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
. . . . . . . . . . . . . . . . . . . 316 5 Principles for premium calculation . . . . . . . . . . .. 294 XI Miscellaneous topics 297 1 The ruin problem for Bernoulli random walk and Brownian motion. . . . . . . . . . . . . . . 350 Bibliography Index 363 383 . . . . . . . . . . 297 2 Further applications of martingales . . . . . . .. . . . . 290 5 Regenerative simulation . 331 A2 WienerHopf factorization . . . . . . 264 5 Finitehorizon ruin probabilities . . . . . . . . 340 A4 Some linear algebra . . . 292 6 Sensitivity analysis . 259 3 The renewal model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271 6 Reservedependent premiums . . . . . . . . . . .. . . . . .. .. . . . . . . . . . . . . . . . The twobarrier ruin problem . . . . 306 4 The distribution of the aggregate claims . . . . . . . .. 251 2 The compound Poisson model . . . . . . . . . . . .. . . . . . .. . . .. . .. . . . . . . . . . . . . . . . . . . ... . . . .. . . . . . . . . . .. . . 261 4 Models with dependent input . . . . . . . . 279 X Simulation methodology 281 1 Generalities . . . . . . . . . . . . . . . . . . . 304 3 Large deviations . . 285 3 Importance sampling via Lundberg conjugation . . . . . . . . . . . . . . . 323 6 Reinsurance . . .. . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . .. . . .CONTENTS vii IX Ruin probabilities in the presence of heavy tails 251 1 Subexponential distributions . . 344 AS Complements on phasetype distributions . . . . . 336 A3 Matrixexponentials . . . . . . . . . . .. . . .. . . . . . . . . . . . . . . 281 2 Simulation via the PollaczeckKhinchine formula . .. . . . . . . . . . ... . . 287 4 Importance sampling for the finite horizon case . . . .. 326 Appendix 331 Al Renewal theory . . . . . . . . . . . . . . . . . . . . . . . . .
This page is intentionally left blank .
it is not by intention. and has been an active area of research from the days of Lundberg all the way up to today. the idea was close to expand these to a short book on the subject. Thus. The course was never realized.Preface The most important to say about the history of this book is: it took too long time to write it! In 1991. and my belief was that this could be done rather quickly. 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. But the pace was much slower than expected. Apart from these remarks. Since I was to produce some handouts for the students anyway. I was invited to give a course on ruin probabilities at the Laboratory of Insurance Mathematics. Let me take this opportunity to thank above all my publisher World Scientific Publishing Co. which has in particular removed one of the standard criticisms of the area. However. As an excuse: many of these projects were related to the book. In particular. it would not be fair not to say that the practical relevance of the area has been questioned repeatedly. I have deliberately stayed away from discussing the practical relevance of the theory. and the result is now that the book is much more related to my own research than the initial outline. if the formulations occasionally give a different impression. and the series editor Ole BarndorffNielsen for their patience. One reason for writing this book is a feeling that the area has in the recent years achieved a considerable mathematical maturity. University of Copenhagen. Risk theory in general and ruin probablities in particular is traditionally considered as part of insurance mathematics. A similar thank goes to all colleagues who encouraged me to finish the project and continued to refer to the book by Asmussen which was to appear in a year which continued to be postponed. It has obviously not been possible to cover all subareas. 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!). the book is basically mathematical in its flavour. and other projects absorbed my interest.
Another interesting area which is not covered is dynamic control. it has not been possible to incorporate more numerical examples than the few there are. see also Schmidli [325] and the references in Asmussen & Taksar [52]. A book like this can be organized in many ways. The present book is in between these two possibilities. 111. see in particular Michna [259].15.lth.6 (to understand the PollaczeckKhinchine formula in 111.2 more properly). IX.maths . read Chapter I.13. Willinger et al. Hojgaard & Taksar [206].1. Finally.45. The rest is up to your specific interests.13 and XI. IV. In the classical setting of CramerLundberg models. IV. I regret that due to time constraints. see e.14.se/matstat / staff/asmus and I am therefore grateful to get relevant material sent by email to asmusfmaths . the standard stochastic control setting of diffusion models has been considered. the first part of 11. More recently.g.13 and IX. For a second reading.4a.g. In addition.lth. some papers not cited in the text but judged to be of interest are included in the Bibliography. The main motivation comes from statistical data for network traffic (e. [381]). For a brief orientation.5. Hojgaard & Taksar [35] and Paulsen & Gjessing [284]. some basic discussion can be found in the books by Biihlmann [82] and Gerber [157]. 111.3. Resnick & Samorodnitsky [303] and references therein. One is by model. VII. incorporate 11. VI. I intend to keep a list of misprints and remarks posted on my web page. VII.x PREFACE field of queueing theory. Chapters IIIVII introduce some of the main models and give a first derivation of some of their properties. X. http:// www.13. Good luck! I have tried to be fairly exhaustive in citing references close to the text. an area which is becoming increasingly important. VIII. Concerning ruin probabilities. e. 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. another by method. Asmussen. for which I apologize to the reader and the authors of the many papers who ought to have been on the list.2. for the effects on tail probabilities.g.89.se Lund February 2000 Soren Asmussen . IV.2. 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). The book does not go into the broader aspects of the interface between insurance mathematics and mathematical finance.
Section VII . of which there are not many at this stage . More substantial remarks.2 by Rafal Kulik . not least the more complicated ones. Lund September 2001 Soren Asmussen Acknowledgements Many of the figures .6. 1 is almost identical to Section 2 of Asmussen [26] and reprinted with permission of Blackwell Publishers. Fig. Fig. 3 is reprinted from Asmussen & Nielsen [39] and parts of IX. Schmidli & Schmidt [47] with the permission from Applied Probability Trust . Section VIII.6 is reprinted from Asmussen & Schmidt [49] and parts of IX. 5 from Asmussen & Kliippelberg [36] with the permission from Elsevier Science . supported by Center for Mathematical Physics and Stochastics (MaPhySto).5 from Asmussen [21] with permission from CRC Press. A number of other figures were supplied by Christian Geisler Asmussen . Parts of II.1 by Bjarne Hojgaard and the table in Example 111.8 .4 from Asmussen.3 are reprinted from Asmussen & Rubinstein [46] and parts of VIII.6 by my 1999 simulation class in Lund. were produced by Lone Juul Hansen . as well as some additional references continue to be at the web page.1 and X. 111 . 5. Aarhus. .PREFACE xi The second printing differs from the first only by minor corrections. Parts of X. IV. many of which were pointed out by Hanspeter Schmidli .
This page is intentionally left blank .
Chapter I Introduction 1 The risk process In this chapter . A risk reserve process { Rt}t>o. (1.Rt.i(u. Letting T(u) = inf {t > 0 : Rt < 0} = inf It > 0 : St > u}.T) = P inf Rt < 0 I . (1. (1.4) O<t<oo O<t<T 1 . it is frequently more convenient to work with the claim surplus process {St}t>0 defined by St = u . The probability O(u) of ultimate ruin is the probability that the reserve ever drops below zero. as defined in broad terms .1) We also refer to t/) ( u) and 0(u. For mathematical purposes. MT = sup St. We denote throughout the initial reserve by u = Ro. is a model for the time evolution of the reserves of an insurance company. results and topics to be studied in the rest of the book. and give a very brief summary of some of the models. t/i(u) = P (infRt < 0) = P (infR t < 0 t>0 t>0 The probability of ruin before time T is t.3) sup St. They are the main topics of study of the present book. respectively.2) (O<t<T Ro=ul. T) as ruin probabilities with infinite horizon and finite horizon . M = (1. we introduce some general notation and terminology.
