h(s) e7yu.0(8) (u+ yu) (6.(s)(u) < C+h(s)e7".12 Let 00)(y) 1 Then info < t<i h(t.w) . we first note that the function fu° exu {w • 3e .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. T) and replace the Lundberg exponent ry by ryy = ay . (6. in our basic example with A = 1. C_h(s)e7u < V.(8) (u.w ) • 7e u{w • 3e3x + ( 1 .13) Elementary convexity arguments show that we always have ryy > Y and ay > ry. e. the proofs are basically the same as in Section 3 and we refer to [44] for details. Theorem 6 .15) The next result improves upon the constant C+) in front of eryu in Theorem 6. e7 ( yx)B(t)(dy) > Then for all $ E [0.11 as well as it supplements with a lower bound. #c( ay) < 0 when y > 1/tc'('y).42 so that 183 tp(8) (u) < 1.
MARKOVIAN ENVIRONMENT attains its minimum 2 /3 for u = oo and its maximum 6 /(7 + 2w) for u = 0.4^ sin 2irs + 16^ cos 41rs .19 I eu. Finally. 0 <'p(8)(u ) . 6c Markovmodulated approximations A periodic risk model may be seen as a varying environment model.9 3 0<8<1 p 27r 47r 167r 161r 2 _ _e.0. This observation motivates to look for a more formal connection between the periodic model and the one evolving in a finite Markovian environment.\ 3 C+ = sup 6 exp { A (.20 •exp { 2n cos 27rs .T) < C+('Yo)h( s.181 s(u) < 1.'Yo)e (6. 14 Let C+('yo) be as in (6. .. Of course. n}.184 CHAPTER VI. Thus. 1) for the environment). and let 8 = er' (Y0). Some of the present proofs are more elementary by avoiding the general point process machinery of [44]. such a deterministic periodic environment may be seen as a special case of a Markovian one (allowing a continuous state space E = [0.1 sin 27rs + 1 cos 47rs .(8)(u.1 sin 2irs + 16_ cos 47rs . and in fact. . 1/i18 1 s (u) > 0. the nth Markovian environmental process {Jt} moves cyclically on {1.16.18) Notes and references The material is from Asmussen & Rolski [44].66.16) with 'y replaced by yo and h(t) by h(t.66. The idea is basically to approximate the (deterministic) continuous clock by a discrete (random) Markovian one with n 'months'. C+ = 1. Then 7oudT ..013. Thus C_ = 2 inf ex cos 2irs .I eu.013.g.cos 27rs . where the environment at time t is (s + t) mod 1 E [0. for A = 1 (where 3 e0.\ = 0 . with s the initial season. but thereby also slightly longer. 1). yo).20).19 } 0 <8<1 8 + cos 21rs Thus e. we have the following result: Theorem 6 . exp 2^ cos 21rs . 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. completing a cycle .L sin 27rs + 1 I cos 47rs . .
but others are also possible. 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. since the settings are equivalent from a mathematical point of view. Let 0j. To this end. so that the intensity matrix is A(n) given by n n 0 ••• 0 0 n n ••• 0 A(n) _ (6. it is desirable to have formulas permitting freely to translate from one setting into the other. and the ruin probability corresponding to the initial state i of the environment is then Y'yn)(t) = F (M(n) > t). This queue is commonly denoted as the Markovmodulated M/G/1 queue and has received considerable attention in the last decade. 7 Dual queueing models The essence of the results of the present section is that the ruin probabilities i/ (u). DUAL QUEUEING MODELS 185 within one unit of time on the average .21) which serves as an approximation to 0(1)(u) whenever n is large and i/n s. Thus. one simple choice is Oni = 0( i . .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. M(n) = Supt>o Stn).jEE of the risk process. We let {Stn)} (6. We want to choose the /3ni and Bni in order to achieve good convergence to the periodic model. (6.20) be the claim surplus process of t>o the nth approximating Markovmodulated model. Bi.1 ((i 1)/n) ) and Bni = B . Notes and references See Rolski [306]. AE= Aii'r?/7ri• The arrival intensity is /3i when Jt = i. z/'i (u. A be the parameters defining the risk process in a random environment and consider a queueing system governed by a Markov process {Jt } ('Markovmodulated') as follows: • The intensity matrix for {Jt } is the timereversed intensity matrix At _ A ())i.7.
I* = i). let T . (7. . .1) over j.186 CHAPTER VI. In particular. The first conclusion of that result then states that the events {T(u) < T. The actual waiting time process 1W1. MARKOVIAN ENVIRONMENT • Customers arriving when Jt = i have service time distribution Bi.n(VT > u. J* = i) for all j.2) Oi(u) = 1. JT = j) = 7rjPj(VT > u..1). Then Pi(T(u) < T.1) follows. For (7. 2 . ii (u) = it /3 P(W > u. • The queueing discipline is FIFO.3. and for (7. JT = i) = P(V > u. J* = i).2). Now let In denote the environment when customer n arrives and I* the steadystate limit.3) 7ri where (V.0i (u .4) where 0* = >jEE 7rj/3j. and the virtual waiting time (workload) process {Vt}too are defined exactly as for the renewal model in Chapter V. Proof Consider stationary versions of {Jt}o<t<T. Proposition 7. T) = 7ri 1 P.T(V > u I J* = i). Proposition 7. I* )3i P(V > u. JT = j} and {VT > u. Jo = i. (7. JJ = i).=1 .oo in u (7.1) 7ri In particular. (7. J* = i) = P. JT = Z).. JT = i) = 'P. {Jt }o<t<T• Then we may assume that Jt = JTt. (7. 0 < t < T and that the risk process {Rt}o<t<T is coupled to the virtual waiting process {Vt}o<t<T as in the basic dualitylemma (Theorem 11. just sum (7. Jt ). Taking probabilities and using the stationarity yields 7riPi(T(u) < T.3).1 Assume V0 = 0. and (7. J*) is the steadystate limit of (Vt. Jo = j..2) and use that limF (VT > u.P(V > u.2 The relation between the steadystate distributions of the actual and the virtual waiting time distribution is given by F(W > u. (VT > u I JT = 2). JT = j) = LjPj (VT > u. JT = i} coincide.
n=1 N However. B(t) have been periodically extended to negative t).4).4) can be found in Regterschot & de Smit [301].T)(T(u) <T) = P(8)(VT > u).I *=i). the dual queueing model is a periodic M/G/1 queue with arrival rate 0(t) and service time distribution B(') at time t of the year (assuming w. .3). and (7. P(. on average /32TP(V > u.8) For treatments of periodic M/G/1 queue. With {Vt} denoting the workload process of the periodic queue. [243]. I*) with the timeaverage . DUAL QUEUEING MODELS 187 Proof Identifying the distribution of (W. Lemoine [242]. J* = i) see W > u.I. >u.1 is from Asmussen [16].3) improving somewhat upon (2.l. see Regterschot & van Doorn [123]. a general formalism allowing this type of conclusion is 'conditional PASTA'.o. see in particular Harrison & Lemoine [186]. In the setting of the periodic model of Section 6.g. and one has PI'>(rr(u) < T) = P('_T)(VT > u). The relation (7.4) and (7. I* = i. that /3(t). P(1')(r(u) < oo) = P(')(00) > u).5) follows from (7. and Rolski [306]. and further references (to which we add Prabhu & Zhu [296]) can be found there.7) of that paper. Taking the ratio yields (7. A more probabilistic treatment was given by Asmussen [17]. with (7.6) (7. if T is large.. T].7) (7. P(W >u. (7.=i) a4. a paper relying heavily on classical complex plane methods. The first comprehensive solution of the waiting time problem is Regterschot & de Smit [301]. on average 0*T customers arrive in [0. Proposition 7. we have 1: I(W.7. u Notes and references One of the earliest papers drawing attention to the Markovmodulated M/G/1 queue is Burman & Smith [84]. N * oo. p < 1 then ensures that V(*) = limNloo VN+9 exists in distribution. and of these.
This page is intentionally left blank .
However . i&(u. and the evolution of the reserve may be described by the equation Rt = u .2) tk(u.. resp . Thus in between jumps. the aggregate claims in [0. with common distribution B and independent of {Nt}. are i.T) = F(T(u) < T). U2. and T(u) = inf {t > 0 : Rt < u} is the time to ruin starting from Ro = u so that '(u) = F(T(u) < oo). Thus.. and that the claim sizes U1.T) = FloinfTRt< OIRo=u1 denote the ruin probabilities with/initial reserve u and infinite.1) (other terms are accumulated claims or total claims). z/i(u) = F IinffRt< 0IRo=u 1 (1.At + p(R8) ds.6.d. . Zt As earlier. {Rt} moves according to the differential equation R = p(R). t] are Nt At = Ui (1.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 . the premium charged is assumed to depend upon the current reserve Rt so that the premium rate is p(r) when Rt = r. 189 . finite horizon.
4 Either i. Another could be the payout of dividends: here the premium paid by the policy holders is the same for all r. oo) is given by i (u + p/S). Example 1. If Ro = v < u. rather than when the reserve itself becomes negative. pi > p2 and p(r) = One reason could be competition. the payout rate of interest is Sx and absolute ruin occurs when this exceeds the premium inflow p.e. when x > p/S. and the probability of absolute ruin with initial reserve u E [p/S.'(u)) > 0 so that V'(v) < 1. Now return to the general model. say at interest rate b. No tractable necessary and sufficient condition is known in complete generality of the model. we get p(r) = p + er. Thus at deficit x > 0 (meaning Rt = x). Hence in terms of survival probabilities. RESERVEDEPENDENT PREMIUMS The following examples provide some main motivation for studying the model: Example 1 . In this situation. That is. Proposition 1. Proof Obviously '(u) < ilb(v) when u > v. or o(u) < 1 for all u. but assume now that the company borrows the deficit in the bank when the reserve goes negative. A basic question is thus which premium rules p(r) ensure that 'O(u) < 1.i(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. dividends are paid out at rate pi .3 (ABSOLUTE RUIN) Consider the same situation as in Example 1.Vi(v) u > e(1 .190 CHAPTER VII. Assume 0(u) < 1 for some u.p2. 1 .p/S) r > p/S p5(p/5r) 0<r<p/5 Then the ruin problem for {Rt } is of the type defined above. However. we can put Rt = Rt + p/S. Example 1. there is positive probability. P(r) _ p + e(r . it seems reasonable to assume monotonicity (p(r) is u . that {Rt} will reach level u before the first claim arrives.2. say e. i.1 Assume that the company reduces the premium rate from pi to p2 when the reserve comes above some critical value v. but when the reserve comes above v. where one would try to attract new customers as soon as the business has become reasonably safe.
let uo be chosen such that p(r) < p = /3µB for r > uo.6 For any T < oo. B and (constant) premium rate p. the probability that Rt < uo for some t is at least tp(0) = 1 (cf. Starting from Ro = uo. Hence ik(u) < 1 for all u by Proposition III. T] i n such a way that the events {r(u) <T} and {VT > u} coincide. if and only if V)(u) < 1 for all u. instead of (1.o(uo) = 1 so that t/'(u) = 1 for all u by Proposition 1.b(u. and hence by a geometric trials argument. That is. This is basically covered by the following result (but note that the case p(r) . {Vt} decreases at rate p(v) when Vt = v (i.2 once more. We next recall the following results.2) we have t Vt = At . appealing to Proposition 111.2(d).6) .e.uo) < 1. (1. In case (a).3µB for all sufficiently large r.4. 296297): Theorem 1. Here {Vt}two is a storage process which has reflection at zero and initial condition Vo = 0. In between jumps. which was proved in 11. cf. one can couple the risk process and the storage process on [0.2(d)).uo) and. hence Rt < uo also for a whole sequence of is converging to oo. say V.1. . (1.4) 0 and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0. However. In case (b). then ?(u) = 1 for all u. In particular.T) = P(VT > u).1.f p(Vs) ds.1. Proof This follows by a simple comparison with the compound Poisson model.. Let Op(u) refer to the compound Poisson model with the same 0. and P(Rt + oo) > 0. (b) If p(r) > /3µB + e for all sufficiently large r and some e > 0.I3IB requires a more detailed analysis and that µB < oo is not always necessary for O(u) < 1 when p(r) 4 oo. let uo be chosen such that p(r) > p = 0ILB + e for r > uo. INTRODUCTION 191 decreasing in Example 1.1 and increasing in Example 1.3. [APQ] pp.5) and the process {Vt} has a proper limit in distribution .5 (a) If p(r) < /. we have z/i(u) <p(u .+ p(r) exists. (1.1. Then if u > no. Theorem 1. { Vt} remains at 0 until the next arrival). that u zPp(u .2) for r sufficiently large so that p(oo) = limr. V = p(V)). obviously infu<uo z/'(u) > 0. Proposition I1I. Then 0(u) = P(V > u). then l/i(u) < 1 for all u.
Jo AX) (1. Then the ruin probability is tp (u) = f' g(y)dy.7 p(x)g(x) = tofB (x) + a f (x .8) Proof In stationarity. t + dt] if and only if Vt E [x.Sx} dx. Considering the cases y = 0 and 0 < y < x separately.Sx}. t + dt] can be neglected so that a path of {Vt} corresponds to a downcrossing in [t. say if p(r) goes to 0 at rate 1 /r or faster as r j 0. x] to (x.Qw(x) . Now obviously. oo). (1.y)g(y) dy. oo) must be the same as the flow the other way. 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. Oeax f x e'Yg (y) dy } = p) eaxa(x) . and the other being given by a density g(x) on (0. Note that it may happen that w (x) = oo for all x > 0.y. of (1. (1. RESERVEDEPENDENT PREMIUMS In order to make Theorem 1. In view of the path structure of {V t }. It is intuitively obvious and not too hard to prove that G is a mixture of two components.6 applicable.h. An attempt of an upcrossing occurs as result of an arrival. and is succesful if the jump size is larger than x . Corollary 1.192 CHAPTER VII. say when {Vt} is in state y. x + p(x)dt]).6x and that w(x) < oo for all x > 0.6w(x) . we thus need to look more into the stationary distribution G. B(x) = e.8 Assume that B is exponential with rate b. the flow of mass from [0. yo ^ 1 + oo Q exp {. we arrive at the desired interpretation of the r.9) Proof We may rewrite (1. where g(x) = p( ^ exp {. It follows in particular that 0(u) = fg(Y)dy.s. for the storage process {Vt}.7) Proposition 1.8) as g(x) = p 1 {yo13e_6x +.h. say. this means that the rate of upcrossings of level x must be the same as the rate of downcrossings.8) is the rate of downcrossings (the event of an arrival in [t.8) as the rate of upcrossings. one having an atom at 0 of size 'yo.s. u Define ^x 1 w(x) Jo p(t) dt. the l. say. of (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
Proof 2 We use Theorem 2. assume that {Wt} is Brownian motion with drift µ and variance v2.200 CHAPTER VII. RESERVEDEPENDENT PREMIUMS u from which (2. and the c. it follows that logH[a] = f 1 c(y)dy = 1 f '(pa/(a +y))dy f 0 0 Ey R/E 1 [pa + )3log 8 .3 is also valid if {Rt} is obtained by adding interest to a more general process {Wt} with stationary independent increments.01'E) + (p/E)al aO l febP/E } IF (0 /0 jF From this (2./3 log(b + a)] = log ePa/f (a + a ) e which shows that Z is distributed as p/E . 13/E). From ic(a) = . where V is Gamma(b.3. RT(u) has an exponential distribution with rate (S) and hence E [H(RT(u))I r(u) < oo] L Pe6'r (P/C .3a/ (5 .pa.2) follows by elementary algebra.pa.V < x)]0 + f P(V > p/E ) + eby fv (p/E .2y +µ ) dy . /^ u Example 2 . As an example. .g.a) . i.5 The analysis leading to Theorem 2.x) dx e.e.2) follows by elementary algebra. with density x(3/e1aQ/e fV (x) _ e 6X ' x > 0.3/E) By the memoryless property of the exponential distribution. H(u) = P(Z r < u) = P(V > u + p/E) = (8(p + Eu)/E.f. then {Rt} is the diffusion with drift function p+Ex and constant variance a2. of Z is IogH[a] = f ytc(y)dy = e fa (0.V. r (j3/E) In particular. The process {St} corresponds to {Wt} so that c(a) or2a2/2 .13 /E) r (.b P/E dx /' P/ ' (p/  x)p/e 150/f I' (/3/E) (6P1'E.
d. Some of these references also go into a stochastic interest rate. [283]. it is also used as basis for a diffusion approximation by these authors. 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).3. [282]. or to nonlinear premium rules p(•). Q2/2E). as in the proof of Proposition 2. and recall Lundberg 's inequality . Corollary 2. see e. write y* for the solution of the Lundberg equation f3(B[ry *] . Z is normal (p/E. Further studies of the model with interest can be found in Boogaert & Crijns [71]. [357]. that the analysis does not seem to carry over to general phasetype distributions. it follows that the ruin probability is Cu) H(u) H(0) 11 Notes and references Theorem 2. se e.4 is classical.8. for a martingale proof. Paulsen [281]. Paulsen & Gjessing [286] and Sundt & Teugels [356].3 is from Harrison [185].. however. [129] and Harrison [185].e. write Vi* (u) for the ruin probability etc.3) was derived by Emanuel et at.g.g. of the form Ei° p"X" with the X„ i. The formula (2.. Goldie & Griibel [167]. 134 (the time scale there is discrete but the argument is easily adapted to the continuous case).1) . 3 The local adjustment coefficient.2 is a special case of a perpetuity. Logarithmic asymptotics For the classical risk model with constant premium rule p(x) . It must be noted. not even Erlang(3) or H3.v.i. [129]. Emanuel et at. A r. Gerber [157] p. and since RT = 0 by the continuity of Brownian motion. Paulsen & Gjessing [286] found some remarkable explicit formulas for 0(u) beyond the exponential case in Corollary 1. Gerber [155]. THE LOCAL ADJUSTMENT COEFFICIENT _ Q2a2 pa 4e E 201 I.Y*p* W*(u) < ery*u = 0.p*. Delbaen & Haezendonck [104].
Let y* < So. and (for simplicity) that inf p(x) > (3µs . as will hold under the steepness assumption of Theorem 3. x * oo. it holds that f3[s] T oo.e.5) which implies inf. x>0 (3. choose c(. Proof of Theorem 3. a) = f3(B[a] .1 Assume that for some 0 < 5o < oo. When u > uo.2) such that p(x) < c(.w (u) J dt > c(3)eeu v 1 p(u+ t) .4) we assume existence of y(x) for all x. we will use the local adjustment coefficient 'y(x)..1. let p* be a in (3.e.i)eex. and that p(x) * oo. a first step is the following: Theorem 3 . (3. which in turn by Lundberg's inequality can be bounded by ery*(1E)" Hence limsup„.C*ef*". RESERVEDEPENDENT PREMIUMS and the CramerLundberg approximation V. i. e(1o+e)2 (x ) u > 00.1 ). If 60 s f 6o.. c(. B(x) > C(2)e(ao+f)x for all x. obviously O(u) can be bounded with the probability that the Cramer Lundberg compound Poisson model with premium rate p* downcrosses level uE starting from u . For the last asssertion .'y ( x)) = 0 where r.E). as solution of the equation n(x. log ?i(u) < < 00 JO . 1) and for a given E > 0. Then we have the following lower bound for the time for the reserve to go from level u to level u + v without a claim: w(u + v) . i.>o 7(x) > 0. then log u (u) In the proof as well as in the remaining part of the section . 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) . Letting first E * 0 and next ry * T 5o yields the first statement of the theorem.*(u) . 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. The steepness assumption and p(x) + oo ensure 'y(x) * So.ap(x). (3. Then lim sup u>oo u and e E''p(r) + 0.202 CHAPTER VII. 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. oo for all E > 0.1.log '(u)/u < ry*(1 . (x.
The first main result in this direction is the following version of Lundberg's inequality: Theorem 3 .3. noting that in many cases the constant C is close to 1. and in particular. or (b) Condition 3. However. For e > 0.1 only presents a first step. 3. u Obviously.(u) > so. Therefore the probability that a claim arrives ( before the reserve has reached level u + v is at least c(.. ruin will occur if the claim is at least u + v. Theorem 3 .2). (3. 2.6) The second main result to be derived states that the bound in Theorem 3.ea°/(ecf1)). 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. Bucklew [81]). the result is not very informative if bo = oo. I. THE LOCAL ADJUSTMENT COEFFICIENT 203 where c. UJU > x cannot have a much heavier tail than the claim U itself. {Rte)} defined as in (1. The rest of this section deals with tail estimates involving the local adjustment coefficient.13 below holds.4)eE" Given such an arrival. the limit is not u + oo but the slow Markov walk limit in large deviations theory (see e. If p(x) = pis constant .v.2 is also an approximation under appropriate conditions. Condition 3.0 are the same. 2 Assume that p(x) is a nondecreasing function of x and let I(u) = fo ry(x)dx.7) CIO Remarks: 1. .3 to be reasonably precise and use e` (u) as approximation to 0 (u). then Rte) = CRtie for all t so that V). 3) = (1 . one can then assume that e = 1 is small enough for Theorem 3. which essentially says that an overshoot r. Then .' (u) < eI("). Theorem 3. let 0e (u) be evaluated for the process only with 3 replaced by /0/e and U. and hence '(u) > c(4)eeuc( 2)e(do+e)u The truth of this for all e > 0 implies lim inf log V. (3.3 Assume that either (a) p(r) is a non decreasing function of r. (u) = O(u/e). Then limelog l/ie (u) = I(u). The form of the result is superficially similar to the CramerLundberg approximation.e.13 is a technical condition on the claim size distribution B.g. the asymptotics u * oo and c . by cU2.
say. J0 ^oo g(x ) dx f AX) lexp IOW (X ) bx + b u 1 exp low(x) .2. the logaritmic form of (3. we show how to rewrite the explicit solution for ti(u) in Corollary 1. RESERVEDEPENDENT PREMIUMS 4. Then y(x) = b .bu}.bx}]o + b /' oo exp low (x) . u .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).(x) dx.bx} dx .6). First. and r j 1 'Yo v(x)dx = bu  a J0 p(x)ldx = Integrating by parts.bx} dx oo exp low(x) bx dx 70 Ju r oo = b J exp low (x) .204 CHAPTER VII.4 Consider again the exponential case B(x) = eax as in Corollary 1.bx} dx fo 00 1 + [exp {/(3w(x) .(iw(x) . we consider some simple examples.exp {/33w(u) .8 in terms of I(u) when the claims are exponential: Example 3 . However. 3.1 + b f e. we get = 1+ J" AX) exp {(3w (x) .bx} dx = 1 + J0 dodx(x) exp {. 5. One would expect the behaviour in 2) to be important for the quantitative performance of the Lundberg inequality (3. it is formally needed only for Theorem 3.3.8.bx} dx 1+0. 3a Examples Before giving the proofs of Theorems 3. rather than eI(u)).3. As typical in large deviations theory.(3/p(x).
fory(x+u)dxdy ( 3. (3. we get r 00 e. oo) with drift µ(x) and variance a2 (x) > 0 at x. > lime log e = 0 and AE * 0.ev 0 O /E) J0 70 70 Yo This implies lim inf A.0. note first that the appropriate slow Markov walk assumption amounts to u. (3.2(X) = ev2(x) so that 7e(x) = 7(x)/e. 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. For Theorem 3. in the definition of AE converges to 0.BE. ry(x /b I u o e f0 °° e  e. the integral is bounded by 1 eventually and hence lim sup AE < lim sup a log 1 = 0. u .2. 191195) that 1P (U) = fu0 eI(v)dy = eI(u) follo e.e.3.10) where AE = e log 000 e. Similarly. 70 > 0 such that 7(x) < 7o for y < yo.3 in the particularly simple case of diffusions: Example 3. applying the inequality 7(x + u) > 7(x) yields immediately the conclusion of Theorem 3. In particular.f y(x)dxd y If 7(x) is increasing . THE LOCAL ADJUSTMENT COEFFICIENT and hence 205 f°° eI(v )dy .I ( u) fool. (u) = I(u) + AE ..fo 7(x)dx/Edy f .1.10 or Karlin & Taylor [222] pp.2.8) 7(x)dxdy 11 We next give a direct derivations of Theorems 3.9 ) 11000 eI(v)dy f000 e.5 Assume that {Rt} is a diffusion on [0. Choosing yo. (X) = µ(x).1/8 .I ( v )dy fo +u) dxdy .3. BE * 0. Be = e log U000 e. 0.fo 7(x) dx /E dy > av 'yo /Edy = E (1 . It is well known that (see Theorem XI.9) yields e log . IE(u) = I(u)/e. and (3.fa 7(x+u)dx/Edy o The analogue of (3.7) follows.5) is infx>o 7(x) > 0 which implies that f °O . and (3. 3..
G.. Of course..6/p* so that u 1 I (U) = bu .0..5. + Gq(u) + o(G9(u))• Gi (u) It should be noted . Ignoring 1/5 in the formula there.6 Assume that B is exponential with rate S. . that the interchange of the slow Markov walk oo is not justified and in fact. RESERVEDEPENDENT PREMIUMS The analogue of Example 3. the results are suggestive in their form and much more explicit than anything else in the literature. however . . 0. ) Note that this expression shows up also in the explicit formula for lk(u) in the form given in Example 3.6) exactly as in Example 3. . so our approach is to determine standard functions Gl (u).7o C 15 I I. Further.5) and 7* = 5 . the slow Markov limit a * 0 and the limit u walk approximation deteriorates as x becomes large.5 for risk processes with exponential claims is as follows: Example 3 . .10) holds if we redefine AE as AE = flog (j °° efo 7(x)dx/edy _ E/5 I and similarly for B. Nevertheless . 7(x) is typically not explicit. _ . I. the slow Markov walk assumption means 5E = b/c.e.206 CHAPTER VII.4. (u) representing the first few terms in the asymptotic expansion of I(u) as u + oo. G./3 1 AX dx.0/e. E+o e*O By (3. lim sup Af < lim sup c log(1 .0) = 0.5. 0 Now (3. As in Example 3. we have 5 > 7o and get lim inf AE > lime log e .. G.+1 (u) = o( 1).5. Thus 7e(x) _7(x)/e and (3. this leads to (3.7) follows just as in Example We next investigate what the upper bound / approximation aI (°) looks like in the case p(x) = a + bx (interest) subject to various forms of the tail B(x) of B. Then the solution of the Lundberg equation is y* = b . .1 3. > .Q/p*.(u) oo. I(u ) = G1(u) + .
u(logu + r7loglogu). say 1 is the upper limit and B(x) . THE LOCAL ADJUSTMENT COEFFICIENT Example 3 .C2p* C2 = (3clr( a))11'.Y . if the phase generator is irreducible ( Proposition VIII.:.8 Assume next that B has bounded support. ry* loge*+ g7loglogp*. 1) and 17 = k + 1 if B is the convolution of k uniforms on (0. Hence (3. in the phasetype case .3. 77 = 1 if B is degenerate at 1. Here B[s] is defined for all s and B[s] . More precisely.1/a). It follows from (3. I(u) Pt.1 =$ f cse8 Sn f e"B(x)dx = e8 Jo s eIB ( 1 . (3. For example.c3 logu a= 1 J 0 a + bx 1/ ( c4ul 1/° a > 1 where c3 = c2 /b.4) or gamma distributions. fu I(u) Su .3cse7*I7(77) . and hence (3. more generally. c4 = c2b 1/'/(1 .g. 2.c2 Su a dx ) Su a<1 Su . u Example 3 . x T 1.x)n1.. .7 Assume that B(x) .1) leads to . phasetype distributions (Example 1. 1. e.clxale5x 207 (3. the typical case is a = 1 which holds . B[s] = 1 + s exB(x)dx = 1 +c1SF(a) ('+o(')) (S .1) leads to (S7T N Ocp a. y = 2 if B is uniform on (0.11) with a > 0.11) that b[s] * co as s f S and hence 7* T S as p* + oo.y/s)dy sn 1 1 f ' evy'7ldy = cse8r(T7) as s T oc.12) with y > 1. This covers mixtures or convolutions of exponentials or.ry*°p*.8).cs(1 . .1/k).s)C' f "o o as s T S.
of U1 + v .u is a nondecreasing function of u. (3. ec78)2/2c7 dx C7 . 1 0 3e. . 7 * . of the increment in a small time interval [0.f.11) and (3. x f oo . g.10 Assume that p(x) is a nondecreasing function of x. this is only possible if 7o(v) 2 7o(u)• .Ul up to the first claim (here ru (•) denotes the solution of i = p (r) starting from ru(0) = u). (3.sp(u).Ote7o( u)(u.1 Cgs o"O 0 esxex2/2c7 dx = cgsec782/2 f .13) We get b[s] . I (u) c8u log u 0 where c8 = 2/c7. (b) 'y(x) <'Yo(x)• Proof That 7(x) is nondecreasing follows easily by inspection of (3..4) of the local adjustment coefficient is not the only possible one: whereas the motivation for (3.9 As a case intermediate between (3. h].log p*.(T1)) > Ee7o(u)(ul+vr»(Ti)).f.14) for the m .g.4). one could also have considered the increment ru (T1) .15) Proposition 3. By convexity of the m .ru(TI)) .r„(Ti).4) is the formula h logEues ( Rhu) .r^.208 CHAPTER VII. RESERVEDEPENDENT PREMIUMS Example 3 . Then: (a) y(x) and 7o(x) are also nondecreasing functions of x. 1 = E.e7o ( u)(ul+u r.1) . (3.2 We first remark that the definition (3. h 10.3 (B[s] .•.B[7o (u)] . Hence for u<V.(t))dt.u .12).c8 log . 3b Proof of Theorem 3. The assumption implies that ru(t) . This leads to an alternative local adjustment coefficient 7o(u) defined as solution of 1 = Ee''o(u)(vi+u .css 2%rc7eC782/2. assume that B(x) CO x2/2c7.
Hence „/.es'Yo(u)Fu(dx)} o0 e fo yo( x)dx j.16) Proof Define 411(n)(u) = P('r(u) < on) as the ruin probability after at most n claims (on = TI + • • • + Tn). We shall show by induction that (' Y'(n) (u) < e fo 'Yo(x)dx (3. we obtain „I. u We prove Theorem 3. The case n = 0 is clear since here To = 0 so that ik(°)(u) = 0. Hence 1 = EeYo(u)(U1+uru(T1)) < E. note that the assumption implies that ru(t) . (3. it is easily seen that fu x7o(y)dy < xyo (u).x)Fu(dx) 00 U efo J = o (y) dYF (dx) )+f I 11 /' / 00 e f oFu fu dx) + of u :7o(Y)dYFu(dx) 00 J u 1 l` Considering the cases x > 0 and x < 0 separately.u[70(u)] fo eyo(x)dx .11 Assume that p(x) is a nondecreasing function of x. Assume (3.(n+1) (u) efo Yo(x)dxI^"Q exyo( I u u)Fu(dx )+ J . Also. Separating after whether ruin occurs at the first claim or not.(n+l) (u) 1 .u > tp(u). THE LOCAL ADJUSTMENT COEFFICIENT For (b).4) considered as function of 7 is convex and 0 for y = 0.17) shown for n and let Fu(x) = P(U1 + u .10(b): Theorem 3.17) from which the theorem follows by letting n + oo. this is only possible if yo(u) > 7(u).Fu(u ) + J ^(n)(u .3.e70(u)(U1P(u)T1) 209 0 + 7o(u)p(u)' 0 <_ 00['Yo( u)] .1) .7o (u)p(u)• Since (3.2 in terms of 7o. Then (u) < efo Yo(x)dx. fa 7o(y)dy < u7o(u) < xyo (u) for x > u. the case of 7 then follows immediately by Proposition 3.ru(T1) < x).
15). given downcrossing occurs. Y*u /E.. define uk. 0 It follows from Proposition 3. P k.n (starting from u/n) without that 2u/n is upcrossed before ruin.2. (3.n AX). op*.n) must first downcross unl.3/e and U.n u k}1.2..E (u/n) Now as e . 0.n = ku. let Op*. y* evaluated for p* = Pk. we used also Proposition 3. the probability that ruin occurs in the CramerLundberg model with p* = pn. {RtE)} (starting from u = un. For these reasons.x/n.. W O .I. and here it is easily seen that yo(u) .e (u) = v'.10(b ) that the bound provided by Theorem 3. in .n = sup p(x).n. for either of Theorems 3.n <Z auk}l.n inf n uk1. Let Ck.11 be reasonably tight something like the slow Markov walk conditions in Theorem 3.nbe C*.3).3). (u) < I(u). ryk. To this end. (u).. However.210 CHAPTER VII. Also. 3. pk n = uk_l.n) > k =1 II v ^k n. Proof For ruin to occur.. and.11 is sharper than the one given by Theorem 3. we have chosen to work with y(u) as the fundamental local adjustment coefficient.E (u/n). in accordance with the notation i/iE (u). C*e where the first equality follows by an easy scaling argument and the approximation by (3.E (u) denote the ruin probability for the classical model with 0 replaced by . RESERVEDEPENDENT PREMIUMS where the last identity immediately follows from (3..12 lim sup4^o f log O.3 is required.n.E (u/ n) Y'E (un . Further. (un2. by €U=. x + x/n] by two classical risk processes with a constant p and appeal to the classical results (3.n.n.10(a) for some of the inequalities.2). Lemma 3.n so that n. yo(u) appears more difficult to evaluate than y(u).n) pn niE (u /n) n n_1 n. 0. the value of {R(E)} at the time of downcrossing is < unl.3 The idea of the proof is to bound { R( f) } above and below in a small interval [x . 3c Proof of Theorem 3. The probability of this is at least n n. resp.E ( u/n) ^•e.: y(u).
a( u)z.7k. B(x) (3. THE LOCAL ADJUSTMENT COEFFICIENT particular.nu /fn( 1 Ck  e. for all x .i.n <X<Uk. /' (u/n) 'T nk. Indeed . in obvious notation one has tC (x) = y(x)/e.3.E (u/n) Op•. 3 now follows easily in case (a).E (urn) < \ *I. since ry' is an increasing function of p'. also ryk. ne7k. 211 Clearly. uk_1. v. In case (b).. 11 Theorem 3.nk=1 limsupelogv).n cE (2u/n) Ck ne7k.n = sup ?'(x). y > 0 it holds that F(U>x +yIU>x) B(x + y) < F (V > y). k=1 k=1 n u _ nE7 k.nu/en(1 + where o(1) refers to the limit e .n..e.n + 0(1). i.E (u/n) OP +^p•. so that Theorem 3.18) (ii) the family of claim overshoot distributions is stochastically dominated by V. 40 Combining with the upper bound of Lemma 3. (3. *p•. 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.! (u/n) n n m 7k.F (2u/n).13 There exists a r. we need the following condition: Condition 3.nu /En) o(1))..n . (u) CIO < Letting n 4 oo and using a Riemann sum approximation completes the proof.. It follows that n log V'C (u) k =1 log Ypk. 0 with n and u fixed.2 gives 7PE (u) < eIi"i/f = lim inf Clog 0E (u) > I (u).12 completes the proof.log Ck. .19) .
^'' = E [ . P (T(E) (u.. let v < u and define T(E) (u.E (u/n .of:>2 in n(x)..V) = e71 nu/Eno(l) (using (3.v.5) and the standard formula for b(0). Write EO. (u/n .nu /EnE [e71. (R. ) (u u /n)) . The probability of ruin in between two downcrossings is bounded by Epp .nu/En0(1) . V < u/En] + P(V > u/En) (u/En .212 CHAPTER VII. infx>2u /n P(x) .^(E) (u.EV) = EiI 1 . N with EN < 1 = infx>2u/nA(x) = 0(1). Ei + E2 < e71.2y 1 ' .( . .E(E) (u. u/n) < oo] .QEU 1 .18) for the last equality). (u/n . RESERVEDEPENDENT PREMIUMS To complete the proof. where El is the contribution from the event that the process does not reach level 2u/n before ruin and E2 is the rest. u/n) < oo] E [OE (u/n .eV) • P (T(E) (u. u/n)) . T(E) (u. v ) = v . u /en 0(i) _n so that E2 < e2ryl nu/En0(1). we first note that the number of downcrossings of 2u/n starting from RoE) = 2u/n is bounded by a geometric r.1 n. u/n)) I T(E) (u.n V. (3.. For E2.n < ery1. Then Y'E (u) ^(E) (u.EV) = e. u/n) < oo) EV).R<) (u v). v ) = inf { t > 0 : R(c ) < v R) = u } .EV) = El + E2. T() (u. u/n) < oo) . u /n) < oo] l = = < E [OE (u/n .. Then the standard Lundberg inequality yields El < E?.E (0) cf.E (2u/n .
) = exp .21) to pass from u to 0. . THE LOCAL ADJUSTMENT COEFFICIENT Hence lim inf e log Ali.r.)ds + Y(R2.20) (with ic(x. s) as in (3.7) for ruin probabilities in the presence of an upper barrier b appears in Cottrell et al.u/n) < oo) CI  > u n n ryi n' i=1 Another Riemann sum approximation completes the proof.g. [89]. where the key mathematical tool is the deep WentzellFreidlin theory of slow Markov walks (see e .3EU) (3. Bucklew [81]). Typically. it might be possible to show that the limits e . one can in fact arrive at the optimal path by showing that the approximation for 0(u.=1 J An approximation similar to (3.T) = P „(info<t <T Rt < 0) via related large deviations techniques. T) is maximized over T by taking T as the time for (3. (u) 40 213 lim inf e log(Ei +E2) + logP (r(`) (u.J y(Rs)dR.7(x)) (3. 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.3. Similarly. Djehiche [122] gives an approximation for tp(u. whereas the most probable path leading to ruin is the solution of r(x) _ k (x. Comparing these references with the present work shows that in the slow Markov walk setup.J r(Rs)p(R. the results are from Asmussen & Nielsen [39]. 0 ) (= p(x) . the approximation (3. s).4) and the prime meaning differentiation w. 0 and b T 00 are interchangeable in the setting of [89].t.7) then comes out (at least heuristically) by analytical manipulations with the action integral. u Notes and references With the exception of Theorem 3.1.21) (the initial condition is r(0) = u in both cases). l o JJJ o . 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.)Ui } . they also discuss simulation based upon 'local exponential change of measure' for which the likelihood ratio is ( /'t /'t Ns Lt = exp S . Whereas the result of [122] is given in terms of an action integral which does not look very explicit.13.
214 CHAPTER VII. We should like. the exponential distribution ). For different types of applications of large deviations to ruin probabilities . see XI. e. RESERVEDEPENDENT PREMIUMS the simplest being to require b[s] to be defined for all s > 0 (thus excluding ..g. . 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.
we write Pv for the case where Jo has distribution v so that Pv = KER viPi• 215 . More precisely. on Eo = E U {A} where A is some extra state which is absorbing. 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<. refers to the case Jo = i. if a problem can be solved explicitly when the relevant distributions are exponentials.1) is 'Here as usual . that is. oo) is said to be of phasetype if B is the distribution of the lifetime of a terminating Markov process {Jt}t>o with finitely many states and time homogeneous transition rates. Typically. Note that since (1. 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. We often write p for the number of elements of E. P. A distribution B on (0. and not in other cases. F (Jt = A eventually) = 1 for all i E E 1 and where all states i E E are transient. A proper knowledge of phasetype distributions seems therefore a must for anyone working in an applied probability area like risk theory.Chapter VIII Matrixanalytic methods 1 Definition and basic properties of phasetype distributions Phasetype distributions are the computational vehicle of much of modern applied probability. if v = (vi)iEE is a probability distribution.
e.1 The phase diagram of a phasetype distribution with 3 phases. an exponential distribution with rate parameter . T is a subintensity matrix2. We now say that B is of phasetype with representation (E.216 CHAPTER VIII.1 Suppose that p = 1 and write . i. T) (or sometimes just (a. the ith component ti gives the intensity in state i for leaving E and going to the absorbing state A. tij > 0 for i 54 j and EjEE tij < 0 . a. E = {i. k}. The initial vector a is written as a row vector. the rows sum to one which in matrix notation can be rewritten as t + Te = 0 where e is the column Evector with all components equal to one. that is.2) The interpretation of the column vector t is as the exit rate vector. i.3. and we have t = Te. 0 2this means that tii < 0. t1 = /3. In particular. Equivalently. j. (1. A convenient graphical representation is the phase diagram in terms of the entrance probabilities ai. Then a = a1 = 1. B(t) = Fa(^ < t ). Thus the phasetype distributions with p = 1 is exactly the class of exponential distributions. the exit rates ti and the transition rates (intensities) tij: tj 3 aj ai i ti tk tjk FkJ ak Figure 1.0 = t11. C is the lifetime sup It > 0 : Jt E E} of {Jt}. and the phasetype distribution is the lifetime of a particle with constant failure rate /3. MATRIXANALYTIC METHODS the intensity matrix of a nonterminating Markov process.T)) if B is the Padistribution of the absorption time C = inf{t > 0 : it = A}.e. Here are some important special cases: Example 1 .
. 0 ••• 0 0 Sp1 0 0 t= 0 0 00 •. p}.1)!e Since this corresponds to a convolution of p exponential densities with the same rate S. .x i=1 Thus E _ Si 0 T 0 S2 0 0 . 00)) S s o .. 6.1. 0 SP 0 and the phase diagram is (p = 2) . . . so that the density is P E ai6ie6. 0 0 0 T= t= 0 ••• S S 0 0 0 0 0 0 ... .. . 0 S 6 Example 1.. 0 •... ..3 The hyperexponential distribution HP with p parallel channels is defined as a mixture of p exponential distributions with rates 51. the EP distribution may be represented by the phase diagram (p = 3) Figure 1..2 corresponding to E = {1... a = (1 0 0 . PHASETYPE DISTRIBUTIONS 217 Example 1. 0 0 0 0 S 6 ...2 The Erlang distribution EP with p phases is defined Gamma distribution with integer parameter p and density bp XP1 6x (p.