1. say. St = E Uk . T3. Thus. per unit time. the ruin probabilities can then alternatively be written as . • Premiums flow in at rate p.pt. 1. and T1 is the time of the first claim. (1.1 ..7) k=1 k=1 The sample paths of {Rt} and {St} and the connection between the two processes are illustrated in Fig. the number Nt of arrivals in [0. However. we see that Nt Nt Rt = u + pt . Figure 1. respectively. That is. and Nt = min {n > 0 : 0rn+1 > t} = max {n > 0: Un < t}• The size of the nth claim is denoted by Un.2 CHAPTER I. the following setup will cover the vast majority of the book: • There are only finitely many claims in finite time intervals.E Uk.6) Sofar we have not imposed any assumptions on the risk reserve process. (1. (1. INTRODUCTION be the time to ruin and the maxima with infinite and finite horizon.b(u) = P (r(u) < oo) = P(M > u).T) = F (MT > u) = P(r(u) < T). .. the time of arrival of the nth claim is an = T1 + • • • + Tn. We denote the interarrival times of claims by T2. t] is finite. Putting things together.i(u.5) i.
• General Levy processes (defined as continuous time processes with stationary independent increments) where the jump component has infinite Levy measure. If 77 > 0. 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.20%. rl= pP P It is sometimes stated in the theoretical literature that the typical values of the safety loading 77 are relatively small. For the purpose of studying ruin probabilities this distinction is. but as an approximation to the risk process rather than as a model of intrinsic merit. a basic references is Gerber [127]. and hence . of course. for example. If 77 < 0. and the basic ruin probabilities are derived in XI. However. one may well argue that Brownian motion in itself could be a reasonable model. since any modeling involves some approximative assumptions. It would appear obvious.e.1.1 the slope of {Rt} should depend also on the level). . however. 1. t * oo.1. We study this case in Ch. (1.) V 0. then M < oo a..1 Assume that (1. say 10% . allowing a countable infinity of jumps on Fig.8) holds. Some main examples of models not incorporated in the above setup are: • Models with a premium depending on the reserve (i. Thus. and hence O(u) < 1 for all sufficiently large u.b(u) = 1 for all u. 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.(.8) The interpretation of p is as the average amount of claim per unit time. 1. on Fig. then M = oo a. not discuss whether this actually corresponds to practice. We shall discuss Brownian motion somewhat in Chapter IV. • Brownian motion or more general diffusions. VII. however.1. and in fact: Proposition 1..s. we shall.s. that the insurance company should try to ensure 77 > 0. The models we consider will typically have the property that there exists a constant p such that Nt a E Uk k=1 p. one could well replace Rt by Rtnr(u) or RtA. though many results are straightforward to generalize from the compound Poisson model. We shall not deal with this case either. immaterial.
T) = i. it is not too difficult to show that p as defined by (1. if {(3(t)} is nonergodic..d.2 (Cox PROCESSES) Here {Nt} is a Poisson process with random rate /3(t) (say) at time t. then this limit is > 0 which implies St a$ oo and hence M = oo a. . St In concrete models.i.s.oo t 0 J (provided the limit exists).11) . However. are i. However. (1.10) hold with p constant.b(u) < 1 for all u when rl > 0.Q (say) and U1. Proposition 1. _ St __ k =1 Uk pt a4. and that .10) is a property which we will typically encounter. and here (1.8) is given by ^t p = EU • lim it (3(s) ds t. namely. t t p  p' t ^ oo. we obtain typically a somewhat stronger conclusion.. . U2. . are i.3 Assume p 54 1 and define Rt = Rt1p.4 CHAPTER I.i. If 77 < 0.d. not all models considered in the literature have this feature: Example 1.s.Tp). where {Nt} is a Poisson process with rate . in connection with risk processes in a Markovian or periodic environment (Chapter VI). This case is referred to as the mixed Poisson process.8). (1. The simplest example is 3(t) = V where V is a r . this needs to be verified in each separate case.v. k=1 (1. M < oo a.6EU (on the average. If u oo.Q claims arrive per unit time and the mean of a single claim is EU) and that also Nt t aoo t lira EEUk = p.8) that F N. If U1. Here it is easy to see that p = . The simplest concrete example (to be studied in Chapter III) is the compound Poisson model. corresponding to the Pdlya process. tb(u) = 1 for all u holds also when rl = 0. rl > 0.i(u. 0(u. Then the connection between the ruin probabilities for the given risk process {Rt} and those ^(u). 0 We shall only encounter a few instances of a Cox process. (1. and independent of {Nt}. Nt)}. Thus p may well be random for such processes.s. then similarly limSt/t < 0. INTRODUCTION Proof It follows from (1.10) Again. zP(u . U2. with the most notable special case being V having a Gamma distribution. and independent of {(0(t)... namely that M = oo a..T) for {Rt} is given by V)(u) = t/i (u).
we shall be able to identify p with the traffic intensity of an associated queue. another important early Swedish work is Tacklind [373]. Some main later textbooks are (in alphabetical order) Buhlmann [82]. in particular. but in probability and applied probability as a whole. Heilmann [191].. while the first mathematically substantial results were given in Lundberg [251] and Cramer [91]. Buhlmann [82]. Schmidt & Teugels [307] and Seal [326]. Insurance: Mathematics and Economics. Some of the main general ideas were laid down by Lundberg [250]. Note that life insurance (e. The term risk theory is often interpreted in a broader sense than as just to comprise the study of ruin probabilities. Sundt [354]. often referred to as collective risk theory or just risk theory. [76]. some main texts (typically incorporating some ruin theory but emphasizing the topic to a varying degree) are Bowers et al. Besides in standard journals in probability and applied probability. CLAIM SIZE DISTRIBUTIONS 5 The proof is trivial. the assumption > 0 is equivalent to p < 1. and in fact p < 1 is the fundamental assumption of queueing theory ensuring steadystate behaviour (existence of a limiting stationary distribution). Daykin. Grandell [171]. Notes and references The study of ruin probabilities. 2 Claim size distributions This section contains a brief survey of some of the most popular classes of distributions B which have been used to model the claims U1. Schmidli. was largely initiated in Sweden in the first half of the century. Embrechts et al. Since { Rt } has premium rate 1. [134]. Pentikainen & Pesonen [101]. U2. and we do not get near to the topic anywhere in this book. Segerdahl [334] and Philipson [289]. the claim arrivals are Poisson or renewal at the same time). the role of the result is to justify to take p = 1. many results and methods in random walk theory originate from there and the area was ahead of related ones like queueing theory. An idea of the additional topics and problems one may incorporate under risk theory can be obtained from the survey paper [273] by Norberg. in a number of models.. see also Chapter XI. Some early surveys are given in Cramer [91]. Hipp & Michel [198]. Cox processes are treated extensively in Grandell [171]. which is feasible since in most cases the process { Rt } has a similar structure as {Rt} (for example. [330]. For mixed Poisson processes and Polya processes. We roughly classify these into two groups . Taylor [364]. The Swedish school was pioneering not only in risk theory. De Vylder [110]. Note that when p = 1. In the even more general area of nonlife insurance mathematics. Mitteilungen der Verein der Schweizerischen Versicherungsmathematiker and the Scandinavian Actuarial Journal. Rolski. Straub [353].g. Daykin et al.2. see e . the research literature is often published in journals like Astin Bulletin . the recent survey by Grandell [173] and references therein.. Gerber [159]) has a rather different flavour. Gerber [157].. [101]. lighttailed distributions (sometimes the term .g.
g.g. a fact which turns out to contain considerable information. then the conditional distribution of X . Example 2 . 2a Lighttailed distributions Example 2. The crucial feature is the lack of memory: if X is exponential with rate 6. where B(bo. (2. regularly varying (see below) or even regularly varying with infinite variance.f. and can also be interpreted as the (constant) failure rate b(x)/B(x). In particular. and heavytailed distributions.1 (THE EXPONENTIAL DISTRIBUTION) Here the density is b(x) = beax (2. On the more heuristical side. for the compound Poisson model with exponential claim sizes the ruin probability . B is heavytailed if b[s] = oo for all s > 0. 6 has density r(p)xPleax b(x) P and m. For example in the compound Poisson model. Equivalently. the exponential distribution is by far the simplest to deal with in risk theory as well.1) The parameter 6 is referred to as the rate or the intensity. P B[s]= (8Is ) . 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'.f. but different more restrictive definitions are often used: subexponential. In contrast. a simple stopping time argument shows that this implies that the conditional distribution of the overshoot ST(u) .B(x) satisfies B(x) = O(e8x) for some s > 0.u at the time of ruin given r(u) is again exponential u with rate 8. the m.e. if 1 °O AB Jbos x B(dx) > 0.O(u) can be found in closed form. As in a number of other applied probability areas. B[s] is finite for some s > 0.2 (THE GAMMA DISTRIBUTION) The gamma distribution with parameters p.x given X > x is again exponential with rate b (this is essentially equivalent to the failure rate being constant). s<8.2) = 0.2 and /LB is the mean of B. INTRODUCTION 'Cramertype conditions' is used).6 CHAPTER I. i.8.3) . Here lighttailed means that the tail B(x) = 1 .
An important property of the hyperexponential distribution is that its s. among others.y i=1 where >i ai = 1.1 Poisson events in [0. .c. P b(x) = r` aibiea. one has r(bx. or just the Erlang(p) distribution..) VarX1 (EX )2 p is < 1 for p > 1.d. then X v Xl + • • • + X. the squared coefficient of variation (s.2) can be considered as the pth power of the exponential density (2. CLAIM SIZE DISTRIBUTIONS 7 The mean EX is p/b and the variance Var X is p/b2. p) = J tPletdt. X2..). Ruin probabilities for the general case has been studied. by Grandell & Segerdahl [175] and Thorin [369]. 0. In particular.3 (THE HYPEREXPONENTIAL DISTRIBUTION) This is defined as a finite mixture of exponential distributions.. u Example 2 . p). The exact form of the tail B(x) is given by the incomplete Gamma function r(x.c.. and exponential with rate d..ate (b2 ): L• i=o In the present text. is > 1. In particular. if p is integer and X has the gamma distribution p.2.. we develop computationally tractable results mainly for the Erlang case (p = 1. > 1 for p < 1 and = 1 for p = 1 (the exponential case).1) (or the 1/pth root if p < 1). p) °° where r (x. . i = 1. u .v.i. x] so that B(x) = r` e. JP 1 B(x) r(p ) XP ie ax In the sense of the theory of infinitely divisible distributions. where X1. B(x) = r(p) Asymptotically. This special case is referred to as the Erlang distribution with p stages. . the Gamma density (2. p. 2. are i.v. 0 < ai < 1. An appealing feature is its simple connection to the Poisson process: B(x) = P(Xi + • • • + XP > x) is the probability of at most p ..
e. B(x) = aeTxe where t = Te and e = (1 .6 (DISTRIBUTIONS WITH BOUNDED SUPPORT) This example (i. (or. equivalently. We give a more comprehensive treatment in VIII. We give some theory for matrixu exponential distribution in VIII.7) q1 b(x) = cjxieWWx + djxi cos(ajx)ea'x + > ejxi sin(bjx)e`ix .8 CHAPTER I. q2 q3 (2. 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 . Example 2 .7) are possibly complexvalued but the parameters in (2. resp.g.. which is slightly smaller but more amenable to probabilistic reasoning. 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.6. T) is called the representation. INTRODUCTION Example 2 . The parameters of a phasetype distribution is the set E of transient states.8) j=1 j=1 j=1 where the parameters in (2. a rational Laplace transform) if B[s] _ p(s)/q(s) with p(s) and q(s) being polynomials of finite degree. B(x) > 0 for x < xo) is of course a trivial instance of a lighttailed distribution.f. 1)' is the column vector with 1 at all entries. of which one is absorbing and the rest transient. the Erlang and the hyperexponential distributions..d. the restriction T of the intensity matrix of the Markov process to E and the row vector a = (ai)iEE of initial probabilities.(2.1 and defer further details to u Chapter VIII. The density and c. a. Equivalent characterizations are that the density b(x) has one of the forms q b(x) j=1 = cjxienbx.4 (PHASETYPE DISTRIBUTIONS) A phasetype distribution is the distribution of the absorption time in a Markov process with finitely many states. Important special cases are the exponential.5 (DISTRIBUTIONS WITH RATIONAL TRANSFORMS) A distribution B has a rational m.6. but the current trend in applied probability is to restrict attention to the class of phasetype distributions. See XI. it is notable from a practical point of view because of reinsurance: if excessofloss reinsurance has been arranged with retention level xo.8) are realvalued. are b(x) = aeTxt. This class of distributions is popular in older literature on both risk theory and queues. The couple (a. T) or sometimes the triple (E. there exists a xo < oo such that B(x) = 0 for x > xo.f.xo)+ is covered by the reinsurer). Example 2 . . However.
9) which is heavytailed when 0 < r < I. the exponential distribution representing the simplest example since here b(x) is constant.8 (THE LOGNORMAL DISTRIBUTION) The lognormal distribution with parameters a2. the mean u is eµ+a /2 and the second moment is e2µ+2o2.9 (THE PARETO DISTRIBUTION) Here the essence is that the tail B(x) decreases like a power of x. However. Writing c = d/r. u Example 2 . CLAIM SIZE DISTRIBUTIONS 9 2b Heavytailed distributions Example 2. and then b(x) = 0. b(x) _ A(1 + (x a The pth moment is finite if and only if p < a .2. It follows that the density is 't (1ogX .13) u .11) ex log logx 2r p 1 1 2 ( a ) f 1 (lox_P)2} (2. (2.u l b(x) = d dx or J ax lor 1 exp Asymptotically. (2.p a 1 (2. (2. p is defined as the distribution of ev where V . b(x) = crx''le`xr. the tail is B (x ) 2 x.7 (THE WEIBULL DISTRIBUTION) This distribution originates from reliability theory. All moments are finite. There are various variants of the definition around.1. one being B(x) (1 + X)b(x) (1 + x)a+1' x > 0. or equivalently as the distribution of a°U+µ where U . we obtain the Weibull distribution B(x) = eCx'. In particular. Here failure rates b(x) = b(x)/B(x) play an important role.N(p.pl = 1 W (logx .10) The loinormal distribution has moments of all orders.1). 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). a2).12) Sometimes also a location parameter a > 0 and a scale parameter A > 0 is allowed.N(0. x < a. Example 2 . a)/A)a+1' x > a.