1)"n! aT"e. Then for i . see A. Theorem 1 . ds^ = ds' = ttlaj + tikpkj.4 (COXIAN DISTRIBUTIONS) This class of distributions is popular in much of the applied literature.aeTxe.d. Proof Let P8 = (p ^) be the sstep EA x EA transition matrix for {Jt } and P8 the sstep E x Etransition matrix for {Jt} . p:. the restriction of P8 to E. (b) the density is b(x ) = B'(x) = aeTxt. The basic analytical properties of phase type distributions are given by the following result . Then: (a) the c. 36) yields s d.t2 yt bP. MATRIXANALYTIC METHODS Figure 1. j E E. [APQ ] p.f.4 For example. dp.3 0 Example 1 . 5 Let B be phasetype with representation (E.e.f is B (x) = 1 . Recall that the matrixexponential eK is defined by the standard series expansion Eo K"/n! 3. the Erlang distribution is a special case of a Coxian distribution.218 CHAPTER VIII. and is defined as the class of phasetype distributions with a phase diagram of the following form: 1 617 ti t2 2 b2.g. the backwards equation for {Jt} (e. E t ikp kj = kEE kEE 3For a number of additional important properties of matrixexponentials and discussion of computational aspects . a. .g.3 . i. B[s] = f0°O esxB (dx) is a(sI T)lt (d) the nth moment f0°O xnB(dx) is (. (c) the m. T).1 tP1 1 Figure 1.
s j# tii i (1. For (c). for n = 1 we may put ki = Ei( and get as in (1.g. we arrive once more at the stated expression for B[s]. Part (d) follows by differentiating the m.B(x) = 1'a (( > x) = P.g.tii tii . Since 1 .s) is the m . of the initial sojourn in state i.f.5) ki = 1 + tii L jj:Ai tii (1.tii we go to A.e.5) as hi(tii + s) = ti  t ij hj.n lt .6) .p. or w.g. i. this means in vector notation that (T + sI)h = t. hj .tii is the rate of the exponential holding time of state i and hence (tii)/(tii . B(n)[0] = _ Alternatively.p. and since b[s] = ah. tij / . . define hi = Eie8S. and since obviously P° = I.5) Indeed .f.T) 't = (. the rule (A.1 ) n +l n ! a (s I + T ) .T) 1t. 1.jEE B'(x) _ cx Pxe = aeTxTe = aeTxt (since T and eTx commute). = aPxe. PHASETYPE DISTRIBUTIONS 219 That is. h = (T + sI)1t. Then h tit ti + ti3 h j . After that.s I . d" dsn a (. d8 P8 = TP8. we i w.tii and have an additional time to absorption either go to state j which has m . ti/ .f.1..12) for integrating matrixexponentials yields B[s] = J esxaeTxt dx = a ( f°°e(81+T)dx ) t a(sI . (Jx E E) = this proves (a). and (b) then follows from 1: aipF. the solution is P8 = eT8. Rewriting ( 1.n1t = (1)nn!aTn1Te (1)nn! aTne. i.. j#i jEE tijhj + his = ti. in which case the time to absorption is 0 with m .g. (1) n+1n!aT . Alternatively.f.
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.220 CHAPTER VIII. making the problem trivial.h. there are some examples where it is appealing to write T on diagonal form.6 Though typically the evaluation of matrixexponentials is most conveniently carried out on a computer. T= 2 111 so that 2 2 Then (cf. This implies that we can compute the nth moment as (1)"n! aT "e 1"n! 1 1 22 9 9 10 70 7 1 10 10 1 9 +6. Example A3. are idempotent.s. we get the density as 9 9 6 (1 1) 10 7 1 0 10 2 aeTyt = e x . see the Appendix. Consider for example 3 9 a= (2 2). MATRIXANALYTIC METHODS which is solved as above to get k = aTle. another the case p = 2 where explicit diagonalization formulas are always available. 0 Example 1."n! ( ( l 2 2 ) 17 9 0 \ 1 / 10 10 32 n! 35 6" +n!353 Similarly. One obvious instance is the hyperexponential distribution.
T) with weight hall and an atom at zero with weight 1 .e 11BIJ = 1laDD < 1.11aDD. 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). 0 Sometimes it is relevant also to consider phasetype distributions.4b for definitions and basic rules): Proposition 1. i.T).1. < 1. 5 and serves at this stage to introduce Kronecker notation and calculus (see A.7) Proof According to (A.7 If B is phasetype with representation (v. or one just lets U be undefined on this additional set. There are two ways to interpret this: • The phasetype distribution B is defective. (1.hall. and in fact one also most often there allows a to have a component ao at A. • The phasetype distribution B is zeromodified. then the matrix m. PHASETYPE DISTRIBUTIONS 1 10 7 10 221 9 6 70 7 9 10 2 +e 6x (1 11 2 2 35ex + 18e6x 35 The following result becomes basic in Sections 4. where the initial vector a is substochastic. hail = E=EE a. T) is then defined to be oo on a set of probability 1. B[Q] of B is f3[Q] = J e'1zB(dx) _ (v (9 I)(T ® Q)1(t ® I).g.f. This is the traditional choice in the literature.4. a random variable U having a defective phasetype distribution with representation (a.e a mixture of a phasetype distribution with representation (a/llall.29) and Proposition A4. i. .
Other expositions of the basic theory of phasetype distributions can be found in [APQ]. . 0 Of course. cf. Neuts. not only in the tail but in the whole distribution.222 CHAPTER VIII.. 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. the result follows (with C = (ah)(ve)). assume that T is irreducible . Schmidli. let . distributions with a rational m. the conditions of Proposition 1. MATRIXANALYTIC METHODS la Asymptotic exponentiality Writing T on the Jordan canonical form.q be the eigenvalue of largest real part of T. Here is a sufficient condition: Proposition 1. where C. let v.hve7x.f. 1. Rolski.1 of the Appendix.g. See in particular the notes to Section 6. (1. Notes and references The idea behind using phasetype distributions goes back to Erlang.8). The Erlang distribution gives an example where k > 0 (in fact. (or Laplace transform) are often used where one would now work instead with phasetype distributions. but in many practical cases. see his book [269] (a historical important intermediate step is Jensen [214]).4c). i is real and positive. x * oo. In Proposition A5. but todays interest in the topic was largely initiated by M. but the relevant T is not irreducible. Schmidt & Teugels [307] and Wolff [384]. T). Then the tail B(x) is asymptotically exponential. 77 > 0 and k = 0. h can be chosen with strictly positive component. h be the corresponding left and right eigenvectors normalized by vh = 1 and define C = ah • ve .Ce7'.f. O'Cinneide [276] gave a necessary and sufficient for a distribution B with a rational m. v. the Erlang case). Example A5. and we have eTx . cf. Using B(x) = aeTxe . it is easily seen that the asymptotic form of the tail of a general phasetype distribution has the form B(x) _ Cxkenx. B(x) . one has k = 0. 2. Lipsky [247]. All material of the present section is standard.. No satisfying . we give a criterion for asymptotical exponentiality of a phasetype distribution B.8 Let B be phasetype with representation (a. here k = p1). In older literature.8 are far from necessary ( a mixture of phasetype distributions with the respective T(') irreducible has obviously an asymptotically exponential tail.g. the text is essentially identical to Section 2 of Asmussen [26].F.8) Proof By PerronFrobenius theory (A.
.f. but nevertheless.i. Let U1. If B is exponential with rate 0. 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. For this reason.t. Then the renewal density exists and is given by u(x) = ae(T+ta)xt. with common distribution B and define4 U(A) = E# {n = 0. RENEWAL THEORY 223 algorithm for finding a phase representation of a distribution B (which is known to be phasetype and for which the m. is Markov and has two types of jumps .2.: U1 + . however.. as the lifetimes of items (say electrical bulbs) which are replaced upon failure. the problem has an algorithmically tractable solution if B is phasetype: Theorem 2. + U0 is 0 .. .r.d.1 of the Appendix. n=O We may think of the U. be i... . +UnEA} 00 = EEI(U1 +. but is in part repeated below. we denote the density by u(x) and refer to u as the renewal density.. U2. 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 +. Jt={Jt?ul}.. The explicit calculation of the renewal density (or the renewal measure) is often thought of as infeasible for other distributions.g. A related important unsolved problem deals with minimal representations: given a phasetype distribution ... known.. the renewals form a Poisson process and we have u(x) = 0.. if U is absolutely continuous on (0. oo) w. . U1<t < U1+U2. JtJt1) Then { 0<t<U1 .T).1 Consider a renewal process with interarrivals which are phasetype with representation (cr.+UnEA). we refer to U as the renewal measure. (2.1. or the density is available ) is.1) Proof Let {Jtk)} be the governing phase process for Uk and define {Jt} by piecing the { J(k) } together.. Lebesgue measure.. and U(A) is then the expected number of replacements (renewals) in A.
e. IIafl < 1. u The argument goes through without change if the renewal process is terminating.224 CHAPTER VIII. 2.T). Proof Just note that { it } is a governing phase process for the lifetime. the density is veTxt = B(x)/µB.T) where vt = ae (T+ta)t . Then: (a) the excess life t(t) at time t is phasetype with representation ( vt. Hence ( 2. see Fig. T).. + Uit_1 where s. This is defined as U1 + . Then the lifetime is zeromodified phase type with representation (a. and the distribution of Jx is ae ( T+t«)x. . i.1 Corollary 2. since Uk = oo with probability 1 .e. that is.3 Consider a renewal process with interarrivals which are phasetype with representation (a. and the jumps of the first type are governed by T. u Returning to nonterminating renewal processes .1) follows by the law of total probability. .2 Consider a terminating renewal process with interarrivals which are defective phasetype with representation (a. the lifetime of the renewal process.U1 U3 U2 U3 U4 Figure 2. However. i.T) where v = aT1 /µB.1) remains valid for that case. which is phase type with representation (v. fi(t) U2 U1 . define the excess life e(t) at time t as the time until the next renewal following t. as the time of the last renewal.IIBII which is > 0 in the defective case.1. is the first k with Uk = 00. MATRIXANALYTIC METHODS of the last type from i to j occurs at rate tiaj .. and hence ( 2. this is welldefined. Corollary 2. B is defective . (b) £(t) has a limiting distribution as t * oo. the phasetype assumptions also yield the distribution of a further quantity of fundamental importance in later parts of this chapter . Hence the intensity matrix is T + ta. which is ti in state i.T + ta). Equivalently. and let µB = aTle be the mean of B. The renewal density at x is now just the rate of jumps of the second type.
q2.2.e. According to Example A3.6. Here are two different arguments that this yields the asserted expression: (i) Just check that aT1/µB satisfies (2. Hence in (b) it is immediate that v exists and is the stationary limiting distribution of it. The renewal density is then aeQtt = (al a2) ( 7i 7"2. Next appeal to the standard fact from renewal theory that the limiting distribution of e(x) has density B(x)/µB. T1 and eTx commute. the unique positive solution of ve = 1. (2. we get B(x) aeTxe aT1eTxTe µB µB PB = veTxt.e. RENEWAL THEORY 225 Proof Consider again the process { Jt } in the proof of Theorem 2. we first compute the stationary distribution of Q. u Example 2 .2) v(T + ta) = 0. 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).T) where vt is the distribution of it which is obviously given by the expression in (a).4 Consider a nonterminating renewal process with two phases. Al. = qz ql (x1 xz) = ql + qz ql + q ' and the nonzero eigenvalue A = ql .2): aT1 e = AB = 1 µB µB a + aT'Tea aT1(T + ta) µB PB a + aea a + a µB µB =0. hence e(t) is phasetype with representation (vt. The time of the next renewal after t is the time of the next jump of the second type.) ( t2 ) .1. i. cf. (ii) First check the asserted identity for the density: since T.
Then Q= 0 55 )+(1o)=( j ad ). Hence 7r = (1/2 1/2). Then _ Q Hence 51 0 0 52 + 51 52 _ 5152 51a2 ) (al a2) 52a1 62a1 Slat + 52a1 51a2 51a2+52a1 A = 51a2 . t1B 0 Example 2 .a27rl) (tl .tl) 7r2t2 + eat (a17r2 .5 Let B be Erlang(2). MATRIXANALYTIC METHODS e. A = 25.4 yields the renewal density as u(t) = 2 (1 .`t (al a2) + C 11 172 ir12 / \ t 2 ) r1 (7r1 7r2) ( t2 7rltl + J + eAt (al a2) ( 71(t2 . The present treatment is somewhat more probabilistic. .6 Let B be hyperexponential.(biaz + aza. and Example 2.t2) 1 + eat (a17r2 . )t (51 .52) 25152 51x2+5251 51a2+5251 Notes and references Renewal theory for phasetype distributions is treated in Neuts [268] and Kao [221].52a1.4 yields the renewal density as u(t) = 5152 e.e2bt) 13 Example 2 .t2) . and Example 2.a27r1) (t1 .226 CHAPTER VIII.
r(u) the time of ruin with initial reserve u. Next. which occurs at rate ti. the Markov processes representing ladder steps can be pieced together to one {my}. T) where a+ is given by a+ = . cf. For (b). (b) V.3. Now just observe that the initial vector of {mx} is a+ and that the lifelength is M.e. T(0) < oo) the ladder height distribution and M = supt>o St. The stars represent the ladder points ST+(k).2. T). G+(. and M is zeromodified phasetype with representation (a+. with 0 denoting the Poisson intensity. T). Corollary 2. and if there is a subsequent ladder step starting in j whic occurs w. Within ladder steps. represent the maximum M as the lifetime of a terminating renewal process and use Corollary 2. Thus the total rate is tip + tia+. and rewriting in matrix form yields the phase generator of {my} as T + ta+. Then each claim (jump) corresponds to one (finite) sample path of the Markov process. Considering the first.(u) = a+e(T+tQ+)u Note in particular that p = IIG+II = a+e. Proof The result follows immediately by combining the PollaczeckKhinchine formula by general results on phasetype distributions: for (a).f3aT1. itself phasetype with the same phase generator T and the initial vector a+ being the distribution of the upcrossing Markov process at time ST+_. . a+j. The essence is contained in Fig. {St} the claim surplus process. 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. Since the results is so basic. use the phasetype representation of Bo.) = F(ST(o) E •. We asssume that B is phasetype with representation (a. B the claim size distribution. marked by thin and thick lines on the figure. T + to+). we shall. however. we see that the ladder height Sr+ is just the residual lifetime of the Markov process corresponding to the claim causing upcrossing of level 0.3.1 on the next page.p. 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. Corollary 3. Here we have taken the terminating Markov process underlying B with two states. i.1 Assume that the claim size distribution B is phasetype with representation (a. add a more selfcontained explanation of why of the phasetype structure is preserved. 3. Then: (a) G+ is defective phasetype with representation (a+.i.
 Figure 3.QaT1.1 .. 3e3x + . T = (3 .M {mx} ST+(2)  S . MATRIXANALYTIC METHODS t .1 This derivation is a complete proof except for the identification of a+ with . This is in fact a simple consequence of the form of the excess distribution B0. 7e7x 2 2 Thus b is hyperexponential (a mixture of exponential distributions) with a (2 2 ).228 CHAPTER VIII. 0 Example 3.3.7)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 .Q = 3 and b(x) = . see Corollary 2...2 Assume that .t t d kkt S.
For the compound Poisson model. Fig. this was obtained in Section 3. T) for some vector a+ = (a+. but there the vector a+ is not explicit but needs to be calculated (typically by an iteration). with A denoting the interarrival distribution and B the service time distribution. cf. 4 The renewal model We consider the renewal model in the notation of Chapter V. see Section 6. 0(8) (u) (recall that z/i(u) refers to the zerodelayed case and iY(8) (u) to the stationary case). but that such a simple and general solution exists does not appear to have been well known to the risk theoretic community.2 are taken from Gerber [157]. see Shin [340]. That is.6). 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 .1): Proposition 4. It is notable that the phasetype assumption does not seem to simplify the computation of finite horizon ruin probabilities substantially. We assume p = PB/µA < 1 and that B is phasetype with representation (a. (a) G+ is of phasetype with representation (a+.4.T).1 can be found in Neuts [269] (in the setting of M/G/1 queues. 3. THE RENEWAL MODEL This is the same matrix as is Example 1. 3.^(u) = a+e( T+ta+)ue = 24eu + 1 e6u 35 35 0 Notes and references Corollary 3.4. We shall derive phasetype representations of the ruin probabilities V) (u). For an attempt.and Markovmodulated models.6.1 which does not use that A is exponential) by noting that the distribution G+ of the ascending ladder height ST+ is necessarily (defective) phasetype with representation (a+. The result carries over to B being matrixexponential. where a+ is the (defective) . The parameters of Example 3. if we define {mz} just as for the Poisson case (cf. the discussion around Fig. see Stanford & Stroinski [351] . For further more or less explicit computations of ruin probabilities. T).1 In the zerodelayed case. the duality result given in Corollary 11. his derivation of +'(u) is different.j). In the next sections. and the argument for the renewal case starts in just the same way (cf. we encounter similar expressions for the ruin probabilities in the renewal.
where B0 is the stationary excess life distribution corresponding to B. In fact. CHAPTER VIII. Nevertheless. Hence by Theorem 11.*} from {St+y . Then .4 Consider the renewal model with interarrival distribution A and the claim size distribution B being of phasetype with representation (a.T). the calculation of the first ladder height is simple in the stationary case: Proposition 4. but with initial distribution a rather than a+. Since the conditional distribution of my given T1 = y is ae4y.Sy} in the same way as {mx} is defined from {St}.T)• Proposition 4.1). with intensity matrix Q given by Q = T + to+.230 distribution of mo. the Palm distribution of the claim size is just B.T). obviously mo = m. Also. G(') = pBo.*'} is Markov with the same transition intensities as {mx}. But by Corollary 2.6. (c) {mx } is a (terminating) Markov process on E. 4.3 a+ satisfies a+ = V(a+). (4.1) Proof We condition upon T1 = y and define {m.1. the form in which we derive a+ for the renewal model is as the unique solution of a fixpoint problem a+ = cp(a+).3.2 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). MATRIXANALYTIC METHODS (b) The maximum claim surplus M is the lifetime of {mx}. Then {m. it follows by integrating y out that the distribution a+ u of mo is given by the final expression in (4. Fig. Proof Obviously.5. which for numerical purposes can be solved by iteration. B0 is phasetype with representation (aT1/µa. where a(8) = aT1/PA. where u w(a +) = aA[T + to+) = a J0 e(T+t+)1A(dy). We have now almost collected all pieces of the main result of this section: Theorem 4 . cf. The key difference from the Poisson case is that it is more difficult to evaluate a+.
1(b).1). The term tf3 in cp(i3) represents feedback with rate vector t and feedback probability vector (3.0) is an increasing function of /3.0. It remains to prove convergence of the iteration scheme (4. (4.^(u) = a+e ( T+ta+)xe.e.•..1/pA.2.1 .M. by a+ = lim a +n) where a+°) . . .4. a+l ) = cp (a+°)) . (4.. a+2) = ^p (a+l)) .3). i y ^ T1= y `•r Figure 4.1 by noting that the distribution of mo is a+. Hence ^p(.3) Proof The first expression in (4. In particular .2 ) follows from Proposition 4. Furthermore . a+) > 0 = a+o) implies a+) _ (a+) > W (a+)) = a+) . i. a+ can be computed by iteration of (4. I {mx} .. the maximum claim surplus for the stationary case has a similar representation as in Proposition 4. only with initial distribution a(*) for mo. and that this is given by Proposition 4. The second follows in a similar way by noting that only the first ladder step has a different distribution in the stationary case.3) (defined on the domain of subprobability vectors . THE RENEWAL MODEL 231 . thus ..^(8)(u) = a ( 8)e(T+ta +) xe.2) where a+ satisfies (4.1) and a(8) _ aT.
7+ ].T)1t.T)'t • A[s] (4. Then e4'h = e82h and hence sh = Qh = (T + taA[Q])h = Th + A[s]tah. the normalization is equivalent to F(s) = 1. Thus . MATRIXANALYTIC METHODS and (by induction ) that { a+ n) } is an increasing sequence such that limn.P[s] = A[s]B[s].4).4) makes sense and provides an analytic continuation of F[•] as long as s ¢ sp(T). which links together the phasetype setting and the classical complex plane approach to the renewal model (see further the notes).1. Obviously. Then (4. To prove the converse inequality. Let Fn = {T1 + • • • + Tn+1 > r+}be the event that {my} has at most n arrivals in [T1. Then s is an eigenvalue of Q = T + ta+ if and only if 1 =.1 arrivals (n arrivals are excluded because of the initial arrival at time T1 ).} can contain at most n . F[s] being interpreted in the sense of the analytical continuation of the m.5 Let s be some complex number with k(s) > 0. Proof Suppose first Qh = sh.T1. s ¢ sp(T).f. To this end. 0 0 We next give an alternative algorithm. . (4. so to complete the proof it suffices to show that &+ < a+) for all n. n) &+n) T a+. Similarly. let F be the distribution of U1 .5) Since s $ sp(T). Then each subexcursion of {St+Tl . 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).T)It.232 CHAPTER VIII.4. In that case.. and hence we may assume that h has been normalized such that ahA[s] = 1. we use an argument similar to the proof of Proposition VI. Assume the assertion shown for n .2. Thus by (4.g. Then F[s] = a(sI . a+ ) exists . Theorem 4. However.4) whenever EeR(S)U < oo.) = P(mTl = i. limn4oo a ) < a+. with B[s]. this implies that ahA[s] # 0.5) yields h = (sI . the corresponding right eigenvector may be taken as (sI . 0 = a+) < a+ yields a+) _ (a+0)) (a+) = a+ (n and by induction that a(n) < a+ for all n . both quantities are just 0 . and let &+".ST. (4. Fn ). For n = 0.
.' that the equation F(s) = 1 has d distinct roots p1.T) = 1 ata+ = a+. . As in Corollary 4. Notes and references Results like those of the present section have a long history... .T)lt = sh.. we have IG_ [s] I < 1 . This gives d roots 'y. pdhd. . Pd with corresponding eigenvectors hl. and the solution is . hd.4.5(c) means that a+(sI T)1t = 1.p1i . Q = CD1 where C is the matrix with columns hl. 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.6... 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. Then G+ is phase. Further. the classical algorithm starts by looking for roots in the complex plane of the equation f3[y]A[ry] = 1. Pd in the domain ER(s) > 0 . D that with columns p1 hl. . Let d denote the number of phases. in turn..6 Suppose u < 0. and define hi = (piI . Given T has been computed.5. we get at a(Q . . letting vi be the left eigenvector of Q corresponding to pi and normalised by vihi = 1 .type with representation (a+. t(ry) > 0... THE RENEWAL MODEL 233 Suppose next F(s) = 1. .. T) with a+ = a(QT)/at. W v M(d) in the notation of Chapter V).T)lt + t = s(sI ...6) i=1 i=1 Proof Appealing to Theorem 4.. Hence with h = (sI T).. In older literature . (4.. This immediately implies that Q has the form CD1 and the last assertion on the diagonal form . 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. hd. the matrix Q in Theorem 2. 0).1 has the d distinct eigenvalues . Q has diagonal form d d Q = dpivi®hi = dpihivi. Corollary 4.T)It. .. and hence by the WienerHopf factorization identity (A..9) we have G+[s] = 1 which according to Theorem 1.lt we get Qh = (T + to+)h = T(sI . yd satisfying R(ryi) > 0.
and appears already in some early work by Wallace [377]. Neuts and his students. The arrival rate in background state i is a.g. with representation say (a(' ).. In queueing theory. The matrix. The exposition here is based upon [18]. T('). the fixpoint problems look like R=Ao+RAI+R2A2+ .type assumption . The distribution of W comes out from the approach but in a rather complicated form .type assumptions are basic. the background Markov process with p states is {Jt}. The solutions are based upon iterations schemes like in Theorem 4.) d (see.exponential form of the distribution was found by Sengupta [335] and the phasetype form by the author [18]. involving . Here phase. similar discussion appears in Kemperman [227] and much of the queueing literature like Cohen [88]. which contains somewhat stronger results concerning the fixpoint problem and the iteration scheme. Asmussen & O'Cinneide [ 41] for a short self. is phasetype. see Neuts [269]. The number of elements of El=> is denoted by q. and the distribution of an arrival claim is B. [119]. For further explicit computations of ruin probabilities in the phasetype renewal case . see Dickson & Hipp [118]. e.4.F. We assume that each B. 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. but the models solved are basically Markov chains and processes with countably many states ( for example queue length processes ).234 then in transform terms CHAPTER VIII.contained derivation). MATRIXANALYTIC METHODS d F 1 + a J e°" ip(u) du = Ee°w = 11(t.. 5 Markovmodulated input We consider a risk process {St } in a Markovian environment in the notation of Chapter VI. an alternative approach (the matrixgeometric method ) has been developed largely by M. E(t)). [270] and Latouche & Ramaswami [241]. For surveys . a pioneering paper in this direction is Tacklind [373]. In risk theory. It turns out that subject to the phase. where R is an unknown matrix. That is . the ruin probability can be found in matrixexponential form just as for the renewal model. whereas the approach was introduced in queueing theory by Smith [350]. the intensity matrix is A and the stationary row vector is ir .. starting around in 1975. Numerical examples appear in Asmussen & Rolski [43].
1. •.5. Section 5b then gives a representation along the lines of Theorem 4. This calculation in a special case gives also the ruin probabilities for the Markovmodulated risk process with phasetype claims. However. (a) 0 0 ♦ o ° tl ♦ • 0 0 o } o o (b) 0 } ♦ • 0 o f o Figure 5.Vt)} obtained by time reversing the I component. The two environmental states are denoted o. 5. say with slope r(i) on intervals where It = i. 5. The version of the process obtained by imposing reflection on the V component is denoted a Markovian fluid and is of considerable interest in telecommunications engineering as model for an ATM (Asynchronuous Transfer Mode) switch. has states o. O.6.4. states . MARKOVMODULATED INPUT 235 some parameters like the ones T or a+ for the renewal model which need to be determined by similar algorithms. and the one E(•) for B. The key unknown is the matrix K. the analysis involves new features like an equivalence with first passage problems for Markovian fluids and the use of martingales (these ideas also apply to phasetype renewal models though we have not given the details). The stationary distribution is obtained by finding the maximum of the Vcomponent of the version of {(It. The connection between the two models is a fluid representation of the Markovmodulated risk process given in Fig. Vt)}t>o such that {It} is a Markov process with a finite state space F and {Vt} has piecewiese linear paths.2.1 In Fig. 5a Calculations via fluid models. Diagonalization Consider a process {(It. for which the relevant fixpoint problem and iteration scheme has already been studied in VI. We start in Section 5a with an algorithm involving roots in a similar manner as Corollary 4.1. the phase space E(°) for B. p = ql = Q2 = 2.
1 A complex number s satisfies 'A+ (f3i(Bi[s] .1) if and only if s is an eigenvalue of E. Recall that in the phasetype case. in the fluid model Eel'. t. 4}.31a(l) (/3i)diag .Vt)} is then obtained by changing the vertical jumps to segments with slope 1. V. A claim in state i can then be represented by an E()valued Markov process as on Fig.1))diag ) a = sa and the eigenvector b = .1))diag + sII = 0 (5. a) of {It}.1(a). F is the disjoint union of E and the Eli). In the general formulation . Bi[s] = a(i)(T(i) + sI)it('). 5. whereas Ee8s' = oo for all t and all s > so where so < oo. 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. F = E U { (i. If s is such a number. Let E denote the matrix .92a(2) 0 0 T(2) 0 0 0 f33a(3) 0 0 T(3) with the four blocks denoted by Ei„ i. 5. < oo for all s. The intensity matrix for { It} is (taking p = 3 for simplicity) I A . MATRIXANALYTIC METHODS 4.1(b) {(It . The fluid model on Fig . •. o. a) = 1. a E E(i) } . we have more martingales at our disposal. i E E. '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.236 CHAPTER VIII. j = 1. Eli) + Proposition 5. consider the vector a satisfying (A + (13i(Bi[ s] . Thus F = {o.(Ni)diag r(i. 4. This implies that in the fluid context. a) : i E E. Second. of E into components indexed by E.A 0 Or 1A/ _ t(i) 0 t(2) 0 0 0 0 0 t(3) 0 T1 0 0 0 . First. r(i) _ 1. resp. corresponding to the partitioning + Epp). 4. 2.
For the assertions on the eigenvectors. E(1) + + E(P).sI. t(1) 0 0 then also 0 t(2) 0 . d = (sI E22)1E21a = E ai(sI . Then E21c+E22d = E21a . and let d = (sI .E22)1 E21) a = 0. d correspond to the partitioning of b into components Proof Using the wellknown determinant identity Ell E12 E21 E22 E22 I ' I Ell .A .sI 0 0 0 T(2) .T('))1t(i) .sI 0 0 0 T(3) . assume that a is chosen as asserted which means (Ell . it follows that if Qla(1) 0 0 . with Eii replaced by Eii . where c.sI 0 0 t(3) 0 0 = 0. resp .E22)1 E21a E21a .sI ()3i)diag .(sI .sI) (sI .E21a + sd = sd. it follows that Ell E12 ( E 21 E22) (d) = s 1 d I .5.E22)1 E21a. Noting that E11c + E12d = se by definition. c = a. iEE (a> of 0* 1 AI. Then (up to a constant) c = a.32a(2) (/3i)diag .1).E22 .A . 0 .sI+ ((3ia(i)(T(i) .Nla(1) 0 0 T 1.sI)1t)) iag I = 0 which is the same as (5.sI + E12 (sI .E12E22 E21 I . MARKOVMODULATED INPUT 237 indexed by E.
v)=inf{t >0:Vtu orVt=. Proof Writing Or'Alb( v) = svb( v) as (AI . define w(u. . I' i( V P2 (w (u) < oo. j. w(u)=inf{t >O:Vtu}..upi(u.j)c v . we first look for the negative eigenvalue s of E = I 0 I which is s = ry with yy = b . .v.( u.. it follows by Proposition II. To determine 0 (u). Example 5 .2 Assume that E = Or 'Al has q = ql + + qp distinct eigenvalues si.. .. Example 5 . a) = Pi (Vw(u.v) = j).3 Consider the Poisson model with exponential claims with rate 5.j. v. MATRIXANALYTIC METHODS Theorem 5...4 that {e"1b(v) is a martingale . v) yields C{V) = e8 . d("))1 e. a)). q.v) = Optional stopping at time w (u. Then .a)d^ ).. v. pi(u. .v) = v) I. j.sv)b(v) = 0. Letting v ^ oo and using Rsv < 0 yields e8'u = Epi(u..v) = (j. s2 are the negative eigenvalues of Al +01 A1 E _ A 2 b1 0 52 A2 +32 0 . u) Iw(u. j. Thus 0(u) = esu/d = pe7 ° as u should be.Q. < 0 and let b(v) = I d(„)) be the right eigenvector corresponding to s. Then we get V)i (u) as sum of two exponential terms where the rates s1.v}.. We can take a = c = 1 and get d = (s + b)16 = 5/(3 = 1/p. v > 0./' u = e' (esiuc ( 1) . . B2 are both exponential with rates 51 i b2. the result u follows.. For u.a Solving for the pi(u. a) = (j.5. j. j..4 Assume that E has two states and that B1. e89uc(e)) (d(1) .. . a )d(a + e8 °vpi (u . sq with $2s.pi(u. j) pi( u .238 CHAPTER VIII.. a). a) and noting that i1 (u) = >I j.. c j. v. a)).v) = = p i( u . Iw(u.O. Here E has one state only. w(u. v = 1.
3) . 8^')IT(j)) where e 3^') =. the Pidistribution of M is phasetype with representation (E(1) + + E(P).( 2.3j eye. j.h. Proof We must show that G+ (i.5 G+(i. (y.33(e = 0 a(j))(K ®T ( j))(ej (9 I).Qj eie 0 f e (j) T(') x T(j)y ej a e dx e e 00 00 eKx ® e T(')' dx (ej (& I)e T(')ye eKa®T(')x dx (ej (9 I)eT(') Ye e(i)eT(')ye. •) is phasetype with representation (E(i). we get the following phasetype representation for the ladder heights (see the Appendix for the definition of the Kronecker product 0 and the Kronecker sum ®): Proposition 5.6 For i E E. MARKOVMODULATED INPUT 239 5b Computations via K Recall the definition of the matrix K from VI.2.y = to B k7 j # k In particular. U) where t(j) + t(j)O(j j = k uja.5. dx)Bj(y . 0 Theorem 5 . j.k. 9('). (5. is 0 /3 f R(i . j.xxej • a 00 oo el . oo)) j)ye. according to VI.b (u) = Pi(M > u) = 9(i)euue. In terms of K.s.x) 00 f ° (') (j) eT (yy)edx . (') a T( However .2) the l. i.
6 Matrixexponential distributions When deriving explicit or algorithmically tractable expressions for the ruin probability.. Associated with each ladder step is a phase process. 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.g. with phase space EU> whenever the corresponding arrival occurs in environmental state j (the ladder step is of type j). which occurs w. An alternative characterization is that such a distribution is matrixexponential. +bn0i1 0n +a10n1 +. For j = k.k y. However. i. a m.) which is rational. Numerical illustrations are given in Asmussen & Rolski [43].p. (6.2) . Bk7 . t) is the representation of the matrixexponential distribution/density): Proposition 6.e.. some square matrix T and some column vector t (the triple (a. say. i. a) to (k.. T. we have sofar concentrated on a claim size distribution B of phasetype. 0 1)'. bn1 bn). Starting from Jo = i. equivalently.5). the initial value of (i. intensity matrix U.y) to occur when j # k. This yields the asserted form of uja. For a transition from (j. Then b*[0] is rational if and only b(x) is matrixexponential. and lifelength M. +aii10+anI then a matrixexponential representation is given by b(x) = aeTxt where a = (b1 b2 . we have the additional possibility of a phase change from a to ry within the ladder step. . oo) and b* [0] = f °O eBxb(x) dx the Laplace transform. a) is obviously chosen according to e(`).1 Let b(x) be an integrable function on [0. and it just remains to check that U has the asserted form.. which occurs at rate t^^7.. which occurs at rate t(i). t = (0 0 .. the ratio between two polynomials (for the form of the density. MATRIXANALYTIC METHODS Proof We decompose M in the familiar way as sum of ladder steps .240 CHAPTER VIII..f.e..2. and a new ladder step of type k must start in phase y. see Example 1. u Notes and references Section 5a is based upon Asmussen [21] and Section 5b upon Asmussen [17]. the current ladder step of type j must terminate. that the density b(x) can be written as aeTxt for some row vector a. Furthermore. Piecing together these phase processes yields a terminating Markov process with state space EiEE E('). if b* [0] = b1 +b20+b302 +.
. 0 1 an an1 an _2 . Example 6 . t= 0 . cannot be phasetype.2). since 1 + 4ir2 03 + 302 + (3 + 47x2)0 + 1 + 47r2 it follows by (6. b(x) > 0 for x > 0.T)1 is so.(6..6.47r2 3 .. s = c/ 2 . we can always obtain a real one (a.2 A remarkable feature of Proposition 6.1 0 .1) as E 1 c. then b*[0] = a(0I T)1t which is rational since each element of (01 . matrixexponentiality implies a rational transform.. 0 0 0 0 1 0 0 .47x2 3 1 0 . For a proof.3) was suggested by Colm O'Cinneide. personal communication).3 A set of necessary and sufficient conditions for a distribution to be phasetype are given in O'Cinneide [276]. . T= 0 0 1 . Namely. Thus. 1 . (6.an_3 an _ 4 . .h.3) 0 0 0 0 0 . t). where c = 1 + 1/47r 2. but as follows from Proposition 6.2). T. One of his elementary criteria. Writing b(x) = c(e( 2ni1 ) y/2 . namely to asssume the roots 6l.s. of (6. ..1. The converse follows from the last statement of the theorem. a2 a1 Proof If b(x) = aeTxt./(0 + bi).. see Asmussen & Bladt [29] (the representation (6. (6. (6.1 is that it gives an explicit Laplace tranform inversion which may appear more appealing than the first attempt to invert b* [0] one would do.3) that we can take 0 1 0 0 a= (1 + 47r2 0 0). bn of the denominator to be distinct and expand the r. . S.1 0 0 )3 = (111). .. u giving b(x) = E 1 ciebiz/bY. MATRIXEXPONENTIAL DISTRIBUTIONS 241 T = 0 1 0 0 0 .e(tai1)x/2 + e'T) it follows that a matrixexponential representation ()3. u Remark 6. s) is given by 27r i . S = f c/2 0 21ri . 0 0 . shows that the distribution B with density b(x) = c(1 cos(21r x))ex.4) 0 0 1 c This representation is complex...
we take as starting point a representation of b* [0] as p( O)/q(9) where p. MATRIXANALYTIC METHODS Example 6 . we have represented ti* [0] as ratio between polynomials (note that 0 must necessarily be a root of the numerator and cancels). and present two algorithms for calculating '(u) in that setting. we shall only consider the compound Poisson model with arrival rate 0 and a matrixexponential claim size distribution B. Corollary 111. 7 + 155ex b(x) Then it is known from O'Cinneide [276] that b is phasetype when 6 > 0.6) The remarkable fact is. t) a phasetype representation with a the initial vector.4 This example shows why it is sometimes useful to work with matrixexponential distributions instead of phasetype distributions: for dimension reasons . then 5(u) = a+e(T+t+)uTle where a+ = /3aT1.5) Thus. We recall (see Section 3. For the second algorithm.1 shows that a matrixexponential representation can always be u obtained in dimension only 3 independently of J. 0.5 (6. that despite that the proof of (6. we use a representation (a.6) holds true also in the matrixexponential case. Then (cf. q are polynomials without common roots.1 to get i (u) = f3esus.1)2 + 6). T the phase generator and t = Te.3. As for the role of matrixexponential distributions in ruin probability calculations. and can use this to invert by the method of Proposition 6. t) of b(x). Consider the distribution with density = 15 ((2e2x . and that the minimal number of phases in a phasetype representation increases to 0o as 5 . . 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. leading to matrix calculus in high dimensions when b is small. T. recall that t = Te) that if B is phasetype and (a.6) in Section 3 seems to use the probabilistic interpretation of phasetype distribution in an essential way. (6.4) the Laplace transform of the ruin probability is /g(e)PO 0*[e] _ /' eeu^G(u)dU = 0 9(/3a0p(9)ap (9)/q(9)) . T. then: Proposition 6.242 CHAPTER VIII. For the first. (6.
1 = ^(T1 + ( 91T)1). b+ = a +(BI . (6.1t du = .1 = A1 .T)1 + (6I .to+)1 = (BI . 519) (A + UBV ).to+)1T .t.T)1 so that b* b** b** a+(9I .T .T). Presumably. the assertion is equivalent to a+(BI .6.B=land V=a+. MATRIXEXPONENTIAL DISTRIBUTIONS 243 Proof Write b* = a(9I .T .1 + b+ = b++ 1 .T . we get (91.a+(9I . we get b+ = 0aT1(9I T).T)1ta+(OI . . Then in Laplace transform formulation .1 + 82 (9I .b* (6.1t = b* .7) 9( cf.T)1T1t.T)1t. Now. (91. (6. From the general matrix identity ([331] p.T)1t ( l .to+)1T .'t. xb(x) dx = aT2t.A .6b* .1BVA1.T)1 (91. this can be verified by analytic continuation from the phasetype domain to the matrixexponential domain .6).T)1 + 1 ib* (91.1UB(B + BVA1UB).1t = f3a (0I T)1T1t . since (91T)1T . U =.5 ). with A = 91T. b+ = a+(9I . but we shall give an algebraic proof.T)1 J0 00 b(x) dx = f aT1t.T)1T 2 = and 1 = AB IT2 + 82T .T)1t)1a +(9I .
8 a(T1 + (01.3a (1 0 T 2 + 1 T 102 (9I + 02 1 T)1) t P + 7. MATRIXANALYTIC METHODS .5 is similar to arguments used in [29] for formulas in renewal theory. premium rate p(r) at level r of the reserve {Rt} and claim size distribution B which we assume to be of phasetype with representation (E. see the Notes to VII. to piece together the phases at downcrossing times of {Rt} (upcrossing times of {St}) to a Markov process {mx} with state space E. From this it is straightforward to check that b**/(b+ .1.7). a key early paper is Cox [90] (from where the distribution in Example 6. A key tool is identifying poles and zeroes of transforms via WienerHopf factorization. see Asmussen & Bladt [29].82b*. 0 Notes and references As noted in the references to section 4. (for some remarkable explicit formulas due to Paulsen & Gjessing [286]. of (6. the ruin probability(u) was found in explicit form for the case of B being exponential. See Fig. T).T)1)t = 8 (1 .la.T)1T2t . cf. /3aT1(0I . 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.1) is the same as the r.8. 3. 7 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate 0. . Lipsky [247] and Asmussen & O'Cinneide [41].1. which is selfexplanatory given Fig.h.1t = /3a (9I . but the argument of [286] does not apply in any reasonable generality).T)1T. Much of the flavor of this classical approach and many examples are in Cohen [88].1. VII.244 CHAPTER VIII. 7a Computing O(u) via differential equations The representation we use is essentially the same as the ones used in Sections 3 and 4. 7. In Corollary VII.3 is taken).1. The proof of Proposition 6. a.b*).s. some key early references using distributions with a rational transform for applied probability calculations are Tacklind [373] (ruin probabilities) and Smith [350] (queueing theory). For expositions on the general theory of matrixexponential distributions.
Note that in general >iEE Vi (U) < 1. 0 < t < u. Figure 7.u)e = A(u)e (7. Ai(t) = P(mt = i).t2) be the matrix with ijth element P (mt2 =j I mtl = i).tl < t2 < u.I] I 80 { tq f Q(v) dvl t1 1 . i. Proof The first statement is clear by definition. Given the v(t) have been computed. Also. we obtain V)(u) = P(m„ E E) = v(u)P(0. t2) = exp where Q(t) = ds [P(t. though still Markov.1) where A(t) = v(u)P(0.e. >iEE Vi (U) is the ruin probability for a risk process with initial reserve 0 and premium function p(u + •). RESERVEDEPENDENT PREMIUMS 245 Rt l0 u . t + s) . t) is the vector of state probabilities for mt. in contrast to Section 3. P(tl. is no longer timehomogeneous.7.t)). By general results on timeinhomogeneous Markov processes. Since v(u) = (vi(u))iEE is the (defective) initial probability vector for {m8}.1z I.1 A(0) = v(u) and A'(t) = A(t)(T + tv(u . Let P(tl. 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. O<.1 The difference from the case p(r) = p is that {m2}. the definition of {m8} depends on the initial reserve u = Ro. the A(t) and hence Vi(u) is available by solving differential equations: Proposition 7. In fact.
t and being followed by a downcrossing. from a computational point of view the remaining problem is to evaluate the v(t). two things can happen: either the current jump continues from u + p(u)dt to u.t). or it stops between level u + p(u)dt and u. Given this occurs. Proposition 7.t)). the probability that level u + p(u)dt is downcrossed for the first time in phase j is vj (u + p(u)dt). we get vi(u) = aidt + (1 .Qdt) vi(u) + vi'(u)p(u)dt + p(u) dt E{tji+tjvi(u)}. vi.(tai + vi(u) E vj(u)tjp (u)  Q + vj (u)tjip ( u). 0 Thus.4) jEE jEE Proof Consider the event A that there are no arrivals in the interval [0.3dt. {mx} has jumps of two types.2 For i E E. those corresponding to state changes in the underlying phase process and those corresponding to the present jump of {Rt} being terminated at level u . Thus. The intensity of a jump from i to j is tij for jumps of the first type and tivj(u . MATRIXANALYTIC METHODS However. jEE . Hence Q(t) _ T + tv(u . dt]. the interpretation of Q(t) as the intensity matrix of {my} at time t shows that Q(t) is made up of two terms: obviously. A'(t) = A(t)Q(t) = A(t)(T + tv(u . In the first case. Given A.t) for the second. the probability of which is 1 . the probability of downcrossing level u in phase i is 8ji(1 + p(u)dt • tii) + (1 . 0 < t < u.246 CHAPTER VIII. given A.(u) p ( u) = . Given A'. whereas in the second case the probability is p(u)dt • tjvi(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. (7.Sj i)p(u)dt • tji = Sji + p(u)tji dt. the probability that level u is downcrossed for the first time in phase i is ai.
This yields v.. consider a modification of the original process {Rt} by linearizing the process with some rate p.2. F" etc. Then pv(r) p(r) r < v p r>v ' and (no matter how p is chosen) we have: Lemma 7.. . Then P(B.)P"(AnB. vi (U) = lim v= (u). P u which implies that v.. refer to the modified process.s. Now since both P(A n Bv) 3 0 and P"(A n Bv) . Rt . after a certain level v. RESERVEDEPENDENT PREMIUMS 247 Subtracting v.) is the tail of a (defective) random variable so that P(Bv) + 0 as v 4 oo. denotes the time of downcrossing level u . From Section 3.) + 0 as v + oo.h.0 as v + 00 we have P(A) P"(A) = P(AnBv)+P(AnBv) P"(AnB.3 For any fixed u > 0. Let p" (t).4) backwards for {va (t)}v>t>o. V . When solving the differential equation in Proposition 7. Thus. (v) is given by the r.5). supRt>v l t<7 I where o.i7rT1/p.7. (u) for any values of u and v such that u < v.00 Proof Let A be the event that the process downcrosses level u in phase i given that it starts at u and let B" be the event By={o. (u) on both side and dividing by dt yields the asserted differential u equation. we can first for a given v solve (7. and similarly P"(Bv) .^ 0. we have p(r) = p = vi (u) 0aTe. To deal with this. we face the difficulty that no boundary conditions is immediately available. say. <oo. of (7. then P(A n Bv) _ P"(A n BV').. Since the processes Rt and Rt coincide under level B.) P"(AnBv) = P(AnB. say. starting from v"(v) = .
x) dx f v(u)eT xt dx . The trapezoidal rule used in [288] gives a precision of 0(n 3). We recall from Propositon VII.v v(u)eTat 1 1 . the evaluation of Vi(u) requires q(u) = 1 . r<v r > v. Let ii'( u) = a+'ie(T+ta +^)"e denote the ruin probability for R't where a+ = a+i) = laT1/pi. Then v(u) is the initial distribution of the undershoot when downcrossing level v given that the process starts at u. T). 0 < u < v.e. < 0}). However. To evaluate p1(u). MATRIXANALYTIC METHODS Next consider a sequence of solutions obtained from a sequence of initial values {v.1.RQ (defined for or < oo only) is defective phasetype with representation (v(u). Recall that q(w) is the probability of upcrossing level v before ruin given the process starts at w < v. the probability of ruin between a and the next upcrossing of 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.V" M 0 . The precision depends on the particular quadrature rule being employed.q(v dx +( ) ) = ( ) ( q( )) vueTva (7...7) equals 01 (v . i.248 CHAPTER VIII. 7b Twostep premium rules We now assume the premium function to be constant in two levels as in VII.7) f o (the integral is the contribution from {R. as well p1(u).6) We may think of process Rt as pieced together of two standard risk processes RI and Rte with constant premiums p1.) are so. Corollary 3. p2.9. typically the complexity in n is 0(n2) for integral equations but 0(n) for integral equations. assuming u > v for the moment. while the fourthorder RungeKutta method implemented in [30] gives 0(n5). > 0} and the last term the contribu tion from {R. such that Rt coincide with RI under level v and with Rt above level v. 2/n. where. v = u. (u)}. 3u etc.10 that in addition to the O'(•). numerically implemented in Schock Petersen [288]) and the present one based upon differential equations require both discretization along a discrete grid 0.1. The algorithm based upon numerical solution of a Volterra integral equation (Remark VII.. The f iin in (7. which is available since the z/i'(. Therefore u pl(vvueTa t 1. let v(u) = a+2ieiT +ta+>)(uv). (7.z51(v)).. p(r) P. cf.1. for u > v the distribution of v .1a. where v = inf It > 0 : Rt < v}.. 2u. say.zp1(u)/(1 . 1/n.
x) dx 1 ^(v) ( 1 .x) dx} V 1 1(v) f V v(u) eTxt.7. (7.01 (v .u .v) + (2^ + 3v2 ea'(u " .v(u)eTVe . p2 < 3. B is hyperexponential corresponding to 3 0 3 a(2 2)' T= ( 0 7 t.24ev .v(u)eTve). Then one gets X20 20 21 f 1ea1(u v) + 1 3 3 ^ A 2(u e . 01(u) _ 24 u + 35 e6u 1 35 e 4(u) _ 35 .21 = ? yields 0(u) = 1.(7 The arrival rate is (i = 3.v) 1eai(u v) + 7 7 1 e\2(u v) 1 3 ^') eA2 (u.4) can be written as (Y(u) ®a+)e(T+t°+>)°1 (T ® (T .and A2 = 3 .8) equals v v(u)eTxta+2) e(T+ta +))( vx)edx which using Kronecker calculus (see A.jl (t ®e) Thus.2V"2.e.to+))11 {e{T®(Ttoy+ ))}„ .^1(v) 1 .2. RESERVEDEPENDENT PREMIUMS 1 . Since µB = 5/21.8) The integral in (7.. Example 7. I.e6u 35 .2.1 from which we see that pl (u) = 1 + 1 249  1 .24e. so we consider the nontrivial case example p2 = 4 and p1 = 1.4 Let {Rt } be as in Example 3. From Example 3.v(u ) eTV e J v(u)eTxtz/)l (v . all quantities involved in the computation of b(u) have been found in matrix form.e6v Let Al = 3 + 2V'2.be the eigenvalues of T + to( 2 ).
21(35e6v .v)esv + 7 4_ 2.7) we see that we can write pi (u) = v(u)V2 where V2 depends only on v. MATRIXANALYTIC METHODS From (7.1.b(v) = 192esv +8 35e6v + 168esv + 7* Thus all terms involved in the formulae for the ruin probability have been exu plicitly derived. The analysis and the example are from Asmussen & Bladt . and one gets 12e5" .250 CHAPTER VIII. ) e sv + ( 2v/2. 21 3 In particular. pi (u) = p12(u)/p1 l(u) where p1i(u) p12(u) 35e6v .24e5v .24e5v .1)' ?.+ it (3 4'I 1 ea2(uv e1\2(u") 7 + ( 32 +4..1 Thus. 192esv + 8 P1 .2 35e6v .1 V2 = 4e5"+6 35e6v . Notes and references [30].24es" ./2) ea 1(u .24es" ./2 ea1(u") .
(b) the lognormal distribution (the distribution of eu where U . 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. for all t > 0. III. 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. see I.2b. x 4 oo. B(x) = L(x)/x" where a > 0 and L(x) is slowly varying. For the definition . a2)) with density 1 e(logyFh) 2/2az . oo ) and say then that B is subexponential (B E S) if 251 . For further examples.46 and at numerous later occasions require a light tail.g. and instead we shall work within the class S of subexponential distributions . the exponential change of measure techniques discussed in II. B(x) = ex0 with 0<0<1.4. For example. we require that B is concentrated on (0. A rough distinction between light and heavy tails is that the m.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.f.N(µ. Some main cases where this lighttail criterion are violated are (a) distributions with a regularly varying tail. L(tx)/L(x) 4 1.B(x). x 2iror2 (c) the Weibull distribution with decreasing failure rate .
P(max(Xi.2B(x). Then X1 +X2 has an Erlang(2) distribution with density yeY so that B*2(x) xex.p.3 Consider the standard exponential distribution. 1/2 'typical' (with distribution B) and w. That is. (b) liminf BB(() ) > 2. HEAVY TAILS B*2\ 2.F(X1 > x.B(x)2 . Thus . the distribution of independent r. X1 is w. (1.'s X1. As contrast to Proposition 1. the behaviour in the lighttailed case is illustrated in the following example: Example 1.'s. Thus the liminf in Proposition 1. We later show: Proposition 1.1 Let B be any distribution on (0.v. X2 > x) = 2B(x) . Proof By the inclusionexclusion formula. one can check that x x where U is uniform on (0. note first the following fact: Proposition 1. Then: (a) P(max(Xi.p. that is. P(Xi <yI Xi+X2>x) 1B(y).2.252 CHAPTER IX.1(b) is oo. and thus the lim inf in (b) is at least lim inf P(max(Xi. 1/2 it has the distribution of X1I X1 > x. X2) > x} C {X1 + X2 > x}. To capture the intuition behind this definition. That is. X2 with distribution B. . B(x) ax. X2) > x) ^' 2B(x).v.v. we have {max(Xi. if X1 + X2 is large . X2) > u x)/B(x) = 2. x 3 00. in the subexponential case the only way X1 + X2 can get large is by one of the Xi becoming large. B(x) Here B*2 is the convolution square. Since B is concentrated on (0. X2 but none of them exceeds x. X2) > x}. proving (a). X2) > x) is P(X1 > x) + P(X2 > x) . oo). 1). In terms of r. In contrast. oo).1) then means P(X1 +X2 > x) 2P(Xi > x). 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. given X1 + X2 > x. the r.2 If B E S. then (with high probau bility) so are both of X1. then P(X1>xI X1+X2>x)* 2.
S)x. Hence lim sup a+oo B*2(x) 2B((1 . 1 < B(x ) B( x) Y) < B( 0). If lim sup B(x .y)/B(x) > 1.xIX > x converges in distribution tooo.B(y)) . This follows since the probability of the overshoot to exceed y is B (x + y)/B(x ) which has limit 1. B( 0 . Let 0 < 5 < 1/2.B E S. [In terms of r. The uniformity now follows from what has been shown for y = yo and the obvious inequality y E [0.B*(n ) B(dz) (1. yo] as X + 00.'s: if X .B*n(x . we therefore get lim sup B*2(x)/B(x) > 1+B(y)+ 1 .5)x)' + 0 _ 2 L(x)l xa (16) Letting S 10.1. a contradiction.yo].2) B(x) B(x ) B(x) Jo with n = 1 and splitting the integral into two corresponding to the intervals [0. SUBEXPONENTIAL DISTRIBUTIONS Here is the simplest example of subexponentiality: Proposition 1. Finally lim inf B(x . we get BZ(x)) > 1 + B(y) + B(B()y) (B(x) . x].B(y) = 2. y] and (y.v.y)/B(x) > 1 since y > 0.5 If B E S. We now turn to the mathematical theory of subexponential distributions. and combining with Proposition u 1. then either one of the Xi exceeds (1 . Using the identity B*(n+1)(x) = 1+ + 1)(x) 1+ 2 1 . 253 Proof Assume B(x) = L(x)/xa with L slowly varying and a > 0.S)x + B(Sx)2 < lim sup B(x) xaoo B(x) lim sup 2L((1 x^oo . Proposition 1.z B(x) .] Proof Consider first a fixed y. If X1 + X2 > x.1(b) we get B*2(x)/B(x) * 2. then the overshoot X .6)x)/((1 . then B(B(x)y) * 1 uniformly in y E [0. we get limsupB*2(x)/B(x) < 2.4 Any B with a regularly varying tail is subexponential. or they both exceed Sx.
x oo.z ) (x ) = 1 + (^ B(x . then e"R(x) * oo. 0 Proof of Proposition 1.5 and dominated convergence. we have by Proposition 1.nI < e for x > y. b[c] = oo for all e > 0. The case n = 2 is just the definition. then for any n B*n(x)/B(x) * n. and this immediately yields the desired conclusions.2). Proposition 1.z) B(dz) 2B(x) o rv 2 0 2 using Proposition 1.z) B(dz).y) sup v>o B(v) B(x) which converges to 0 by Proposition 1. B*(n+1) (x I xy + Jxx y) W.z) B(dz) (n + O(e)) ^x JO B(x) (n + 0(0) I B (x) .z) B(dz) _y B(x) 111 Lx B .B(x .5 that B(n) > e6B(n . Then by (1.1) for all large n so that B(n) > cle6n for all n. its intuitive content is the same as discussed in the case n = 2 above. This implies B(x) > c2e5x for all x.B*2 (x) B(x) (x .6 If B E 8. Given e > 0.7 If B E S. B(x) \Jo _ B(x .. choose y such that IB*n(x)/B(x) . The first integral is y B(x . so assume the proposition has been shown for n.254 CHAPTER IX. 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 . Proof We use induction. HEAVY TAILS Corollary 1. Proof For 0 < 5 < e.2.z) B(x) Here the second integral can be bounded by B*n(y) B(x) . O The following result is extremely important and is often taken as definition of the class S.5 and the induction hypothesis.
3) Using the necessity part in the case Al = A2 = B yields f xv B(x . 0 Proposition 1.5 easily yields P(X1 + X2 > x.(al + a2)B(x).i). Then Al * A2(x) = P(X1 + X2 > x).2).8 If B E S.z) B(x . A2 be distributions on (0. Since P(X1+X2 > x. choose T such that (B(x)B*2(x))/B(x) < 1 + b for x > T and let A = 1/B(T).4) .'s such that Xi has distribution Ai. Proposition 1.z) B(dz) < 1 + A + an(1 + d) .1. Proof Let X1.X2 > xv) < A1(xv)A2(x v) . it follows that it is necessary and sufficient for the assertion to be true that JX_VA (x . X2 be independent r.y)Ai(dy) = (x)o(1) (1.v.ajB(x)Ai(v) = ajB( x)(1+o„(1)) (j = 3 . an = supx>o B*n(x)/B(x).z) B(dz) x . an+1 fX B*n( *n(x . 0 Lemma 1. Proof Define 5 > 0 by (1+5)2 = 1+e. Then Al * A2 (x) .X1 > xv.9 Let A1.B(x .y)Ai(dy) v) f o . Then by (1.0 completes the proof.z) B(x) < 1 + A + an sup f x B(x . e > 0. Combining these estimates and letting a 4. For any fixed v. oo) such that Ai (x) _ aiB(x) for some B E S and some constants al. 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.y)B(dy) = B(x)ov (1)• v (1.ala2B(x)2 which can be neglected. Xi <= v Ai (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) . a2 with a1 + a2 > 0.y))/B(x).z) B(dz ) + sup < 1 + sup f x<T B ( x) x>T 0 B(x . then there exists a constant K = KE such that B*n(x) < K(1 + e)nB(x) for all n and x.
a2 = 1. if q(x) aB(x) for some B E S and some constant a > 0. V (1. the l.Bl (x) + B2 (x) follows precisely as in the proof of Proposition 1. A2 = B so that a1 = 0. L2 are slowly varying. of (1.Ai(x .v)Ai(v) . A(x) = o(B(x)).s. Proof Taking Al = A2 = A. That is.y)B(dy). HEAVY TAILS 'VV B(x .y)Ai(dy) = B(x)o„(1). Then B E S provided fo "O exA(x) b(x) dx < oo.11 Let B E S and let A be any distribution with a ligther tail.Bl (x) + B2 (x) when B1. L2 slowly varying. We next give a classical sufficient (and close to necessary) condition for subexponentiality due to Pitman [290]. That is.h. u It is tempting to conjecture that S is closed under convolution.4). it should hold that B1 * B2 E S and B1 * B2 (x) .10 The class S is closed under tailequivalence.256 Now (1. In the regularly varying case. with a > 0 and L1.5) Here approximately the last term is B(x)o„(1) by ( 1. B1 * B2 (x) . it is easy to see that if L1. However.12 Assume that Bi(x) = Li(x)lxa. B2 E S.(x) is decreasing for x > x0 with limit 0 at oo.B(x) Proof Take Al = A.3) follows if CHAPTER LX.5) becomes x B(x .v)B(v) + _'U Aq(x . Then L = L1 + L2 is slowly varying and B1 * B2(x) sim L(x)/x«.2. . then so is L = L1 + L2.2A(x). f " By a change of variables. Recall that the failure rate A(x) of a distribution B with density b is A(x) = b(x)/B(x) Proposition 1. i = 1. whereas the two first yield B(x)(Ai(v) . B1 * B2 E S does not hold in full generality (but once B1 * B2 E S has been shown.9). Hence Corollary 1.13 Let B have density b and failure rate A(x) such that .2aB(x) . then A E S. Then A * B E S and A * B(x) . u Corollary 1. u Corollary 1.aiB(v)) = B(x)o„(1). a1 = a2 = a yields A*2(x) .
16 For L(x) slowly varying and a > 1.1)xcl1 .2). we can use the same domination for the second integral but now the integrand has limit 0 . Further. replace B by a tail equivalent distribution with a failure rate which is everywhere decreasing). Example 1. and exa(x)b(x) = (3x01e(10)x9 is integrable.`(x)b(x) is integrable. By (1.1 has limit 1 + 0. a(x) = ax01. Thus B*2(x )/ B(x) . subexponentiality has alrady been proved in Corollary 1. the u lognormal distribution is subexponential.y) dy. L(x) y° (a . The middle bound shows that it converges to b(y) for any fixed y since \ (x . Define A(x) = fo . x . proving B E S.12.(y) dy. Thus. Then b(x) = Ox0lexp.y) < A (y) for y < x/2.A(y)\(y) dy + fox/ 2 eA(x ).13 works in this setting. Goldie & Teugels [66]): Proposition 1. Example 1.3 < 1.14 Consider the DFR Weibull case B(x) = ex0 with 0 <.U) /or) v 2x This yields easily that ex.A(x . Since ) (x .1 B(x) eA( x)A(xv )A(y)A(y) dy f B(x . the DFR Weibull distriu bution is subexponential..A(y)a(y ) = ev'(y) b(y). Thus A(x) is everywhere decreasing. we first quote Karamata's theorem (Bingham.y) < yA(x . the first integral has limit 1 .15 In the lognormal distribution. To illustrate how Proposition 1.. 0 A(x) .y ) b(y)dy = B (x) o ox _ J = ox/2 eA( x)A(xy ). B*2(x) . Thus. Then B(x) = eA(x).A(xy)A ( y). elementary but tedious calculations (which we omit) show that A(x) is ultimately decreasing.y) * 0.e009xv)2/2a2/(x 2irv2) logx ( ) 't ((logx . In the regularly varying case. Jo For y < x/2. SUBEXPONENTIAL DISTRIBUTIONS 257 Proof We may assume that A(x) is everywhere decreasing (otherwise. Thus by dominated convergence .1. an integrable function by assumption. f ' L(y) dy .y) y\(y)• The rightmost bound shows that the integrand in the first integral is bounded by ey"(v).(x .
L(x)/x" and )t(x) .xjX > x. Thus exa(x)b(x) .1)]) xa L(x) (x[1 + y/(a .1) and P(X (. we get (1.6) EX(x) . Then: Proposition 1. then 7(x) x/(a . Proof ( a): Using Karamata's theorem. let X W = X .1))^ ' (b) Assume that for any yo )t(x + y/A(x)) 1 A(x) uniformly for y E (0.258 From this we get CHAPTER IX. (c) Under the assumptions of either ( a) or (b).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 .)/Y(x) > y) (1 + y/(a .ea b(x) is integrable.1)])a 1 1 . the monotonicity condition in Proposition 1. However.4 is necessary in full generality. More precisely.y(x)B(x).17 If B has a density of the form b(x) = aL(x)/x°+1 with L(x) slowly varying and a > 1. Then 7(x) . then B(x) . f O B(y) dy . . yo] .E(X . (1 + y/(a . 'y(x) = EXix>. We conclude with a property of subexponential distributions which is often extremely important: under some mild smoothness assumptions.1/A(x) and P(X ixil'Y (x) > y) * e'.13 may present a problem in some cases so that the direct proof in Proposition 1.1)] I X > x) L(x[1 + y/(a .a/x.1)X(x)/x > y) = P(X > x[1 + y/(a .18 (a) If B has a density of the form b(x) = aL(x)/xa with L(x) slowly varying and a > 1. the overshoot properly normalized has a limit which is Pareto if B is regularly varying and exponential for distributions like the lognormal or Weibull.1))a . HEAVY TAILS Proposition 1.
Then P(Y1 + • • • + YK > u) . Kliippelberg & Mikosch [134]. i. 2 The compound Poisson model Consider the compound Poisson model with arrival intensity /3 and claim size distribution B. then Vi(u) P Bo(u)..EK G(u).2.2 Let Y1. be i.n0 1•P(K= n)•n = EK. 1.. 0 G(u) L G(u) . Theorem 2 .1 If Bo E S. .15. Let St = Ei ` Ui .f yl 0 0 = exp {y (1 + 0(1))} 0 fY A( x + u /A( x)) a(x) du } The property (1.(x). .nn.t be the claim surplus at time t and M = sups>0 St. The remaining statement (1. THE COMPOUND POISSON MODEL 259 We omit the proof of (c) and that EX (x) . cf.8) in (b) then follows from P (A(x)X (x) > y) = F(X > x + y/.A(x + y/A(x))} =a(x) a(x + x) dx = ex p ex P .+YK> u) = ^•P(K = n)G* n(u ) . Notes and references A good general reference for subexponential distribution is Embrechts. We get p(yl+.. d. Y2. St > u}.and lognormal distributions . Recall that B0 denotes the stationary excess distribution. Lemma 2. P The proof is based upon the following lemma (stated slightly more generally than needed at present). with common distribution G E S and let K be an independent integervalued r.7) is referred to as 1/A(x) being selfneglecting. It is trivially verified to hold for the Weibull.14. u a oo. nG(u). 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. Bo(x) = f0 B(y) dy / µB.v.1/. with EzK < oo for some z > 1.A(x) I X > x) = exp {A(x) . Examples 1. We assume p = /3µB < 1 and are interested in the ruin probability V)(u) = P(M > u) = P(r(u) < oo). r(u) = inf it > 0..
Borovkov [73] and Pakes [280]. Proof of Theorem 2.x400 PBB(x) PB Leta+oo.. mathematically one must note that there exist (quite intricate) examples where B E S. in our three main examples (regular variation .µB(01 .p) and EzK < oo whenever pz < 1.13). we have fx B(y)dy = a B0 (x) > lim inf lim inf x+oo B(x) . see Abate. u x+a Notes and references Theorem 2.1) In particular .1 is essentially due to von Bahr [56].x^ ) B(x) _ f or ( lox . and for the lognormal and Weibull cases it can be verified using Pitman 's criterion (Proposition 1. Bo is more heavytailed than B .260 CHAPTER IX. Since EK = p/(1.3 If B E S. 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].µ J ) .1. For some numerical studies. . (2.1 is notoriously not very accurate..?(xµ 8 (x). HEAVY TAILS u using dominated convergence with >2 P(K = n) Dz" as majorant. P(K = k) = (1. However. Bo ¢ S.p)p'.18. as well as examples where B ¢ S. u The condition Bo E S is for all practical purposes equivalent to B E S. r(1/Q) xlQexp B(x) = ex' From this .2. The PollaczeckKhinchine formula states that (in the setup of Lemma 2. _ B(x^sx Bo(x) µ8 I aoB(y )dy = (^) . Bo E S.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 ) . lognormal . In general: Proposition 2. The approximation in Theorem 2. then Bo(x)/B(x) + 00. Bo E S is immediate in the regularly varying case. Weibull) one has Bo(x ( B(x) . Proof Since B(x + y)/B(x) * 1 uniformly in y E [0. x 4 00. Note that in these examples . a]. The problem is a very slow rate of convergence as u ^ oo. the result follows immediately from Lemma 2.2) M = Yl + • • • +YK where the Yt have distribution Bo and K is geometric with parameter p.
Asmussen & Binswanger [27] suggested an approximation which is substantially better than Theorem 2. 1 Assume that (a) the stationary excess distribution Bo of B is subexponential and that (b) B itself satisfies B(x .. 195 there are numerical examples where tp(u) is of order 105 but Theorem 2. Kalashnikov [219] and Asmussen & Binswanger [27]. Snd) = Xl +.1. The main result is: Theorem 3 . T1 the ith interarrival time and Xi = U.g. G+ (A) = P(Sq+ E A. i=1 . one may have to go out to values of 1/'(u) which are unrealistically small before the fit is reasonable.. + Xn.y + as usual denotes the first ascending ladder epoch of the continuous time claim surplus process {St}. there is a representation of M similar to the PollaczeckKhinchine formula. Based upon ideas of Hogan [200]. let t9+ = i9(0) be the first ascending ladder epoch of {Snd> }. .e. To Bo(u) u + 00.y)/B (x) > 1 uniformly on compact y internals. THE RENEWAL MODEL 261 Choudhury & Whitt [1].} Then ik(u) = F ( M > u) = P(i9 (u) < oo).1 gives 1010.3.9+ < oo) = P(S.. p = iB /µA < 1.Ti. Thus G+ is the ascending ladder height distribution (which is defective because of PB < PA). also a second order term is introduced but unfortunately it does not present a great improvement. Then l/i(u) 1 P P [Note that (b) in particular holds if B E S. 3 The renewal model We consider the renewal model with claim size distribution B and interarrival distribution A as in Chapter V. Then K M=EY... (3. t9(u) = inf {n : Snd> > u} . M = sup s$ .. [279].] The proof is based upon the observation that also in the renewal setting. E. We assume positive safety loading. T+ < oo) where r+ = T1 + • • • + T.1 when u is small or moderately large.. Define further 0 = IIG+II = P(r9+ < oo). i. This shows that even the approximation is asymptotically correct in the tail.+ E A. in [219] p. {n= 0. In [1]. Somewhat related work is in Omey & Willekens [278]. Let U= be the ith claim .1) this end .
9)9'' and Y1. with distribution G+/9 (the distribution of S.262 CHAPTER IX.d)) E A) denote the pre19+ occupation measure and let and U_ = Eo G'_" be the renewal measure corresponding to G_..Y2. we will use the fact that the proof of (3. 0] to the integral is O(F(x)) = o(FI(x)). Then 0 0 F( x . the contribution from the interval (. 0] normalized by IPG_ I so that we should have to G+(x) . Let F denote the distribution of the Xi and F1 the integrated tail.Ti). x + oo. (b) and does not rely on the structure Xi = Ui . 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. FI (x) _ fz ° F(y) dy.1) is equivalent to P(M > u) " .2)..g+ > x. Let further 19_ _ inf {n > 0: S^d^ < 0} be the first descending ladder epoch.2 F(x) . cf.y_ E A) the descending ladder height distribution (IIG II = 1 because of PB < P A) and let PG_ be the mean of G_.y+ given r+ < oo).1) holds for a general random walk satisfying the analogues of (a).3 G+ (x) .oo. HEAVY TAILS where K is geometric with parameter 9.. whereas for large y . are independent of K and i.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_.FI(u).3) and we will prove it in this form (in the next Section. The heuristics is now that because of (b).y) dy = 1 Pi (X) oo IPG_ I .d. A. As for the compound Poisson model. U_ (dy) is close to Lebesgue measure on (. d+ < oo). x > 0. (3. 0 The lemma implies that (3. x * oo. A(dy) = 1. P(K = k) = (1 . Write G+( x) = G+ ( x.N.FI(x) /IPG_I. and hence FI(x) . B(x) _ J O° B(B(x)y) A(dy) f 1 .i. Lemma 3 . oo) = F(S.B(x). G_(A) = P(S. Proof Let R+(A) = E E'+ ' I(S.PBBo(x). Lemma 3 . Proof By dominated convergence and (b). u a 00.1 IPG_ I / F(x .(.
I n=N (1 E)2 r00 F(x + y) dy + e) lim sup .1. Hence using dominated convergence precisely as for the compound Poisson model.3.3.UG_ I x.0)0k k I(u) A. We then get lim sup G+(x) xro0 Fj(x) < lim sup X)00 o F(x .(dy) fN FI ( x) + lim sup ZY00 N F(x . n] < (1 + e)/1µc_ I for n > N.9) 1 . If G_ is nonlattice.y) U_ (dy) 00 FI (x) < lim sup F(x) U(N.1. By Lemma 3.2).O[s])(1 . the proof is complete. THE RENEWAL MODEL 263 We now make this precise. choose N such that F(n . 0] x+00 FI(x) 00 + lim up 1 x) E F(x + n) U_ (n .1 I .=1 BIp G_ I (1. and in the last that FI is asymptotically proportional to Bo E S.oo Fj(x) N J (1 +6)2 I {IC_ I lim sup X400 FI(x + N) _ (1 + e)z (x) I Pi µ G_ I Here in the third step we used that (b) implies B(x)/Bo(x) + 0 and hence F(x)/FI(x) 4 0.e) z lim inf G+(x)  FI (x) Ip G_ I Letting a 10. Given e. n] F1 ( n=N _1 1+e E F(x+n) 0 + limsup xr00 FI(x) FAG. F(Y= > x) FI(x)/(OIp _ 1).9)IpG_ I Differentiating the WienerHopf factorization identity (A.1.F[s] = (1 .2) yields 00 F F I (u) P(M > u) _ E(1 . Similarly. In the lattice case. we can assume that the span is 1 and then the same conclusion holds since then U(n .G+[s]) . n] + 1/I µG_ I. n] is just the probability of a renewal at n. (3.1. u Proof of Theorem 3.1.1)/F(n) < 1 + e for n > N (this is possible by (b) and Lemma 3. and that U_(n . > (1 .y) U. then by Blackwell 's renewal theorem U_ (n .
S+9(u) .AB iP We conclude by a lemma needed in the next section: Lemma 3 . Note that substantially sharper statements than Lemma 3.(1 . u)). on the set {M > u.. and {Su. allowing also for possible dependence between the arrival process and the claim sizes.SS(u)}n=o. must attain a maximum > 0 so that P(M > u.a.Sty(u)_I < a} we have w(u) < oo.Se(u)_1 < a) = o(Fj(u))..l. FJ(u) UBBO(U) PBo(u) N = (10)Ipc_I JUA .264 and letting s = 0 yields CHAPTER IX. P(M > u.IIG+II)µc_ = (1 . see Asmussen & Kliippelberg [36]. Proof Let w(u) = inf {n : Sid) E (u . 4 Models with dependent input We now generalize one step further and consider risk processes with dependent interclaim times.So(u)) are available.4 on the joint distribution of (S.0)ua_ . with roots in von Bahr [56] and Pakes [280]. Notes and references Theorem 3.u)) = o(P (M > u)) = o(FI(u)).1 is due to Embrechts & Veraverbeke [136]. .a.a.a. Therefore by Lemma 3.2. Sty(u) . 10(0) But since P(M > u . S+q(u) .4 For any a < oo.. Then P(M E (u . we have P(M E (u . u).1)6+[0] .(u)+n .So( u)_1 < a) < P (w(u) < oo)j/i(0) < 0(0) P(M E (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.yiui_1. u)) > P(w(u) < oo)(i lp (0))• On the other hand. Mn < u}.a) N P(M > u). HEAVY TAILS µF = (1 .
2. Figure 4.. MODELS WITH DEPENDENT INPUT 265 Various criteria for this to be true were recently given by Asmussen.. M* = max S. See Fig. 4.1 based upon a regenerative assumption. G(x) (4. Schmidli & Schmidt [47]..i. see [47]. Define S..X1 is the generic cycle. examples and counterexamples.1) .. +1... Assume that the claim surplus process {St}t>o has a regenerative structure in the sense that there exists a renewal process Xo = 0 < Xl <.4.1) is assumed.. M = sup St. assume pp.. 0o(u) etc. E0. The zerodelayed case corresponds to Xo = Xl = 0 and we write then F0. The idea is now to observe that in the zerodelayed case. {SX1+t . Thus the assumption .1 Note that no specific sample path structure of {St} (like in Fig..n n=0.4 below.X2 < . Theorem 4.1 where the filled circles symbolize a regeneration in the path..1 except for the first one) is a random walk.d. For further approaches.1 = max k=0. 4. We give here one of them..1. ... = Sx. M. We return to this point in Example 4. and the distribution of {Sxk+t .Sxk}o<t<xk+1xk is the same for all k = 1. {Sn}n=o. (corresponding to the filled circles on Fig. We let F* denote the Podistribution of Si. (viewed as random elements of the space of Dfunctions with finite lifelengths) are i.. t>0 S.. < 0 and EoX < oo where X = X2 . and apply it to the Markovmodulated model of Chapter VI.F*(X) = P0(Si > x) .1... such that {SXo+t  SXo}0<t< X 1Xo ...Sxi}0<t<x2Xl . 4..
.266 CHAPTER IX. it suffices by (4.1 Assume that (4. (4.2..S. Then '00 (u) = Fo(M > u) .3) where Mnx) = sup o<t<xn +1 X.3) applicable so that F(M* > u) 141 F*(u).2) Imposing suitable conditions on the behaviour of {St} within a cycle will then ensure that M and M* are sufficiently close to be tail equivalent. jF11 F* (U). Since clearly M(x) > Sl .2) to show F(M* > u) > 1. Sxn +t . the assumption means that Mix) and Sl are not too far away. u p 00.. See Fig. 4. HEAVY TAILS for some G such that both G E S and Go E S makes (3.Sxn = sup Sxn+t .4) liminf u>oo F(M > u) . The one we focus on is Fo (Mix) > x) .* i o<t<xn+1x. N N Xi=0 N Figure 4.3) hold.2 Theorem 4. Proof Since M > M*. (4.. (4. Fo(Si > X).1) and (4.
4).1 can be rewritten as 00 (U) (4. 2.6) 1 p pBo(u) u where B is the Palm distribution of claims and p . Let a > 0 be fixed. S.2.5) which follows since Po (M > u. M^xu)+l > a) ..Mn +1 >aV(u n=1 00 S.: Sn > u} .e)Po (MMX> > x). u)} < P(M* E (u ..a. 0 yields (4.(u) . u))/P(M* = 0) = o(Po(M* > u)).+Mn+1>u} 267 (note that {M> u} = {3(u) < oo}).(1 ..1 = limti00 St/t. Mn+l > a V (u .. Po(M* > u) . Given e > 0. Letting first u + oo and next e . Theorem 4.: S. .a. choose a such that Po(Si > x ) > (1 . /3(u) = inf{n=1.S..( u)1 > a) 00 1: Po(Mn<u. MW O(u)+1 < a) IN ( U n=1 A1.Po (M* > u.Sn+1Sn>aV(uSn*)) n=1 00 > (1E)EPo(Mn<u.4. We shall use the estimate Po(M > u) Miu^+ 1 < a) = o(Po (M > u)) (4...Sn 0<t<x„+j ( 1 . assume the path structure Nt St = EUit+Zt i=1 . )) > (1 .4. MODELS WITH DEPENDENT INPUT Define 79* (u) = inf {n = 1 .E) Po ( n max St u. x > a. Under suitable conditions . Then by Lemma 3.e)Po (M > U). E (u . To this end.e)Po (M > u.
Assume further that (i) both B and Bo are subexponential.7). cf. independent of {> CHAPTER IX. the proof of Lemma 4. oX (see Proposition A1.v.p) Ju P Bo(u) 1p 0 . we get 00 (u) 1 IPF. Corollary 4. (iii) For some o field Y.268 with {Zt} continuous. a4' 0. Then the Palm distribution of claims is B(x) = E N Eo 0 I( U1 < x) .Q = EoNx/EoX.6 below.} and satisfying Zt/t N.6) holds with p = .I u J Po(Sl > x) dx 1 EoNxB(x) dx EoX(1 . are Fmeasurable and NX Po J:U=>x i=1 (iv) Po sup Zt > x / (0:5t<x o(B(x)) Then (4. and the rest is just rewriting of constants: since p = 1+tlim St = 1+ . HEAVY TAILS N` U. and ENX Ui .4). Mix) < > UE + i=1 o<t<x Thus Theorem 4.3PB.X both have tails of sup Zt. The same is true for Sl.8) x Write . X and N. Proof It is easily seen that the r. since the tail of Zx is lighter than B(x) by (iv). i=1 (4. and also for Mix) since Nx FNX U.'s order EoNx • B(x). (ii) EozNX < oo for some z > 1.1 is in force.2 Assume that {St} is regenerative and satisfies (4.
The number N. We now return to the Markovmodulated risk model of Chapter VI with background Markov process {Jt} with p < oo states and stationary distribution 7r. Then (4. X2 = 1.6) holds. and for some constants ci < oo such that cl + • • • + c.0 (thus (iv) is trivial).9). we will assume that lim B2(x) = ci x+oo G(x) for some distribution G such that both G and the integrated tail fx°O G(y) dy are subexponential .6) u holds. X2 = 1.. . (iii) is obvious. Bo E S. we conclude just as in Example 4. Again .t} is standard compound Poisson and {Zt} an independent Brownian motion with mean zero and variance constant a2. Theorem 4. i=1 B = >2 7riaiBi i=1 and we assume p = 014 B = Ep ri/3ipB.. The regenerative assumption is satisfied if we take Xo = Xi = 0. Assume that B E S..e.5 Consider the Markovmodulated risk model with claim size distributions satisfying (4. The arrival rate is /3i and the claim size distribution Bi when Jt = i. MODELS WITH DEPENDENT INPUT 269 Example 4 . In particular. Example 4 . More precisely. (i) holds. i. Thus we conclude that (4.4 Assume that St = Zt . 3 The average arrival rate / and the Palm distribution B of the claim sizes are given by P P Q = ir i/i.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). 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.. then (iv) holds since the distribution of supo<t<i Z(t) is the same as that of I Zl 1. We consider the case where one or more of the claim size distributions Bi are heavytailed. in particular lighttailed. > 0.(NX). < 1. and taking F = o. X3 = 2. we assume that B E S.6) holds.t + EN'I Ui where {>N`1 Ui .4. Taking again Xo = Xi = 0.3 As a first quick application.. consider the periodic model of VI. Zt . . . 1) is Poisson with rate /3 = fo /3(s) ds so that (ii) holds. The key step of the proof is the following lemma. of claims arriving in [0.3 that (4. X3 = 2. Bo E S. .
Let {Fi}t=1 P be a family of distributions on [0. . we can define the regenerations points as the times of returns to i. . It follows by a slight extension of results from Section 1 that P P(Yo > x I Y) G( x) ci Ni.. An easy conditioning argument then yields the result when Jo is u random. The same dominated convergence argument completes the proof. For lighttailed distributions."+Np . 1. oo) such that G E S and some c1. Markovmodulation typically decreases the adjustment coefficient y and thereby changes the order of magnitude of the ruin .ciG(x).. . Assume EzN1+"'+Np < oo for some z > 1 and all i. X > 0 a r. 6 Let (N1. . If Jo = i. .v.F) = P(Yo > X+x I •^) G (x +x)>2ciNi i=1 . HEAVY TAILS Lemma 4 . oo) and define p Ni Yx = EEX'i . i =1 P(Yo > x I ^ ) < CG(x)zN1+ +Np for some C = C(z) < oo.Fmeasurable. P P P(YX and > x I. NP ) be a random vector in {0.}P. and the rest of the argument is then just as the proof of Corollary 4. as x a oo. Then P P(Yx > x) .. u Proof of Theorem 4. cp with cl + > 0 it holds that Fi(x) . Thus dominated convergence yields ( P(Yo>x P(Yo>x .2.. NP ) and X are . .X i=1 j=1 where conditionally upon F the Xi. i=1 P(Yx > x ^) < P(Y0 > x I.F) < CG(x)zn'1+. are independent with distribution Fi for Xij..^•) G(x) P ^ E ciNi = C.270 CHAPTER IX. i1 = E\ G(x) In the general case.. i=1 Proof Consider first the case X = 0. and F a aalgebra such that (N1. 2 . .c'(x) where c = ciENi . and that for some + cp distribution G on [0..5.G( x ) > ciNi ..