16) 11 Example 2. A = 1 and X is standard exponential.10 CHAPTER I.12) (here L (x) * 1) and ( 2. { s () 1s+3s29s3log(1+2s I p=3. x 4 oo (any L having a limit in (0.L( x ). INTRODUCTION Example 2. The simplest examples correspond to p small and integervalued. i. u Example 2 .2). (2. x + 00. another standard example is (log x)'). the loggamma distribution is a Pareto distribution. Thus. (2.11 (PARETO MIXTURES OF EXPONENTIALS) This class was introduced by Abate. u . examples of distributions with regularly varying tails are the Pareto distribution (2.12 (DISTRIBUTIONS WITH REGULARLY VARYING TAILS) The tail B(x) of a distribution B is said to be regularly varying with exponent a if B(x) . Choudhury & Whitt [1] as the class of distributions of r.(1 + Zx + $ p = 3. 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). in particular. The density is 8p(log x)pi b(x) .10 (THE LOGGAMMA DISTRIBUTION) The loggamma distribution with parameters p.'s of the form YX. the loggamma distribution (with exponent 5) and a Pareto mixture of exponentials.15) x2 + 16x3 ) a3x/2) 3 (1 . oo) is slowly varying .17) where L (x) is slowly varying.(1 + 2x + 2x2)e2x) p = 2 (2. the density is { 3 (1 . 6 is defined as the distribution of et' where V has the gamma density (2.14) The pth moment is finite if p < 5 and infinite if p > 5.v. In general. in particular.x6+lr(p) (2. B(x) = O(xP). satisfies L(xt)/L(x) 4 1.13).1)/p. where Y is Pareto distributed with a = (p .e. For p = 1.
but the model also admits a natural interpretation : a large portfolio of insurance holders . When studying ruin probabilities. Also.1) that any distribution with a regularly varying tail is subexponential. However. We return to a closer study in IX.. and so is the Weibull distribution with 0 < r < 1.3.4) or even to completely different applied probability areas like extreme value theory: if we are using a Gaussian process to predict extreme value behaviour. 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. the claim size distribution represents of course only one aspect (though a major one). though the proof of this is nontrivial. The reason is in part mathematical since this model is the easiest to analyze. From a practical point of view. Namely. but can never be sure whether this is also so for atypical levels for which far less detailed statistical information is available. one may argue that this difficulty is not resticted to ruin probability theory alone..13 (THE SUBEXPONENTIAL CLASS OF DISTRIBUTIONS) We say that a distribution B is subexponential if xroo lim B `2^ = 2. it will be seen that we obtain completely different results depending on whether the claim size distribution is exponentially bounded or heavytailed. Thus. 3 The arrival process For the purpose of modeling a risk process . By far the most prominent case is the compound Poisson (CramerLundberg) model where {Nt} is Poisson and independent of the claim sizes U1. THE ARRIVAL PROCESS 11 Example 2. (2.. Similar discussion applies to the distribution of the accumulated claims (XI. 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. which each have a ( timehomogeneous) small rate of experiencing a . U2.1. this phenomenon represents one of the true controversies of the area. the knowledge of the claim size distribution will typically be based upon statistical data. 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. for example the lognormal distribution is subexponential (but not regularly varying). the subexponential class of distributions provide a convenient framework for studying large classes of heavyu tailed distributions.18) B(x) It can be proved (see IX..
This model . Nevertheless . IV (and. see 11. 5. 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. we study this case in VI . radioactive decay (a huge number of atoms each splitting with a tiny rate ) and many other applications. The point of view we take here is Markov dependent random walks in continuous time (Markov additive processes ).6. 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 . however.. such that 8(t) = . too general and one neeed to specialize to more concrete assumptions . A more appealing way to allow for inhomogeneity is by means of an intensity . in particular to allow for certain inhomogeneities. T2. To the author 's knowledge . I. The compound Poisson model is studied in detail in Chapters III. e. In others. found the Poisson distribution to be inadequate and suggested various other univariate distributions as alternatives .g. The one we focus on (Chapter VI) is a Markovian environment : the environmental conditions are described by a finite Markov process {Jt }too. An obvious example is 3(t) depending on the time of the year (the season). to be studied in Chapter V. epidemics in life insurance etc. not many detailed studies of the goodnessoffit of the Poisson model in insurance are available . Cox processes are. This applies also to the case where the claim size distribution depends on the time of the year or . has some mathematically appealing random walk features . it is more questionable whether it provides a model with a similar intuitive content as the Poisson model.i..12 CHAPTER I. in Chapter VII). when Jt = i. Some of them have concentrated on the marginal distribution of NT (say T = one year ). and also that the ruin problem may be hard to analyze .(3. Mathematically. are i. gives rise to an arrival process which is very close to a Poisson process. the periodic and the Markov modulated models also have attractive features .8 (t) is a periodic function of t. This model can be intuitively understood in some simple cases like { Jt} describing weather conditions in car insurance .3(t) fluctuating over time. However . Another one is Cox processes. which facilitate the analysis. with a common term {Nt} is a Markovmodulated Poisson process . but with a general not necessarily exponential distribution ). its basic feature is to allow more variation (bursty arrivals ) than inherent in the simple Poisson process. with the extension to premiums depending on the reserve. in just the same way as the Poisson process arises in telephone traffic (a large number of subscribers each calling with a small rate). the first extension to be studied in detail was {Nt } to be renewal (the interarrival times T1 . getting away from the simple Poisson process seems a crucial step in making the model more realistic. In order to prove reasonably substantial and interesting results . where {/3 (t)}too is an arbitrary stochastic process . the negative binomial distribution.. so that .e. Historically..d. INTRODUCTION claim .
it is desirable to have a set of formulas like (4. point processes and so on. A SUMMARY OF MAIN RESULTS AND METHODS 13 the environment (VI.4. interacting particle systems.1) where V is the limit in distribution of Vt as t + oo. however.0 (u.v. reliability. queueing theory.T) = P(VT > u). that quite often the emphasis is on computing expected values like EV. (4. Some of these have a certain resemblance in flavour and methodology. The study of the steady state is by far the most dominant topic of queueing and storage theory. A stochastic process {Vt } is said to be in the steady state if it is strictly stationary (in the Markov case. 4 A summary of main results and methods 4a Duality with other applied probability models Risk theory may be viewed as one of many applied probability areas. and here (4. it is a recurrent theme of this book to stress this connection which is often neglected in the specialized literature on risk theory. 0(u) = P(V > u). Mathematically. It should be noted.'s like V is available. methods or modeling ideas developed in one area often has relevance for the other one as well. ruin probabilities for risk processes with an input process which is renewal.1). The M/G/1 workload process { Vt } may also be seen as one of the simplest storage models. More generally.1) permitting to translate freely between risk theory and the queueing/storage setting.6) . A general release rule p(x) means that {Vt} decreases according to the differential equation V = p(V) in between jumps. with Poisson arrivals and constant release rule p(x) = 1. and the limit t 4 oo is the steadystate limit. genetics models. stochastic differential equations. dam/storage processes. the classical result is that the ruin probabilities for the compound Poisson model are related to the workload (virtual waiting time) process {Vt}too of an initially empty M/G/1 queue by means of . others being branching processes. Similarly. Markovmodulated or periodic can be related to queues with similar characteristics.1) holds as well provided the risk process has a premium rule depending on the reserve. this amounts to Vo having the stationary distribution of {Vt}). In fact. this gives only f0 O°i (u)du which is of limited . Thus. In the setting of (4. and a lot of information on steadystate r. time series and Gaussian processes. stochastic geometry. others are quite different. The ones which appear most related to risk theory are queueing theory and dam/storage processes. and which seems well motivated from a practical point of view as well. extreme value theory. R = p(R) in between jumps.
. as is typically the case. see Corollary VII.1 .3.. have to some extent a different flavour. e . which gives a sample path version of (4. Example VIII.'s like the environmental process {Jt} in a Markovmodulated setting. The qualifier 'with just a few phases ' refers to the fact that the diagonalization has to be carried out numerically in higher dimensions. • The compound Poisson model with premium rate p(x) depending on the reserve and exponential claim size distribution B. • The compound Poisson model with constant premium rate p = 1 and B being phasetype with a just few phases . Vi(u. Thus . which can be expanded into a sum of exponential terms by diagonalization (see.2).1). though overlapping. Here ?P(u) is explicit provided that . 3. • The compound Poisson model with a claim size distribution degenerate at one point. p = 0/8 and y = 8 . A prototype of the duality results in this book is Theorem 11. The fact that Theorem H. see Boxma & Cohen [74] and Abate & Whitt [3]. the functions w x f d 1 exdx () .T). Here Vi(u) is given in terms of a matrixexponential function ( Corollary VIII. B(x) = ebx.3. INTRODUCTION intrinsic interest . the two areas.1 is a sample path relation should be stressed : in this way the approach also applies to models having supplementary r. Here O(u) = peryu where 3 is the arrival intensity. 4b Exact solutions Of course . the ideal is to be able to come up with closed form solutions for the ruin probabilities 0(u). 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.14 CHAPTER I.3.g. 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 ). see Corollary III. Similarly.6.1. • The compound Poisson model with some rather special heavytailed claim size distributions.3.v.p(y) y^ Jo p(x) can be written in closed form. The infinite horizon (steady state ) case is covered by letting T oo.8. 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.