3.e. 5 Finitehorizon ruin probabilities We consider the compound Poisson model with p = /3pB < 1 and the stationary excess distribution Bo subexponential. I T(u) < oo). Within the class of risk processes in a Markovian environment.T2. In contrast.d. As usual. 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. see Schlegel [316]. this then easily yields approximations for the finite horizon ruin probabilities (Corollary 5.2 and Example 4. > 0) matter for determining the order of magnitude of the ruin probabilities in the heavytailed case. Schmidli & Schmidt [47].. cf. 5 was first proved by Asmussen.. and the final reduction by Jelenkovic & Lazar [213]. It follows from Theorem 4. Notes and references Theorem 4.. VI.4. FINITEHORIZON RUIN PROBABILITIES 271 probabilities for large u.6) to hold in a situation where the interclaim times (T1.1 is from Asmussen. Theorem 5.. The present approach via Theorem 4.4.i. An improvement was given in Asmussen & Hojgaard [33]. For further studies of perturbations like in Corollary 4. ) form a general stationary sequence and the U. Theorem 2. in particular Proposition 2. We start by reviewing some general facts which are fundamental for the analysis. for lighttailed distributions the value of the adjustment coefficient y is given by a delicate interaction between all B.4. ). IV. this should be compared with the normal limit for the lighttailed case. Theorem 4.1. cf.. i. The main result of this section. 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). states that under mild additional conditions. Combined with the approximation for O(u). Essentially. Floe Henriksen & Kliippelberg [31] by a lengthy argument which did not provide the constant in front of Bo(u) in final form.5. and independent of (T1. we let PN"N = P(.7.5 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. Then O(u) . this is applied for example to risk processes with Poisson cluster arrivals.4. That paper also contains further criteria for regenerative input (in particular also a treatment of the delayed case which we have omitted here).p)Bo(u).5 shows that basically only the tail dominant claim size distributions (those with c. as well as a condition for (4. r(u) is the time of ruin and as in IV. i.pl(1 .. m is a (orfinite) .7).T2. the discussion provides an alternative point of view to some results in Chapter IV. cf.
y = 0).2) means ffh(a. Sw(F. Rt is distributed as x + t . to consider only the case Px(w(F`) = 0) 0.00). say. For the present purposes it suffices .t. and starting from So = y.z. (note that we allow x. and (5.2) with t = 1 means m. Proof Starting from Ro = x.>N` Ui. Let G denote the distribution of ENt U. Thus.t + EI U. a familiar case is time reversion (here m is the stationary distribution). t. however . k on E. . 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} . y to vary in. x = 0+ and F = (0. where we can take h.1 A compound Poisson risk process {Rt} and its associated claim surplus process {St} are in classical duality w . but the example of relevance for us is the following: Proposition 5. The equality of the l. resp.2) for all bounded measurable functions h. {St} and {Rt} are in classical duality w. Say {St} is reflected Brownian motion on [0. oo). Lebesgue measure. for states i. We let QS be the corresponding distribution and Qx. The simplest example is a discrete time discrete state space chain.s. m.s=j are the transition probabilities for {St}. u For F C E. r. {Rt}.). a main difficulty is to make sense to such excursions also when Px(w(F°) = 0) = 1.r.)k(x .y = Qx (.s. Then (5. w(Fc) < oo ) 'In general Markov process theory. in the terminology of general Markov process theory. the whole of R and not as usual impose the restrictions x > 0.h. to the r. St is distributed as y .t. .rij = mjsji where r13. HEAVY TAILS measure on E such that L for all measurable A C E and all t > 0.= y. k as indicator functions. 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. follows by the substitution y = x . j.h.z) dx G(dz) = ffh(y + z) k(y)dy G(dz).272 CHAPTER IX.
itt) = P Px(w(Fc) < 00. Sn+1 E Fc) nx.y = Qy Q. Sn = in = y.13AB < 1] Proof of Theorem 5.SS(F... when p = . 5. FINITEHORIZON RUIN PROBABILITIES 273 y E F (in discrete time. Qx y is the distribution of an excursion of {St} conditioned to start in x E F and terminate in y E F.5. The theorem is illustrated in Fig . But in the risk theory example (corresponding to which the sample paths are drawn). . io = x.1 The sample path in (a) is the excursion of {St} conditioned to start in x = 0 and to end in y > 0. In particular: Corollary 5.s. and we let Qy y refer to the time reversed excursion .y(2p21 . in = y. Sw(Fo)_ should be interpreted as Sw(F^)_1). /^s x (S1 = Z1. . The theorem states that the path in (b) has the same distribution as an excursion of {Rt} conditioned to start in y < 0 and to end in x = 0. . We can then view Qy.2 Qy. oo) = r(0) x= St y (a) Figure 5. S. Sw(F)1 = y) . That is. z > 0. [note that w(z) < oo a. i1...1 for the case F = (oo. Thus.y() = P ({SW(F`)t} 0<t<w(F °) E So = x.= y) Theorem 5 . the one in (b) is the time reversed path. x = 0. 0]. in with i0.2. Qx.. w(0. QR and QRy are defined similarly.(0)_ = y < 0 is the same as the distribution of w(y) where w(z) = inf It > 0 : Rt = z}... in E F..3 The distribution of r(0) given r(0) < oo. .y as a measure on all strings of the form i0i1 . 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). this simply means the distribution of the path of {Rt} starting from y and stopped when 0 is hit.).
.ii ..gilt' k=1 ii .. R Qy x(2p21 .ik_1EF Sxin_1 ...... 2n) = Qx. . Rn = in = x. in with 20... i0) Q x. in) = Qx.... . S.(F<)1 = Y) S S and Qx y( ipil . R . in = x. .. .274 note that Fx(w(Fc) < 00..i„_iEF Similarly. in = x. = in = y. (Fc)1 = y) 00 CHAPTER IX.. in)  Pt' (R1 = ii. 21 ..... 2p) when 20..TI( 2n2n _1 .. Si11 S 1 .. To show Q y x (i0 i 1 . Si l io E mjSjx. Rn+1 E FC) TioilTili2. Silt' E SO k=1 i1. MY Thus Qx(ioii .. Rn = in = x. .. . in) = oo jEF^ Sxin1 . .ik_1EF . Si1y 00 jEF° E E 5xik_ 1 .. i0 = y..... Sn+1 E Fc) n=1 i1..y(inin _ E SYj jEF` 00 Sxik _1 .... Silt' E Sxik_1 .rin_1in E Txj jEFC m21 s2120 m2252221 m in Ssn n1 mjSjx Mx m2p mil min1 jEF` 1 Sinin _ 1 .. 20 = y.. in E F.. t' y and Qy x are measures on all strings of the form ipi l . note first that Pt' (R l = il.. Si1y k=1 i1 . ..... Rn+1 E F`) F (w(Fc) < 00... HEAVY TAILS E E Px (Si = 21i ....in1 .ik1EF Similarly but easier Sxin_1 ... S..in E F.. 2p)..... .....J (i.
That is.')density of Y is B(y)/[.2. 7(0) < oo. Z follows the excess distribution B(Y) given by B(Y) (x) _ B(y + x)/B(y). y > u. the P(u. Bo") is also the P(u. the distribution w. 5.p. 1 w . P(") = P(. Now the P(u.')distribution of Z since P(Z>aIY>u) = 1 °° B(y) B(y + a) dy FLBBo(u) B (y) J°° (z) dy . S.t.2.')distribution of Yu is Bo").'s are defined w. that is. The formulation relevant for the present purposes states that Y has distribution Bo and that conditionally upon Y = y. 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 .B(a) +a PBBo(u) .5.v. Let Y = Yl = Sr+( 1) be the value of the claim surplus process just after the first ladder epoch .t. P(o) ). ST(o) > y. see Fig.r.UBBo(u)]. To clarify the ideas we first consider the case where ruin occurs already in the first ladder segment . We are interested in the conditional distribution of T(u) = T(0) given {T(0) < oo.r. the case r (O) < oo. U T(O) = T (u) Y Figure 5. Y > u). that is. Z = Zl = ST+( 1)_ the value just before the first ladder epoch (these r.(o) > y} = {T(0) < oo.2 The distribution of (Y.2. Z) is described in Theorem 111. Y > y} .
.. . Yn_1 'typical'.p) in Pi"'')distribution... Hence Z.. Bo") ).r+ (n . i. Since w(z)/z a$. must be large and Z1.e.. and distributed as (Y.o be defined by w(z) = inf It > 0: Rt = z} where {Rt} is is independent of {St}. We let K(u) = inf In = 1.p) then yields the final result T(u)/y(u) + W/(1 . Recall the definition of the auxiliary function y(x) in Section 1. .p)..1. then by the subexponential property Yn must be large. Zn_1 'typical' which implies that the first n1 ladder segment must be short and the last long. > u with high probability. more precisely. Z). Since the conditional distribution of Z is known (viz. z ^ oo. the duration T+ (n) . Then Corollary 5. in particular of Z.3) where the distribution of W is Pareto with mean one in case ( a) and exponential with mean one in case (b). are i. 2. 5.T+(2)..e.p). Y1 + • • • + Yn > u} denote the number of ladder steps leading to ruin and P("'n) = P(• I r(u) < oo. Then 7(u)/y(u) ^ W/(1 . Z1).. this in principle determines the asymptotic behaviour of r(u). K(u) = n). . i. .i. cf. That is . However.: r+ (n) < oo. Now Bo E S implies that the Bo ")(a) + 0 for any fixed a. let r+(1) = T(0). We now turn to the general case and will see that this conclusion also is true in P(")distribution: Theorem 5 . In the proof. It is straightforward that under the conditions of Proposition 1. Z = ZI but relative to the kth ladder segment. r(u)/Z 4 1/(1 .p) in F(u) distribution. HEAVY TAILS Let {w(z)}Z^. we get the same asymptotics as when n = 1.276 CHAPTER IX.1) of the last ladder segment can be estimated by the same approach as we used above when n = 1. and YI. 1/(1 . The idea is now to observe that if K(u) = n. it therefore follows that T(u)/Z converges in Pi"'')probability to 1/(1 .18(c) Bo")(yY (u)) + P(W > y) ( 5. Zn). a slight rewriting may be more appealing. .d. . 4 Assume that Bo E S and that (5. the random vectors (YI... P(Z < a I Y > u) 3 0.. denote the ladder epochs and let Yk. (Y. Zk be defined similarly as Y = Y1. conditionally upon r+ (n) < oo... and since its dominates the first n .3) holds..3 implies that the P("'1)distribution of T(u) = r(0) is that of w(Z). Fig. Z/'y(u) * W in Pi "' ')distribution ..3. Then.
u) E • I A'(u)) = Bo (n1) ®Bou) .3 In the following. I A'(u)) = P(u. suitably adapted). Y„1. Proof We shall use the easily proved fact that if A'(u).u) II 0. A"(u) are events such that P(A'(u) AA"(u)) = o(F (A'(u)) (A = symmetrical difference of events). FINITEHORIZON RUIN PROBABILITIES 277 16 Z3 Z1 r+(1) T+(1) T+(1) Figure 5. . P(. ...u) E •) .5. Further.. > u}... A"(u) _ {K(u)=n} = {Y1+ P(. II ' II denotes the total variation norm between probability measures and ® product measure.Bo (ri1) ®B( . +Yn1<u.5 Ilp(u.n).Yn1iYn . the condition on A'(u) A A"(u) follows from Bo being subexponential (Proposition 1.n) (y1.. I A"(u ))II + 0.2.Yl+ +Yf1>u}. Yn . Lemma 5. . .. P (Yj. then IIP( I A'(u)) Taking A'(u) = {Y.
Y'. n .. The same calculation as given above when n = 1 shows then that the marginal distribution of Zn is Bou). + Y" > u) Flul (K (u ) = n) _ Cu) P"F(1'i +. Proof of Theorem 5. Z' are arbitrary random vectors. . Zn) E •) .P) Bo(u) for n = 1.P(Y' E •)II * 0.P(Z' E •)II > 0 (here Y. copies of {w(z)}.. see Fitzsimmons [144]).. It therefore suffices to show that the P(u'")distribution of T(u) has the asserted limit. .t.. . then 11P(Z E •) . in our example Y = (Y1..4). 2.u has distribution Bout That is. Z.' = y is BM. y > u.d. ..Bo (n1) ®Bo' 0. in particular his Proposition (2.. wk(Zk) has a proper limit distribution as u + oo for k < n. whereas wn(Zn) has the same limit behaviour as when n = 1 (cf... +wn(Z n))l7( u ) > 1y) ^' P(u'n)(wn (Zn)/7(u) > y) 4 NW/(1 . {wn(z)} be i.P) > y) Corollary 5.. Similarly (replace u by 0).4. . Proof Let (Y11. (Y. . Then according to Section 5a.. the F'distribution of r(u) is the same as the P'distribution of w1(Zl) + • • • + wn(Zn). .. The first step is to observe that K(u) has a proper limit distribution w. 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 •) . n_1 < u.6 IIPIu'n ) CHAPTER IX... Let {wl(z)}..7 O (u...1 P PBo(u) • P(W/(1 .p) < y).1.6..i.+y 1 p"F(Yn > u) P)Pn1 P/(1 .2.. Zn are independent.1).. . the marginal distribution of Zk is Bo for k < n. Notes and references Excursion theory for general Markov processes is a fairly abstract and advanced topic...278 Lemma 5 . Z11). . .). . k = 1. Thus F(u'n)(T(u) /7(u) > y) = F(u'n)((wl (Z1) + .. By Lemma 5. Y1 +. n. be independent random vectors such that the conditional distribution of Zk given Y. Y") u etc.r. the discussion just before the statement of Theorem 5.. the density of Yn is B(y)/[IBBO(u)].. P(u) since by Theorem 2. and clearly Zi.y(u)T) . Zn). . and that Yk has marginal distribution B0 for k = 1..1 and Y„ ... HEAVY TAILS ((Z1'. For Theorem 5..
T) when T + oo with u fixed. 3. the results only cover the regularly varying case.e. i. p(Y) and the result follows. however. V.(u) = P(V > u) = f f (y) dy . 6 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate /3. nontrivial and we refer to Asmussen [22]. 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). 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 . cf. Proof of Theorem 6. Corollary II.1 Assume that B is subexponential and that p(x) > 00.y) .. = supo<t<0.1) The key step in the proof is the following lemma on the cycle maximum of the associated storage process {Vt}. Asmussen & Teugels [53] studied approximations of i (u. one expects the level y form which the big jump occurs to be 0(1). the probability that is exceeds u is then B(u . Extensions to the Markovmodulated model of Chapter VI are in Asmussen & Hojgaard [33].(3 u u J B(y) dy .B(u). Assume for simplicity that {Vt} regenerates in state 0 . that fo p(x)1 dx < oo. that MQ becomes large as consequence of one big jump. max VB>0I Vo=0^ o<s<t J11JJJ Lemma 6 . Then P(MT > u) . We will show that the stationary density f (x) of {Vt} satisfies f (x) /B(x) r(x) We then get V.2 Define M. Then 0 (u) Qf "O ^) dy.2. x > oo. .1. The heuristic motivation is the usual in the heavytailed area.6./3Ea B(u). and define the cycle as a = inf{t>0: Vt=0. Theorem 6 . u (6. The rigorous proof is. More precisely. and premium rate p(x) at level x of the reserve. claim size distribution B.
Then D(u) = f(u)p(u) and.P(MT > u) $B(u) Ft µ(1 .q(u) Now just use that p(x) * oo implies q (x) + 0. where also the (easier) case of p(x) having a finite limit is treated . D(u) = DQ(u)/µ. 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). Hence f (u)r(u) = D(u) = Do(u) .280 CHAPTER IX. by regenerative process theory. u Notes and references The results are from Asmussen [22]. It is also shown in that paper that typically. there exist constants c(u) 4 0 such that the limiting distribution of r(u)/c(u) given r(u) < oo is exponential.q ( u)) 1 . . HEAVY TAILS Define D(u) as the steadystate rate of downcrossings of {Vt} of level u and Da (u) as the expected number of downcrossings of level u during a cycle.
. We shall be brief concerning general aspects and refer to standard textbooks like Bratley. 4Z). Hence 1. 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. z) 2 = Zit NE ii ii According to standard central limit theory . Fox & Schrage [77].Chapter X Simulation methodology 1 Generalities This section gives a summary of some basic issues in simulation and Monte Carlo methods . topics of direct relevance for the study of ruin probabilities are treated in more depth. Rubinstein [310] or Rubinstein & Melamed [311] for more detail . 281 . where a2 = Var(Z ). Ripley [304].. estimating z by the empirical mean (Z1 + • • + ZN)/N and the variance of Z by the empirical variance N s2 = E(Z{  N 2. replicates Zl. . ZN. and this is the form in which the result of the simulation experiment is commonly reported.i.2) is an asymptotic 95% confidence interval .96s z f (1.z) 4 N(0. vrN(z . The crude Monte Carlo ( CMC) method then amounts to simulating i.d..
Conditional Monte Carlo Let Z be a CMC estimator and Y some other r . We survey two methods which are used below to study ruin probabilities. We mention in particular ( regression adjusted) control variates and common random numbers. The situation is more intricate for the infinite horizon ruin probability 0(u). However. writing Var(Z) = Var(E [Z I Y]) + E(Var[Z I Y]) . Letting Z' = E[Z I Y]. conditional Monte Carlo and importance sampling. there are others which are widely used in other areas and potentially useful also for ruin probabilities. v. 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.b(u. SIMULATION METHODOLOGY In the setting of ruin probabilities. Z = I inf Rt < 0 (0<t<T = I('r(u) < T). 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. and a longer CPU time to produce one replication. one can argue that unless Var(Z') is considerable smaller than Var(Z). and in most cases this modest increase of N is totally unproblematic. an added programming effort. Typically variance reduction involves both some theoretical idea (in some cases also a mathematical calculation). Then replacing the number of replications N by 2N will give the same precision for the CMC method as when simulating N' = N replications of Z'. Further. Say that Var(Z') = Var(Z)/2.282 CHAPTER X. variance reduction is hardly worthwhile. so that Z' is a candidate for a Monte Carlo estimator of z. Sections 24 deal with alternative representations of Vi(u) allowing to overcome this difficulty. This is a classical area of the simulation literature. generated at the same time as Z. it is straightforward to use the CMC method to simulate the finite horizon ruin probability z = i. we then have EZ = EZ = z. Therefore. typically by modifying Z to an alternative estimator Z' with EZ' = EZ = z and (hopefully) Var(Z') < Var(Z). lb Variance reduction techniques The purpose of the techniques we study is to reduce the variance on a CMC estimator Z of z. and many sophisticated ideas have been developed.
1.. Thus we cannot compute L = Z/z (further. it appears that we have produced an estimator with variance zero. L such that z = EZ = E[LZ].zrs. LN) from P and uses the estimator N zrs = N > L:Zj i=1 and the confidence interval zrs f 1. GENERALITIES 283 and ignoring the last term shows that Var(Z') < Var(Z) so that conditional Monte Carlo always leads to variance reduction. and the problem is to make an efficient choice.v. L = z/Z (the event {Z = 0} is not a concern because P(Z = 0) = 0). the obvious possibility is to take F and P mutually equivalent and L = dP/dP as the likelihood ratio. one would try to choose P to make large values of Z more likely. . i.96 sis v^ N 2 1 where srs = N j(LiZi .zrs) = 2 1 N 2 2 2 i=1 i=1 N > Lt Zi . Nevertheless. . it may often be impossible to describe P in such a way that it is straightforward to simulate from P). In order to achieve (1. (1.3) Thus.. even if the optimal change of measure is not practical. using the CMC method one generates (Z1. L1). Then z Var(LZ) = E(LZ)2 . However.E [Z Z]2 = z2 .3). Thus.[E(LZ)] = E Z2 Zz . To this end.z2 = 0.e. 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. it gives a guidance: choose P such that dP/dP is as proportional to Z as possible. a crucial observation is that there is an optimal choice of P: define P by dP/dP = Z/EZ = Z/z. Variance reduction may or may not be obtained: it depends on the choice of the alternative measure P. This may also be difficult to assess . but tentatively. the argument cheats because we are simulating since z is not avaliable analytically. . (ZN.
This leads to the equation 1. say of the order 103 or less.1.z) 1001. Two established efficiency criteria in rare events simulation are bounded relative error and logarithmic efficiency. Thus. i.e.284 CHAPTER X. u + oo. To introduce these.100 . in terms of the halfwidth of the confidence interval. z I. large sample sizes are required. assume that the A(u) are rare in the sense that z(u) * 0. the optimal P is the conditional distribution given A. For each u. Another way to illustrate the problem is in terms of the sample size N needed to acquire a given relative precision .B = iP(AB) = P(BIA). a confidence interval of width 10 4 may look small. and further it is usually not practicable to simulate from P(•IA).96 2 z2 z increases like z1 as z . if z is small.1. Again.e.e. 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. as is the case of typical interest. We then .. let z(u) = P(A(u)).5 or even much smaller . I.96oz /(zV) = 0. N . assume that the rare event A = A(u) depends on a parameter u (say A = {r(u) < oo}). say 10%. Z = I(A) and A is a rare event.96 2Z ( 1 . it does not help telling whether z is of the magnitude 104. but if the point estimate z is of the order 105. the issue is not so much that the precision is good as that relative precision is bad: oZ z(1 . we may try to make P look as much like P(•IA) as possible. In ruin probability theory.z) which tends to zero as z ^ 0. The optimal change of measure ( as discussed above) is given by P(B) = E [ Z] i. SIMULATION METHODOLOGY 1c Rare events simulation The problem is to estimate z = P(A) when z is small . and let Z(u) be a Monte Carlo estimator of z(u). However.0. An example where this works out nicely is given in Section 3. We shall focuse on importance sampling as a potential (though not the only) way to overcome this problem. A = {T(u) < T} or A = {r(u) < oo} and the rare events assumption amount to u being large.z) 1 > 00. However.. Z z V5 In other words . 10 . The CMC method leads to a variance of oZ = z(1 .
The algorithm gives a solution to the infinite horizon problem .4). 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.p)pk.. . 2 Simulation via the PollaczeckKhinchine formula For the compound Poisson model. Notes and references For surveys on rare events simulation. and in practice. If M > u. However.. let Z +.. . it is not efficient for large u . This allows Var(Z(u)) to decrease slightly slower than z(u)2. let Z +.1. with common density bo(x) = B(x)/µB and K is geometric with parameter p. logarithmic efficiency is almost as good as bounded relative error. this means that the sample size N = NE(u) required to obtain a given fixed relative precision (say a =10%) remains bounded. Thus. F(K = k) = (1 . see Asmussen & Rubinstein [45] and Heidelberger [190]. but as a CMC method .log z(u) of (1. Var(Z(u)) hm sup U+00 z (u) 2E < oo (1. SIMULATION VIA THE POLLACZECKKHINCHINE FORMULA 285 say that {Z(u)} has bounded relative error if Var(Z(u))/z(u)2 remains bounded as u 3 oo. it is appealing to combine with some variance reduction method .d. The term logarithmic comes from the equivalent form . XK from the density bo(x). O (u) = z = EZ. the PollaczeckKhinchine formula III.p)pk. P(K = k) = (1 .0. 3. i. which gives a logarithmically efficient estimator .2..i.e. where Z = I(M > u) may be generated as follows: 1. the mathematical definition puts certain restrictions on this growth rate. According to the above discussion.4) for any e > 0. . so that NE (u) may go to infinity. Let M . are i. We shall here present an algorithm developed by Asmussen & Binswanger [ 271. 2.log Var(Z(u)) lim inf > 2 u+oo .1) may be written as V) (u) = P(M > u). Therefore . where M = X1 + • • • + XK. X2. Generate X1. Otherwise.X1 + + XK. where X1.(2. Generate K as geometric.
. .X(n_1)) Bo(X(„_l) V X) Bo(X(n1)) . X1 + + XK_ 1 > x when X1 > x... . compute Y = u . Xl > x. . For the simulation.1) V)(u) .... note first that To check the formula for the P(X(n) > x I X(1)....p/(l . y < 0). SIMULATION METHODOLOGY when the claim size distribution B (and hence Bo) has a regularly varying tail.X(K1)) . 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..Xl . and let Z(2)(u) = _ P (SK B0((u > u I X(l).X(2).2. and considering only the remaining ones.. 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. just note that EZ(1)(u ) 2 > E[Bo (x . asymptotically it presents no improvement : the variance is of the same order of magnitude F(x)..S( K_1)) V X(K1)) / Bo(X(K 1)) where S(K_l) = X(1) + X(2) + • • • + X(K_1). form the order statistics X(1) < X(2) < .. The idea of [27] is to avoid this problem by discarding the largest X.+XK > uIXl. . So. and that Bo(y) = 1.XK_1). assume in the following that Bo(x) . we generate only X1. we thus generate K and X1i . Theorem IX.. However.p)Bo(x). As a conditional Monte Carlo estimator ..SK1)2.X(2). A first obvious idea is to use conditional Monte Carlo: write i. Z(1)(u) is defined as 0)... Z(1) (u) has a smaller variance than Zl (x)..b(u) = P (Xl +•••+XK>u) = EF[Xl + ... conditional probability. Then (cf. XK... XK1.X1 . K > 2] = P2p(Xl > x) = P2Bo(x) (here we used that by positivity of the X.L(x)/x`' with a > 0 and L(x) slowly varying.XK_1 and let Z( 1)(u) = Bo (Y) (if K = 0. Thus.XK1] = EBo(uX1 . To see this. < X(K) throw away the largest one X(K).286 CHAPTER X.......
. use the the CramerLundberg approximation so that z(u) = '(u) = e7"ELe7E(") where ^(u) = ST(") . . and we refer to [27]. . Asmussen . B. Notes and references The proof of Theorem 2.1) . . . IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 287 We then get P(S" > x I X( 1)._1) > P(X(n) > _ X X(1). that is. X(2).1 is elementary but lengty. The algorithm is sofar the only efficient one which has been developed for the heavytailed case. Binswanger and HOjgaard of [28] give a general survey of rare events simulation for heavy tailed distributions . X(2). . Thus. X (.3.. X(2). For practical purposes. 1 Assume that Bo (x) = L(x)/x° with L(x) slowly varying.Khinchine formula and importance sampling . and define )3L. BL by I3L = /3B[y]. and simulate from FL. BL instead of 0.u is the representation 0(u) = e7sr(u) overshoot (cf. 3 Importance sampling via Lundberg conjugation We consider again the compound Poisson model and assume the conditions of Ce7". X(n1)) P(X(TZ) + S(. it must be noted that a main restriction of both algorithms is that they are so intimately tied up with the compound Poisson model because the explicit form of the PollaczeckKhinchine formula is crucial (say.. the algorithm for generating Z = Z(u) is: 1.5). 111. for the purpose of recording Z(u) = erysr(u). 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.S(n_1) I X(1).y. the continuoustime process {St} is simulated by considering it at the discrete epochs {Qk} corresponding to claim arrivals. l)) . l)) BO(X(n1)) Theorem 2 . However .S (n1)) V X (.. BL(dx) = e7sB(dx)/B[y]. X(n1)) Bo((x . Then the algorithm given by { Z (2) (u) } is logarithmically efficient. and that paper contains one more logarithmically efficient algorithm for the compound Poisson model using the Pollaczeck.. using 13L.modulated model P(r+ < oo) and G+ are not explicit ).. . Compute y > 0 as solution of the Lundberg equation 0 = K(y) = )3(B[y] . X . in the renewal or Markov.
3. We formulate this in a slightly more general random walk setting '.QL.i. Let X1. and we have: Theorem 3. to deal with the infinite horizon problem . b) # (/3L. In detail . cf. = X1 + . Generate T as being exponential with parameter . with distribution F. BL). let S. X2. The answer is no. More precisely. If S > u.288 2.. B) is not logarithmically efficient when (/3. r(u) < oo) and FL (both measures restricted to. Proof Just note that EZ(u)2 < e . Let S . one must restrict attention to the case 4µB > 1. . 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. .. SIMULATION METHODOLOGY 3. return to 3. The algorithm generalizes easily to the renewal model .d.1 The estimator Z(u) = e'rs* "u) (simulated from FL) has bounded relative error...2ryu _ z (u)2/C2.l3 and U from B. P'[y] < oo for some ry > 0.S+U . In fact: Theorem 3. so that changing the measure to FL is close to the optimal scheme for importance sampling .. Xi = U. The proof is given below as a corollary to Theorem 3.F. 4. A > AL as in Chapter V.r(u) < oo) = 1. be i.(u)) are asymptotically coincide on {r(u)} < oo.T. We may expect a small variance since we have used our knowledge of the form of 0(u) to isolate what is really unknown.1) (simulated with parameters ^3. Otherwise. M(u) = inf {n : S„ > u}. There are various intuitive reasons that this should be a good algorithm.. and the change of measure F r FL corresponds to B > BL. b different from . Let FL (dx) = 'For the renewal model. It resolves the infinite horizon problem since FL(. and assume that µF < 0 and that F[y] = 1.. Ti.2 The estimator (3. u It is tempting to ask whether choosing importance sampling parameters . let Z F e_'s.7 tell that P(. BL could improve the variance of the estimator . and avoid simulating the known part e7". namely ELery£("). Let Sf0 CHAPTER X. + X. the discussion at the end of Section 1b.Q. the results of IV.
where e' = EL Iog dFL (Xi) > 0 by the information inequality. IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 289 e7yF(dx). + KM(u))} = exp {ELM(u)(E . write W(F IF) _ F(XI). Jensen's inequality and Wald's identity yield EpZ(u)2 > exp {EL(K1 + . More generally. it thus follows that for 0 < e < e'/ELXi. When F # FL. where Kl og (X) (j) 2 ) = log dFL (Xi) ... The importance sampling estimator is then Z( u) = e'rSM( )..yu+elu u +oo etry' 1 > lim up C2e2. For the second.. is not logarithmically efficient.i.2) (simulated with distribution F of the X3 has bounded relative error when .. Here ELK.2'X1 .d. . Since ELM(u)/u + 1/ELXi. are i. let F be an importance sampling distribution equivalent to F and M(u) dF Z(u) _ I (Xi) .. EFZ(u)2 EFZ(u)2 lim sup z(u)2eeU = lim cop C2e2.3 The estimator (3.. By the chain rule for RadonNikodym derivatives...3. = c'.2) dF Theorem 3. K2. Proof The first statement is proved exactly as Theorem 3 . EFZ(u)2 = EeW2(FIF) = Ep [W2(FIFL)W2(FLIF)] = EL [W2 ( FIFL)w(FLIF)] = ELexp {Kl+. . (3.2 > 0.2ryELXi)} .+KM(u)}. 1. Since K1.yu = G.2ryELXi. F(XM(u)).P = FL.
e. with the present (shorter and more elementary) proof taken from Asmussen & Rubinstein [45]. Consider compound Poisson risk process with intensities /3'.T".4 indicate that we can expect a major difference according to whether y < 1/r. The extension to the Markovian environment model is straightforward and was suggested in Asmussen [ 16]. U' . T". Further discussion is in Lehtonen & Nyrhinen [245]. /3".2. U' .4.1 is from Lehtonen & Nyrhinen [244].T' has a left exponential tail with rate /3' and U" .T". The results of IV. optimality is discussed in a heavy traffic limit y 10 rather than when u + oo. 4 Importance sampling for the finite horizon case The problem is to produce efficient simulation estimators for '0 (u. This immediately yields / = 3".'(y) or y > 1/r. see e.1 If y > 1/ic'('y). yu) is close to zk(u). The queueing literature on related algorithms is extensive .290 which completes the proof.T' = U" . U". CHAPTER X. In [13]. BL) has bounded relative error. the references in Asmussen & Rubinstein [45] and Heidelberger [190]. 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 .i. /3'eQ'YR'( x + y) dy = . then the estimator Z(u) = e7Sr(°)I(r(u) < yu) (simulated with parameters /3L. T) with T < oo.3. The easy case is y > 1/k'(y) where O(u. The optimality result Theorem 3.3"eQ x 0 J eQ zB (z) dz x (x > 0) and /3' = /3". Next.g. SIMULATION METHODOLOGY u Proof of Theorem 3. Then according to Theorem 3.T" has a left exponential tail with rate /3'.T" > x) J /3"e0 yB (x + y) dy = . we conclude by differentiation that Bo(x)=B' (x)forallx > 0. we write T = yu. so that one would expect the change of measure F 4 FL to produce close to optimal results. claim size distributions B'. First by the memoryless distribution of the exponential distribution . from 3' P(U'T'>x) ^ = ^ eQ'zB (z) dz. all that needs to be shown is that if U' .T' D U" . .B'=B". then /3' B' = B". In fact: Proposition 4. As in IV. generic interarrival times T' . B" and generic claim sizes U'.'(y).3'eO'x f f P (U" .
3 Theorem IV. IMPORTANCE SAMPLING FOR THE FINITE HORIZON CASE 291 Proof The assumption y > 1/n'(y) ensures that 1fi(u.yy> 2 .4. yu)/z.yu. yu) is of order of magnitude a71. lim inf uoo 27yu . Bay) is logarithmically efficient. yu) in the sense that .4. (4.1).3) (simulated with parameters /gay.4.2) Since the definition of ay is equivalent to Eay r(u) . Proof Since ryy > y.yk(ay) determines the order of magnitude of z'(u.to g x ( u ) u u so that (1.5) follows.(ay). the result follows as in the proof of Theorem 3. Further .1. we have ic(ay ) > 0 and get Eay Z(u)2 = Eay [e  2aySr( u)+2r(u )r. T(u) < yu] .2 The estimator (4.8 has a stronger conclusion than (4. (4.1) which is all that is needed here can be showed much easier .log Var(Z(u)) _ .log 4')u) 4 u (Theorem IV. and that ryy > ry. one would expect that the change of measure F Pay is in some sense optimal.8). and in fact. Remark 4 .(u) * 1 (Theorem IV.1). Bounding u ELZ(u)2 above by a7u. 7y (4. T( u) < yu] e2ryyuEay le 2ay^(u). We next consider the case y < 1/r.O(u.'(7). The corresponding estimator is Z(u) = eavS' ( u)+T(u)K (ay)I(T( u) < yu). T(u) < yu] e Hence by (4. that ryy = ay . We recall that ay is defined as the solution of a'(a) = 1/y.4.3) and we have: Theorem 4. Let Qy2 = . .log Var(Z(u)) l im of . yu) = eayu Eay Leay^(u)+r(u)K(ay).1) so that z(u) = zP(u.
N(0.o. we believe that there are examples also in risk theory where (5. 0 Notes and references The results of the present section are new.1) (see Proposition IV.b(u.2).a vt(u). '%(u) = P I info Rt < 0) = P(VV > u).3: for many risk processes {Rt }.4. Z (u)2 above..7y x(u) > hm inf u+Oo U . related discussion is given in a heavy traffic limit q J.o . yu .Qyu1/2 < T(u) C yu e.uaoo U That lim sup < follows similarly but easier as when estimating En.2) .4. there exists a dual process { V t} such that i.u1/2 < r(u) < yu Le ] l = e. yu . One main example is {Vt} being regenerative (see A. 0 rather than when u 3 oo.(av)Eav l e. However.1): then by Proposition A1.1) is used to study Voo by simulating {Rt} (for example. > u) = E f I(VV > u) dt 0 (5.T) = P O<t<T inf Rt < 0 = P(VT > u).292 CHAPTER X.ryyu +oy u1/2K'(av)Eo l v 1/2) where the last step follows by Stam's lemma (Proposition IV. In most of the simulation literature (say in queueing applications).1) where the identity for Vi(u) requires that Vt has a limit in distribution V.3. Hence lira inf log ryyu + vyu 1/2 tc(ay) .a yu +l/ur' (av)Ei`av reav^(u)+(T(++)(U) yu . the object of interest is {Vt} rather than {Rt}. the algorithm in Section 3 produces simulation estimates for the tail P(W > u) of the GI/G/1 waiting time W). SIMULATION METHODOLOGY Vara„ (r(u))/u so that (T(u) . zi(u) = INV.yu)/(uyu1/2) . Then z(u) = Eay Z(u) > Eay avS'(u)+T( u)k(av 1 ).yu1/2 <1 T(u) < yu l r > e7vu +avul/ 2r.4). and (5. In Asmussen [13]. 5 Regenerative simulation Our starting point is the duality representations in 11. (5.1) may be useful.
For the ith cycle. j = 1. record Zi'i = (Z1').t(u)) 4 N(0. Z2'> the time during the cycle where { Vt} exceeds u and zj = EZJ'). oh) for h : R2 ^ R and Ch = VhEVh. + Z2N)) .3) .. EZ1'i = z1 = Ew.E). the regenerative estimator z%(u) is consistent.. To derive confidence intervals . For details . z2) z2/z1 yields Vh = (z2/z2 1/zl). Z1 = (Zl1i +. Z2 . 2. Z2 = N (X21' + . )). EZ2'i = z2 = E Thus. i. Thus.5. The method of regenerative simulation. Therefore . Vh = (8h/8z1 8h/ 8z2)....+Z(N) z 1. Then Z(1).. provides estimates for F ( V. Z(N) are i .. (u) ?2 = E fo I(Vt > u) dt = 0( u ) zl Ew as N > oo.d. Zl the LLN yields Z1 a$' Z(1) +. Then (Z1z1i Z2z2 ) 4 N2(0. consider first the case of independent cycles . z2)) > N(O. REGENERATIVE SIMULATION 293 where w is the generic cycle for {Vt}.h (zl. . which we survey below .. . > u) (and more general expectations Eg(V. is the cycle length. i (^(u) . letting J0 'o I (Vt > u) dt .. Z2 a4* z2. Thus the method provides one answer on to how to avoid to simulate { Rt} for an infinitely long time period. + Z1N>) . Simulate a zerodelayed version of {V t } until a large number N of cycles have been completed. 02) (5.. a standard transformation technique (sometimes called the delta method) yields 1 V 2 (h (Zi. and Z2'>) where Zi'i = w. Taking h(zl.. let E denote the 2 x 2 covariance matrix of Z(').
96s/v"N.C)dx = f w(x) d( f ( x.5) Z1 Z1 Z1 and the 95% confidence interval is z1 (u) ± 1.Z) ^Z(=) . Z of the form Z = ^p(X) where X is a r . 6 Sensitivity analysis We return to the problem of 111 .294 where 01 2 CHAPTER X. We here consider simulation algorithms which have the potential of applying to substantially more complex situations. () dx f Ax) (dl d()f (x' () f ( z.2. consider an extremely simple example . to evaluate the sensitivity z/i( (u ) = (d/d() 0(u) where ( is some parameter governing the risk process . () depending on C. () dx so that differentiation yields zS d( fco(x)f(x. 9. Before going into the complications of ruin probabilities . However . () dx = E[SZ] f(X.g S12 (5. Then z(() = f cp(x) f (x. v. Here are the ideas of the two main appfoaches in today 's simulation literature: The score function ( SF) method . Rubinstein [310] and Rubinstein & Melamed [311].g. Notes and references The literature on regenerative simulation is extensive.v. asymptotic estimates were derived using the renewal equation for z /i(u). the expectation z = EZ of a single r. Let X have a density f (x.0 .9. The regenerative method is not likely to be efficient for large u but rather a brute force one. with distribution depending on a parameter (. in some situations it may be the only one resolving the infinite horizon problem .2 E1 2 z1 z1 Z2 The natural estimator for E is the empirical covariance matrix N S = N 1 12 (ZW . say risk processes with a complicated structure of the point process of claim arrivals and heavy tailed claims . There is potential also for combining with some variance reduction method. SIMULATION METHODOLOGY 2 Eli = Z2 z1 + 2 E22 .z^ i=1 so a2 can be estimated by 2 2 = 72 S11+ 12 S22 . see e. In 111.
one can take h (U. ()) is 0 w . p. just take cp as an indicator function . if f (x. nonpathological examples where sample path derivatives fail to produce estimators with the correct expectation. SENSITIVITY ANALYSIS where 295 S = (d/d()f (X. () where U is uniform(0. () = . for some Co = (o(U). giving h( (U.log U/(. ()) d( hc(U. For example . Thus. however . IPA will estimate zS by 0 which is obviously not correct. ()). For IPA there are. C) f(X. = E [`d (h(U. one.r. cp' (h(U. () can be generated as h(U. C). () _ (eSx.1). ()) h((U. ()) is 0 for C < Co and 1 for C > Co so that the sample path derivative cp'(h(U.1 Consider the sensitivity tka(u) w. this phenomenon is particularly unpleasant since indicators occur widely in the CMC estimators . /3 is 0. The likelihood ratio up to r(u) for two Poisson processes with rates /3. For the SF method. In the setting of ruin probabilities . To see this. r(u) = Tl + • • • +TM(u)). A related difficulty occurs in situations involving the Poisson number Nt of claims: also here the sample path derivative w.v. () = d log f (X. So assume that a r. 11 /3oeOoT. C)). zc = E [d co(h(U. /3o is M(u) Oe (3T: < oo) . say W(x) = I(x > xo) and assume that h(U.r. the Poisson rate /3 in the compound Poisson model. cp(h(U.t. Then z(() = Ecp(h(U. () is increasing in C. () Thus. this is usually unproblematic and involves some application of dominated convergence . Let M(u) be the number of claims up to the time r(u) of ruin (thus. () = (8/8()h (u. ( where h( (u.t. The following example demonstrates how the SF method handles this situation. Thus . I(r(u) . Infinitesimal perturbation analysis (IPA) uses sample path derivatives. () is an unbiased Monte Carlo estimator of zS. SZ is an unbiased Monte Carlo estimator of z(. Then . () = log U/(2.6. with density f (x. Example 6 . The derivations of these two estimators is heuristic in that both use an interchange of expectation and differentiation that needs to be justified.() d( is the score function familiar from statistics .
4) that V5. BL).296 CHAPTER X. change the measure to FL as when simulating tp(u).t. for different models and for the sensitivities w. a relevant reference is VazquezAbad [374]. There have been much work on resolving the difficulties associated with IPA pointed out above.3 (u) is of the order of magnitude ue7u. whereas for the SF method we refer to Rubinstein & Shapiro [312]. we get 1 M(u) 00(u) = E (_Ti)I(T(U)<) E [(M(u) . since ELZp(u)2 < (M(U) _T(u) \ 1 2 a2ryu = O(u2)e27u.0(1) so that in fact the estimator Zf(u) has bounded relative error.t. .r. the risk process should be simulated with parameters .T(u)) I(T(u) < co) ] .r.1 is from Asmussen & Rubinstein [46] who also work out a number of similar sensitivity estimators. 0 Notes and references A survey of IPA and references is given by Glasserman [161] (see also Suri [358] for a tutorial). in part for different measures of risk than ruin probabilities. In the setting of ruin probabilities. We then arrive at the estimator ZZ(u) = (M(u) . We recall (Proposition 111.T(u)) e7uerVu) for ?P. However. Thus. the estimation of z(ip(u) is subject to the same problem concerning relative precision as in rare events simulation . ) we have VarL(ZQ(u)) ZO(u)2 O(u2)e2 u2e2ryu yu . differentiating w. j3 and letting flo = 0. Example 6. SIMULATION METHODOLOGY Taking expectation. To resolve the infinite horizon problem .3L.9 .3 (u) (to generate Zp (u). different parameters.
. T+(a) = inf It > 0 : Rt > al. . Oa(U ) can also be a useful vehicle for computing t/i(u) by letting a * oo. R„ = u+X. '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 ). are i.d.. Y'a(U) = P(T (u) = r+(a)) = 1 .g.(u) is defined as the probability of being ruined (starting from u) before the reserve reaches level a > u. }).. That is.. Besides its intrinsic interest .. and {1. T(u.1. a) = r(u)).+• • •+X.1}valued . a) = r(u) A T+(a).(u) = 0 ) = 0) or it is trivial to translate from one setup to the other. in most cases ...i. 297 . X2. defined as Ro = u (with u E {0. as e...P(•r(u. either this makes no difference (P(R. with P(Xk = 1) = 9. wherel T(u) = inf {t > 0 : Rt < 0} .. The twobarrier ruin problem The twobarrier ruin probability 0.Chapter XI Miscellaneous topics 1 The ruin problem for Bernoulli random walk and Brownian motion. Consider first a Bernoulli random walk. in the Bernoulli random walk example below. where X1.
. = z°Va(u) + za(1  .. MISCELLANEOUS TOPICS Proposition 1. zu = EzRO = EzRT(u.. By optional stopping.1.. one elementary but difficult to generalize to other models. Conditioning upon X1 yields immediately the recursion 'a(1) = 19+00a(2)..(4. u + 1. and the other more advanced but applicable also in some other settings. (1. The martingale is then {zuzXl+•••+X„ } = {zR° }.a(u)).o)'t/1a(a .1 For a Bernoulli random walk with 0 0 1/2.4) by ea(u+Xl+. Proof 1. where a is any number such that Ee°X = F[a] <oo. u Proof 2. The Lundberg equation becomes 1=F[ry]=(19)+9z..r(u.(1B)u oJ 0.0)/0.1) is solution. C1_0\a. then 'Oa(u) _ au a We give two proofs . Wald's exponential martingale is defined as in 11. In a general random walk setting . and in view of the discrete nature of a Bernoulli random walk we write z = e7.+Xn) F[ a]n n=0. We choose a = ry where ry is the Lundberg exponent.1) o If 0 = 1/ 2.o» = z°P (RT ( u.2) Oa(a .(u) I\ e = 1 oa ' ()i a = u.2).y] = 1.298 CHAPTER XI.o)T/la (1) + 8z/'u(3). z and the solution is z = (1 . tba(2) _ (1 . the solution of F[.. 7/la(a .1) = (19)4/'0(a3)+9ba(a1). = (1 .e.a) Y.a) = 0) + zap ( R. and insertion shows that ( 1... i.
1h (u) = a el u \1 If 9 < 1/ 2. then Vi(u) = 1.2 For a Bernoulli random walk with 9 > 1/2.3 Let {Rt} be Brownian motion starting from u and with drift p and unit variance .4 For a Brownian motion with drift u > 0. If 9 = 1/2.1 If p = 0.1) for p # 0. (1.. If p<0.1 yields 't/la(u) = (a .5) . } yields e7u = Ee7R° = e°Wa(u) + e7a(1 . thenz1 (u)=1. {Rt} is itself a martingale and just the same calculation as in the u proof of Proposition 1. u Proposition 1. {R.} 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)).u) = et(a2 /2 +aµ) the Lundberg equation is rye/2'yp = 0 with solution y = 2p. (1.1). pa( u) _ u Corollary 1.2) is trivial (z = 1).zu)/(za . Corollary 1.• a2µa e2µu . If p = 0.0a(u)).ba(u) = e2µa . i1(u) = e211 . Then for p 0 0.1). RANDOM WALK. BROWNIAN MOTION. Proof Let a+ oo in (1. then Proof Since 'Oa (U)  au a Eea(R°.e7u)/(e7° . and solving for 9/la(u) yields Z/)a(u) = (e 76 . TWO BARRIERS 299 and solving for 4/la(u) yields t/ia(u) = (za .1. Applying optional stopping to the exponential martingale {e7R. . However.u)/u.
.7/la(u)). 1 .a) < 0) + e7°P (R(u. a) I R(u a ) < 0] P (R(u . say. valid if p < 1 (otherwise .vi(a) Proof By the upwards skipfree property.a) = a ) + e ' ° ( 1 .a ) < 0) + e 7aF ( R (u.e7a (u) = 6 /0 .616).7) . 7/'(u) = 1). 7O(u) = 7/la(u) + (1 . (1.7).e7a Again .0 (u) (where u p =. Here the undershoot under 0 is exponential with rate 5.4). (u) _ O(u) .5a). we obtain 'Oa a7u .3.6 If the paths of {Rt} are upwards skipfree and 7//(a) < 1. VIII.5 Consider the compound Poisson model with exponential claims (with rate. For most standard risk processes .a) = a) = 5 y = P (R (u.300 Proof Let a * oo in (1. and thus one encounters the problem of controlling the undershoot under level 0. passing to even more general cases the method quickly becomes unfeasible (see.. and hence e7u = Ee7Ro E [e7R(. however.0(a) 0 < u < a. 5). this immediately yields (1.+^a(u))^(a) If 7k(a) < 1. implying R(u. the paths are upwards skipfree but not downwards.a) = r+ (a)} and similarly for the boundary 0. letting a * oo yields the standard expression pe7u for . 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. CHAPTER XI. Here is one more case where this is feasible: Example 1. However. Ic 5ry 'pa(u) Using y = 6 . It may then be easier to first compute the onebarrier ruin probability O(u): Proposition 1. 0.a) = a on {r (u.
Then the density and c.)_ _( u)µ2 /2. For the symmetric (drift 0) case these are easily computable by means of the reflection principle: Proposition 1. RANDOM WALK.µ%T (1. the density dPµ / dP0 of St is eµsttµ2/2.9) = 2P(ST > u). T(u) E dT..µ so that {St} is Brownian motion with drift µ . Corollary 1. and hence Pµ('r(u) E dT) = Eo [e µsr(. 0(u. of r(u) are ( U2 Pµ (T(u ) E dT) = 2^T 3/2 exp µu .ST>U).11 ) is the same as (1.d. ( 1. we have ili(u. For µ # 0. (i).8 Let {Rt} be Brownian motion with drift . We now return to Bernoulli random walk and Brownian motion to consider finite horizon ruin probabilities. and (1 . P(MT > u) = P(ST > u) + P(ST < u.f. MT > u) = P (ST > u) + P (ST > u.Rt}. + µ2T) } .7 For Brownian motion with drift 0. Hence P(MT>u. TWO BARRIERS 301 Note thas this argument has already been used in VII.8 ). MT > U) = P(ST > u) + P(ST > u) (1..T) P(MT > u) where MT = maxo<t<T St. (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 . 10) follows then by straightforward differentiation. (1.r(u).2 ..1a for computing ruin probabilities for a twostep premium function.4) I = .11) VIT ) Proof For p = 0.1. = 1 .ST<u) = P(MT>u. Here {St } is Brownian motion with drift 0 (starting from 0). BROWNIAN MOTION..10) Pµ (T(u) < T) !. T ) = P(T(u) < T ) = 241.µ T I + e2µ"4) ( .8) Proof In terms of the claim surplus process { St} = {u .
g. Here {2T( (v}TT)/2) v=T. S(oo) = f c s(y)dy.h.T (1. oo) with drift µ(x) and variance a2 (x) at x. If this assumption fails. We assume that u(x) and a2 (x) are continuous with a2 (x) > 0 for x > 0. the behaviour at the boundary 0 is more complicated and it may happen. Breiman [78] or Karlin & Taylor [222] p.s.T)dx. that 0(u).12) P(ST = v) = 0 otherwise..T) = P(ST = u) + 2P (ST > u). Proof The argument leading to ( 1.13) with 0 as lower limit of integration.13) The following results gives a complete solution of the ruin problem for the diffusion subject to the assumption that S(x). 226). MISCELLANEOUS TOPICS which is the same as (1.T+2.T2. as defined in (1.9 For Bernoulli random walk with 9 = 1/2.12) 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. u Small modifications also apply to Bernoulli random walks: Proposition 1. and (1. e..10).10 Consider a diffusion process {Rt} on [0. close to x {Rt} behaves as Brownian motion with drift µ = u(x) and variance a2 = a2(x). and in a similar spirit as in VII. such that the drift µ(x) and the variance a2(x) are continuous functions of x and that a2(x) > 0 .. 0 0 (1.g. (1. oo).3 we can define the local adjustment coefficient y(x) as the one 2µ(x)/a2(x) for the locally approximating Brownian motion. is (1. see e.8 also applies to the case 9 54 1/2. The same argument as used for Corollary 1.302 CHAPTER XI.10) and that the value at 0 is 0. Theorem 1. We finally consider a general diffusion {Rt} on [0. S(x) = f x s(y)dy. is finite for all x > 0. whenever u.9) goes through unchanged..11) then follows by checking that the derivative of the r. but we omit the details. Let s(y) = ef0 ry(. as defined above as the probability of actually hitting 0. Vi(u.9). Thus. T are integervalued and nonnegative. The expression for F ( ST = v) is just a standard formula for the u binomial distribution.
i.b(Rdt). [117].e LVa.(u) < 1 for all u > 0 and ^ S^ Conversely. . RANDOM WALK.S(u)/S(a).1.11 Let 0 < b < u < a and let t&0.S(u) (1.16) S(a) . b = 0.b(u) be the probability that {Rt} hits b before a starting from u. The obvious boundary conditions '0a.ba.b = a+/3S. If b < u < a.14) S(oo) < 00.ba.16).10. the function S(x) is . 1'. if (1. If (1. Wa. Lemma 1. Using s'/ s = 2p/a2. O.16) yields 4b (u) = 1 . see in particular pp. Notes and references All material of the present section is standard. Then YIa.b = 0 implies that VQ b/s is constant.14) fails.17) Hence L. then. then 0 < 2l.b(u) + L. TWO BARRIERS 303 for x > 0. elementary calculus shows that we can rewrite L as Lq(u) d 1a2 (u)s(u)d [ s (u) ? ] .b (Rdt) = Oa.b('u) = Eu &0. where Lq(u) = 0'22u) q "(u) + p(u)q(u) is the differential operator associated with the diffusion. Letting a T oo and considering the cases S(oo) = oo.b(b) = 1. Letting b J. BROWNIAN MOTION. E„ q(Rdt) = q(u)+Lq(u)dt. so that Y)n. (1 . For generalizations of Proposition 1. 0 in (1. Further references on twobarrier ruin problems include Dickson & Gray [116]. 191195 for material related to Theorem 1.b(a) = 0 then yield the result.b(u)dt. 0 Proof of Theorem 1.b('u) = Eu . we can ignore the possibility of ruin or hitting the upper barrier a before dt. (1. In view of (1. Assume further that S (x) as defined in (1.13) is finite for all x > 0.b(u) = S(a) .6 to Markovmodulated models .0(u) = 1 for all u > 0. S(oo) < oo separately u completes the proof.S(b) Proof Recall that under mild conditions on q.e. A good introduction to diffusions is in Karlin & Taylor [222]. and we get Wo. 15) i.10. A classical reference for further aspects of Bernoulli random walks is Feller [142]. see Asmussen & Perry [42].
yu) where W (ay) = y. Markovmodulated Brownian models .. variance 0. with the drift and the variance depending on an underlying Markov process .(.(7) .aRo . IV.aR.9 ) and optional stopping applied to the stopping time r(u) A T. They all use the fact that ( tx(a) l ( eaRt = eau + aSttx(a) < e7yu. one works instead with a lower limit 5 > 0 of integration in (1.13)).4) I.5.t&(u. but by duality.6) .2.(T(u)AT) r. yu) '+/1(u) . MISCELLANEOUS TOPICS referred to as the natural scale in the general theory of diffusions (in case of integrability problems at 0. 1 y < k (y).4. (2. Lo is a martingale (cf.304 CHAPTER XI. Lo I. much of the literature dels with the pure drift case.2) C_e7u < t(u) < C+e _7u. is currently an extremely active area of research. (2. 111 . The emphasis is often on stationary distributions . (2.)AT .o•K(a) = Ee . Remark 11. See Asmussen [20] and Rogers [305] for some recent treatments and references to the vast literature.1.6. and here are alternative martingale proofs of the rest . yielding eau = Ee.17).5) A martingale proof of (2.(a) (2.1 ) was given already in II. correponding to piecewise linear paths or . defined by the density 1/va(u)s(u) showing up in (1. information on ruin probabilities can be obtained .3. 7y = ay . y > .4. where C_ = B(x) _ B(x) sup 2no fy° e7(Y )B(dy)' f2e7(Y2)B(dy)' C+ i/i(u. equivalently. Another basic quantity is the speed measure M .1) (2.3) < e 7yu.ytc (ay).5): _ z/'(u) < e 7u. which is motivated from the study of modern ATM (asynchronous transfer mode ) technology in telecommunications. 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. (2.
Hence E [e7Rr (u) Jr(u) < ool ^00 H( dt.3). dr) denote the conditional distribution of (T(u).6) below by 1 E Le7Rr(. we have tc(ay) > 0 and we can bound (2.(u. it follows easily from (2.4). A claim leading to ruin at time t has c.6) with = 'y that eyu . and the proof of the lower inequality is similar. (2. Equivalently. Let H(dt.yuk (ay)(u&(u.. (B(y) . u Proof of (2.7R. .( u ) I T(U) < 00] .2): As noted in Proposition II. Proof of ( 2.E [e.f. so that i/1(uL yu) < eayu . y > r.yu))• b(u.yu) Y Similarly for (2.d.2.)r(u)r.6). eyuk (ay) = e7yu e > eyu"(ay ) ij(u. yu))• Letting T + oo yield e_ayu > eyur4ay)(0(u)  Notes and references See II. Rt has distribution B(r + dy)/B(r). 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).3). For (2. when Rt_ = r.T(u)K(ay) I yu < r(u) < T] F(yu < r(u) < T) > e. dr JO Zoo ) f e7'B(r + dy) B(r) Jo ^00 ^00 H(dt.T)  V.1 .4): We take a = ay in (2. dr) e 7( yr)B(dy) B(r) f oo o 0 r > H(dt.1. RT(u)_) given r(u) < oo. we have ic(ay ) < 0 and use the lower bound E [e7Rr („). dr) 1 = 1 I0 /o C+ C+ From this the upper inequality follows.(ay)I T(u) < yu] P(r(u) < yu) (using RT(u) < 0).B(r))/B(r).1.
Accordingly. The limit result (3. The advantage of the large deviations approach is. + X. which in the setting of (3. However . og For sequences fn. The classical result in the area is Cramer's theorem. For example. the parameter will be u rather than n).1 We will go into some more detail concerning (3. we will write fn 1.1) is an example of sharp asymptotics : . Cramer considered a random walk Sn = X1 + .gn if nioo lim 109 fn = 1 log gn (later in this section.. however . logarithmic asymptotics is usually much easier to derive than sharp asymptotics but also less informative . 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. in being capable of treating many models beyond simple random walks which are not easily treated by other models .2). then P C S. nroo n n /// Note in particular that (3. . its generality.means (as at other places in the book) that the ratio is one in the limit (here n * oo).the correct sharp asymptotics might as well have +. if x > EX1.3na with a < 1. gn with fn + 0 . large deviations results been. such that the cumulant generating function r.1). (3.1) where we return to the values of 0. and that a considerable body of theory has been developed.. Thus .?n typically only give the dominant term in an asymptotic expression .1) does not capture the \ in (3.nn or C2e. cle . not quite so much in insurance risk. 1) but only the dominant exponential term . (3. Example 3. MISCELLANEOUS TOPICS 3 Large deviations The area of large deviations is a set of asymptotic results on rare event probabilities and a set of methods to derive such results. and gave sharp asymptotics for probabilities of the form P (S.1) amounts to the weaker statement lim 1 log P I Sn > x I = 17. ri.2) can be rewritten as F (Sn/n > x) 1g a'fin.306 CHAPTER XI. v2 later. e..^ e nn 1 > x n 0o 2xn (3.(B) = log EeOX 1 is defined for sufficiently many 0. ..g./n E I) for intervals I C R. Thus. gn 4 0. logarithmic asymptotics .
sseo f which in conjunction with (3.rc(0) where 0 = 0(x) is the solution of x = rc'(0).(e)i XI E dx]. P with mean nx and variance no.e. 2 where o2 = o2(x) = rc"(0). since Sn is asymptotically normal w.q = rc* (x). More precisely. if we replace Sn by nx + o / V where V is N(0. i.4) immediately yields (3. In fact.(0)) e 307 (other names are the entropy.r. Most often.3) is put equal to x. (3. LARGE DEVIATIONS Define rc* as the convex conjugate of rc. 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. S rtn > x 1.the mean rc'(0) of the distribution of X1 exponentially tilted with 0.r. V > 0 e. Define .3.t.960/) * 0.2).96o /] > 0. the LegendreFenchel transform or just the Legendre transform or the large deviations rate function). we have P(nx < Sn < nx + 1. Since P nn > x) = E {e_8 ' ( 9).9S„+n' ( 9).425.1). we get P(Sn/n > x) E [e9nx +nK(9)9" '. exponential change of measure is a key tool in large deviations methods.4 enn +1. of P(X1 E dx) = E[e9X1K.tin f o') o e9o^y 1 1 ey2/2 dy 21r = etin 1 Bo 27rn .4) n Next. the sup in the definition of rc* can be evaluated by differentiation: rc*(x) = Ox . and hence for large n P(Sn/n > x) > E [e. nx < Sn < nx + 1. rc*(x) = sup(Ox .
In the application of large deviations to ruin probabilities. .'s.s. 1].....e < 8 < y + e.dxn) where Fn is the distribution of (X1i ..'(u) )Ng a"u. For the proof.p > 7 < zn. (iii) #c (8) = limn.h. . to be made rigorous. .. X2. Xn given by Fn(dxl. Mogulskii's theorem which gives path asymptotics./^ >7 < zn n for n n0.e < 8 < y + e. We shall need: Lemma 3 . (iv) tc(ry) = 0 and r.. Assume that there exists 'y. . Ee9X n < oo for e < 0 < e. asymptotics for probabilities of the form P ({S[nti/n}o<t<l E r) for a suitable set r of functions on [0. be a sequence of r. which is of similar spirit as the dicussion in VII. The substitution by V needs. see Jensen u [215] or [APQ] p. 260 for details. however. (ii) lim supn. Sanov's theorem which give rare events asymptotics for empirical distributions. .v. We further write µ = tc'(ry). and the WentzellFreidlin theory of slow Markov walks.. r(u) = inf {n : Sn > u} and o(u) = P('r(u) < oo). that is.. MISCELLANEOUS TOPICS which is the same as (3.3 For each i > 0. there exists z E (0... and write Sn = X1 + • • • + Xn..o log Ee9Sn /n. e > 0 such that (i) Kn (0) = log Ee°Sn is welldefined and finite for 'y .. is differentiable at ry with 0 < K'(y) < 00. which is a version of Cramer's theorem where independence is weakened to the existence of c(O) = limn. Pn Sn1 .. we shall concentrate on a result which give asymptotics under conditions similar to the GartnerEllis theorem: Theorem 3 . Further main results in large deviations theory are the GartnerEllis theorem.. integrates to 1 by the definition of Icn).3.. we introduce a change of measure for X1. Xn) and sn = x1 + • • • + xn (note that the r. n Icn(0) exists and is finite for ry .2 (GLYNN & WHITT [163]) Let X1.308 CHAPTER XI.dxn) = 05nKn(7)Fn(dx1. commonly denoted as is the saddlepoint approximation. Then i/. 1) and no such that Sn .1)..
h.> ..n > u ) = [ Em [em Em 1e. The corresponding claim for Pn(Sn/n < µ .71 < e and jq9j < e.s. for Sn. S. is of order .077 n^oo n and by Taylor expansion and (iv ).Bµ . This establishes the first claim of the lemma .77) follows by symmetry (note that the argument did not use µ > 0).ne(p+ 17). h.2. the r . S. > 1 +17] m(7).YS. can be chosen strictly negative by taking p close enough to 1 and 0 close enough to 0.r (7) n = e. mµ Sm > u] km e7Sm+n. Then V.]1/q = e. Since I EeqOX „ ] 1/q is bounded for large n by (ii).2. it is easy to see that the r.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 +. 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.W.h.. can be chosen strictly negative by taking 9 small enough.3. P n(Sn/n > {c+77) < e no(µ 309 +n)Enees n +n)elcn(B +7).y) . log zl'(u)/u > 'y.. 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.. 0..n e(µ +o)w"(7) [Eep(B +7)Sn]1 /p [EegoX. The rest of the argument is as before. in particular the r.m(7).+r7) < zn for n > no. ( U) P(S. u Proof of Theorem 3. We first show that lim inf„_.s. Clearly.+r. This proves the existence of z < 1 and no such that Pn (Sn/n > µ.s.n m µ 1 + rl . LARGE DEVIATIONS Proof Let 0 < 9 < e where a is as in Theorem 3.ne(µ limsup 1 log Pn (Sn/n > µ + 17) < ic(9 + ry) .91) + o(O ) as 0 J.µ?7 . Let r7 > 0 be given and let m = m(77) = [u(1 + 77)/µ] + 1.
(iv) and Lemma 3..310 ]Em I e.3.n(ry)/u 4 0andm/u* (1 + r7)/µ. 0 yields liminfu __. Sn > u] < eYu+Kn(7)pn(Sn > u) (3. Obviously. P(T(u) = n) < P(Sn > u) = En [e7S.6) for some z < 1 and all n > n(E)...log z) /2 and Sn Fn\ n >lb+S) <Zn. logO(u)/u > ry.+wn(7).. I2 = F(T(u) = n).. n=1 ..7) so that n(b) I1 < e'Yu E en. 3. MISCELLANEOUS TOPICS (7).. I > IL exp `S. and since Ic.(Y). Pn \ > la+ 8 I < zn (3. 14 = = E Lu(16)/aJ+1 Lu(1+6)/µJ+l = n) and n(S) is chosen such that icn('y )/n < 6 A (. For lim supu.I < µl1 1+77 I M 1_ 1+277 S.(•) goes to 1 by Lemma 3.0 log i'(u )/u < 'y. n=1 n=n(b)+1 00 Lu(1 +6) /µJ 13 F( T (u) = P(T(u) n).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). we get lum inf z/i(u) 1 +12r7 >_ ry + 77 Letting r7 J.n Yµ 1 + m + r ('Y) } U n \ 77 m µ µ7 1 < 1+ 77 ) Here E. this is possible by (iii).
Sn1 C U. eryu en logz/2p n nt n. LARGE DEVIATIONS Lu(16)/µJ 311 I2 < e"u n=n(6)+1 e'n(Y)P(Sn > u) < Lu(16)/µJ ^. C 26u `p / +1 I e6u(1+6)/µ (3. > u) Lu(1+6) /µJ +l 00 )^n 'YSn+kn (7) .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).' 1 + b) n e7u x 1 /2 1 n x n / 2x (3. S. we get lim sup log u/00 O (U) < y + b(1 + b) U Letbl0.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. u .3. Sn > U] [ e(u(1+6)/µJ+l < eYu (u(1+6)/µJ+1 7u r 0 0 e L^ en('Y ) fPn (I Sn 1 . µ n=n(6)+1 \ 1u(16)/µ1 00 1 zn < e7u E Z n/2 < e(U xn/2 E n=n(6)+1 n=0 eYu = 1 .10) 00 I4 < E F(Sn_1 < u.
say n n1. this is straightforward since the last inequality in (3.Q is so small that w = 1 . IV.8) by P(S. (7 + a) < 2arc'(7).4. cf.u(1+b)/rc'(7)).2. e'. ryue«iu .(u) = I1+I2+I3+I4'^ ery( u). For 14. 13 = P(T (u) E (u(1 b)l^ (7). 2.4/3rc'(y) > 0.312 CHAPTER XI. we replace the bound P(Sn > u ) < 1 used in (3. Letting c11 = maxn<n.11 ) can be sharpened to x 4 [u(1+6)/µJ /2 1 .4 Under the assumptions of Theorem 3. we need to redefine n(b) as L. Then for n large.3ui where .z 1/z For I1. For 12. I2.('+'Y).. MISCELLANEOUS TOPICS The following corollary shows that given that ruin occurs. the last steps of (3. we get Lou] E exp {( 7 + a)u + Kn(a +7)} n=1 Il Lou] exp {(y + a)u} { 111 + exp {4narc'(7)} n=1 exp {('y + a)u} c1 exp {4/3uarc'(7)} = clewhere a1 = aw. For I. 4 there is an aj > 0 and a cj < oo such that Ij < c3e.9) can then be sharpened to x LQuJ /2 I2 < e7u 1 ..7' a"ju. Corollary 3. u . u(1 + b)/i(7)) Proof Since V.b)/i(7). it suffices to show that for j = 1. we have rcn (a + 7) < 2n^c(7 + a) < 4narc' (7).xl/2 to give the desired conclusion. the typical time is u/rc'(7) just as for the compound Poisson model. > u) < e"' E eIsn = ectueKn (a+'Y)Kn(7) where 0 < a < e and a is so small that r. it holds for each b > 0 that 0(u) 1' g F(T(u) E (u(1 .
An event occuring at time s is rewarded by a r. for the ruin probability z/'h(u) of any discrete skeleton {Skh}k=0.3(s) at time s..e. Obviously many of the most interesting examples have a continuous time scale. i. 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.5 Assume the Xn form a stationary Gaussian sequence with mean p < 0. If {St}t> 0 is the claims surplus process.v. and in fact. and we conclude that Theorem 3 .2 shows that the discrete time structure is used in an essential way.. Theorem 3. whether P ( sup St > u ltg a ^" 0<t<oo // (3. the key condition similar to (iii). t] is Rt = E V (Un) n: o„ <t .LARGE DEVIATIONS 313 Example 3 .1. The reader not satisfied by this gap in the argument can easily construct a discrete time version of the models! The following formula (3..12) k=0.. 2 is in force with y = 2p/wz..13) One would expect this to hold in considerable generality. The problem is whether this is also the correct logarithmic asymptotics for the (larger) ruin probability O(u) of the whole process.2 then immediately yields the estimate log F( sup Skh > u) a7u (3. 11 Inspection of the proof of Theorem 3. 09(9). (iv) becomes existence of a limit tc(9) of tct(9) _ log Ee8S° It and a y > 0 with a(y) = 0.'(y) > 0. V(s) with m.. Thus the total reward in the interval [0. To verify these in concrete examples may well present considerable difficulties.f.. but nevertheless.3.g. we shall give two continuous time examples and tacitly assume that this can be done.1. Let {Nt}t>0 be a possibly inhomogeneous Poisson process with arrival rate .(O) = 9µ+02 for all 9 E R. Assuming that the further regularity conditions can be verified.14) is needed in both examples . r. Xk+l) k=1 00 naoo n provided the sum converges absolutely. Hence z z\ 2 z nrn(9) _ n Cn0p+BZn/ * .. criteria are given in Duffield & O'Connell [124].
the above discussion of discrete skeletons). We let ic (9) = 3(EeWU° . but that a claim is not settled immediately. is At . we conclude that Cu) log e7 u (cf. the best estimator of /3µB based upon Ft.1) ds rt (3. one would take p(t) = (1 + rt)At/ t.15) . (3. = U„ ( t . if the nth claim arrives at time a.v. Example 3. this is not realistic . we have rct (9)/t 4 ic (9). . 0 Example 3 . Since the remaining conditions of Theorem 3.9t. Un represents the total payment for the nth claim).g. If the nth claim arrives at time Qn = s. leading to St = At(1+77) Joo t S8 ds.1) ds .14). e. a differential equation in t). Un(s).. Most obviously.d.Lundberg model has the larger ruin probability. We further assume that the processes {U1(s)}8>0 are i. assuming a continuous premium inflow at unit rate. we have S. Thus. More precisely. n: o. derive .t. At = . <t which is a shotnoise process. the Cramer..s). 0 and since EeOUn(8) + Ee°U^ as s * oo.9t = /3 J t (Ee8U° i8l . (9) < oo for 9 < 'y + C. Kt (0) t (Ee9U"it8i J0 . Of course .314 where the an CHAPTER XI.2 are trivial to verify. e > 0 such that ic('y) = 0 and that r.. i.1) . MISCELLANEOUS TOPICS are the event times.It..6 We assume that claims arrive according to a homogeneous Poisson process with intensity 0 .1) ds . Then logEeOR° = J0 /3(s)(^8(9) .noise model is the same as the one for the Cramer Lundberg model where a claim is immediately settled by the amount Un. It is interesting and intuitively reasonable to note that the adjustment coefficient ry for the shot .14) (to see this . Of course. nondecreasing and with finite limits Un as s T oo ( thus.'`1 U. it contributes to St by the amount Un(t . where Ft = a(A8 : 0 < s < t). Thus. Thus by (3. the CramerLundberg model implicitly assumes that the Poisson intensity /3 and the claim size distribution B (or at least its mean µB) are known.0 and assume there are y. O'n +S] is a r . An apparent solution to this problem is to calculate the premium rate p = p(t) at time t based upon claims statistics .Q„) . then the payments from the company in [on. 7 Given the safety loading 77.
d. we conclude that t.e.17) K(a) f o 1 O (a[I + (1 + 77) log u]) du )3.16) i=1 o i=1 Let ict (a) = log Eeast . and since the remaining conditions are trivial to verify.18) Thus (iii) of Theorem 3.1) or . Thus./3. LARGE DEVIATIONS With the Qi the arrival times.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 .b(u) IN a'Yu (cf. (3.i. To see this . i.3. standard exponential . Indeed.2 hold. It then follows from (3. rewrite first rc as te(a) _ /3E 1 1 +(1+77)aUJ eau 1 .20) (3. uniform (0.i.(1 + 17)0µB = 0.d. (3.(1 +i) f > i= 1 s ds = E Ui 1 . we have Nt t N. the solution of /3(Eelu . 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. 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 . Ui Nt / t 01i 315 St = Ui .(1 + r7) log t (3. typically the adaptive premium rule leads to a ruin probability which is asymptotically smaller than for the CramerLundberg model . one has y > y' (3.1) . the Vi = .19) with equality if and only if U is degenerate.14) that rt _ 13 Jo _ (a [1_( i+77)log]) ds_flt = t (a) (3. equivalently.log Oi are i.
and since tc(s). and k(x) < 0. Dembo & Zeitouni [105] and Shwartz & Weiss [339]. a* (s) are convex with tc'(0) < 0 . For notational simplicity. 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. the function k(x) = e7*x . rc*' (0 ) < 0. 4 The distribution of the aggregate claims We study the distribution of the aggregate claims A = ^N' U. . using that Ek(U) = 0 because of (3. k'(0) < 0. with common distribution B and independent of Nt. MISCELLANEOUS TOPICS Next. the study is motivated from the formulas in IV.i. Further. 0 < x < x0. [257] and Nyrhinen [275]. are i. x > x0. In addition to Glynn & Whitt [163]. so there exists a unique zero xo = xo(r7) > 0 such that k(x) > 0.2 expressing the finite horizon ruin probabilities in terms of the distribution of A. Further.20) is due to Tatyana Turova. at time t.xo.7.. the proof of (3. k(0) = 0. This implies n(y*) < 0.1 . we are interested in estimating P(A > x) for large x. we then take t = 1 so that p. Further applications of large deviations idea in risk theory occur in Djehiche [122]. = P(N = n) = e(3an However. 11 Notes and references Some standard textbooks on large deviations are Bucklew [81]. [245].19). Therefore e7'U _ k(U) E [1+(1+77)y*U] . assuming that the U.(1 + ri)y*x is convex with k(oo) = 00.d. see Nyrhinen [275] and Asmussen [25]. though we do not always spell this out. For Example 3. see also Nyrhinen [275] for Theorem 3.2. this in turn yields y > y*. y = y* can only occur if U .1 E [1+(1+77)y*U] 0 k (+ *y B(+ 1 + (1(+71)y*y B(dy) L xa 1 + f + (1 + rl) Y* xo jJxo k(y) B(dy ) + f' k(y) B(dy) } = 0. much of the analysis carries over to more general cases. say one year. Lehtonen & Nyrhinen [244]. MartinL6f [256]. The main example is Nt being Poisson with rate fit.316 CHAPTER XI. In particular.
3e(bo[a] .1 Assume that lim8T8.4. Proposition 4.2) implies that the limiting Pedistribution of (A .1. 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 . Then Ee"A = e'(") where x(a) _ 0(B[a] .e.x)//3B"[9] is standard normal. only with 0 replaced by a9 and B by B9. e9x+K(°) P(A > x) B 2ir /3 B" [9] Proof Since EBA = x. (4.1). The analysis largely follows Example 3. For a given x.3B"[9]. K'(0) _ ic'(9) = x. no(a) = logE9e'A = rc(a + 9) . we define the saddlepoint 9 = 9(x) by EBA = x."(0) = . B"[s] = oo.1) where )30 = . 818' where s' = sup{s : B[s] < oo}. A > x)] = eex+K( e)E9 [e .ic(9) = .9(Ax). B"' [s] lim (B". This shows that the Pedistribution of A has a similar compound Poisson form as the Fdistribution. Hence P(A > x) = E e [e9A+ ic(9). The exponential family generated by A is given by Pe(A E dx) = E [eeA K(9). i.3B[9] and Be is the distribution given by eox B9(dx) = B [9] B(dx). A E dx] . Vare(A) = s.[s])3/2 = 0. In particular.1). Then as x * oo. THE DISTRIBUTION OF THE AGGREGATE CLAIMS 317 4a The saddlepoint approximation We impose the Poisson assumption (4.
1).ycix °ie6x B. In fact. large x. B covers distributions with finite support or with a density not too far from ax° with a > 1. either of the following is sufficient: A. For example.e.318 CHAPTER XI. 4b The NP approximation In many cases . 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.2) is often referred to as the Esscher approximation. For a rigorous proof. then P(A > x) . i. some regularity of the density b(x) of B is required. b is gammalike.(D X .l3pB.2 If B is subexponential and EzN < oo for some z > 1. [138].2i and that (A . Y satisfies 9(u) ti eu2/2(1 + ibu3) (4. Furthermore 00 b(x)Sdx < oo for some ( E (1. A covers the exponential distribution and phasetype distributions. Thus . In particular. where q(x) is bounded away from 0 and oo and h (x) is convex on an interval of the form [xo.3) The result to be surveyed below improve upon this and related approximations by taking into account second order terms from the Edgeworth expansion.EN B(x). more generally. 1 . and (4.x') where x' = sup {x : b(x) > 0}. MISCELLANEOUS TOPICS It should be noted that the heavytailed asymptotics is much more straightforward.1 goes all the way back to Esscher [141]. it holds that EA = . Var(A) _ ^3p. leading to P(A > x) :.Q{AB (4. Remark 4 . For example. Notes and references Proposition 4. just the same dominated convergence argument as in the proof of Theorem 2. the distribution of A is approximately normal .v. Jensen [215] and references therein. 2). For details. see Embrechts et al.4) .3) and related results u for the case of main interest .1 yields: Proposition 4. bounded with b(x) . it is quite questionable to use (4. The (first order) Edgeworth expansion states that if the characteristic function g(u) = Ee"`}' of a r. b(x) = q(x)eh(z).(3µB)/(0µB^))1/2 has a limiting standard normal distribution as Q ^ oo. or. b is logconcave. under the Poisson assumption (4.
K3 = E(Y .. u5.equantile in the distribution of Y.5) may be negative and is not necessarily an increasing function of y for jyj large. K4 . and so as a first approximation we obtain a1_E = EA + yle Var(A) . the density of Y is 1 °° _ eiuy f(u) du 2x _. (4. are the cumulants .i 6 r 1 3 so that we should take b = ic3/6 in (4. the standard normal distribution. Let Y = (A . s.s. .. (4.2K3 + 4i 64 + ..2 2 . Remark 4. one expects the u3 term to dominate the terms of order u4. THE DISTRIBUTION OF THE AGGREGATE CLAIMS where b is a small parameter. which is often denoted VaR (the Value at Risk). If the distribution of Y is close to N(0. and from this (4.6) .h.. defined as the the solution of P(A < yle) = 1 . in particular.2X2 . however. where Kl . as u2 u3 u4 9(u) = Ee'uY = exp {iuci . zl_e be the 1 .. .5).y2)^P(y)• 319 Note as a further warning that the r. the CLT for Y = Y6 is usually derived via expanding the ch. In concrete examples .3!).c2i. . A particular case is a.5 (y3 . are small.6(1 .f.5) is obtained by noting that by Fourier inversion.EA)/ Var(A) and let yl_E.EY)3.: EA + zl_E Var(A) . Thus if EY = 0. then P(Y < y) 4(y) . the NP (normal power) approximation deals with the quantile al_E.. K2 = Var (Y).5) follows by integration. Heuristically.. of (4.4. f °o 9(y) = 1 e'uye u2/2(1 + iSu3) du 27r _ cc(y) .i 3 K3 } Pt^ exp . ylE should be close to zl_E (cf.e. one needs to show that 163. If this holds .l = EY. so that 1(u) 3 exp { .2 ^ \1 ... Var(Y) = 1 as above .3& (y). resp.1). Rather than with the tail probabilities F(A > x).99.
3ni /3 . Note.1)^ 2) µ'E Notes and references We have followed largely Sundt [354]. that [101] distinguishes between the NP and Edgeworth approximations.zlE )w(zl _E) = which combined with S = EY3/6 leads to q^ 1 Y1 . this holds with a = 0. b such that EN 1 U%.6 (1 .zlE)W(zlE) 1 ..S(1 .7) as 1 (3) a1E = Qµa +z1 .S(1 ..(y) terms dominate the S(1 .E = z1E + S(zi_E . and assume that there exist n ) Pn_i .zl E)V(zl_E) . this yields the NP approximation 6(Z1 _E . This leads to t( yl E) . For example.E )Azl E) 4(z1E) + ( ylE .EA ) / Var(A).1)EY3.y2)cp( y) term.k = /3µB^1 / (. In particular . We can rewrite (4. let pn Pn = (a+ = P(N = n). Another main reference is Daykin et at. 21 . K5 .5(1 .1)! n ^eQ . however.6pBki) d/2.E(/3PB^1 )1^2 + s(z1E .zi.5) by noting that the 4. . as required .. k3 is small for large /3 but dominates 1c4.E + (yl.. b = /3 for the Poisson distribution with rate /3 since Pn = Pn1 n! n (n .320 CHAPTER XI. [101].. Using Y = (A .zl E)^o(zl E) .1) E (A .E)A1 l E) 1 E 4)(yl E) ^' ...EA)3 a1_E = EA + z1_E(Var (A))1/2 + 1 Var(A) Under the Poisson assumption (4.E . 4c Panjer 's recursion Consider A = constants a. MISCELLANEOUS TOPICS A correction term may be computed from (4.1). the kth cumulant of A is /3PBk' and so s. n = 1.yi.
k . Since the sum over i is na + b.13) Namely. fj = E (a+ b k =1 )9kfi_k . . . the value of (4.14) is independent of i = 1..12) where g*n is the nth convolution power of g.4 is that the algorithm is much faster than the naive method..4. j = 0.10) f o = po.} and write gj = 2 . n. .. . the complexity (number of arithmetic operations required) is O(j3) for (4.. . j1 g.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 . (4. which would consist in noting that (in the case go = 0) fj = pn9jn n=1 (4. 2. . fj = P(A = j). 2.. then j (a + b!) 1ag k_1 3 gkfj.4. j = 1. .9). By symmetry.. . j = 1. and calculating the gj*n recursively by 9*1 = 9j.12).13) but only O(j2) for Proposition 4. (4.4. if go = 0. u Proof of Proposition 4.. E[a +bU=I >Ui =j l i=1 J (4.14) is therefore a + b/n. n = k=n1 9k(n1 )9j k • (4.4 Assume that B is concentrated on {0. THE DISTRIBUTION OF THE AGGREGATE CLAIMS 321 Proposition 4..1. .5 The crux of Proposition 4.. (4...11) Remark 4. Then fo = >20 9onpn and fi = 1 E In particular. Hence by (4.. (4. The expression for fo is obvious. 2. 1.
322
00 J
CHAPTER XI. MISCELLANEOUS TOPICS
EE (a + bk I gkg3 _ k lieni n=ik=0 (a+bk l gkE g j'`kpn = E (a+b!)9kfi_k n=0 k=0 k=0 ^I 1 E(a+b. agofj+ k Jgkfjk, k=i /
and and (4.9) follows . (4.11) is a trivial special case.
u
If the distribution B of the Ui is nonlattice , it is natural to use a discrete approximation . To this end, let U(;+, U(h) be U; rounded upwards, resp. downwards , to the nearest multiple of h and let A}h) = EN U. An obvious modification of Proposition 4.4 applies to evaluate the distribution F(h) of A(h) letting f( ) = P(A() = jh) and
g(h) gkh+
= P (U(h2 = kh) = B((k + 1)h)  B(kh ), k = 0, 1, 2, ... , = P (U4;+ = kh) = B(kh)  B (( k  1)h) = gk  l,, k = 1, 2, ... .