However.2) is the natural scale. relevant references are Grubel [179]. but are somewhat out of the mainstream of the area . where Furrer [150] recently computed ii(u) as an infinite series involving the Mittag.1) are so complicated that they should rather be viewed as basis for numerical methods than as closedform solutions. the second best alternative is a numerical procedure which allows to calculate the exact values of the ruin probabilities.b(u)du . Embrechts. [s. • An astable Levy process with drift . f {eXp U LX 2. A notable fact ( see again XI. see VIII. 191). Here are some of the main approaches: Laplace transform inversion Often. esuTb( u. T) du dT 0 TO 00 in closed form than the ruin probabilities z/'(u).Lef$er function. Grubel & Pitts [132] and Grubel & Hermesmeier [180] (see also the Bibliographical Notes in [307] p. 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 {.7. 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 . Also Brownian models or certain skip free random walks lead to explicit solutions (see XI . T) can then be calculated numerically by some method for transform inversion.f f 2µ(y)/a2(y) dy} dx  (4. a2 (x): Ip (u) = where S(u) = f °O exp {.u(y)/a2(y) dy} 4c Numerical methods Next to a closedform solution. . T) themselves. Ab(u).ff 2µ(y)/a2(y) dy} dx . the formulas ( IV. T). 1).4. it is easier to find the Laplace transforms = f e8 . Given this can be done. 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. (u.1) is the explicit form of the ruin probability when {Rt} is a diffusion with infinitesimal drift and variance µ(x). For the finite horizon ruin probability 0(u. O(u. We don't discuss Laplace transform inversion much. Abate & Whitt [2].
For the compound Poisson model with p = 1 and claim size distribution B with moment generating function (m.and integral equations The idea is here to express 'O(u) or '(u.16 CHAPTER L INTRODUCTION Matrixanalytic methods This approach is relevant when the claim size distribution is of phasetype (or matrixexponential). dt] most often leads to equations involving both differential and integral terms. One example where this is feasible is the renewal equation for tl'(u) (Corollary III.3. An example where this idea can be carried through by means of a suitable choice of supplementary variables is the case of statedependent premium p(x) and phasetype claims.) B[s].p)/(13B'[ry] .3) where C = (1 . T) as the solution to a differential. and carry out the solution by some standard numerical method. which can equivalently be written as f3 [7] = 1 +13 . u * oo. and in particular the naive idea of conditioning upon process behaviour in [0.7. (4.1) and y > 0 is the solution of the Lundberg equation (4.g.f.or integral equation. U is explicit in terms of the model parameters. it states that i/i(u) . 4d Approximations The CramdrLundberg approximation This is one of the most celebrated result of risk theory (and probability theory as a whole). 0(u) is then given in terms of a matrixexponential function euu (here U is some suitable matrix) which can be computed by diagonalization. as the solution of linear differential equations or by some series expansion (not necessarily the straightforward Eo U'u/n! one!). Differential. and in quite a few cases (Chapter VIII).Ce"u. whereas for the renewal arrival model and the Markovian environment model U has to be calculated numerically.3) in the compound Poisson model which is an integral equation of Volterra type. In the compound Poisson model with p = 1. most often it is more difficult to come up with reasonably simple equations than one may believe at a first sight.4) 00['Y]1)'Y = 0. . However. see VIII. 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.
For example. in such cases the evaluation of C is more cumbersome. T) for large u are available in most of the models we discuss. other approaches are thus required. the exact solution is as easy to compute as the CramerLundberg approximation at least in the first two of these three models. often for all u > 0 and not just for large u. u > oo. the claim size distribution should have an exponentially decreasing tail B(x). in some cases the results are even more complete than for light tails. Large claims approximations In order for the CramerLundberg approximation to be valid.2.6) are by far the best one can do in terms of finite horizon ruin probabilities '(u. a Markovian environment or periodically varying parameters. See Chapter IX. some further possibilities are surveyed in 111 . In the case of heavytailed distributions. In fact. It has generalizations to the models with renewal arrivals. but typically not very precise in their first naive implementation. 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. Approximations for O(u) as well as for 1(u. T). 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.7 and IV. The CramerLundberg approximation is renowned not only for its mathematical beauty but also for being very precise. This list of approximations does by no means exhaust the topic. J B dx. . However. for the compound Poisson model ^(u) p pu In fact . (4. 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. corrected diffusion approximations (see IV. In particular.4. incorporating correction terms may change the picture dramatically.6) 4e Bounds and inequalities The outstanding result in the area is Lundberg's inequality (u) < e"lu. However. Diffusion approximations are easy to calculate.
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. and to plot the empirical mean residual life 1 N . it has the advantage of not involving approximations and also. How do we produce a confidence interval? And. T]. UNT are i. .i. empirical evidence shows that the general principle holds in a broad variety of settings.. For example. We return to various extensions and sharpenings of Lundberg's inequality (finite horizon versions.U(k)) i =k+ i . they have however to be estimated from data. which is a standard statistical problem since the claim sizes Ui. one expects a model with a deterministic claim size distribution B. .. . as a general rule. . UNT may be viewed as an interpolation in or smoothing of the histogram). . This procedure in itself is fairly straightforward.) at various places and in various settings. of being somewhat easier to generalize beyond the compound Poisson setting. For example. to have smaller ruin probabilities than when B is nondegenerate with the same mean m. When comparing different risk models. .3). However.8. This is proved for the compound Poisson model in 111. In practice. say degenerate at m.. The standard suggestion is to observe that the mean residual life E[U .x U > x] = B(x) f '(yx)B(dx) typically has a finite limit (possibly 0) in the lighttailed case and goes to oo in the heavytailed case.g. given NT. fitting a parametric model to U1. INTRODUCTION Compared to the CramerLundberg approximation (4. in the compound Poisson model. though not too many precise mathematical results have been obtained. the difficulty comes in when drawing inference about the ruin probabilities. e. more importantly. can we trust the confidence intervals for the large values of u which are of interest? In the present author's opinion. obtained say by observing the risk process in [0. it is a general principle that adding random variation to a model increases the risk. However. one may question whether it is possible to distinguish between claim size distributions which are heavytailed or have an exponentially decaying tail.k (U(`) . lower bounds etc.18 CHAPTER I. 4f Statistical methods Any of the approaches and results above assume that the parameters of the model are completely known. it splits up into the estimation of the Poisson intensity (the estimator is /l3 = NT/T) and of the parameter(s) of the claim size distribution.d.
. to observe whether one or the other limiting behaviour is apparent in the tail. Truncation to a finite horizon has been used. this is a straightforward way to estimate finite horizon ruin probabilities. . respectively. reference [14].3 (or just VI.(5.2.3) and Section VI. A main problem is that ruin is typically a rare event (i. the more typical situation is to perform a Monte Carlo experiment to estimate probabilities (or expectations or distributions) which are not analytically available. UN..v. formula VI. The problem is entirely analogous to estimating steadystate characteristics by simulation in queueing/storage theory. However. and also discuss how to develop methods which are efficient in terms of producing a small variance for a fixed simulation budget. in this book referred to as [APQ]). See further Embrechts. Simulation may be used just to get some vague insight in the process under study: simulate one or several sample paths. claims U1. (or a functional of the expectation of a set of r. We look at a variety of such methods in Chapter X. The infinite horizon case presents a difficulty. but is not very satisfying. good methods exist in a number of models and are based upon representing the ruin probability zb(u) as expected value of a r. having small probability) and that therefore naive simulation is expensive or even infeasible in terms of computer time. in all other chapters than VI where we just write ... The chapter number is specified only when it is not the current one. formula (5. Klnppelberg & Mikosch [134].e.2. because it appears to require an infinitely long simulation.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. 4g Simulation The development of modern computers have made simulation a popular experimental tool in all branches of applied probability and statistics.. . and look at them to see whether they exhibit the expected behaviour or some surprises come up.5.i. where U(1) < . Still.3) or Section 3 of Chapter VI are referred to as Proposition VI.d. For example. Thus Proposition 4. CONVENTIONS 19 as function of U(k). and in fact methods from that area can often be used in risk theory as well . < U(N) are the order statistics based upon N i.3). and of course the method is relevant in risk theory as well.4.
In particular.g.d. i. B(x) = P(X < x) = fx.g. n! 27r nn+1/2en.h. say a heuristic approxi1 + h + h2/2. If.3) or Section 3.r. w. i. formula (5. then for 1s < 5).i. . n i oo. infinitely often l. Abbreviations c.s. cumulant generating function.B(x) = P(X > x) of B. if B(x) .o. b[s] is defined always if Rs < 0 and sometimes in a larger strip (for example.g. oo).g.4.f. mation.2.20 CHAPTER L INTRODUCTION Proposition 4. (A.s. The Laplace transform is b[s]. A different type of asymptotics: less precise.v. independent identically distributed i. right hand side (of equation) r. cumulative distribution function P(X < x) c. with probability Mathematical notation P probability.The same symbol B is used for a probability measure B(dx) = P(X E dx) and its c. E.h. with respect to w.t.g. log E[s] where b[s] is the m. B is concentrated on [0.f. random variable s. moment generating function.d. .c. B[s] the m. B(dy). EX2/(EX)2.Used in asymptotic relations to indicate that the ratio between two expressions is 1 in the limit. h + 0.f. left hand side (of equation) m.e. or a more precise one like eh .f. r. squared coefficient of variation. for a probability distribution IIGII = 1.d.p. IIGII the total mass (variation ) of a (signed ) measure G .ceax.f. see under b[s] below. and for a defective probability distribution IIGII < 1.29) refer to the Appendix. References like Proposition A. (moment generating function) fm e82B(dx) of the distribution B. E expectation.f. B(x) the tail 1 .v. as for typical claim size distributions.
of numbers. . (the dimension is usually clear from the context and left unspecified in the notation). i. Thus. Usually. I(A) the indicator function of the event A. only have finitely many jumps in each finite interval. E[X. 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. oo) the space of Rvalued functions which are rightcontionuous and have left limits. all stochastic processes considered in this book are assumed to have sample paths in this space. intensity interpretation.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). xa.e. R(s) the real part of a complex number s. 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. Notation like f3i and 3(t) in Chapter VI has a similar . Usually.e. i.A] means E[XI(A)]. row vectors have lowercase Greek letters like a. 0 marks the end of a proof. and column vectors have lowercase Roman letters like t. Unless otherwise stated. CONVENTIONS {6B the mean EX = f xB(dx) of B ABA' the nth moment EXn = f x"B(dx) of B. though slightly more complicated. . the ith entry is 1 and all other 0.5. A. a2) the normal distribution with mean p and variance oa2. 21 D [0. often the term 'cadlag' (continues a droite avec limites a gauche) is used for the Dproperty. Xt_ the left limit limstt X8f i. 7r. (xi)diag denotes the diagonal matrix with the xi on the diagonal (xi)row denotes the row vector with the xi as components (xi). a. the processes we consider are piecewise continuous.. . In the Frenchinspired literature. the value just before t. matrices have uppercase Roman or Greek letters like T. N(it. an example or a remark. F o r a given set x1. Matrices and vectors are denoted by bold letters.e.. the ith unit row vector is e'i.
or quantities with a similar time average interpretation.g. I.22 CHAPTER L INTRODUCTION B the claim size distribution. .5. interpretation.1. e. J the rate parameter of B for the exponential case B(x) = eby. p the net amount /3pB of claims per unit time. I. cf. ry The adjustment coefficient. cf. VI.5. 111.1. though slightly more complicated. Notation like BE and B(t) in Chapter VI has a similar. FL. EL the probability measure and its corresponding expectation corresponding to the exponential change of measure given by Lundberg conjugation. 'q the safety loading . cf.
used in Chapter VI on risk processes in a Markovian (or periodic) environment. Due to the generality of the theory. More precisely. 5) are. Sections 4. When encountered for the first time in connection with the compound Poisson model in Chapter III. in most cases via likelihood ratio arguments. Sections 4. somewhat more advanced than in the rest of the book. The topic is. the relevance for the mainstream of exposition is the following: The martingale approach in Section 1 is essentially only used here. however. The general theory is. The reader should therefore observe that it is possible to skip most of the chapter. a parallel selfcontained treatment is given of the facts needed there. fundamental ( at least in the author' s opinion) and the probability involved is rather simple and intuitive. however. in particular at a first reading of the book. in part. strictly speaking. The likelihood ratio approach in Section 2 is basic for most of the models under study. not crucial for the rest of the book. however. The duality results in Section 3 (and.Chapter II Some general tools and results The present chapter collects and surveys some topics which repeatedly show up in the study of ruin probabilities. the level of the exposition is. 23 . 5 on random walks and Markov additive processes can be skipped until reading Chapter VI on the Markovian environment model. All results are proved elsewhere .
T(u) < oo] + 0 = eryuE [e7Vu). however. As usual. We get 1 = Ee7So = E e'Y S(. Proposition 1. SOME GENERAL TOOLS AND RESULTS The ladder height formula in Theorem 6.5 can be skipped. Example 1 . T(u) > T] .u denote the overshoot. T) = P(T(u) < T). the time to ruin r(u) is inf It > 0 : St > u}. {e'YS° }t>0 is a martingale. f1 . Let e(u) = ST(u) . claim size distribution B and p = . StUit.T(u) < cc] = e7uE {e7f(u) I T(u) < cc] z/. and in the limit (1. 1 Martingales We consider the claim surplus process {St} of a general risk process.2) takes the form 1 = E [e'ys().1 is basic for the study of the compound Poisson model in Chapter III. Thus N.)AT = E [e7ST(°). using r(u) A T invokes no problems because r(u) A T is bounded by T).24 CHAPTER II.QµB < 1. V) (u. (1.1 Assume that (a) for some ry > 0.(u). and the ruin probabilities are ip(u) = P (T(u) < oo). The more general Theorem 6. (b) St a$ oo on {T(u) = oo}.0. 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. T(u) < T] + E [eryST . Our first result is a representation formula for O(u) obtained by using the martingale optional stopping theorem ..2 Consider the compound Poisson model with Poisson arrival rate . the second term converges to 0 by (b) and dominated convergence (e7ST < eryu on {r(u) > T}).2) As T > oo.