Then the error on the tail probabilities (which can be taken arbitrarily small by choosing h small enough ) can be evaluated by
00 00
< P(A > x ) f (h) j=Lx/hl j=Lx/hl
Further examples ( and in fact the only ones , cf. Sundt & Jewell [355]) where (4.9) holds are the binomial distribution and the negative binomial (in particular, geometric ) distribution . The geometric case is of particular importance because of the following result which immediately follows from by combining Proposition 4.4 and the PollaczeckKhinchine representation: Corollary 4.6 Consider a compound Poisson risk process with Poisson rate 0 and claim size distribution B. Then for any h > 0, the ruin probability zb(u) satisfies 00 00
f^,h) Cu) < E ff,+, j=Lu/hJ j=Lu/hJ (4.15)
f! h)
5. PRINCIPLES FOR PREMIUM CALCULATION
where f^ +, f^ h) are given by the recursions
(h) 3 (h) (h)
323
fj,+ = P 9k fjk,+ ' I = 17 2, .. .
k=1 3 (h)
(h)
=
P
(h)
f9,  (h) gk,fAk, e 1  ago, k=1
j = 1+2,
starting from fo + = 1  p, f(h) = (1  p)/(1  pgoh) and using 07
g(kh) 1 (k+1)h
=
Bo((k + 1 ) h)  Bo(kh ) =  f
AB
kh
B(x) dx, k = 0, 1, 2, ... , k = 1,2 .....
gkh+
Bo(kh )  Bo((k  1 ) h) = 9kh)1 ,
Notes and references The literature on recursive algorithms related to Panjer's recursion is extensive, see e.g. Dickson [115] and references therein.
5 Principles for premium calculation
The standard setting for discussing premium calculation in the actuarial literature does not involve stochastic processes, but only a single risk X > 0. By this we mean that X is a r.v. representing the random payment to be made (possibly 0). A premium rule is then a [0, oo)valued function H of the distribution of X, often written H(X), such that H(X) is the premium to be paid, i.e. the amount for which the company is willing to insure the given risk. The standard premium rules discussed in the literature (not necessarily the same which are used in practice!) are the following: The net premium principle H(X) = EX (also called the equivalence principle). As follows from the fluctuation theory of r.v.'s with mean, this principle will lead to ruin if many independent risks are insured. This motivates the next principle, The expected value principle H(X) = (1 + 77)EX where 77 is a specified safety loading. For 77 = 0, we are back to the net premium principle. A criticism of the expected value principle is that it does not take into account the variability of X which leads to The variance principle H(X) = EX+77Var(X). A modification (motivated from EX and Var(X) not having the same dimension) is
324
CHAPTER XI. MISCELLANEOUS TOPICS
Var(X).
The standard deviation principle H(X) = EX +rl
The principle of zero utility. Here v(x) is a given utility function, assumed to be concave and increasing with (w.lo.g) v(O) = 0; v(x) represents the utility of a capital of size x . The zero utility principle then means v(0) = Ev (H(X)  X); (5.1)
a generalization v(u) = Ev (u + H(X)  X ) takes into account the initial reserve u of the company. By Jensen 's inequality, v(H(X)  EX) > Ev(H(X)  X) = 0 so that H(X) > EX. For v(x) = x, we have equality and are back to the net premium principle. There is also an approximate argument leading to the variance principle as follows. Assuming that the Taylor approximation
v(H(X)  X) ^ 0 +v'(0)(H (X)  X) + v 0 (H(X)  X)2 ,/2
is reasonable , taking expectations leads to the quadratic v"H(X )2 + H(X) (2v'  2v"EX) + v"EX2  2v'EX = 0 (with v', v" evaluated at 0) with solution
H(X)=EXv^±V( ^ )2Var(X).
Write
( vI ) 2 \
Var(X) v^  2v^Var(X)/ I  (
, Var(X) )2
If v"/v' is small, we can ignore the last term. Taking +f then yields H(X) ,:: EX 
2v'(0) VarX;
since v"(0) < 0 by concavity, this is approximately the variance principle. The most important special case of the principle of zero utility is The exponential principle which corresponds to v(x) = (1  e6x)/a for some a > 0. Here (5.1) is equivalent to 0 = 1  e0H(X)EeaX, and we get
H(X) = 1 log Ee 0X .
a
5. PRINCIPLES FOR PREMIUM CALCULATION
325
Since m.g.f.'s are logconcave, it follows that H,, (X) = H(X) is increasing as function of a. Further, limQyo Ha (X) = EX (the net premium princiHa (X) = b (the premium ple) and, provided b = ess supX < oo, lim,, H(X) = b is called the maximal loss principle but is clearly not principle very realistic). In view of this, a is called the risk aversion The percentile principle Here one chooses a (small ) number a, say 0.05 or 0.01, and determines H(X) by P(X < H(X)) = 1  a (assuming a continuous distribution for simplicity). Some standard criteria for evaluating the merits of premium rules are 1. 77 > 0, i .e. H(X) > EX. 2. H(X) < b when b (the ess sup above ) is finite 3. H(X + c) = H(X) + c for any constant c
4. H(X + Y) = H(X) + H(Y) when X, Y are independent
5. H(X) = H(H(XIY)). For example , if X = EN U= is a random sum with the U; independent of N, this yields
H
C^
U; I = H(H(U)N)
(where, of course, H(U) is a constant). Note that H(cX) = cH(X) is not on the list! Considering the examples above, the net premium principle and the exponential principle can be seen to the only ones satisfying all five properties. The expected value principle fails to satisy, e.g., 3), whereas (at least) 4) is violated for the variance principle, the standard deviation principle, and the zero utility principle (unless it is the exponential or net premium principle). For more detail, see e.g. Gerber [157] or Sundt [354]. Proposition 5.1 Consider the compound Poisson case and assume that the premium p is calculated using the exponential principle with time horizon h > 0. That is,
N,,
Ev I P  E U;
i =1
= 0 where
v(x) = 1(1  e°x
a
Then ry = a, i.e. the adjustment coefficient 'y coincides with the risk aversion a.
326
Proof The assumption means
CHAPTER XI. MISCELLANEOUS TOPICS
0 a (1  eareo (B[a11)
l
i.e. /3(B[a]  1)  ap = 0 which is the same as saying that a solves the Lundberg u equation. Notes and references The theory exposed is standard and can be found in many texts on insurance mathematics, e.g. Gerber [157], Heilman [191] and Sundt [354]. For an extensive treatment, see Goovaerts et al. [165].
6 Reinsurance
Reinsurance means that the company (the cedent) insures a part of the risk at another insurance company (the reinsurer). Again, we start by formulation the basic concepts within the framework of a single risk X _> 0. A reinsurance arrangement is then defined in terms of a function h(x) with the property h(x) < x. Here h(x) is the amount of the claim x to be paid by the reinsurer and x  h(x) by the the amount to be paid by the cedent. The function x  h(x) is referred to as the retention function. The most common examples are the following two: Proportional reinsurance h(x) = Ox for some 0 E (0, 1). Also called quota share reinsurance. Stoploss reinsurance h(x) = (x  b)+ for some b E (0, oo), referred to as the retention limit. Note that the retention function is x A b. Concerning terminology, note that in the actuarial literature the stoploss transform of F(x) = P(X < x) (or, equivalently, of X), is defined as the function
b * E(X  b)+ =
f
(s  b)F(dx) _ f
6 00
(x) dx.
An arrangement closely related to stoploss reinsurance is excessofloss reinsurance, see below.
Stoploss reinsurance and excessofloss reinsurance have a number of nice optimality properties. The first we prove is in terms of maximal utility: Proposition 6.1 Let X be a given risk, v a given concave nondecreasing utility function and h a given retention function. Let further b be determined by E(X b)+ = Eh(X). Then for any x,
Ev(x  {X  h(X)}) < Ev(x  X A b).
6. REINSURANCE
327
Remark 6 .2 Proposition 6.1 can be interpreted as follows. Assume that the cedent charges a premium P > EX for the risk X and is willing to pay P1 < P for reinsurance. If the reinsurer applies the expected value principle with safety loading q, this implies that the cedent is looking for retention functions with Eh(X) = P2 = P1/(1 + 77). The expected utility after settling the risk is thus
Ev(u + P  P1  {X  h(X)})
where u is the initial reserve . Letting x = u + P  P1, Proposition 6.1 shows that the stoploss rule h (X) = (X  b)+ with b chosen such that E(X  b)+ u = P2 maximizes the expected utility. For the proof of Proposition 6.1, we shall need the following lemma: Lemma 6 .3 (OHLIN'S LEMMA) Let X1, X2 be two risks with the same mean, such that Fj(x) < F2 (x), x < b, Fi(x) ? F2(x), x > b for some b where Fi(x) = P(Xi < x). Then Eg(X1) < g(X2) for any convex function g. Proof Let Yi=XiAb, Zi=Xivb.
Then
P(Yl < x) _ Fi(x) <_ F2 (x) = P(Y2 < x) x < b 1=P(Y2<x) x>b so that Y1 is larger than Y2 in the sense of stochastical ordering . Similarly, P(Zl < x) _ 0 = P(Z2 < x) x < b Fi(x) > F2(x) = P(Z2 < x) x > b
so that Z2 is larger than Zl in stochastical ordering. Since by convexity, v(x) = g(x)  g(b)  g'(b)(x  b) is nonincreasing on [0, b] and nondecreasing on [b, oo), it follows that Ev(Y1) < Ev(Y2), Ev(Zi) < Ev(Z2). Using v(Yi) + v(Zi) = v(Xi), it follows that
0 < Ev(X2)  Ev(Xi) = Eg(X2)  Eg(X1),
using EX1 = EX2 in the last step. u
Proof of Proposition 6.1. It is easily seen that the asssumptions of Ohlin' s lemma hold when X1 = X A b, X2 = X  h(X); in particular, the requirement EX1
328
CHAPTER XI. MISCELLANEOUS TOPICS
= EX2 is then equivalent to E(X  b)+ = Eh(X). Now just note that v is convex. u
We now turn to the case where the risk can be written as N
X = Ui
i=1
with the Ui independent; N may be random but should then be independent of the Ui. Typically, N could be the number of claims in a given period, say a year, and the Ui the corresponding claim sizes. A reinsurance arrangement of the form h(X) as above is called global; if instead h is applied to the individual claims so that the reinsurer pays the amount EN h(Ui), the arrangement is called local (more generally, one could consider EN hi(Ui) but we shall not discuss this). The following discussion will focus on maximizing the adjustment coefficient. For a global rule with retention function h* (x) and a given premium P* charged for X  h* (X), the cedents adjustment coefficient y* is determined by
1 = Eexp {ry*[X  h*(X)  P*]},
for a local rule corresponding to h(u) and premium P for X look instead for the ry solving
J _f
(6.2) N 1 h (Ui), we
[ X_P_^
1 = Eexp
[ Ei  h(Ui)] P [U
= Eexp{ry
h(Ui)]
l (6.3) This definition of the adjustment coefficients is motivated by considering ruin at a sequence of equally spaced time points, say consecutive years, such that N is the generic number of claims in a year and P, P* the total premiums charged in a year, and referring to the results of V.3a. The following result shows that if we compare only arrangements with P = P*, a global rule if preferable to a local one. Proposition 6.4 To any local rule with retention function h(u) and any
N
J}
P > E X  N h(Ui)
4 =1
(6.4)
there is a global rule with retention function h* (x) such that
N
Eh*(X) = Eh(U1)
i=1
and 'y* > ry where ry* is evaluated with P* = P in (6.3).
then (6. This follows by taking Xl = U A b. Assuming for simplicity that the Ui are i. it suffices to show that Eexp {ry ii 'UiAb. however.h(Ui)] . that 01[ry] < 0[y] where 0[y] = Ee'r(U^') .P } < 1 = Eexp E[Ui.3). ry* > 0 because of (6.5 Because of the independence assumptions .6).h(Ui)P JJJ l:='l {ry ] or.5) holds trivially. the excess ofloss rule hl (u) = (u . expectations like those in (6. we get EX = EN • EU. as often local as global.h * (X) . (6.P > EexP{7[X .6. Local reinsurance with h(u) = (u .h( UU) = EN • E[U .P]}.6) u where C[ry] = Ee'r(u4(u)). Applying the inequality Ecp(Y ) > EW(E (YIX )) (with W convex ) to W(y ) = eryy. y = Ei [Ui .P I = EC [7]N.h(Ui)] . u But since ry > 0.d.h(U) (as in the proof of Proposition 6.h(Ui)] .b)+ = Eh(U) (and the same P) satisfies 71 > ry.4).P. REINSURANCE Proof Define N 329 h* (x) = E > h(Ui) X = x . Then for any local retention function u . Remark 6. Eexp 7 [E [Ui . Proof As in the proof of Proposition 6.4.5) reduce quite a lot.h(U)]. i.6 Assume the Ui are i. this implies 7* > 7.4). appealing to (6.4).. ' ii (6. (6.h(u) and any P satisfying (6. . X2 = U .b)+ with b determined by E(U .b)+ is referred to as excessofloss reinsurance and plays a particular role: Proposition 6.i. and so on. we get N 1 = Eexp ry E[Ui ii . N E X . The arrangement used in practice is.4) and u g(x) = e7x in Ohlin's lemma.d.
330 CHAPTER XI. See further Hesselager [194] and Dickson & Waters [120]. . Bowers et at. The present proof is from van Dawen [99]. Heilman [191] and Sundt [354]. The original reference for Ohlin's lemma is Ohlin [277]. see also Sundt [354]. [76].g. e.many texts on insurance mathematics. MISCELLANEOUS TOPICS Notes and references The theory exposed is standard and can be found in.
The number max k : Tk_j < t of renewals in [0. Technically. The point process is called a renewal process if Yo.1) (here U(t) = U([0. . all have the same distribution.. Lebesgue measure for some n > 1). .e. = T„ . some condition is needed: that F is nonlattice.. i.Appendix Al Renewal theory la Renewal processes and the renewal theorem By a simple point process on the line we understand a random collection of time epochs without accumulation points and without multiple points. . If Yo = 0. The renewal theorem asserts that U(dt) is close to dt/µ.. of interarrival times and the time Yo = To of the first arrival (that is.t. The associated renewal measure U is defined by U = u F*" where F*" is the nth convolution power of F. of epochs or the set Y1. t 00 (A. 2h. then Stone 's decomposition holds : U = U... the renewal process is called zerodelayed. note in particular that U({0}) = 1. not concentrated on {h. The mathematical representation is either the ordered set 0 < To < T1 < . + U2 where U1 is a finite measure and U2(dt) = u(t)dt where 331 . denoted by F in the following and referred to as the interarrival distribution. Y2. If F satisfies the stronger condition of being spreadout (F*' is nonsingular w . t]) so that U(t + a) .U(t) is the expected number of renewals in (t. Y.. are independent and Y1..T„_1). U(A) is the expected number of renewals in A C R in a zerodelayed renewal process.. stating that U(t+a)U (t) ^ a. That is.. .. t +a]). when t is large. Lebesgue measure dt normalized by the mean to of F.} for any h > 0. Y2.. Then Blackwell 's renewal theorem holds. the distribution of Yo is called the delay distribution. Y1. t] is denoted by Nt.r.
5) 2This condition can be weakened considerably . but suffices for the present purposes .e. wee shall need the following less standard parallel to the key renewal theorem: Proposition A1. Under weak regularity conditions (see [APQJ Ch.EN(t) .1 if F is nonlattice and z (u) is directly Riemann integrable (d. (A. i. z(x) = 0. then Z(u) i f0 z(x)dx . Then Z(u) 4 z(oo). (A. (A. oo). u u PF 4 00. the asymptotic behavior of Z(u) is given by the key renewal theorem: Proposition A1. resp.2) Z(u) = J0 u z(x)U(dx). then it suffices for (A. stating that U(t)/t > 1/p.9.a.R. ENt 4 1 lb Renewal equations and the key renewal theorem The renewal equation is the convolution equation Z(u) = z(u) + f where Z(u) is an unknown function of u E [0 . U Z(u . A weaker (and much easier to prove) statement than Blackwell's renewal theorem is the elementary renewal theorem. IV).i". IV).2). Equivalently. z(u) a known function. in convolution notation Z = z + F * Z. see [APQ] Ch. and F(dx) a known probability measure . Both result are valid for delayed renewal processes.4) that z is Lebesgue integrable with limZ. Note in particular that F is spreadout if F has a density f.2) has the unique solution Z = U * z.2 Assume that Z solves the renewal equation (A.4) If F is spread.3) Further.out. (A.i. and that F has a bounded density2. In 111. that z(u) has a limit z(oo) (say) as u 4 oo..x)F(dx). the statements being EN(t + a) .332 APPENDIX u(t) has limit 1/µ as t 4 oo. µF (A.
. 0 PF µF 11 In risk theory. . The property of independent cycles is equivalent to the postTk process {XTk+t}t>0 being independent of To.. A stochastic process {Xt}t>0 with a general state space E is called regenerative w. refer to the zerodelayed case.r. This program has been carried out in III. z(x) = e7xz(x). Z(u) U = 1 u 1 u f z(u . However. T1. the postTk process {XT. {Tn} if for any k. T1. where the Tn are the instants where a customer enters an empty system (then cycles = busy cycles). Eo etc. is called the cycle length distribution and as before. .APPENDIX 333 Proof The condition on F implies that U(dx) has a bounded density u(x) with limit 1/µF as x * oo... i. We let FO. . Tk (or. The distribution F of Y1. that the existence of y may fail for heavytailed F.d. Y2.t. results from the case fo F(dx) = 1 can then be used to study Z and thereby Z. that F is a probability measure. Assuming that y can be chosen such that f °° Ox F(dx) = 1. F(dx) = e7xF(dx). equivalently.. . Yk ). Hence by dominated convergence.. we let µ denote its mean.5a. this expression is to be interpreted as a random element of the space of all Evalued sequences with finite lifelengths. of Yo. A regenerative process converges in distribution under very mild conditions: .i.x)u(x) dx = z(u( 1 . To this end.} be a renewal process. 1c Regenerative processes Let {T. a basic reason that renewal theory is relevant is the renewal equation II. Here the relevant F does not have mass one (F is defective). however.2) by e7x to obtain Z = z +P * Z where Z(x) = e'Y'Z(x). Tk and {Xt }o<t<Tk • For example. . The kth cycle is defined as {XTk+t}o<t<Yk .3) satisfied by the ruin probability for the compound Poisson model. the present more general definition is needed to deal with say Harris recurrent Markov chains.. However.(3. cycles. Y1 . and its distribution does not depend on k.t))u(ut) dt 0 0 J f z(oo) • 1 dt = z(OO)... asymptotic properties can easily be obtained from the key renewal equation by an exponential transformation also when F(dx) does not integrate to one.e.k+t }t>o is independent of To. or many queueing processes. multiply (A. this covers discrete Markov chains where we can take the Tn as the instants with Xt = i for some arbitrary but fixed state i. • . The simplest case is when {Xt} has i. Note..
(A. If p = oo.r. P(C ( t) < a) 4 0 for any a < oo) and ij (t) * oo. under the condition of Blackwell's renewal theorem. An example is Zt = fo f (X8) ds where {Xt} is regenerative w. where the distribution of X.3.r... oo).4 Let {Zt}t^.. are i. assume that p < 00 and define Un = ZT}1 . in total variation. Y1) le Residual and past lifetime Consider a renewal process and define e ( t) as the residual lifetime of the renewal interval straddling t. oo). Then {e(t)}. i.t : t < Tk}. {Tn}. but in fact.ZT }0<t<Y„+.d.ZTOI < 00.i...ZT Then: (a) If E sup I ZTo+t . Then it (ii. for n = 1.e.0 be cumulative w. e(t )) . just the same proof as there carries over to show: Proposition A1.e. resp .t.v. This is the case considered in [APQ] V. Otherwise . then e (t) . Then {Zt}t^..+ X. r. µ 0 If F is spreadout. then Xt .t. fi (t) = inf {Tk ..oo (i. 0<t<Yi then Zt /t a$• EU1/µ.r.i. 2.6) id Cumulative processes Let {Tn} be a renewal process with i.334 APPENDIX Proposition A1.d. and q(t) = sup It . {i7(t)} are Markov with state spaces (0. C).Tk : t < Tk} as the age. Then Xt Di X. {Tn} if the processes {ZT +t .. then (Zt .'s by e. cycles (we allow a different distribution of the first cycle).t.. C(t) and ij (t) both have a limiting stationary distribution F0 given by the density F (x)/p. and we have: holds more generally that (rl(t). We denote the limiting r..tEU1/µ)/f has a limiting normal distribution with mean 0 and variance Var(Ui) + (!)2Var (Yi)_ 2EU1 Cov(U1.3 Consider a regenerative process such that the cycle length distribution is nonlattice with p < oo. {Tn}.. (b) If in addition Var(Ul ) < oo. is given by Eg(Xoo) = 1 E0 f Ylg (Xt)dt. [0..0 is called cumulative w.
U(x) < U( 1)). Yo > 0] + f Eo^ (t . = z is Foz) The proof of (a) is straightforward by viewing {(r.d. In IV. and the conditional distribution of ri given l. (1 V)W) where V. 1) and W has distribution Fw given by dFw/dF(x) = x/pF. Y1i Y2. assume first the renewal process is zerodelayed.APPENDIX 335 Theorem A1.t. the first statement follows.y)P(Yo E dy) .. 0 If Markov renewal theory By a Markov renewal process we understand a point process where the interarrival times Yo . we used: Proposition A1. For the second. the sum is o(t) so that Eo£(t)/t + 0 . r.'s with finite mean satisfies Mn/n a$• 0 (BorelCantelli). In the general case.dy )z(y) < c ^ l z(k) Eoe(t 0 0 k=o where c = sup.i. Yl > t]. are not i. (c) the marginal distribution of q is FO.i. Since the maximum Mn of n i. W are independent. the joint distribution of (rl.d.U(x) (c < oo because it is easily seen that U(x + 1) . and the conditional distribution of given 17 = y is the overshoot distribution R0(Y) given by FO(Y) (z) = Fo (y+z)/Fo(y). Then Eo^(t) satisfies a renewal equation with z(t) _ E[Y1 .(t).4. but governed by a Markov chain {Jn} (we . Y1 > t] 4 0. ^ > y) = 1 f +Y (z)dz.5 Under the condition of Blackwell's renewal theorem. Hence for t large enough. and the equivalence of (a) with (b)(d) is an easy exercise. Since z ( k) < E[Yi .t. l:) is the same as the distribution of (VW.y) = f U(t . Proof The number Nt of renewal before t satisfies Nt/t a4' p. V is uniform on (0. EC(t)/t + 0. ^) is given by the following four equivalent statements: (a) P (77 > x.6 Consider a renewal process with µ < oo. we can bound e(t) by M(t) = max {Yk : k < 2t/p}. Then fi(t)/t a4' 0 and. if in addition EYo < oo. use t E^(t)/t = E[Yo . Hence t t lt ) = f U(dy)z(t . (d) the marginal distribution of ^ is FO.. U(x + 1) . .. (b) the joint distribution of (ri.^(t))} as a regenerative process.v.
+ < x. the Markov renewal process if for any n. = io for some arbitrary but fixed reference state io E E. These facts allow many definitions and results to be reduced to ordinary renewal. . G+(x) = P(S..r.t.. r+ < oo). Further: Proposition A1.. A Markov renewal process {Tn} contains an imbedded renewal process. G_(x) = P(ST_ < x. Jn_1. Assume that uj = EjYo < oo for all j and that {J„} is irreducible with stationary distribution (v3)jEE... 0] or (0 . distribution ofjXt}t>o itself where Pi refers to the case Jo = i. Jn = i is the same as the P. Jn +1=j} where J = a(JO. in [APQ]. A stochastic process {Xt}t>o is called semiregenerative w.}.i . . . J1 i . oo).. ... Notes and references Renewal theory and regenerative processes are treated. . Sn = X1 + • • • + Xn the associated random walk. A2 WienerHopf factorization Let F be a distribution which is not concentrated on (oo. Then Xt 4 Xo. the semiregenerative process is called nonlattice if {T. < yIJ) = Fij( y) on {Jn= i. and define r+=inf{n>0: Sn>0}. . be i. where the distribution of X. IT.7 Consider a nonlattice semiregenerative process.T_ < oo). T_=inf{n>0: Sn<0}. .) and (Fij )i.. X2.. Let X1.} is nonlattice (it is easily seen that this definition does not depend on i). Yn. Y1.jEE is a family of distributions on (0..g. with common distribution F. Alsmeyer [5] and Thorisson [372].t.. We call r+ (T_) the strict ascending (weak descending) ladder epoch and G+ (G_) the corresponding ladder height distributions. .and regenerative processes.d.336 APPENDIX assume here that /the state space E is// finite) in the sense that P(Y. oo). is given by Eg(X00) = 1 YO vjEj f g(Xt) dt µ jEE o where p = ujEEViAj. Jo. For example. The semiregenerative process is then regenerative w.r. ... the conditional distribution of {XT„+t}t>o given Yo.. e. . namely {Twk } where {Wk } is the sequence of instants w where Jo.
n 0 R_(A) = E I(Sn E A). (A. oo) (A.=EGn.T_=n} = {S. Sr_ _1 is at its minimum .7). A C (0. More rigorously. . 0] and (0. G_.. F(A) is the contribution from the event {T_ = 1} = {X1 < 0}. define w as the time where the preT_ path S1. 0]).r. Proof Considering the restrictions of measures to (oc. the renewal measures U+=>G+. U.1 (a) F = G+ + G_ .APPENDIX 337 Probabilistic WienerHopf theory deals with the relation between F.7) follows since G+(A) = 0 when A C (oo.7) (A. we may rewrite (a) as G_ (A) = G+(A) = F(A) + (G+ * G_)(A). 0<j<m. S. On {T_ > 2}.8) (e. oo). 0].S. . A C (oo. u . n=0 The basic identities are the following: Theorem A2. >0.x)R_ (dx).. we consider the last such time (to make w unique) so that {w=m.. (e) R_ = U+. F(A .x)R+(dx).S. (c) G+(A) = f °. 0).and r_ preoccupation measures T+1 r_1 R+(A) = E E I(Sn E A). In (A. A C (oo.G+ * G_: (b) G_ (A) = f °° F(A .g. oo).=n w=m i Figure A. F(A) + (G+ * G_)(A). A C (0.1 .>0. G+. (d) R+ = U_. n=0 n=0 00 00 and the T+. m<j<n}.
+ E du) E P(S. m < j <n.0<k<ri . (b) follows from 00 G+ (A) _ E F(Sn E A. ST_ E A) P(T+ = m.XnEAx) 00 f 0 f 0 00 00 1: F(A . ST+Edu).Sn_1Edx. clearly (Sj Sm>0..du) (G+ * G)(A)• C llecting terms. m it follows (see Fig.u) f0m m=1 n=m+1 00 J0 OO P(S. SnEAIS. It follows that for n > 2 F (7.3 8 APPENDIX Reversing the time points 0._ = n .1) that P(Sj Sn. ST_ E A . 0<j<m. 0 < k < n. Sn1 E dx) n=1  F(A . E du) = P(T_=nm. .x)R+(dx)..m.+ E du)P(S.8) is similar. r+ = n) n=1 n=1 0  C0 E fF(Sk< 0._ E A .F(r_n_mSrEA_u). and the proof of (A. A.7) follows. Aso. and reversing the order of summation yields P(T_ > 2. A.1). SmEdu) = P(T+=m.1.= n. m=1 f S mming over n = 2.. S. (A. .>0... S._ E A) n1 f P(r_=nw=m Sm EduSrEA) m=1 n1 F(r+=mSr+Edu)..x)P(Sk < 0. ... Sr_ E Adu) (s ee again Fig .3.
H+ (s) = 1G+[s] is defined and bounded in the halfplane Is : ERs < 0} and nonzero in Is: Rs < 01 (because IIG+lI _< 1). In discrete time. we can rewrite (a) as 1 .g. this holds always on the line its = 0. see e. Summing over n yields R+ (A) = U_ (A).g.2 In terms of m. E. is based upon representing G+ as in (b). and the proof of (e) is similar.G_[s]) (A.G_ [s] is defined and bounded in the halfplane is : ERs > 01 and nonzero in Is : ERs > 0}. Again.0+[s])(1 . then T+ = inf It > 0 : St = 0} is 0 a.Sn_k. Sk = X1 + • • • + Xk = Sn . such developments motivate the approach in Chapter VI on the Markovian environment model. For (d).O<k<n. a number of related identities can be derived. the analogue of a random walk is a process with stationary independent increments (a Levy process. G_ are trivial. see for example Bingham [65]. In this generality of.6. Since G+ is concentrated on (0.0<k<n.g.9) whenever F[s]. Nevertheless.1. there are direct analogues of Theorem A2. 11. In continuous time. the survey [15] by the author and the extensive list of references there.f. and sometimes in a larger strip. 0]. it serves as model and motivation for a number of results and arguments in continuous time.s.1).T+> n) = P(Sk < O.APPENDIX 339 and the proof of (c) is similar.SnEA) = P(Sn<Sk. there is no direct analogue of Theorem A2. The classical analytical form of the WienerHopf problem is to write 1 . cf. 6+ [s].. and G+.0<k<n. . G_ [s] are defined at the same time. u Notes and references In its above discrete time version. if {St} is Brownian motion.P as a product H+H_ of functions with such properties. However. oo). consider a fixed n and let Xk = Xn_k+l. which is basic for the PollaczeckKhinchine formula. Then for A C (oo.1.O<k<n.SnEA) = P(Sn<Sk.1(a) is from Kennedy [228]. For example.4).SnEA) is the probability that n is a weak descending ladder point with Sn E A. P(SnEA .'s.SnEA) = P(SnSn_ k. Another main extension of the theory deals with Markov dependence. and using timereversion as in (d) to obtain the explicit form of R+ (Lebesgue measure). being concentrated at 0. The present proof of Theorem A2.F[s] = (1 . the derivation of the form of G+ for the compound Poisson model (Theorem 11. u Remark A2. WienerHopf theory is only used at a few places in this book. and similarly H_ (s) = 1 ..
Thus.12) eA'AO = Ale AA (A. 0 . _I 0 (A.10) d dteAt = AeAt = eAtA (A. ere A is the eigenvalue of largest absolute value. Some fundamental properties are the following: sp(eA) = {e' : A E sp(A)} (A. JAI = max {Jjt : µ E sp(A)} and sp(A) is the set of all eigenvalues of A (the spectrum).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. 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). one needs to compute matrix inverses Q1 and matrix exponentials eQt ( r just eQ ). if m is s fficiently large. To circumvent this. write eQ = (eK)m where = Q/m for some suitable integer m (this is the scaling step). Eo Kn/n! converges rapidly and can be evaluated without p oblems.11) A f eAtdt = eA. It is seen from Theorem VIII. Here it is standard to compute matrixinverses by GaussJordan el imination with full pivoting . however . and eQ can then be computed as the mth power (by squaring if = 2).13) henever A is a diagonal matrix with all diagonal elements nonzero. 1. Here are. three of the c rrently most widely used ones: xample A3. whereas there is no similar single established a proach in the case of matrix exponentials.1 (SCALING AND SQUARING) The difficulty in directly applying t e series expansion eQ = Eo Q"/n! arises when the elements of Q are large.5 that when handling phase type distributi ons.
Zo = a (Z = QZ. .. the procedure consists in choosing some suitable i > 0.e.]t)n (A.3 (DIFFERENTIAL EQUATIONS) Letting Kt = eQt. and we may consider a new Markov process {Xt} which has jumps governed by P and occuring at epochs of {Nt} only (note that since pii is typically nonzero .. we have k = QK (or KQ) which is a system of p2 linear differential equations which can be solved numerically by standard algorithms (say the RungeKutta method) subject to the boundary condition Ko = I. p different eigenvalues Aj i . The approach is in particular convenient if one wants eQt for many different u values of t.. Let vi.e. what is needed is quite often only Zt = TreQt (or eQth) with it (h) a given row (column) vector.4 (DIAGONALIZATION) Assume that Q has diagonal form.14) holds is therefore that the tstep transition matrix for {fft} is eQt = E ent (.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}. i. construction of {Xt} by realizing the jump times as a thinning of a Poisson process {Nt } with constant intensity 77. However . some jumps are dummy in the sense that no state transition occurs ). 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. assume that Q is the intensity matrix for {Xt} and choose q with rt > max J%J = max qii• 1. In practice. letting P = I + Q/i and truncating the series in the identity = e17t 00 Pn(. Here is a further method which appears quite appealing at a first sight: Example A3 .3 i (A. . Zo = h). One then can reduce to p linear differential equations by noting that k = ZQ. vp be the corresponding left .. To this end.. condition upon the number n of Poisson events in [Olt])  Example A3.15) Then it is easily checked that P is a transition matrix .APPENDIX 341 Example A3. The probabilistic reason that (A. i.7t) n=0 n! u °O n Pn (to see this.2 (UNIFORMIZATION) Formally. Ap.
Then vihj = 0. The phenomenon occurs not least when the dimension p is large. hi have been computed.18) Namely. In view of this phenomenon alone care should be taken when using diagonalization as a general tool for computing matrixexponentials.. we have an explicit formula for eQt once the A j. and vihi ¢ 0. this last step is equivalent to finding a matrix H such that H1QH is a diagonal matrix. Then P P Q = > Aihivi = E Aihi (9 vi. not all ai are real..16) (A. under the conditions of the PerronFrobenius theorem).5 If Q= ( 411 ( q21 q12 q22 is 2 x 2. Complex calculus : Typically. and hence A2 is so because of A2 = tr(Q). two serious drawbacks of this approach: u Numerical instability : If the A5 are too close.342 APPENDIX (row) eigenvectors and hl. say A = (Ai)diag. of largest real part is often real (say. and writing eQt as eQt = He°tH1 = H (e\it)di. hp. D = ) 2 2 . Nevertheless.. the eigenvalue. vi. v5Q = Aivi... say Al. i # j. (A.17) eQt = E e\`thivi = E ea:thi ® vi. (A. and we need to have access to software permitting calculations with complex numbers or to perform the cumbersome translation into real and imaginary parts. i= 1 i=1 P P (A. some cases remain where diagonalization may still be appealing. Qhi = vihi. we can take H as the matrix with columns hl. There are. i=1 i=1 Thus.g H1. however.18) contains terms which almost cancel and the loss of digits may be disasterous.... hp the corresponding right (column) eigenvectors. Everything is nice and explicit here: 411+q2+D' )12_g11+q2^^ where (411422z + 4412421. Example A3. and we may adapt some normalization convention ensuring vihi = 1.
Then 7r = (ir1 7r2 ) = a (q21 Al . where (A. However.APPENDIX 343 Write 7r (= v1) for the left eigenvector corresponding to a1 and k (= hl) for the right eigenvector. b are any constants ensuring//Irk = 1. i.Q2i and after some trivial calculus one gets eQt = 7r 1 112 + eat 7r1 7r2 / (7fl 7r2) = ( 7r2 1r2 7r1 IF. replacing ai by A2.e.20) ir = q2 ql qi +q 2 9l +q2 (A.7 Let 3 9 2 14 7 11 2 2 . The other eigenvalue is A = A2 = q1 .21) Here the first term is the stationary limit and the second term thus describes the rate of convergence to stationarity. l ab (g12g21 + (A1  411) 2) = 1. eqt = eNlt ( ir1ki i2k1 \ ir1 k2 72 k2 + e azt 7r2k2 i2k1 7ri k2 7r1 k1 (A. Then Al = 0 and the corresponding left and right eigenvectors are the stationary probability distribution 7r and e. h2 = Thus. it is easier to note that 7rh2 = 0 and v2k = 1 implies v2 = (k2 . u Example A3. k  C k2 ) =b ( A1 q 1 Q11 / where a .q.6 A particular important case arises when Q = q1 qi ) q2 q2 J is an intensity matrix. 1) . Of course.k1).19) Example A3 . v2 and h2 can be computed in just the same way.
23) . Generalized inverses play an important role in statistics. APPENDIX x1 3/2 . but only that dimensions match .satisfying AAA = A. (A.11/2 . (AA+)' = AA+. (A..11/2 + 5 1. e_6u A4 Some linear algebra 4a Generalized inverses A generalized inverse of a matrix A is defined as any matrix A. (A+A)' = A+A.5 . 2 2 1=ab(142+(1+2)2 ) = tab.6. A2 = 3/2 . for example AA+A = A. They are most often constructed by imposing some additional properties . A+AA+ = 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_.344 Then D= 2+ 11)' 7 T4 2 =52.22) Note that in this generality it is not assumed that A is necessarily square.. and a generalized inverse may not unique.
Assume that a unique stationary distribution w exists .. (A. are ordered such that Al > 0.1 goes under the name fundamental matrix of the Markov chain). = 0 where m < p is the rank of A. and can define /ail 0 0 0 0 0 0 A+ = C A' 0 0 0 C' .e. if A is a possibly singular covariance matrix (nonnegative definite).23) is called the MoorePenrose inverse of A. E.25) .g...1Q = Q(Q . Am > 0. 0 01 In applied probability. one then works with Q = (Q .g.eir )1.P + e7r)1 (here ( I .ew. (I . Here is a typical result on the role of such matrices in applied probability: Proposition A4..P + e7r ).D + O(ebt). ( Q .24) = te7r .eir ).= (I . .eir)1 = I . These matrices are not generalized inverses but act roughly as inverses except that 7r and e play a particular role .1 Let A be an irreducible intensity matrix with stationary row vector it..I) (A. _ A. Then for some b > 0.. and exists and is unique (see for example Rao [300]).P). .e ® 7r)1.APPENDIX 345 A matrix A+ satisfying (A. most often either an intensity matrix Q or a matrix of the form IP where P is a transition matrix. Am+1 = . one is also faced with singular matrices . then there exists an orthogonal matrix C such that A = CDC' where 0 0 D = AP Here we can assume that the A . lt o eAx dx = te7r + D(eAt . Rather than with generalized inverses . and define D = (A .
I. resp.D + D2 + O(ebt).I) .e. . the rows are proportional to it.26) follows by integration by parts: t f t /' xeAx dx = [x {xe7r + D(eAx .91a(2) .2 Let it be a row vector with m components and h a column vector with k components. of (A. Equivalently. o Finally. . Interpreting 7r.J {xe^r + D(e .I)} dx.. respectively. then the Kronecker (tensor) product A(') ®A(2) is the (k1 x k2) x (ml x m2) matrix with (il i2) (jl j2)th entry a. (A.27) Proof Let A(t).2e7r . and in fact any rank 1 matrix can be written on this form.s. ()®(6 f 6/ 7f 8^ 7 8 )=! ^)( 6 7 8 )=(6^ 7^ 8^) \ u Example A4. Note that h ® it has rank 1. (A. see below.I) (A. B'(t) = e7r + DAeAt = eir + (I .26) 2 = 2 e7r + tD . the r. it follows that h ® it is the k x m matrix with ijth element hi7rj .346 t APPENDIX 2 xe Ax dx = eir + t(D + e7r) + D(eAt .h.24).s. 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. in block notation i2h A®B= ( a11B a21 B a12B a22 B Example A4. B(t) denote the l.3 Let 2 A= 4 3 Vf' N7 5 )' B= ( 8 ). the formulas involving O(e6t) follow by PerronFrobenius theory.DZ(ent .h. h ® it reduces to hit in standard matrix notation. h as 1 x m and k x 1 matrices. Then A(O) _ B(O) = 0.eir)eAt = eAt = A'(t). and the columns to h.I)}. For example.
31).3V8. (A B)' = eA®B e! L 1=0 0 .4vf. and v1B1h1 • v2B2h2 = v1B1h1 ® v2B2h2 = ( v1(&v2 )( B1(&B2 )( h1(&h2 ) . and the number of such factors is precisely given by the relevant binomial coefficient.30) eA+B = eAeB function generalizes to Kronecker notation (note that in contrast typically only holds when A and B commute): Proposition A4. (AED B)1 = (A®I+I(9 B)l is the sum of all products of t factors.50 6 7 6 4f 4. each of which is A ® I or I ® B. Using (A.3v'6.28) In particular.5v'8 5vf9 11 A fundamental formula is (A1B1C1) ®(A2B2C2) = (A1 (9 A2)(B1 (9 B2)(C1®C2). then v1B1h1 and v2B2h2 are real numbers.31) Indeed. C2 = h2 are column vectors.(A. (A.29) If A and B are both square (k1 = ml and k2 = m2). (A.A9.APPENDIX 347 Then A®B = 2 f 20.5v/.3f 4v/. 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.4 eA® B = eA ®eB. then the Kronecker sum is defined by A(1) ®A(2) = A(1) ®Ik2 + k ®A(2). A2 = v2 are row vectors and C1 = h1. such a factor is Ak (&B 1k according to (A. if Al = vi. it follows that e® ® e B An _ 0o oo oo Bn 7 I F n! = ` k! (I .29).3vV/72f 20. if A ® I occurs k times.k)! ( n0 n=0 t=0 k=0 J _ ® Ak ®Blk r ^.
we have P8 = Pal) ® p(2). Yt(2 ) }. Thus . P(t) Yt(2) }.s.348 APPENDIX Remark A4. { 1't(1) }. From what has been said about matrices of {Yt( 1). P8 = Pal ) ® P82) exp {Q ( 1) ® Q(2)1 = eXp {Q( 1) } ® exp {Q(2) } Also the following formula is basic: B are both square such that a +.32). h.32) is the intensity matrix of the bivariate continuous Markov process {Yt(1). and Q = Q(1) ® Q (2) = Q(1) ® I + I ® Q(2) (A. in the definition (A. the {Yt(2) } transitions in the {Yt(1) } component and the second transitions in the component . (A. the same time. p = P(1) ® {X }. independent Markov chains. v whenever a is an eigenvalue of A and 0 is an eigenvalue be any row vectors and h. n2 n1 ) {X(2) } are independent Markov chains with transition matrices P(1). P(2).3 < 0 Lemma A4 .6 Suppose that A and of B. P8 = exp {sQ} = exp {s (Q(1) ®Q(2)) } . X ) }. {Yt(1). Let P8f P(Sl). Ps 1) = exp {sQ ( 1) } > p(2 ) = exp {sQ(2) } can therefore be rewritten as Taking s = 1 for simplicity . Q(2). and the form of the bivariate intensity matrix reflects the fact that Yt(2) } cannot change state in both components at due to independence .5 Many of the concepts and results in Kronecker calculus have p(2) is the intuitive illustrations in probabilistic terms. { On the other hand.I)(h ® k). first term on the r .33) . where transition matrix of the bivariate Markov chain {X n1). Let further it. k any column vectors. Yt(2) where independent Markov processes with intensity matri{y(2) } are {Y(1) }. Then 2 0 ire At h • ve Bt kdt = (^®v)(A®B)1(e A®Ba . A special case of Proposition A4. represents ces Q( 1). resp .4 can easily be obtained by probabilistic be the sstep transition reasoning along the same lines .
in such that io = i. (A. ao). Here is the PerronFrobenius theorem. Then: (a) The spectral radius Ao = max{JAI : A E sp(A)} is itself a strictly positive and simple eigenvalue of A. . j = 1. and the corresponding left and right eigenvectors v. E (0. f o r each i..3 whenever a is an eigenvalue of A and 3 is an eigenvalue of B. Now note that the eigenvalues of A ® B are of the form a +. and if we normalize v.. i.. then An = Aohv+O(µ") = Aoh®v+O(µ") for some u. which can be found in a great number of books. h can be chosen with 3By this. and the corresponding left and right eigenvectors v. . il. . we have AO = 1. 4c The PerronFrobenius theorem Let A be a p x pmatrix with nonnegative elements. Similarly. = j and atk_li. p there should exist io. A is called aperiodic if the pattern of zero and nonzero elements is the same as for an aperiodic transition matrix. . see e.. then IN < Ao for all A E sp(A).29). the integrand can be written as ( 7r (9 v)( eAt ® eBt )(h ®k ) = ( 7r ®v)(eA (DBt)(h (& k). [APQ] X. h such that vh = 1.APPENDIX 349 Proof According to (A. so that by asssumption A ® B is u invertible. . .. > 0 for k = 1. h = e and v = 7r (the stationary row vector).g. we mean that the pattern of nonzero offdiagonal elements is the same as for an irreducible intensity matrix.1 and references there (to which we add Berman & Plemmons [63]): Theorem A4.8 Let B be an irreducible3 p x pmatrix with nonnegative offdiagonal elements. and appeal to (A.12). . (b) if in addition A is aperiodic. We call A irreducible if the pattern of zero and nonzero elements is the same as for an irreducible transition matrix.7 Let A be a p x pmatrix with nonnegative elements.. . n.The PerronFrobenius theorem has an analogue for matrices B with properties similar to intensity matrices: Corollary A4. .. h can be chosen with strictly positive elements. Then the eigenvalue Ao with largest real part is simple and real.34) Note that for a transition matrix. That is.