6a for details) that typically a solution to the Lundberg equation K(y) = 0 exists. the ruin probability is O(u) = pe.Q(B[a] . Under the conditions of Proposi Proof Just note that C(u) > 0. MARTINGALES 25 where {Nt} is a Poisson process with rate .= e"(') where K(a) = ./3.a.x.2 and drift p > 0. and thus Ee'rs° = 1. Now at the time r(u) of ruin {St} upcrosses level u by making a jump .1 are satisfied. From this it is immediately seen that the solution to the Lundberg equation ic(y) = 0 is y = 2p/a2. Eeas° = e"(°) where n(a) = a2a2/2 .5 For the compound Poisson model with B exponential. and thus Ee7s° = 1.1) .3 Assume that {Rt} is Brownian motion with variance constant o.4 (LUNDBERG ' S INEQUALITY ) tion 1 .f. Thus.1. the conditions of Proposition 1. 1.(„)_ = x is that of a claim size U given U > u . Thus.2(c)). with common distribution B (and independent of {Nt}). and (b) follows from p < 1 and the law of large numbers (see Proposition III. Since {St} has stationary independent increments. Proof Since c(a) = /3 (B[a] . u Corollary 1. and p =. the martingale property now follows just as in Example 1.u + x is again just exponential with rate S.Q.. By standard formulas for the m. the conditional distribution of the overshoot e(u) = U . Then {St } is Brownian motion with variance constant o2 and drift p < 0.g.2. From this it is readily seen (see III. are i.1.1 is satisfied.i.a = a .ap. A simple calculation (see Proposition III. it follows that E [e7st+v I J] = e"rstE [e7(st+vSt) I Ft] = e7StEe"rs° = elst where .3/6 < 1.Q and the U.r" where y = S .d.Ft = a(S" : v < t). Since {St} has stationary independent increments. of the normal distribution. and thus by the memoryless property of the exponential distribution . condition (a) of Proposition 1.1) .1. Thus 00 E [e'rt (") I T(u) < oo] = I e5e  dx = f 5edx .a it is immediately seen that y = S . The available information on this jump is that the distribution given r(u) = t and S. Example 1 . B(x) _ edx. u Corollary 1. O(u ) < e7".1) shows that Eels.
P are then singular (concentrated on two disjoint measurable sets).6 N S = { lim Nt I t +00 t gJ are both in F. hence so is Nt = limfyo N2`i. which we equip with the natural filtration {. A E Ft. and by the law of large numbers for the Poisson process . we look for a process {Lt} (the likelihood ratio process) such that P(A) = E[Lt. P on (DE. as shown by the following example this setup is too restrictive: typically'. F(S) = P(S) = 1. A somewhat similar u argument gives singularity when B $ B. [172]. and in analogy with the theory of measures on finite dimensional spaces one could study conditions for the RadonNikodym derivative dP/dP to exist. The number Nt F) of jumps > e before time t is a (measurable) r. .26 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Corollary 1. More recent references are Dassios & Embrechts [98]. (2. on (DE..v. oo).F). the parameters of the two processes can be reconstructed from a single infinite path. F). Theorem 2. The interesting concept is therefore to look for absolute continuity only on finite time intervals (possibly random. Delbaen & Haezendonck [103] and Schmidli [320]. Grandell & Schmidli [131]. then z/'(u) = e7" where 'y = 21A/a2. and F. 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. and is further exploited in his book [157]. Example 2 . Thus the sets S = I tlim +oot t =. Two such processes may be represented by probability measures F. B. then S and S are disjoint . cf.1) 'though not always: it is not difficult to construct a counterexample say in terms of transient Markov processes. u Notes and references The first use of martingales in risk theory is due to Gerber [156].1 Let F. But if a $ ^ . Embrechts. 0 and claim size distributions B. P correspond to the claim surplus process of two compound Poisson risk processes with Poisson rates /3. A].3 below). Proof Just note that ^(u) = 0 by continuity of Brownian motion.e.6 If {Rt} is Brownian motion with variance constant a2 and drift p > 0.Ft}too and the Borel afield F. Grandell [171]. However. I.
then {Lt} is a nonnegative martingale w. ELt = 1 follows by taking A = DE in (2. G C {T < oo}.F such that (2. Conversely. G l ] = E [_I(G)E[LTIFT] ] = E { _I(G)Lr ] = P(G). (ii) Conversely. A] = E[Lt. Then P(A) = E[Lt. Then Ft (A) = E[Lt. then there exists a unique probability measure Pon .A] = EE[LtI(A)IF8] = EI(A)E[LtIFB] = EI(A)L8 = PS(A). ({.3 Let {Lt}.2) Proof Assume first G C {T < T} for some fixed deterministic T < oo.2 Let {Ft}t>o be the natural filtration on DE.1) holds. if for some probability measure P and some {. Hence the family {Pt} is t>o consistent and hence extendable to a probability measure F on (DE. (i) If {Lt}t> o is a nonnegative martingale w. P) such that LLt = 1. A].1) holds. A E F. F) such that ELt = 1.t. that the restriction of P to (DE. under the assumptions of (ii) we have for A E rg and s < t that A E Ft as well and hence E[L8. The truth of this for all A E Y.2(i). the restriction of P to (DE.F8] = L8 and the martingale property. define P by Pt (A) = E[Lt. we have E [ LTIFT]1 = LT on {T < T}. Then Lt > 0 and ELt = 1 ensure that Pt is a probability measure on (DE. A E F8. Lt < 0] can only be nonnegative if P(A) = 0. . 1 J (2. This proves (i).r.Pt)) The following result gives the connection to martingales. Proposition 2. then { 1 P(G) = EG .Pt}adapted process {Lt}t>o (2.Tt) is absolutely continuous w. A]. A E Ft .Y) such that P(A) = Pt(A). 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 . using the martingale property in the fourth step. _. Lets < t. implies that E[LtI.F). Ft). Hence E [_ .2.e. Finally. F the Borel o•field and P a given probability measure on (DE.. . ({Ft} . If r is a stopping time and G E PT. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 27 (i.r.Ft}. G ] = E [LT .t. By the martingale property. P be as in Proposition 2.t. Proof Under the assumptions of (i).1) and nonnegativity by letting A = {Lt < 0}.r.
applying (2. t.. (2.28 CHAPTER II. SOME GENERAL TOOLS AND RESULTS In the general case .3) to G of{r < T} we get 1111 F(Gn {r <T}) = E[ 1 .1) is that it seems in general easier to deal with the (unconditional) expectation E[eryVu). (2.r.1).Ft} .h. St). say. u From Theorem 2.2) by noting that 1 = ersr(„) = e1'ue7Ou).3 we obtain a likelihood ratio representation of the ruin probability V) (u) parallel to the martingale representation (1..t.O(u) = eryuE[e'YC(u).1) in Proposition 1. {St} = {u . For the definition. where {Rt} is the risk reserve process. each F.r. is nonnegative and has Ey Lt = 1 for all x. A change of measure is performed by finding a process {Lt} which is a martingale w. in continuous time (the discrete time case is parallel but slightly simpler). Consider a (timehomogeneous) Markov process {Xt} with state space E. The problem is thus to investigate which characteristics of {Xt} and {Lt} ensure a given set of properties of the changed probability measure. Xt = (Jt.s. The crucial step is to obtain information on the process evolving according to F. of (2. we have F(G) = V )(u). one would typically have Xt = Rt. we need the concept of a multiplicative functional. 1 Since everything is non negative. Xt = St.Rt} the claim surplus process and {Jt} a process of supplementary variables possibly needed to make the process Markovian. 5) for processes with some randomwalklike structure. r(u) < oo] occuring there than with the (conditional) expectation E[e'r{(u ) Jr(u) < oo] in (1. and this problem will now be studied. T(u) < oo]. we assume for simplicity that {Xt} has Dpaths.1: Corollary 2. the natural filtration {.t.2) follows by monotone convergence. first in the Markov case and next (Sections 4.4 Under condition (a) of Proposition 1. Now just rewrite the r. In the context of ruin probabilities. . Rt) or Xt = (Jt.1. is Markov w.4) Proof Letting G = {r(u) < oo}. and letting T * oo. First we ask when the Markov property is preserved.Gn {r <T} .4) compared to (1. Lr(u) 11 The advantage of (2. To this end. both sides are increasing in T.