350 APPENDIX strictly positive elements. Ao). The next result gives a condition for asymptotical exponentiality. the phasetype distribution B(a) with representation (.2). we have A0 = 0.(ti)diag where Q = T + (ti)diag is a proper intensity matrix (Qe = 0). but is an easy consequence of the PerronFrobenius theorem. it was shown that under mild conditions the tail of a phasetype distribution B is asymptotical exponential.1 Let Q be a proper irreducible intensity matrix with stationary distribution a. h such that vh = 1. let {Yti°i } be a Markov process with initial distribution a and intensity . T(°)) is asymptotically exponential with parameter t* _ r EiEE aiti as a 4 oo. I. h = e and v = 7r (the stationary row vector). not only in the tail but in the whole distribution. note that we can write the phase generator T as Q .8 is most often not stated explicitly in textbooks.e. let t = (ti)iEE # 0 have nonnegative entries and define T(°) = aQ .. relate the eigenvalues of B to those of B via (A. the condition is that t is small compared to Q. 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). then eBt = ea0thv + O(eµt) = eA0th ® v + O(et t) (A. the analogy of this procedure with unformization. For example.(ti)ding.35) for some p E (oo. The content is that B is approximately exponential if the exit rates ti are small compared to the feedback intensities tij (i # j).(3. if we normalize v.8. Then for any (3. Corollary A4. Furthermore. Note that for an intensity matrix. Proposition A5.n t AL n=0 n! (cf. Example A3.1. To this end. 10) and use the formula me at e Bt = e 00 Antn = e . Bi° (x) + at*x Proof Let { 4 } be the phase process associated with B(a) and (°) its lifelength. A5 Complements on phasetype distributions 5a Asymptotic exponentiality In Proposition VIII.
J(()) _ = i) + at•x t tt' . and this easily yields a(x)/x a' 1/t*. In addition to the asymptotic exponentiality. Proof Assume first ti > 0 for all i and let I. J^O)_ = j) Pi (v(aaV) > x. from which the phase process is terminated . Let further V be exponential with intensity V and independent of everything else. a .Yj(av) = j f . and that Yt(a) = Yat for all t. a'/a + 1.YQ(av) = j) Pi ( ci(a'V) > x. We shall . 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}. We can think of ( ( a) as the first event in an inhomogeneous Poisson process ( Cox process ) with intensity process matrix aQ . fo tY dv/t a$' t*.(a) > x . has a limit distribution: Proposition A5. Hence O ((a) aa. {t Y( a) } v>0 .1.)_ = Y(a) = 1'aS(a) = Ya(av)^ it follows that Pi ((. a' = a . = YQ(x). and write Yt = Yt(1). v/ t. t < (a). we get dx F (Idx = j) = (1 + qij t )Sij + qij dt.a' + oo (e. dx/ti] or not.9.g. Since JJ(.bij) Hence the intensity matrix of { Ix} is (qij/ti)i.x (1 . Then a(a'V)/a (aV) a' 1. in fact .jEE. Then {Ix} is a Markov process with to = Yo. 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. By the law of large numbers for Markov processes . Conditioning upon whether { Yt} changes state in [0. We can assume that Jta) = Yt(°).2 Pi (c(a) > x. prove a somewhat more general result which was used in the proof of Proposition VI. it states that the state.APPENDIX 351 ((1) etc.aE where 0 < e < 1).
.4 Any discrete distribution B with finite support.. Keilson [223]. Penev & Turbin [238]. zkbk is za(I .p)k1 p..} is said to be discrete phasetype with representation (E. a = b = (bk)k=1. k>1. Gnedenko & Kovalenko [164] and Glasserman & Kou [162]).zP)'p. (c) the nth moment k 1 k"bkis 1)"n!aP"p. is discrete phasetype. so is the geometric distribution. the simplest discrete phasetype distribution: here E has only one element. and thus the parameter p of the geometric distribution u can be identified with the exit probability vector p. a) if B is the lifelength of a terminating Markov chain (in discrete time) on E which has transition matrix P = (p. (b) the generating function b[z] _ E' ..+ at*x • a't' L ` at t* t* J Reducing the state space of {Ix } to {i E E : t. 2. say bk = 0.352 rr Ia(a'V) Ei I ( > x) P APPENDIX L at (Yo (aV) .. See also Korolyuk. Then: (a) The point probabilities are bk = aPklp.x k > K. . . an easy modification of the argument yields finally the result for the case where t. K}. = 0 for one or more i..1 and A5.. so we shall be brief. A distribution B on {1. 2. Example A5. Then P is substochastic and the vector of exit probabilities is p = e ..2 do not appear to be in the literature.. with point probabilities bk = (1 . these results are in the spirit of rare events theory for regenerative processes (e.g.Pe. Example A5. 5b Discrete phasetype distributions The theory of discrete phasetype distributions is a close parallel of the continuous case.j) and initial distribution a. k = 1. Indeed. Et II I a(a^V) > x) at' . a). 1 k=1 1 0 otherwise. P.. . let E and Pkj j=k1..3 As the exponential distribution is the simplest continuous phasetype distribution. > 0}. u Notes and references Propositions A5. ' pk 0 k>1 11 Theorem A5. However.5 Let B be discrete phasetype with representation (P.
Jt t > U1 + U2.T(2)). 11 Example A5. T= ( 0 T(2) ) (A. and hence the negative binomial distribution is discrete phaseu type.{ 0. r . initial distribution a and phase generator T.T(1)). and a=1).APPENDIX 353 5c Closure properties Example A5. . resp. Then {Jt} has lifetime U1 + U2 . a' .a(1). (E(2). a. resp.. The discrete counterpart is the negative binomial distribution with point probabilities bk k1) (1 k = r. A. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2). A reduced phase diagram (omitting transitions within the two blocks) is am E(1) t(1) a(2) (2) t(2) Figure A.36) in blockpartitioned notation (where we could also write a as (a (1) 0)). B2 be phasetype with representations (E(1)..r + 1. U2.a(2). _ i E E(1) T(1) t(1)a(2) i E E(2) .2 The form of these results is easily recognized if one considers two independent phase processes { Jt 1) }. { Jt 2) } with lifetimes U1 . Then the convolution B = B1 * B2 is phasetype with representation (E. as is seen by minor modifications of Example A5.6.7 (THE NEGATIVE BINOMIAL DISTRIBUTION) The most trivial special case of Example A5.1 This corresponds to a convolution of r geometric distributions with the same parameter p.6 (CONVOLUTIONS) Let B1.6 is the Erlang distribution Er which is the convolution of r exponential distributions.. and piece the processes together by it = 41) 0<t<U1 U1 < t < U1 + U2 2U..
Example A5.T. if U1.. this means that a = (Oa(1) (1 . T) and C = EO°_1(1 .10 (GEOMETRIC COMPOUNDS) Let B be phasetype with representation (E. a mixture of more than two phasetype distributions is seen to be phasetype. i E E(2) 0 T(2) =IT (in blockpartitioned notation. Let B(") be the corresponding phasetype distribution.p)pn1B*n. are i. U2.37) (1) (1 . Example A5. A reduced phase diagram is 0a(1) E(1) A .d. we need to restart the phase process for B w. with common distribution and N is independent of the Uk and geometrically distributed with parameter p.p)pn1. Equivalently. B2 be phasetype with representations (E(1).T(1))..O)B2 (0 < 0 < 1) is phasetype with representation (E.E) where a(°) = fAa(a)v(da). and o'i Oa. (E(2). one obvious interpretation of the claim u size distribution B to be a mixture is several types of claims.'). and consider B(") = fA B(a) v(da) where v is a probability measure on A.a(2). i E E(1) T 0 I (A. To obtain a phase process for C. Then the mixture B = 9B1 + (1 .9 (INFINITE MIXTURES WITH T FIXED) Assume that a = a(°) depends on a parameter a E A whereas E and T are the same for all a. P(N = n) = (1 . In risk theory.T(2)).0)a(2) E(2) Figure A. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2).354 APPENDIX Example A5.. resp.8 (FINITE MIXTURES) Let B1.a(1).4 . a. p at each termination.0)ai2).p.0)a(2))). a reduced phase diagram is f a E t Figure A. then C is the distribution of Ul + • • • + UN. a. Then it is trivial to see that B(") is u phasetype with representation (a(").3 In exactly the same way.i. Thus.
U2. then C is the distribution of U1 + • • • + UN.. Thus the representation is (E(1) x E(2). It is zeromodified phasetype with representation (E.T + pta). i.11 (OVERSHOOTS) The overshoot of U over x is defined as the distribution of (U . X independent of U.g..9) that (U .APPENDIX 355 and C is phasetype with representation (E. Example A5 . v. { 4 } as exit of {Jt}. P).T) where F[T] = J0 "o eTx F(dx) u is the matrix m. v. U2. Then the minimum U1 A U2 and the maximum U1 V U2 are again phasetype. cf.TWWW). Equivalently. a. Corollary VIII. Example A5 . with common distribution B and N is independent of the Uk with P(N = n) = f. T + ta. j E F}. T) and C = F. are i. then U1 +• is phasetype with representation (E.T) if U is phasetype with representation (E. T)..1. T(1) ® T(2)). say with distribution F. f2. a(2). +UN 2. resp. Proposition VIII. To obtain a phase representation for C . if B is defective and N + 1 is the first n with U„ = oo. Example A5. if {Jt} is a phase process for U. . E). but the same T. let {Jtl)}. resp.. To see this. Minor modifications of the argument show that 1.v.°_1 f„ B*?l. let B be a continuous phasetype distribution with representation (F. a.. then U1 + • • + UN is zeromodified phasetype with representation (a. 12 (PHASETYPE COMPOUNDS ) Let fl. of F. it follows by mixing (Example A5. Indeed. if U1. { Jt2) } be independent with lifetimes U1.. a. a(1) ® a(2 ).°.2.x)+. 13 (MINIMA AND MAXIMA ) Let U1.aeTx.aF[T]. we then let the governing phase process be {Jt} _ {(411 Jt2))} 2) interpreting exit of either of {4 M }. then Jy has distribution aeTx.7. T(2) ). T + pta). B2 of phasetype with representations (E(')..f.a(1).. say v.2. cf.d. Note that this was exactly the structure of the lifetime of a terminating renewal u process.X)+ is zeromodified phasetype with representation (E. If U1 has a different initial vector. U2 be random variables with distributions B1. let the phase space be E x F = {i j : i E E. If we replace x by a r. For U1 A U2. be the point probabilities of a discrete phasetype distribution with representation (E. . (E(2). let the initial vector be a ® v and u let the phase generator be I ® T + P ® (ta).
elementary) Let {bk} be any dense sequence of continuity points for B(x).B(bk) I < 1/n for n > k. and the closedness of the class of phasetype distributions under the formation of finite mixtures. we need to allow { Jt.} of phasetype distributions such that Bn 3 B as n + oo. Then from above.(bk) + B(bk) for all k. 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 .(bk) + B(bk) for all k as n * oo.2) } to go on (on E(2)) when { i 1) } exits. Let the support of Dn be {xl(n). The mean of B„ is n/Sn = b and the variance is n/Sn = b2/n. Example A5. i= 1 C. That is.. Hence it is immediate that Bn 4 B. q(n) q(n) pi(n)a . we can assume that ID.. Here are the details at two somewhat different levels of abstraction: (diagonal argument .n = I:pi(n)Er v ( __ ) n) ) a= 1 . The general case now follows easily from this. By the diagonal argument (subsequent thinnings). and let Bn be the Erlang distribution E. Then we must find phasetype distributions Bn with B.. the fact that any distribution B can be approximated arbitrarily close by a distribution with finite support...(n) = D. Proof Assume first that B is a onepoint distribution. oo). the initial vector is (a(1) (& a (2) 0 0). any distribution B on (0.. Thus the state space is E(1 ) x E(2) U E(1) U E( 2). with weight pi(n) for xi(n). see Neuts [269] (where the proof. there is a sequence {B. Now we can find first a sequence {Dm} of distributions with finite support such that D.(Sn) with Sn = n/b. cf. r # oo.xq(n)(n)}.(bk)'.14 To a given distribution B on (0. 5d Phasetype approximation A fundamental property of phasetype distributions is denseness .8. and vice versa. oo) can be approximated 'arbitrarily close' by a phasetype distribution B: Theorem A5. relies more on matrix algebra than the probabilistic interpretation exploited here). however.356 APPENDIX For U1 V U2.. say degenerate at b..
oo) such that f (x) = O(e«x). For a general Bo.( dx) * f r f{(x)B(dx). and we can take Bn = Cr(n). It should be noted. oo).n( b k ) . however.. for some a < oo.r. and that cp is known to be continuous. k < n. Since PET is closed under the continuous operation of formation of finite mixtures.. the topology for weak convergence) PET of the class PET of phasetype distributions contains all onepoint distributions. k < n. PIT contains all finite mixtures of onepoint distributions. 2. 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. u 2 (abstract topological ) The essence of the argument above is that the closure (w. ..e.t. one would use the B given by some statistical fitting procedure (see below). oo) * [0. Let E be the class of functions f : [0.15 To a given distribution B on (0 . then it is immediate that WI(B) = p2(B) for all distributions B on [0. replications). f2.. In particular. But To is the class G of all distributions on [0. Hence G C PET and L = PIT..APPENDIX 357 Hence we can choose r(n) in such a way that ICr( n).. there is a sequence {Bn} of phase type distributions such that Bn Di B as n 4 oo and f ' f. we can then approximate Bo by a phasetype B. the class CO of all discrete distributions. if information on Bo is given in terms of observations (i.(x)Bf.. that this procedure should be used with care if ^p(B) is the ruin probability O(u) and u is large.14 is fundamental and can motivate phasetype assumptions. oo) and any fl. compute W(B) and use this quantity as an approximation to cp(B0). i. If Cpl (B) and ^02(B) are weakly continuous. Then ICr( n ).B(bk )I < . x 4 oo.n. oo) approximation Assume that we can compute a functional W(B) when B is phasetype. say on the claim size distribution B in risk theory.D(bk)I < n.d.. u Theorem A5. i = 1.i. Corollary A5. E E.n (bk) .
n(dx) + f 0 fi(x)Dn(dx). for each i.38 ). there is a sequence {Bn} of phase type distributions such that Bn Di B as n + oo and all moments converge. i = 1.38) We first show that for each f E E.  APPENDIX B implies that 00 o o 00 n.2 .n(dx) < 1+. liminf B. and the case of a general f then follows from the definition of the class E and a uniform integrability argument.f (z) = f = 1 1 1 1n/ o .... By (A.. .. then cc f (x)Bn ( dx) = (?!c ) e'= . i = 1... \\ 0 Corollary A5. n. Now returning to the proof of (A.. and hence we may choose r(n) such that L 9l) f (x)Cr(n).14 Dn has been chosen such that 00 1 °° f fi(x)D n(dx ) < 1++ '  o \ n o f fi(x)B(dx).. n B=az. we may assume that in the proof of Theorem A5. if f (x ) = e°x.f ' f (x)B(dx). . ...39) Indeed. TO (A. 2.(dx) > J fi(x)B(dx). . . and hence it is sufficient to show that we can obtain limsup n4oo fi(x)Bn(dx) < Jo 0 f fi( x)B(dx ). n.39).f (x)B(dx). oo). f° xtBn(dx ) * f °° x`B( dx).. Bn=En z f f (x)Bn(dx) fof (x)B(dx) = ° (A.oo J fi(x)B.358 Proof By Fatou' s lemma. f00 fi(x)Cr.. i = 1. i=1.16 To a given distribution B on (0 .
0 as i * oo. oo) with B[y +e] < oo for some e > y = 7(B.16. 5e Phasetype fitting As has been mentioned a number of times already.14 is classical. the loggamma or the Weibull have been argued to provide adequate descriptions of claim size distributions. the adjustment coefficient 'y = 7(B.. . one can obtain 7(Bn. the problem thus arises of how to fit a phasetype distribution B to a given set of data (1./3). and therefore the following result is highly relevant as support for phasetype assumptions in risk theory: Corollary A5. there is a sequence {B. (N or a given distribution Bo. . lim sup ryn < 7.. However. . then Bn['Y + ei] * B[y + ei] > 1 + 7 Q implies that 'yn < ry + ei for all sufficiently large n .} of phasetype distributions such that Bfz + B as n * oo and Yn 4 ry where ryn = y(Bn.e. .> y for some sequence {ei} with ei E (0.l3µb < 1. /3) = ry for all n. The present section is a survey of some of the available approaches and software for inplementing this. there is substantial advantage in assuming the claim sizes to be phasetype when one wants to compute ruin probabilities.18 In the setting of Corollary A5. We shall formulate the problem in the slightly broader setting of fitting a phasetype distribution B to a given set of data (1i . If ei > 0.3). e ) and ei J. from a more conceptual . Notes and references Theorem A5. lim inf > is proved similarly. The adjustment coefficient is a fundamental quantity. For practical purposes.. .17 To a given /3 > 0 and a given distribution B on (0. This is motivated in part from the fact that a number of nonphasetype distributions like the lognormal./3) is defined as the unique solution > 0 of B[y] = l+y/j3. .APPENDIX 359 In compound Poisson risk processes with arrival intensity /3 and claim size distribution B satisfying . . but are certainly not unexpected. the remaining results may be slightly stronger than those given in the literature. (N. Proof Let fi(x) = el'r+E. I. and in part from the fact that many of the algorithms that we describe below have been formulated within the setup of fitting distributions. O We state without proof the following result: Corollary A5.
The earliest such reference is Bux & Herzog [85] who assumed that the Erlang distributions have the same rate parameter.f.g . and this is what matters when using phasetype distributions as computational vehicle in say renewal theory.. the L1 distance between the c . Asmussen & Nerman [38] implemented maximum likelihood in the full class of phasetype distributions via the EM algorithm . Of course. The constraints were the exact fit of the two first moments and the objective function to be minimized involved the deviation of the empirical and fitted c. at a a number of selected points . cf.g. g. . where more than two Erlangs are allowed and in addition to the exact matching of the first three moments a more general deviation measure is minimized (e.g. and we next describe two such approaches which also have the feature of being based upon the traditional statistical tool of like maximum likelihood. It seems therefore a key issue to develop methods allowing for a more general phase diagram. and used a nonlinear programming approach . The characteristics of all of these methods is that even the number of parameters may be low (e.. . B„ The problem is that the constructions of {B„} are not economical : the number of phases grows rapidly.} of phasetype distribution such that Bo. for some suitable large n. e . [70]) restrict attention to acyclic phase type distributions .'s). d. [202]. Johnson & Taaffe considered a mixture of two Erlangs (with different rates ) and matched (when possible ) the first three moments . giving mass 1 /N to each S=. (N is the empirical distribution Be. A number of approaches restrict the phase type distribution to a suitable class of mixtures of Erlang distributions . and as fitted distribution we may take B. a program package written in C for the SUN workstation or the PC is available as shareware. risk theory. reliability or queueing theory..f. one could argue that the results of the preceding section concerning phasetype approximation contains a solution to our problem : given Bo (or Be). [317] ) has considered an extension of this setup. [216] ). In a series of papers (e. Schmickler (the MEDA package..360 APPENDIX point of view the two sets of problems are hardly different : an equivalent representation of a set of data (1 . the number of phases required for a good fit will typically be much larger. and in practice this sets a limitation to the usefulness (the curse of dimensionality . The likelihood function is maximized by a local linearization method allowing to use linear programming techniques. we have constructed a sequence { B. A method developed by Bobbio and coworkers (see e. three for a mixture of two Erlangs ).d. The observation is that the statistical problem would be straightforward if the whole ( EAvalued) phase process { Jtk)} o<t<( k associated with each observa . we do not not want to perform matrix calculus in hundreds or thousands dimensions). defined by the absence of loops in the phase diagram .g.
g. (N) tJk Ea ( n).. . The general idea of the EM algorithm ([106]) is to replace such unobserved quantities by the conditional expectation given the observations. N Ti = I(J= i) dt.. . (n+1) _ Ea (n). then the estimators would be of simple occurenceexposure type. e.. In fact.(k] (Ti is the total time spent in state i and Nii is the total number of jumps from i to j).(N) = E Ea(n). eieT(n)((k.. Nii = = ..T(n) (Ti ^^ 1.. .g. jEEA. it seems open whether the restriction to the acyclic case is a severe loss of generality.APPENDIX 361 tion Sk was available.. . Thus. it is easy to see that N (k Ea(n).. In practice. EN where ai = N 1 I ((k) = i) tii=i iEE.T (n)(TiI(1. = j) f k=1 k =1 tE[0. since this is parameterdependent.x)t(n) 1 and this and similar expressions are then computed by numerical solution of a set of differential equations.T(n) k=1 I (Jti) dt o \f a(n)eT(n )(kt(n) N f:i a(n)eT(n)xei .T(n) (Nik IC1. one is lead to an iterative scheme. (N ) (^ 54 k )+ and similarly for the cn+1) The crux is the computation of the conditional expectations. . the methods of [70] and [38] appear to produce almost identical results.. E.
This page is intentionally left blank .
In: Limit Theorems and Related Problems (A. 101116. 345368. Teubner.A. Arndt (1984) On the distribution of the supremum of a random walk on a Markov chain . Aktuar Tidskr. B. 3157 . [11) S. Skand. Arfwedson ( 1955) Research in collective risk theory.). [8] G. [14] S. Arfwedson ( 1954) Research in collective risk theory. Stegun ( 1972 ) Handbook of Mathematical Functions (10th ed. Asmussen (1985) Conjugate processes and the simulation of ruin problems. [3] J. Abramowitz & I. Proc. 213229. [6] V. 363 . 587. Act. Probab. 57. Dover. Whitt ( 1992) The Fourierseries method for inverting transforms of probability distributions . Anantharam (1988 ) How large delays build up in a GI/GI11 queue. AT&T. Stuttgart. pp. Stoch. 143170. 1985. John Wiley & Sons. Asmussen (1987) Applied Probability and Queues. 253267. Transactions XVth International Congress of Actuaries. ed. Queueing Systems 10. Preprint. 219229. [5] G. ibid. Choudhury & W. [2] J. II. New York. Abate & W. 311338. Borokov. Appl. 53100. J. Asmussen (1989a) Aspects of matrix WienerHopf factorisation in applied probability. [13] S. 14. Scand. Alsmeyer (1991) Erneuerungstheorie. Chichester New York. Appl.G. Aktuar Tidskr. Abate. New York. Asmussen ( 1982 ) Conditioned limit theorems relating a random walk to its associate. Whitt (1994) Waitingtime tail probabilities in queues with longtail service time distributions . 37. Queueing Systems 5 . G. [12] S. 191223.L. Optimizations Software. Whitt (1998) Explicit M/G/1 waitingtime distributions for a class of longtail service time distributions . Adv. 1984 . [4] M. [7] E. Queueing Systems 16. [10] K. Abate & W. with applications to risk reserve processes and the GI /G/1 queue.Bibliography [1] J. [15] S.). The Mathematical Scientist 14. 20. Sparre Andersen (1957) On the collective theory of risk in the case of contagion between the claims . New York. The case of equal risk sums. Asmussen (1984) Approximations for the probability of ruin within finite time . 38. [9] G. Skand.
Asmussen (1992c) Stationary distributions for fluid flow models and Markovmodulated reflected Brownian motion. Ann. 37. J. Stochastic Models 11. Appl. 772789. stationary distributions and first passage probabilities. Scand. [22] S. eds. Binswanger (1997) Simulation of ruin probabilities for subexponential claims. 8. Asmussen (1992a) Phasetype representations in random walk and queueing problems. 137168. J. Act. Asmussen & B. 27.). CRC Press. MatrixAnalytic Methods in Stochastic Models (A. [19] S. 189201. [27] S. SIAM Review 40. 69100. Florida. Act. Asmussen. 303322. Ann. Frey. Asmussen & K. Methods V Problems (J. Advances in Queueing: Models. Act. 2943. Scand. Scand. L. 106119. Kalashnikov & A. Asmussen & M. 315. Binswanger & B. Random Processes 2. Bernoulli 6. 2149. New York. Proc. 297318. 1999 . Asmussen & B. Statist. [29] S. [25] S. Hojgaard (1999) Approximations for finite horizon ruin prob abilities in the renewal model. Asmussen (1998c) A probabilistic look at the WienerHopf equation.M. K. 354374. A. Th. Schmidt (1995) Does Markovmodulation increase the risk? ASTIN Bull. [24] S. 79102. 313326. 96107. Appl. Asmussen (1995) Stationary distributions via first passage times. [31] S. Probab. Astin Bulletin 27. J. [33] S. 54. eds. [26] S. Asmussen (1998b) Extreme value theory for queues via cycle maxima. Dshalalow ed. [32] S. Alfa & S. 555574. Asmussen (2000) Matrixanalytic models and their analysis. T. Asmussen (1989b) Risk theory in a Markovian environment. Boca Raton. 4966. Stoch. 1989 .). 193226. Appl. Rolski & V. Hojgaard (1996) Finite horizon ruin probabilities for Markovmodulated risk processes with heavy tails. [23] S. 313341. Probab. [18] S. Proc. Scand. Asmussen (1999) On the ruin problem for some adapted premium rules.S. [34] S. Andronov. Marcel Dekker. J. Appl. Asmussen. Ann. Floe Henriksen & C. Asmussen (1991) Ladder heights and the Markovmodulated M/G/1 queue. Kliippelberg (1994) Large claims approximations for risk processes in a Markovian environment. [20] S.K. 20. . Stoch. Extremes 1 . Asmussen (1992b) Light traffic equivalence in single server queues. 2. Chakravarty. Asmussen (1998a) Subexponential asymptotics for stochastic processes: extremal behaviour. [30] S. 25. Hojgaard (2000) Rare events simulation for heavytailed distributions. [28] S. Probab. [17] S. Riga Aviation University. [21] S. Asmussen. Probabilistic Analysis of Rare Events (V. Bladt (1996) Renewal theory and queueing algorithms for matrixexponential distributions. 1936. 1996. Asmussen & M. Bladt (1996) Phasetype distributions and risk processes with premiums dependent on the current reserve.364 BIBLIOGRAPHY [16] S.).
Proc. Boca Raton. Hojgaard & M. [43] S. Asmussen & S. [44] S. Math. Opns. Appl. Stoch.BIBLIOGRAPHY 365 [35] S.Y. 20. Management Science 45. 421458. Prob 32 . Asmussen & C. Sc. Asmussen & R. [52] S. Probab. Appl. H. [49] S. O'Cinneide (2000/2001) Matrixexponential distributions [Distributions with a rational Laplace transform] Encyclopedia of Statistical Sciences. Oper. 736755. Asmussen & C. Appl. 365372. 410433. Marked point processes as limits of Markovian arrival streams. [37] S. Nerman & M. Asmussen & H.). Advances in Queueing: Models. with applications to insurance risk Stoch. 913916. Probab. Fitting phasetype distributions via the EM algorithm. O'Cinneide (2000/01) On the tail of the waiting time in a Markovmodulated M/G/1 queue.M. [39] S. Johnson. Probab. 58. Asmussen & C. 10. Insurance: Mathematics and Economics 10. [38] S. Florida. Example of excessoffloss reinsurance for an insurance corporation. 115. 30. Nielsen (1995) Ruin probabilities via local adjustment coefficients . Res. Kliippelberg (1996) Large deviations results for subexponential tails. Res. Appl. Rolski (1994) Risk theory in a periodic environment: Lundberg's inequality and the CramerLundberg approximation. Rubinstein (1995) Steadystate rare events simulation in queueing models and its complexity properties. [40] S.A. Statist. J. [41] S. Schmidli & V. O. Statistica Neerlandica 47. CRC Press. Eng. 429466. Inf. 23. 299324. Asmussen & G. Asmussen & R. Appl. 422447. 31.). 419441. first passage problems and extreme value theory for queues. Asmussen. Wiley. (to appear). 120. Asmussen & V. [50] S. Finance and Stochastics 4. 11251141. Asmussen. [48] S. Proc. Olsson (1996). Scand. Asmussen & V. Asmussen . Schock Petersen (1989) Ruin probabilities expressed in terms of storage processes. Probab. Rubinstein (1999) Sensitivity analysis of insurance risk models. [45] S. B. Th. Asmussen & T. [47] S. Sigman (1996) Monotone stochastic recursions and their duals. Methods 8 Problems (J. [42] S. 259274. 105119. [36] S. Supplementary Volume (Kotz. 19. Taksar (1997) Controlled diffusion models for optimal dividend payout. Schmidt (1999) Tail approximations for nonstandard risk and queueing processes with subexponential tails. [51] S. Perry (1992) On cycle maxima. 64. Asmussen & K. Asmussen & D. . Insurance: Mathematics and Economics 20. Stochastic Models 8. Koole (1993). Appl. Rolski (1991) Computational methods in risk theory: a matrixalgorithmic approach. J.A. Asmussen & M. 103125. Schmidt (1995) Ladder height distributions with marks. J. Schmidt (1993) The ascending ladder height distribution for a class of dependent random walks.Y. Adv. [46] S. Read eds. Taksar (2000) Optimal risk control and dividend distribution policies . Dshalalow ed. Adv. Asmussen & T.
L. 33.H. 221232.M. Berman & R. Beekman (1985) A series for infinite time ruin probabilities. Insurance: Mathematics and Economics 9. Bingham (1975) Fluctuation theory in continuous time. Probab. Boogaert & V. Probab. 610. Appl. Scand. Bjork & J. Teugels (1987) Regular Variation.A. Oxford. Plemmons (1994) Nonnegative Matrices in the Mathematical Sciences. Actuaries 21. von Bahr (1974) Ruin probabilities expressed in terms of ladder height distributions. . Trans. Wiley.R. von Bahr (1975) Asymptotic ruin probabilities when exponential moments do not exist. J. Beekman (1969) A ruin function approximation. Bjork & J. Soc. [64] P. J. Boogaert & A. Athreya & P. Scand. BarndorffNielsen (1978) Information and Exponential Families in Statistical Theory. [56] B. J. Scand.J. Cox (1972) A low traffic approximation for queues.H. [71] P. 7. Schmidli (1995) Saddlepoint approximations for the probability of ruin in finite time.B . 4148. [70] A. Clarendon Press. Act. [60] J. 1975. Wiley. Bingham.366 BIBLIOGRAPHY [53] S. New York. Cambridge University Press. 11811190. [69] P. [55] B. 832840. 705766. Bobbio & M. J. J. [65] N. Asmussen & J. Goldie & J. J. Appl. Ney ( 1972) Branching Processes . de Waegenaere (1990) Simulation of ruin probabilities. 275279. Act. Stochastic Models 10. 1974. [73] A. 9599. C. Appl. [58] O. Act. [59] O. Scand. Grandell (1988) Exponential inequalities for ruin probabilities in the Cox case. 1985 .T. SpringerVerlag. BarndorffNielsen & H. 9. [54] K. 169186. 1995. Adv. Billingsley (1968) Convergence of Probability Measures. [61] J. Springer . Scand. Cambridge. [63] A. Act.A. Grandell (1985) An insensitivity property of the ruin probability. 661667. Teugels ( 1996) Convergence rates for M /G/1 queues and ruin problems with heavy tails . SIAM. [62] J. Insurance: Mathematics and Economics 4. 129134. Baker (1977) The Numerical Solution of Integral Equations. Insurance: Mathematics and Economics 6.H. 190204. [68] T. New York. 148156. 1988 . 77111. [66] N. Bloomfield & D. [72] P. [67] T. Halsted Press. Berlin. Telek (1994) A bencmark for PH estimation algorithms: results for acyclic PH. Act. Crijns (1987) Upper bounds on ruin probabilities in case of negative loadings and positive interest rates. J. New York. [57] C. Borovkov (1976) Asymptotic Methods in Queueing Theory. Probab.L. Beekman (1974) Two Stochastic Processes.
Gerber. Cottrell. 14331453.W. J. New York San Francisco London. Amsterdam. Area Commun. 749763. Tech. Boxma & J.Y . Proc.A. Cramer (1930) On the Mathematical Theory of Risk.R.C. [80] F. Probab. 34. Queueing Systems 33. Cohen (1999) Heavytraffic analysis for the GI/G/1 queue with heavytailed service time distributions . Tweedie (1982) Storage processes with general release rule and additive inputs . AddisonWesley. [93] K. [81] J. Geb. Skandia Jubilee Volume. [92] H. Academic Press. [75] O. D. Simulation and Estimation. Chung (1974) A Course in Probability Theory (2nd ed. [87] E.C. Burman & D .I. [76] N. Hickman. 3539. Breiman ( 1968) Probability. Syst. Bucklew ( 1990) Large Deviation Techniques in Decision. 105119. verve. New York. Reiser eds. [91] H. 127130. Z. Herzog (1977) The phase concept : approximations of measured data and performance analysis. Fort & G. IEEE Trans. Cox (1955) Use of complex probabilities in the theory of stochastic processes. Insurance : Mathematics and Economics 5. Aut. [83] D. Fox & L. Wahrscheinlichkeitsth. Itasca.L. Adv. Cohen (1998) The M/G/1 queue with heavytailed service time distribution. Smith ( 1983) Asymptotic analysis of a queueing system with bursty traffic. Boxma & J.R. Brockwell .J. 24. Bowers . Wiley.). Malgouyres (1983) Large deviations and rare events in the study of stochastic algorithms . Oper.R. Springer. Goovaerts & F. Chandy & M. Cambr. Qinlar (1972) Markov additive processes. Illinois. SpringerVerlag. Smith ( 1986) An asymptotic analysis of a queueing system with Markovmodulated arrivals . IEEE J.L. NorthHolland. [86] K. Schrage (1987) A Guide to Simulation. Stockholm.J. [89] M. [84] D. Control AC28. [82] H.A. [78] L. Stockholm.W. Burman & D. H. [79] P. Reading. Sel. 93121. Computer Performance (K.L.M. Jones & C. 907920. 313319.W.). [88] J.L. [90] D. 177204. The Society of Actuaries . 14. B. The Jubilee volume of Forsakringsbolaget Skandia. Croux & N. Broeck . Bratley.Y. Cramer (1955) Collective risk theory. Cohen (1982) The Single Server Queue (2nd ed. [85] W.BIBLIOGRAPHY 367 [74] O. 16. II. Bell. Resnick & R.) NorthHolland. Amsterdam. Soc. Nonparametric estimators for the probability of ruin. De Vylder (1986) Ordering of risks and ruin probabilities. [77] P. Res. Nesbitt (1986) Actuarial Mathematics . M. Bux & U. Jr. Appl. 62.. 51. 392433. Insurance : Mathematics and Economics 9. S. J.U. J. Philos. Biihlmann ( 1970) Mathematical Methods in Risk Theory. Berlin. New York. 2338. . Veraverbeke (1990).
Stochastic Models 5. 9 . [96] D. Embrechts (1989). Scand. Insurance: Mathematics and Economics 7. 114119. Comp. De Vylder (1977) A new proof of a known result in risk theory. [112] F. New York. Martingales and insurance risk. Res. Rolski (1991) Light traffic approximations in queues. Scand. van Dawen (1986) Ein einfacher Beweis fur Ohlins Lemma. Davidson (1946) On the ruin problem of collective risk theory under the assumption of a variable safety loading (in Swedish). Dassios & P. [104] F. Haezendonck (1987) Classical risk theory in an economic environment. Dempster. Dembo & O. [103] F. 135159. De Vylder (1996) Advanced Risk Theory. Pesonen (1994) Practical Risk Theory for Actuaries. [100] A. Oper. [101] C. Steinebach ( 1991 ) On the estimation of the adjustment coefficient in risk theory via intermediate order statistics . 22. Appl. Actuarial J. J. Aktuar. Insurance: Mathematics and Economics 6. Delbaen & J. Daley & T. 624628. J.B. Res. Insurance: Mathematics and Economics 4. 435436. T. 27. Daley & T.. Insurance: Mathematics and Economics 10. Suppl. [110] F. Tidskr. 201206. Teugels ( 1990) Empirical Laplace transform and approximation of compound distributions .. 88101. Steinebach (1990) On some alternative estimators of the adjustment coefficient in risk theory. [107] L. Probab. Editions de 1'Universite de Bruxelles. [102] P. Rolski (1984) A light traffic approximation for a single server queue. Math. Math. 121131.P. Oper. Csorgo & J. Math. De Vylder (1978) A practical solution to the problem of ultimate ruin probability. [95] M. Blotter der deutschen Gesellschaft fur Versicherungsmathematik XVII. J. 227229. Boston. Goovaerts (1984) Bounds for classical ruin probabilities. . 583602. [109] F. Soc. 7083 (1969). Goovaerts (1988) Recursive calculation of finitetime ruin probabilities. [97] D. Insurance: Mathematics and Economics 3.368 BIBLIOGRAPHY [94] M. N. Delbaen & J. [111] F. [106] A. De Vylder & M. [105] A. 1990 . Rubin (1977) Maximum likelihood from incomplete data via the EM algorithm. [98] A. Chapman & Hall. English version published in Skand. De Vylder & M. Laird & D. Statist. Zeitouni (1992) Large Deviations Techniques and Applications. Devroye (1986) Nonuniform Random Variate Generation. 3. Deheuvels & J. J.L. Daykin. Act. Forsiikringsmatematiska Studier Tillagnade Filip Lundberg. Stockholm. Csorgo & J. 16. Pentikainen & E. Appl. 17. 85116. 181217. 5771. [99] R. [108] F.M.D. Roy. Springer. Haezendonck (1985) Inversed martingales in risk theory. Jones and Bartlett. 3750.
5159. 6180. Dickson & C. Act. 1993 . Duffield & N.U.C. Oper.W. Dickson & J. Dickson & C. [115] D. Shiu (1991) Risk theory with the Gamma process. [116] D.U.C. 2000 . 1.M. Act. Gray (1984) Exact solutions for ruin probability in the presence of an upper absorbing barrier. Ruin estimates for large claims . 269274. Scand. J.R. Taylor ( 1975) A diffusion approximation for the ruin probability with compounding assets. 251262. Dickson (1995) A review of Panjer's recursion formula and its applications.M. Scand. Emanuel .R.M.R. Gerber & E. Gerber (1988) The surpluses immediately before and at ruin.R. [129] D . 191207. . [126] F. Gray (1984) Approximations to the ruin probability in the presence of an upper absorbing barrier. Insurance: Mathematics and Economics 7. J. Dufresne & H. Proc. Djehiche (1993) A large deviation estimate for ruin probabilities. 105115. [114] D. Harrison & A. Act. Res. 3745. and the amount of claim causing ruin. Act. Springer . [118] D.M. 147167. 4259. Astin Bulletin 21. Dickson & H.M. Dickson (1994) An upper bound for the probability of ultimate ruin. Dynkin (1965) Markov Processes I.J. Scand. 107124. Waters (1996 ) Reinsurance and ruin . [127] F. Dickson & H. [122] B. 177192. 118. Insurance : Mathematics and Economics 10.M.C. O'Connell (1995) Large deviations and overflow probabilities for the general singleserver queue. [119] D. Insurance : Mathematics and Economics 7.B. Math. Insurance: Mathematics and Economics 25.C. Scand. Gerber (1991) Risk theory for the compound Poisson process that is perturbed by a diffusion. [128] E. 174186. Insurance: Mathematics and Economics 19. Philos. Act. Camb. Dufresne & H. 229232. Embrechts . 1984 . [121] D. 4962. Soc. J. Insurance : Mathematics and Economics 11. [130] P.BIBLIOGRAPHY 369 [113] D. Embrechts ( 1988 ).K.C.M. with applications.C. Dickson & J. Schmidli (1993) Finitetime Lundberg inequalities in the Cox case . H. 1975 . J. Hipp (1999) Ruin problems for phasetype(2) risk processes. 1984 . J. Waters (1999) Ruin probabilities wih compounding.C. Act.M.G. Insurance : Mathematics and Economics 22. Grandell & H. 193199. Act. Letters 7. [131] P.M. Scand. J. J. 131138.U. British Actuarial J.M. [120] D.C. 1994 .J. Dufresne. 363374. J. [117] D.S. [123] E. Scand. Dickson (1992) On the distribution of the surplus prior to ruin . 1741.C. Regterschot (1988) Conditional PASTA.C. J. Hipp (1998) Ruin probabilities for Erlang (2) risk processes. Berlin Gottingen Heidelberg. [125] F. [124] N. 1993 . van Doorn & G. Scand.
5974. Schmidli (1994) Ruin estimation for a general insurance risk model. Probab. 159178. Skand. [143] W. Schmidt (1992) An insensitivity property of ladder height distributions. Furrer (1998) Risk processes perturbed by astable Levy motion. Embrechts & J. Embrechts. 1998 . Mikosch (1991). first passage times and moments of ladder variables in Markov random walks. Insurance : Mathematics and Economics 1.D. Adv. Insurance: Mathematics and Economics 10. Griibel & S. R. 269274. Approximations for compound Poisson and Polya processes.W. Adv. 26. [137] P. Probab. 17.K. Frenz & V. Tidsskr. 14661474. Trans. Embrechts & H. J. with applications.). Appl.D. 695712. Probab. Franken. 131137. [136] P. Veraverbeke (1982) Estimates for the probability of ruin with special emphasis on the possibility of large claims .370 BIBLIOGRAPHY [132] P. F.M. [134] P. [142] W. Insurance: Mathematics and Economics 7. [146] E. Appl. 35. Embrechts & N. U. J. Kliippelberg & T.D. New York. Nyt Tidsskrift for Matematik B20. Fuh & T. Fields 75. [148] C. Probab. C. Embrechts & T. 38. Th. 8190. 5572. Wiley.L. Konig.J. Maejima & J. [133] P. Math. [139] A. Embrechts. [147] M. Esary. 181190. 404422. Reprinted 1948 as 'The theory of probabilities and telephone conversations' in The Life and work of A. 29. 5975. Embrechts. Statist. Rel. 175195. Wiley. Springer. Adv.A. Pitts (1993) Some applications of the fast Fourier transform in insurance mathematics . 566580. D. Ann. Erlang 2. Proschan & D. Appl. Fuh (1997) Corrected ruin probabilities for ruin probabilities in Markov random walks. Appl. [145] P. Scand. On the excursions of Markov processes in classical duality. Fitzsimmons (1987). [140] J. Schmidt (1982) Queues and Point Processes. Frees (1986) Nonparametric estimation of the probability of ruin. Probab.W. Heidelberg. Lai (1998) Wald's equations. [141] F. Walkup (1967) Association of random variables. A bootstrap procedure for estimating the adjustment coefficient. . Mikosch (1997) Modelling Extremal Events for Finance and Insurance. [135] P. 29. M. Sci.L. Danish Academy Tech. Arndt & V. Probab.). Feller (1966) An Introduction to Probability Theory and its Applications I (3nd ed. [150] H. [144] P. Feller (1971) An Introduction to Probability Theory and its Applications II (2nd ed. J. Act. Villasenor (1988) Ruin estimates for large claims. Akt. [149] C. 623637. [138] P. Statistica Neerlandica 47. Teugels (1985). 616624. Appl. Wiley. 3339.K. Esscher (1932) On the probability function in the collective theory of risk. New York. Jensen. Erlang (1909) Sandsynlighedsregning og telefonsamtaler . Astin Bulletin 16. J.
Mitt. Adv. Gerber (1970) An extension of the renewal equation and its application in the collective theory of risk. Life Insurance Mathematics. Act. 28. NorthHolland. Gerber (1979) An Introduction to Mathematical Risk Theory.J. J.71. Kaas. Gnedenko & I.U. Gerber (1986). [156] H. 5. 1981 . [155] H. Schweiz. ETH Zurich. Z. [167] C. Ann. 97114. Dordrecht. Appl. A. Grandell (1977) A class of approximations of ruin probabilities. Scand.BIBLIOGRAPHY 371 [151] H. Michna & A.U. 73. Kluwer. [162] P. Probab.W. 463480. [153] H. Goldie & R. Insurance: Mathematics and Economics 20. 7778. Springer.J. London. R.U. J. Schweiz. Unpublished. Gerber (1971) Der Einfluss von Zins auf die Ruinwahrscheinlichkeit. Gerber (1973) Martingales in risk theory. Boston. 424445. [166] M. [158] H. Gerber (1981) On the probability of ruin in the presence of a linear dividend barrier. de Vylder & J. 131156. Skand. [159] H. Basel. Griibel (1996) Perpetuities with thin tails. Appl.U. 6370.E. Ver. 3752. J. Haezendonck (1984) Insurance Premiums. Probab. Scand. Act. . [161] P. [164] B.N. [168] J. van Heerwarden & T. [160] H. Insurance : Mathematics and Economics 7.. Aktuarietidskrift 1970 . [152] H. Furrer & H.. Gani.U. Act. 1977. Schmidli (1994) Exponential inequalities for ruin probabilities of risk processes perturbed by diffusion. Glasserman (1991) Gradient Estimation via Perturbation Analysis. Bauwelinckx (1990) Effective Actuarial Methods.U. Grandell (1978) A remark on 'A class of approximations of ruin probabilities'. Kou (1995) Limits of first passage times to rare sets in regenerative processes.S. Furrer (1996) A note on the convergence of the infinitetime ruin probabilities when the weak approximation is astable Levy motion. Insurance: Mathematics and Economics 15. Math. Glasserman & S. Glynn & W. 31A. Weron (1997) Stable Levy motion approximation in collective risk theory. [169] J. Mitt. 205216. 2336. Math. Suppl. F. 105115.M. J. Gerber (1988) Mathematical fun with ruin theory. Goovaerts.U. Birkhauser . [157] H. Kovalenko (1989) Introduction to Queueing Theory (2nd ed. Vers. Vers. [165] M. 205210. [163] P. Galambos & J. Whitt (1994) Logarithmic asymptotics for steadystate tail probabilities in a singleserver queue. Amsterdam. eds.).). Ver. Huebner Foundation Monographs. In Studies in Applied Probability (J. [154] H. contained in the authors PhD thesis. NorthHolland. Appl. Probab. Amsterdam. 1523.V. Furrer. Scand. S.G. Goovaerts. 1978 . University of Pennsylvania.
I. Appl. Resnick (1977) The recurrence classification of risk and storage processes . 3 . 143158. Harrison & S. [172] J. [180] R. Math. Astin Bulletin 29. Preprint. Insurance: Mathematics and Economics 26. Winter School on Stochastic Analysis and Appl . Harrison & S. 14. . M. Grandell (1999) Simple approximations of ruin functions. 3041.131135. M. Harrison & A.I. Grandell (1979) Empirical bounds for ruin probabilities. Gusak & V. 5766. Chapman & Hall. Resnick (1976 ) The stationary distribution and first exit probabilities of a storage process with general release rule . Tinbergen Institute Research Series 20. Grandell (1990) Aspects of Risk Theory. KTH. [178] B. Silvestrov (2000) Cram€rLundberg approximations for nonlinearly perturbed risk processes . [186] J. Math. 566576. Applied Statistics 40. Theory and Applications. Grandell (1997) Mixed Poisson Processes. Statistical Algorithm 265. Grigelionis ( 1992) On Lundberg inequalities in a Markovian environment. 332. 347358.S. AkademieVerlag. 14. ed. Opns .. Gut (1988 ) Stopped Random Walks . 197214. E. 6779. Proc. SpringerVerlag. 1. Probab. 7590. Grigelionis ( 1993) Twosided Lundberg inequalities in a Markovian environment. Skand. Grnbel (1991) G/G/1 via FFT. Appl. [175] J. Matem. Res. Lemoine (1977) Limit theorems for periodic queues. Hadwiger (1940) Uber die Wahrscheinlichkeit des Ruins bei einer grossen Zahl von Geschiiften . [187] J. Ibragimov. Archiv v. [188] J. Gyllenberg & D. Res. [182] A.V. 5 . mathematische Wirtschaft .). Harrison (1977) Ruin problems with compounding assets . 8. 355365.J. Grandell & C. Amsterdam. Korolyuk (1969) On the joint distribution of a process with stationary increments and its maximum . Berlin. Segerdahl (1971) A comparison of some approximations of ruin probabilities. Grigelionis (1996) Lundbergtype stochastic processes. Probability Theory and Mathematical Statistics (I. [184] H . M. Stoch. Griibel & R. [181] D. J. [176] B. Grandell ( 1992) Finite time ruin probabilities and martingales . [179] R. [183] M. Liet. Hermesmeier (1999) Computation of compound distributions I: aliasing errors and exponential tilting.and Sozialforschung 6. van Heerwarden (1991) Ordering of Risks: Theory and Actuarial Applications. 400409. [185] J. Appl. Aktuar Tidskr. New York.372 BIBLIOGRAPHY [170] J. [174] J. M. 33. [189] A. Proc.A. Stoch.243255. Proc . [173] J. Appl. Rink. [177] B. 167176. Opns .O. [171] J. Informatica 2. SpringerVerlag. Probab. Th. 54.
Scand. Michel (1990) Risikotheorie: Stochastische Modelle and Statistische Methoden.BIBLIOGRAPHY 373 [190] P.M. Mandl & M. Asmussen & 0. Ann. Staalhagen (1999) Waiting time distributions in M/D/1 queueing systems. Geb. Act. Hipp (1989b) Estimators and bootstrap confidence intervals for ruin probabilities. Heidelberger (1995) Fast simulation of rare events in queueing and reliability models. Hesselager (1990) Some results on optimal reinsurance in terms of the adjustment coefficient.V. Math. Haggstrom. Herkenrath (1986) On the estimation of the adjustment coefficient in risk theory by means of stochastic approximation procedures. verve. Statist. The ruin problem for finite Markov chains. Nerman (1992) EMPHT . 377381. 19. [199] R. Wiley. 11631174. Hill (1975) A simple general approach to inference about the tail of a distribution. J. 1990 . 4385. [194] 0. Probab. Chalmers University of Technology and the University of Goteborg. 305313. 123151. Wahrscheinlichkeitsth. 5770. J. No. ACM TOMACS 6. 29. 8996. [192] U. Hogg & S. 25. 3. Foruth Prague Symp. J. Proc. on Asymptotic Statistics (P. [196] C. Hogan (1986) Comment on corrected diffusion approximations in certain random walk problems. Hoglund (1991).A. Heilmann (1987) Grundbegriffe der Risikotheorie. Verlag Versicherungswirtschaft e. Ann. eds. New York. 285293. [195] B. 6. 18. 259268. J. Hojgaard & M. Hermann (1965) Bin Approximationssatz fiir Verteilungen stationarer zufalliger Punktfolgen. Probab. 8095.L. [193] U. Taksar (2000) Optimal proportional reinsurance policies for diffusion models. [204] T. [200] M. [197] C. [203] T. [201] L. Probab. . Astin Bulletin 19. Horvath & E. 285292. Z.V. Ann.a program for fitting phasetype distributions. [205] T. Insurance: Mathematics and Economics 5. Hipp (1989a) Efficient estimators for ruin probabilities. 1998 . Huskova. Karlsruhe. Karlsruhe. [202] 0.. Hoglund (1990) An asymptotic expression for the probability of ruin within finite time. [198] C. 166180. 12981310. Act. Hipp & R. Iversen & L. Prob.). [208] V. 378389. Versicherungswirtschaft. Klugman (1984) Loss Distributions.L. Insurance: Mathematics and Economics 5. Willekens (1986) Estimates for the probability of ruin starting with a large initial reserve. Electronic Letters 35. Appl. Scand. Nachrichten 30. [191] W. S. Iglehart (1969) Diffusion approximations in collective risk theory. Appl. [207] D. Mathematical Statistics. Studies in Statistical Quality Control and Reliability 1992 :4.R. Hoglund (1974) Central limit theorems and statistical inference for Markov chains. [206] B. 23.B.
Kalashnikov (1997) Geometric Sums: Bounds for Rare Event with Applications. Jagerman (1985) Certain Volterra integral equations arising in queueing. Copenhagen. Princeton. Stochastic Models 5. Soc. ibid. 63.J. ibid. Kalashnikov (1996) Twosided bounds of ruin probabilities. Kemperman (1961) The Passage Problem for a Markov Chain. Act. J. 61. 6. [215] J. [220] V. N. [225] J. Karlin & H. 711743. Keilson & D. . [221] E. 37. Keilson & D. Addenda to for processes defined on a finite Markov chain. Appl. [223] J. 31. J. 60. [216] M. [227] J. Proc. NEC Research Institute. 11. Goovaerts (1986) General bound on ruin probabilities. Jagerman (1991) Analytical and numerical solution of Volterra integral equations with applications to queues. J. Wishart (1964) A central limit theorem for processes defined on a finite Markov chain. [219] V. 123133. University of Chicago Press. 1996. Cambridge Philos. Wishart (1964). [224] J.M. Jensen (1995) Saddle Point Approximations.G. 547567. Soc.M.M.P. Jelenkovic & A. Insurance: Mathematics and Economics 5. Keilson & D.G. [222] S. Keilson (1966) A limit theorem for passage times in ergodic regenerative processes. Kao (1988) Computing the phasetype renewal and related functions. Janssen (1980) Some transient results on the M/SM/1 special semiMarkov model in risk and queueing theories. Kaas & M. Math. Kluwer. [211] J. [217] R. Statist. Cambridge Philos. 325247. Scand. Kennedy (1994) Understanding the WienerHopf factorization for the simple random walk. Reinhard (1985) Probabilities de ruine pour une classe de modeles de risque semiMarkoviens.J. Probab. Taylor (1981) A Second Course in Stochastic Processes.C. 239256.B. [213] P. Munksgaard. 118. Probab. 35.374 BIBLIOGRAPHY [209] D. 283306. Ann. 259281. Astin Bull. Proc. Academic Press. 561563. Proc.L. Cambridge Philos. 15.A. [226] J. 116129. Appl. [228] J. Amer. 165167. Wishart (1964) Boundary problems for additive processes defined on a finite Markov chain. 173190.G. [212] J. Kalashnikov (1999) Bounds for ruin probabilities in the presence of large claims and their comparison. 187193. Act. [214] A.M. Johnson & M. Technometrics 30. N. 4151. Stochastic Models 1.R. Taaffe (1989/90) Matching moments to phase distributions. New York. Astin Bull. Lazar (1998) Subexponential asymptotics of a Markovmodulated random walk with a queueing application. 8793.M. J. 866870. 6. [218] V. Clarendon Press. Soc. 3. [210] D.H. Janssen & J. Manuscript. Chicago. Jensen (1953) A Distribution Model Applicable to Economics. Oxford.
[233] C. [236] C. Philos. Cambr. 6075. [239] H. SpringerVerlag. Probab. Adv.F. Soc. [231] J. Latouche & V. Probab. 18. Scand. Probab. J. Insurance : Mathematics and Economics 12. I. Mikosch (1995) Explosive Poisson shot noise with applications to risk retention.S. Penev & A. 48. Lucantoni (1991) New results on the single server queue with a batch Markovian arrival process.F. J. 4958. Klnppelberg (1993) Asymptotic ordering of risks and ruin probabilities. [245] T. [248] D. [243] A. Prob. Quart. Appl. 383392. [247] L. 1998 . 259264. 25. Klnppelberg (1989) Estimation of ruin probabilities by means of hazard rates. Lipsky (1992) Queueing Theory . [246] D. Turbin (1973) An asymptotic expansion for the absorption time of a Markov chain distribution. Lehtonen & H. 60 . Korolyuk. [240] U. 125147. SpringerVerlag. Lindley (1952) The theory of queues with a single server. Kliippelberg (1988) Subexponential distributions and integrated tails. 283284. 133135 (in Russian).C. B24. Ramaswami (1999) Introduction to MatrixAnalytic Methods in Stochastic Modelling. 279285. Stadtmiiller (1998) Ruin probabilities in the presence of heavytails and interest rates . Camb. Scand. 24. Appl. [238] V. Lemoine (1981) On queues with periodic Poisson input.a Linear Algebraic Approach. 359361. Klnppelberg & T. Klnppelberg & T. Quart.P. Act. Scand. Proc. 132141. Proc.J. SIAM. Kliippelberg & U.BIBLIOGRAPHY 375 [229] J. Insurance: Mathematics and Economics 8. [230] J. Appl. Soc. [232] C. J. . J. Bernoulli 1 . Act.C. Kunita (1976) Absolute continuity of Markov processes. J. Macmillan. 858874. Lecture Notes in Mathematics 511. Math. Kingman (1962) On queues in heavy traffic. Mikosch (1995) Modelling delay in claim settlement. Philos.C. 390397. [234] C. Soc. [242] A. Actuarial J. [237] C. [241] G. 889900.F. Nyrhinen (1992a) Simulating levelcrossing probabilities by importance sampling.J.F. 4477. Sorensen (1997) Exponential Families of Stochastic Processes. Seminaire de Probabilties X . Kingman (1964) A martingale inequality in the theory of queues. [244] T. J. J. Stochastic Models 7. New York. Roy.. Nyrhinen (1992b) On asymptotically efficient simulation of ruin probabilities in a Markovian environment. (235] C. 146. 277289. 26. Lemoine (1989) Waiting time and workload in queues with periodic Poisson input. Appl. Lehtonen & H. Oxford 12. 1995 . Kuchler & M. Statist. Kingman (1961) A convexity property of positive matrices . Cybernetica 4. 154168.
Malinovskii ( 1994) Corrected normal approximation for the probability of ruin within finite time. J. Malinovskii ( 1996) Approximation and upper bounds on probabilities of large deviations of ruin within finite time. [250] F. 29 . Soc. [257] A. Lundberg (1903) I Approximerad Framstdllning av Sannolikhetsfunktionen. 1994. J. 124147.F. Makowski ( 1994) On an elementary characterization of the increasing convex order . [256] A. Probability and Mathematical Statistics (A. Ann. Act. Meier ( 1984). Uppsala. J. MartinL3f (1986) Entropy. Hoist eds. 561592. Gut & J. Ney & E. Act. MeierHellstern & M . Adv. Appl.S. Appl. [264] G. [265] S. 268285. [262] H . Englunds Boktryckeri AB. Probab. [254] V . J. Probab. Miller (1962 ) A matrix factorization problem in the theory of random variables defined on a finite Markov chain . Ann. 36. 370377. Almqvist & Wiksell. Mammitzsch ( 1986 ) A note on the adjustment coefficient in ruin theory. Res. [260] H. a useful concept in risk theory. 58 . F. Probab . Cambridge Philos. 3. .K. Appl. [252] A.D. 676705. Soc.376 BIBLIOGRAPHY [249] D. 1996 .V. Cambridge Philos.). 31. 147149. Schmidt ( 1993) On ladder height distributions of general risk processes . Eigenvalue properties and limit theorems . K. J. Proc. [259] Z . [266] P. 22. with an application . 1986 . Th. 161174. Moustakides ( 1999) Extension of Wald 's first lemma to Markov processes. [251] F. Opns. Stoch. Probab. Miller ( 1961 ) A convexity property in the theory of random variables defined on a finite Markov chain. 32. 378406. MartinL6f (1983) Entropy estimates for ruin probabilities . Miyazawa & V. 1261 (0?)1270. 286298. Act. J.V. Math. J. Nummelin (1987) Markov additive processes I. 429448. Probab. Insurance : Mathematics and Economics 5. 58. Math.M.K. 2. Statist. [255] V. Scand. II Aterforsdkring av Kollektivrisker . Ann. Lundberg (1926) Forsdkringsteknisk Riskutjdmning. 4859. Stockholm. Anal. Michna ( 1998) Selfsimilar processes in collective risk theory. Proc. 11 . Lucantoni . Europ. 2939. [253] V. 223235. [258] K .S. [263] M . [261] H. Probab. Scand. 763776. 834841. 15. Neuts ( 1990) A single server queue with server vacations and a class of nonrenewal arrival processes . . Scand.D. A fitting algorithm for Markovmodulated Poisson processes having two arrival rates .D. Nagaev (1957) Some limit theorems for stationary Markov chains. Appl. Appl. Miller ( 1962 ) Absorption probabilities for sums of random variables defined on a finite Markov chain . Appl.
Paulsen (1998) Ruin theory with compounding assets . Neuts ( 1978) Renewal processes of phase type. Neuts ( 1989) Structured Stochastic Matrices of the M/G/1 Type and their Applications. Pakes (1975) On the tail of waiting time distributions. 316. 555564. [272] J . IEICE Trans. 16. Johns Hopkins University Press. 135148. Appl. 46. London. Proc. 215223.K. Appl. Stoch. Neuts (1977) A versatile Markovian point process. [282] J . [281] J. [283] J. Omey & E. [275] H. Stoch. Appl. [270] M . Norberg ( 1990) Risk theory and its statistics environment . Appl.K.A. [278] E. 273299. Models 6. 30. . 36. Marcel Dekker. [273] R. O'Cinneide (1990) Characterization of phasetype distributions. 123144.F. Proc. [284] J. 764779. 12. 249266. Neuts ( 1981 ) MatrixGeometric Solutions in Stochastic Models. 75. 12551265.BIBLIOGRAPHY 377 [267] M . Baltimore . 71. Insurance: Mathematics and Economics 22. [269] M . 339353. E75B . [280] A. Proc. [277] J.F. New York. Commun. [276] C. [279] E. 445454. Probab. 157. Astin Bulletin 5 . Probab. 21. Stoch.F. [271] M. Ohlin (1969) On a class of measures for dispersion with application to optimal insurance. [268] M . Paulsen & H. Abh. Appl. Stoch. Sem. [274] H .a survey. Nyrhinen (1999) Large deviations for the time of ruin. Appl. Willekens (1986) Second order behaviour of the tail of a subordinated probability distributions. Hamburg. Neveu ( 1961 ) Une generalisation des processus a accroisances independantes. 311342. [285] J. Gjessing (1997a) Optimal choice of dividend barriers for a risk process with stochastic return on investments. Willekens (1987) Second order behaviour of distributions subordinate to a distribution with finite mean . Naval Research Logistics Quarterly 25. Neuts (1992) Models based on the Markovian arrival process. Statistics 21.G. J. Insurance: Mathematics and Economics 20.F. J. Stoch. 10081026. Math. Stochastic Models 3 . Omey & E. Paulsen & H. Appl. Appl. 733746. Proc. J. Adv. Paulsen (1993) Risk theory in a stochastic economic environment. Nyrhinen ( 1998 ) Rough descriptions of ruin for a general class of surplus processes. 327361. Gjessing (1997b) Present value distributions with applications to ruin theory and stochastic equations. Paulsen ( 1998) Sharp conditions for certain ruin in a risk process with stochastic return on investments.F. Probab. Probab.
Appl. Adv.R.A. Appl. [289] C. Insurance: Mathematics and Economics 3. 222261. [291] S. Ramaswami (1980) The N/G/1 queue and its detailed analysis. [287] F. 691707.J. Prabhu (1980) Stochastic Storage Processes.K. Probab. Astin Bulletin XIV. Appl. Insurance Risk. Probab . 1989 .M. Wiley. Queues. [300] C. Scand. Math. Probab.U. Math. 4568. 568576.378 BIBLIOGRAPHY [286] J. New York. Pellerey (1995) On the preservation of some orderings of risks under convolution. 965985.J. 30. [305] C. [290] E. Math.. Berlin. 147159. [294] N. R. Rogers (1994) Fluid models in queueing theory and WienerHopf factorisation of Markov chains. [306] T. . 46. Math. Insurance: Mathematics and Economics 16. Tidskr. Act. Paulsen & H. [295] N. J. Samorodnitsky (1997) Performance degradation in a single server exponential queueing model with long range dependence.M. Appl. [299] C. Philipson (1968) A review of the collective theory of risk. Statist.U. 757764. [304] B. Adv.U. Res. 390413. and Dams. Statist. 32. Probab. Math. Pyke (1959) The supremum and infimum of the Poisson process. Queueing Systems 5. Wiley. Prabhu (1961). Pitts (1994) Nonparametric estimation of compound distributions with applications in insurance. [303] S. 117133. Ann. 337347. 4. 19. de Smit (1986) The queue M/G/1 with Markovmodulated arrivals and services. Austr. [292] S.G. 61. J. New York. Adv. [297] R. Springer. Ripley (1987) Stochastic Simulation . Wiley. [288] S. [301] G. [302] J. Appl. Probab. Ann. Rao (1965 ) Linear Statistical Inference and Its Applications. Resnick & G. Stat.G. Oper. Aktuar. J. 215246. Opns. 465483. [293] N.L. 29A. Ann. 537555. 12. Ramsay (1984) The asymptotic ruin problem when the healthy and sick periods form an alternating renewal process. Heidelberg. 2330. Embrechts (1996) Confidence bounds for the adjustment coefficient. Ann.U.M. 11. Pitts. Prabhu & Zhu (1989) Markovmodulated queueing systems. 29. Inst. 235243. Schock Petersen (1989) Calculation of ruin probabilities when the premium depends on the current reserve. [298] V.K. Skand. Res. Reinhard (1984) On a class of semiMarkov risk models obtained as classical risk models in a markovian environment. Regterschot & J. Pitman (1980) Subexponential distribution functions. [296] N. Soc. Gjessing (1997c) Ruin theory with stochastic return on investments. 2343.H. On the ruin problem of collective risk theory. Prabhu (1965) Queues and Inventories. Rolski (1987) Approximation of periodic queues. 139143. 820827. 28. Griibel & P.
Astin Bulletin 29. Lecture Notes of the 8th Summer School on Probability and Mathematical Statistics ( Varna). Schmidli ( 1995 ) CramerLundberg approximations for ruin probabilities of risk processes perturbed by a diffusion . [313] M. [325] H . Probab . [315] T. 687699. 5. Stochastic Models 10. Schmidli ( 1997a) Estimation of the Lundberg coefficient for a Markov modulated risk model . 4857. Statist. Rolski. Appi. Schmidli. 131156. Comp. Wiley. Schmidli ( 1996) Martingales and insurance risk. J. Wiley. Schmidli ( 2001 ) Optimal proportional reinsurance policies in a dynamic setting . Data Anal. Adv. [321] H. Ryden (1996) An EM algorithm for estimation in Markovmodulated Poisson processes.Y. 227244. Schmidli ( 1999b) Perturbed risk processes: a review . Ann. 7. [309] H.L. 145165. J. [312] R. Probab. 365388. Rubinstein & B. Singapore. Appl. Proc. [322] H. Insurance: Mathematics and Economics 16. Ross (1974) Bounds on the delay distribution in GI/G/1 queues. [308] S. Operationsforsch. Ryden (1994) Parameter estimation for Markov modulated Poisson processes. . 431447 [316] S. Schmidt & J. 262286. 1997. Seal ( 1969) The Stochastic Theory of a Risk Business.BIBLIOGRAPHY 379 [307] T. [311] R. Schmidli ( 1994) Diffusion approximations for a risk process with the possibility of borrowing and interest. Schmidli ( 1997b) An extension to the renewal theorem and an application in risk theory.M. Act. [317] L. Melamed (1998) Modern Simulation and Modelling. Scand. 2001 . Stochastic Models 10 . [323] H. 155188. 5. Shapiro (1993) Discrete Event Systems: Sensitivity Analysis and Stochastic Optimization via the Score Function Method. Act. 21. 121133. Schmickler ( 1992 ) MEDA : mixed Erlang distributions as phasetype representations of empirical distribution functions . Wiley.J. Schlegel ( 1998) Ruin probabilities in perturbed risk models. [314] T. 5. Rubinstein & A. [320] H. [318] H. 11. 4068. Rudemo (1973) Point processes generated by transitions of a Markov chain. 93104. [319] H. J. Schmidli ( 1999a) On the distribution of the surplus prior and at ruin. [326] H. Stochastic Models 6.Y. Wiley. Statist. Math. V. Th.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. Teugels (1999) Stochastic Processes for Insurance and Finance. Insurance: Mathematics and Economics 22. Probab. [324] H . [310] R. Science Culture Technology Publishing . Wiley. H. 795829. 135149. Stoch. Scand. Rubinstein (1981) Simulation and the Monte Carlo Method. 417421. Appl.
Scand.L. Mitt. 157166. Seal (1974) The numerical calculation of U(w. [347] K. Segerdahl (1955) When does ruin occur in the collective theory of risk? Skand. Math. [329] H. 279299. Aktuar Tidsskr. Probab. Siegmund (1979) Corrected diffusion approximations in certain random walk problems. Siegmund (1975) The time until ruin in collective risk theory. Seal (1978) Survival Probabilities.S. Segerdahl ( 1942) Uber einige Risikotheoretische Fagestellungen . 21. [328] H. Siegmund ( 1985) Sequential Analysis . 2236. [348] K . Shin (1987) Convolution of uniform distributions and ruin probability. [336] B. 914924. Shaked & J. 171178.G. Stochastic Models 6. Act. Sengupta (1990) The semiMarkov queue: theory and applications. L. t). Grenander). [338] E. Insurance: Mathematics and Economics 8. Mitteil. Shin (1989) Ruin probability by operational calculus. 191197. Risk teory and the single server queue . [343] D. Wiley. [345] D. J. Ann.380 BIBLIOGRAPHY [327] H. [346) D . 383413. Probab.A. [342] D. Shwartz & A.S. Versich. Probab. [333] C: O. Adv. [332] C : O.L. 429442. J. Shtatland ( 1966) On the distribution of the maximum of a process with independent increments. 61. 4. [341] E. 1955. Appl. Seber (1984) Multivariate Observations. Verein Schweiz. Sigman ( 1994 ) Stationary Marked Point Processes: An Intuitive Approach. Statist. [334] C: O. Math. .the Harald Cramer volume (ed. Chapman & Hall. SpringerVerlag. New York. Mitt. Seal (1972) Numerical calculcation of the probability of ruin in the Poisson/Exponential case. [340] E. Math.L. Chapman and Hall. t). Wiley. Verein Schweiz. [339] A. 72. Appl. Shantikumar (1993) Stochastic Orders and Their Applications. Seal ( 1972). 121139. 701719. Almqvist & Wiksell. Siegmund (1976) Importance sampling in the Monte Carlo study of sequential tests.S. 11. Segerdahl (1959) A survey of results in the collective theory of risk. . 673684. pp. Versich. Queueing Systems 11. 4383. Ann. the probability of nonruin in an interval (0. Verein Schweiz Versich. [335] B . Adv. Aktuar Tidsskr. Probab . Appl. Sigman (1992) Light traffic for workload in queues.W. Th. 159180. Act.F. Stockholm. [330] H. 1974. 4. [331] G. Sengupta ( 1989) Markov processes whose steadystate distribution is matrixgeometric with an application to the GI/PH/1 queue. [344] D. 483487. 1987. 11.W. Weiss (1995) Large Deviations for Performance Analysis. [337] M. 77100. 75. 72. Siegmund (1976) The equivalence of absorbing and reflecting barrier problems for stochastically monotone Markov processes . In Probability and statistics . Academic Press. Skand. 243249. Scand.
J. [357] B. Springer. 163175. Taylor (1979) Probability of ruin with variable premium rate. Act. 1980 .). [354] B.). Scand. 118. Thorin (1977). Sundt & W. 197208. 114137. Scand.V. Daley ed.L. [352] D.C.C. Sundt & J. 21. Taylor (1979) Probability of ruin under inflationary conditions or under experience ratings . Karlsruhe.. [368] O. [359] L. Hoesman ( 1989) Moderate. [350] W. John Wiley & Sons. 5776. [361] G. 8199. 8594. Sundt ( 1993) An Introduction to NonLife Insurance Mathematics (3rd ed. Stanford & K. [351] D . 2739. Takhcs (1967) Combinatorial Methods in the Theory of Stochastic Processes. Teugels (1997) The adjustment coefficient in ruin estimates under interest force . [365] J. [358] R. Smith ( 1953) Distribution of queueing times. [356] B . Insurance: Mathematics and Economics 1. 1982 . Verlag Versicherungswirtschaft e. 1974. NorthHolland. 235254. Act. Scand. Astin Bulletin 16 . J. 65102. Stoyan (1983) Comparison Methods for Queues and Other Stochastic Models (D. 722. Astin Bulletin 12. Sundt & J. [364] G. Adv.J. Taylor (1978) Representation and explicit calculation of finitetime ruin probabilities.L.A.S. 149162. Ruin probabilities prepared for numerical calculations. Thorin (1982) Probabilities of ruin. [366] O. Astin Bulletin 24. Insurance: Mathematics and Economics 16. Act.BIBLIOGRAPHY 381 [349] E. Scand. New York. 1978. J. [353] E. Jewell ( 1981 ) Further results on recursive evaluation of compound distributions . Proc.C. Act. [367] O.L. [369] O. Thorin (1974) On the asymptotic behaviour of the ruin probability when the epochs of claims form a renewal process. 65102. Slud & C. New York.C. J. J.L. Actuarial J. Wiley. Stroinski ( 1994) Recursive method for computing finitetime ruin probabilities for phasedistributed claim sizes. Appl. Taylor (1976 ) Use of differential and integral inequalities to bound ruin and queueing probabilities ..J. Scand. 1976. Suri ( 1989) Perturbation analysis : the state of the art and research issues explained via the GI/G/1 queue. Proceeedings of the IEEE 77. Thorin (1986) Ruin probabilities when the claim amounts are gamma distributed. Soc 49. [363] G. [360] G. Unpublished manuscript. Cambridge Philos. 725741.and largedeviations probabilities in actuarial risk theory. Insurance: Mathematics and Economics 19. 449461. Straub (1988) NonLife Insurance Mathematics.C. Probab. Scand. Teugels (1982) Estimation of ruin probabilities . [355] B . . [362] G. Teugels ( 1995 ) Ruin estimates under interest force . Taylor (1986) Claims reserving in NonLife Insurance. Act.
M. Wallace (1969) The solution of quasi birth and death processes arising from multiple access computer systems. Goovaerts. Dordrecht Boston Lancaster. M. Waters (1983) Probability of ruin for a risk process with claims cost inflation. SpringerVerlag. Taqqu. 148164. [374] F. Thorin & N. Willmot (1994) Refinements and distributional generalizations of Lundberg's inequality. VazquezAbad (1999) RPA pathwise derivatives of ruin probabilities.382 BIBLIOGRAPHY [370] 0. Submitted. J. J. Wald (1947) Sequential Analysis. Unpublished Ph. J. University of Michigan. Tacklind (1942) Sur le risque dans les jeux inequitables. Aktuar. In: Premium Calculation in Insurance (F. PrenticeHall. Willinger. Van Wouve. Thorisson (2000) Coupling. [373] S. Wikstad (1977) Calculation of ruin probabilities when the claim distribution is lognormal. Reidel. Statistical Science 10.).E.E. [383] G. Appl. [378] H. De Vylder. [380] W. Oper. 37. 5762. Res. De Vylder & M. 743756. [384] R. [381] W. Insurance: Mathematics and Economics 15. Tidskr. 1983 . [382] G. 137153. [371] 0. [376] A. Insurance : Mathematics and Economics 13. 231246. [375] N. Whitt (1989) An interpolation approximation for the mean workload in a GI/G/1 queue. Act. F. Wilson (1995) Selfsimilarity in highspeed traffic: analysis and modeling of ethernet traffic measurements. Goovaerts (1983) The influence of reinsurance limits on infinite time ruin probabilities. Astin Bulletin VII. thesis. Astin Bulletin IX. Lin (1994) Lundberg bounds on the tails of compound distributions. D.R. [372] H. Willmot & X. Haezendonck eds. Skand. [377] V. Thorin & N. 31. 936952. Scand. Wikstad (1977) Numerical evaluation of ruin probabilities for a finite period. Stationarity and Regeneration. Probab. [379] M. Wolff (1990) Stochastic Modeling and the Theory of Queues.W. 1942. R. 6785. Veraverbeke (1993) Asymptotic estimates for the probability of ruin in a Poisson model with diffusion. Sherman & D. . 142. Wiley. 4963.
271274.269. 119. 97129.86. 341. 9293. 111117.307312 compound Poisson model 4. 2425.292293 Edgeworth expansion 113. 37. 14. 205.160167. 40. 5.6779. 91. 217. 3032.4447. 141144. 15.314316.178184. 135.328330. 360 excursion 155156. 227229.318320 change of measure 2630.308 CramerLundberg model: see compound Poisson model cumulative process 334 dams: see storage process differential equation 16.346349 383 . 323 Coxian distribution 147. 278 gamma distribution 67.135. 3839. 189.and sum 221.137141.285292.228229. 7179. 8283 hyperexponential distribution 7. 2526. 201 Brownian motion 3 . 7079.293294. 162164. 97. 122.299.226. 361 diffusion 3. 9496. 239. 12 CramerLundberg approximation 1617.259261. 5796.100. 180182. 248 WienerHopf 144 interest rate 190. 3436. 110113.249250 integral equation 16 Lindley 143 renewal 64.281. 170173.359 aggregate claims 103106. 79. 1415. 117127 corrected 121127 duality 1314. 86.217. 308. 332333 Volterra 192194.272.242. 218 Cox process 4. 17. 89. 117128.251280 heavy traffic 76.203. 283.200201.185187. 1112. 226. 201214.121129. 138139. 9396.287292. 17. 7475. 302303 diffusion approximation 17. 39.182. 245248.203. 7879.301 central limit theorem 60 .249. 301 Kronecker product. 207 heavytailed distribution 6. 316323 Bessel function 102.9899. 196201 inverse Gaussian distribution 76. 318319 Erlang distribution 7.150. 80 81. 1819.Index adjustment coefficient 17. 4851. 3334.
6162. 306316 Levy process 3. 2730. 134. 203204. 185187 GI/G/1 141144 M/D/1 6667 equation 16.339 large deviations 129. 9899. 14.336339 Laplace transform 15. 37. 133. 179 NP approximation 318320 Palm distribution 5253. 16.302. 52 53. 157.234.240244. 269 PerronFrobenius theory 4142. 149. 100.315 inequality 1718. 15. 32. 133. see also sensitivity analysis phasetype distribution 8.340350 multiplicative functional 2830.285287 queue 14 .139141.128129. 227228.304 process 2830. 142 likelihood ratio : see change of measure lognormal distribution 9. 295. 65. 25.350361 Poisson process Markovmodulated 12 periodic 12. 35. 137139.297299.287291 INDEX matrix equation . 145187.259261.146148.218221.215250. 261264. 144.178182. 213214. 44. 38. 7179. 230.269271.234240. 141144. 108109. 134135.288290. 304305 random walk 3336. 176185. 113114. 80. 42.148. 154. 16.180.238. 4446.161. 171. 162. 44. 175 light traffic 8183 Lindley integral equation 143 process 3334.336339 . 38. 35. 71.261264.201. 245 M/G/1 13. nonlinear 155. 271274.152160. 106108.349 350 perturbation 172173. 178 modulation 12. 6970. 99. 3639. 39.287. 3947.134135.123. 132133.298299. 96. 25.384 ladder heights 4756.108.227230. 234 matrixexponential distribution 240244 matrixexponentials 14. 203 Markov additive process 12. 59.160161. 112113.161164. 7576.275278.348 terminating 215216. 260 Lundberg conjugation 6979 .174. 176185 nonhomogeneous 60 PollaczeckKhinchine formula 6167. 257. 108 life insurance 5.161. 3947. 5758. 138. 86 periodicity 12. 251. 229 M/M/1 101 Markovmodulated 185187 periodic 187 martingale 2426. 267269 Panjer's recursion 320323 Pareto distribution 910. 41.
191192. 292294. 3032. 11. 89. 120 statistics x.244250. 233. 168172 storage process 13. 12. 229234.279280 Rouche roots 158. 251280 time change 4. 60. 332333 model 12. 223226. 317318 semiMarkov 147.359361 stochastic control x stochastic ordering 18. 7475. 338 utility 324. 222.INDEX 385 waiting time 141. 240. 327 . 261264 reservedependent premiums 14.314. 160. 1819. 244. 146. 87.273274. 256258. 253. 333334 regular variation 10.154157. 152. 279280 subexponential distribution 11. 257. 294296 shotnoise process 314 simulation 19. 189214. 238 saddlepoint method 115117. 260 reinsurance 8. see also matrixexponential distribution regenerative process 264 268. 307308.186. 141144. 172173. 9693. 174. 162. 37. 4950. 251. 131144. 251. 326330 Weibull distribution 9. 5455. 281296 stable process 15. 147.336339 workload 13. 8386. 213. 335336 sensitivity analysis 8693. 186187 renewal process 131. 186187 virtual: see workload rational Laplace transform 8. 280. 233234. 177 timereversion 14. 31. 331336 equation 64. 260 WienerHopf theory 144.262263. 123. 107.
worldscientific. Markovmodulation or periodicity.Advanced Series on Statistical Science & Applied Probability . Some i (l I JL I J r of the topics are Lundberg's inequality. "This book is a must for anybody working in applied probability. y finite horizon ruin probabilities.. the ^W A l \ i l ' ''' CramerLundberg approximation. P'i yfliother approximations (e. I 1! Ruin Probabilities .g. extensions of the classical compound Poisson model to allow f o r reservedependent premiums. for heavytailed claim size distributions).com 2779 he 9 "789810ll22293211 . exact solutions. phasetype distributions as a computational vehicle and the connection to other applied probability areas like queueing theory.Vol. Special features of the book are the emphasis on change of measure techniques." Short Book Reviews ISBN 9810222939 mi u inn i nun I I I I I I i in u www.. 2 A I 11 JjVb l' i  i Yj . It is a comprehensive treatment of the known results on ruin probabilities.T [Ail i The book is a comprehensive treatment of  I i I \ classical and modern ruin probability theory.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue reading from where you left off, or restart the preview.