since Ext [L8Y8] = E[(Y8 o et)(L8 o 8t)I.6) implies (2. t. Proof Since both sides of (2. let {Lt} be a nonnegative martingale with Ex Lt = 1 for all x.(Xtitl) with all t(i) < t + s. nonnegative and Lt+8 = Lt•(Lso9t) (2. for all x.'s of the form fl' f. or.Ftmeasurable.2.7).Pt+8measurable r. o 9tI.7) for any Ftmeasurable Zt and any . oo).5) is equivalent to Ex[Lt+8Vt+8] = E8[Lt • (L8 o 91)Vt+8] for any . t] * [0. 0 .v.5 Let {Xt} be Markov w.[Y. The converse follows since the class of r. Lt has the form Lt = 'Pt ({x }0<u<t) for some mapping cot : DE[O.Ft] = Ex.Ft+8measurable. the natural filtration {Ft} on DE. since Zt • (Y8 o Ot ) is . the Markov property can be written E.(A) = Ex [Lt.Ft].6) for any .5) Pxa.. Then the family {Px}xEE defines a timehomogeneous Markov process if and only if {Lt} is a multiplicative functional. Zt. Indeed. Vt+e. By definition of Px. o 9t = V. Y8. ({Xt+u} 0<u<8) Theorem 2. (2.r.5) are Tt+e measurable.v. and then L. which is the same as Ex[Zt(Y8 o Bt)] = E8[ZtEx. which in turn is the same as Ex[Lt+8Zt • (V8 o Bt)] = Ex[Lt • (L8 o 91)Z1 • (Y8 o et)] (2. this in turn means Ex[Lt+8Zt(V8 oet)] = Ex[LtZtExt[L8Y8]]. Ex[Lt+8Zt(Y8 o et)] = Ex[LtZt(Y8 o 0t)(L8 o Bt)].Ft }.v. (2.YB] for any Ftmeasurable r.T9measurable Y8. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 29 on DE and define {Lt} to be a multiplicative functional if {Lt} is adapted to {.7).t. The precise meaning of this is the following: being .Y8f t < s. (2. t and let Px be the probability measure given by t.'s of the form Zt • (Y8 o 0t) comprises all r. Similarly.s.v..F8measurable r. s.8) which is the same as (2.v. (2. A]. where Ot is the shift operator.
the arrival epochs are Qi. aN where or* = T UN_k+l. R = p(R)). then Remark 2. The formulation has applications to virtually all the risk models studied in this book.. it xEE suffices to assume that {Lt} is a multiplicative functional with Ex Lt = 1 for all x. we shall establish a general connection between ruin probabilities and certain stochastic processes which occurs for example as models for queueing and storage.. t. the premium rate is p(r) when the reserve is r (i. (using the Markov property in the second step) so that the martingale property is automatic. . {Vt} . < aN < T. t] = LtExt L8 = Lt.. 3 Duality with other applied probability models In this section. and just after time or* {Vt} makes an upwards jump of size UU = UN _k+l. (3.. CN. see Dynkin [128] and Kunita [239].. The storage process {Vt }o<t<T is essentially defined by time reversion.5 can be found in Kuchler & Sorensen [240].At where At = k: vk <t U. Indeed. u Notes and references The results of the present section are essentially known in a very general Markov process formulation. .6 For {u . The result is a sample path relation.30 CHAPTER H. and the time to ruin is 7(u) = inf {t > 0: Rt < 0}. 0 < vl < . .1) The initial condition is arbitrary. . The corresponding claim sizes are Ul. In between jumps.Ft] = LtE[L8 o 9t I. reflection at zero and initiar condition Vo = 0. Ro = u (say). More precisely .. with a proof somewhat different from the present one... .. Thus R = Ro + f p(R8) ds . SOME GENERAL TOOLS AND RESULTS to define a timehomogeneous Markov process. We work on a finite time interval [0. and thus for the moment no parametric assumptions (on say the structure of the arrival process) are needed.e. In between jumps. T] in the following setup: The risk process {Rt}o<t<T has arrivals at epochs or. UN.. A more elementary version along the lines of Theorem 2.. } E[Lt+B I.
. . :. {Vt} remains at 0 until the next arrival).3. Theorem 3._: 1} 0 011 =T01N ^N3 To 0 011 014 01N Figure 3. (3... Note that these definitions make {Rt} rightcontinuous (as standard) and {Vt} leftcontinuous.2) k: ok <t and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0. That is. 3.__.T) = inf Rt < 0 P (O<t<T P(r(u) < T) be the ruin probability.11 4.3) (u) Proof Let rt' denote the solution of R = p(R) subject to r0 = u..1 Define r(u) = inf It > 0: Rt < 0} (r(u) = oo if Rt > 0 for all t < T) and let ii(u._.e..____•_. (3.. instead of (3..1) we have Vt = At  f P(Vs)ds where A= U= AT . The sample path relation between these two processes is illustrated in Fig..1 The events {T(u) < T} and {VT > u} coincide.1. DUALITY WITH OTHER APPLIED PROBABILITY MODELS 31 decreases at rate p(r) when Vt = r (i. V = p(V)).. In particular.__.T) = P(VT > u).x.___ .AT_t. Then rt°) > rt°) for all t when u > v. V)(u.
and a general premium rule p(r) when the reserve is r. The arrival epochs correspond to rainfalls. Historically. say of water in a dam though other interpretations like the amount of goods stored are also possible. Nevertheless.and left continuity is immaterial because the probability of a Poisson arrival at any fixed time t is zero). Hence RQ„ > 0 for all n < N. Some further relevant more general references are Asmussen [21] and Asmussen & Sigman [51]. and in between rainfalls water is released at rate p(r) when Vt (the content) is r. Then the time reversibility of the Poisson process ensures that {At } and {At } have the same distribution (for finitedimensional distributions. Then Vo. Suppose next VT < u (this situation corresponds to the broken path of {Rt} in Fig. Hence if n satisfies VVN_n+1 = 0 (such a n exists. the connection between risk theory and other applied probability areas appears first to have been noted by Prabhu [293] in a queueing context. 3. u Notes and references Some main reference on storage processes are Harrison & Resnick [187] and Brockwell. the distinction between right. and then '0 (u) = P(V > u).1 and its proof is from Asmussen & Schock Petersen [50].3 and being i. We get: Corollary 3. Corollary 3. and since ruin can only occur at the times of claims.2 Consider the compound Poisson risk model with a general premium rule p(r). say V. we have RQ„ < 0 so that indeed r(u) < T. Theorem 3. we have r(u) > T.U1 = Rol. . Resnick & Tweedie [79].i.32 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Suppose first VT > u (this situation corresponds to the solid path of {Rt} in Fig. = r(VT) .2 from Harrison & Resnick [188]. Historically. and so on. with distribution B. The results can be viewed as special cases of Siegmund duality. 3. this represents a model for storage.1 with Ro = u = u2). see Siegmund [344]. if nothing else n = N). one may feel that the interaction between the different areas has been surprisingly limited even up to today.U1 > roil . If VaN > 0.1 with Ro = u = ul). we can repeat the argument and get VoN_1 > Ra2 and so on. Proof Let T ^ oo in (3. Then the storage process {Vt} has a proper limit in distribution. Thus we may think of {Vt} as having compound Poisson input and being defined for all t < oo. u A basic example is when {Rt} is the risk reserve process corresponding to claims arriving at Poisson rate . Then similarly VVN = r0.3). if and only if O(u) < 1 for all u.Ul < roil  Ul = RQ„ Va1V_1 < RQ2.d.T l .
WN = YN ... where the Yi are i . N min (Y1 + • • • + YNn) n=0.. 1} is often referred to as simple random walk or Bernoulli random walk). and is reset to 0 once the r.. the Lindley process Wo..1. Here F is a general probability distribution on R (the special case of F being concentrated on {1. W1.Yl min (YN .2 The following assertions are equivalent: (a) 0(u) = P(r(u) < oo) < 1 for all u > 0. has a proper limit W in distribution as n + oo.YNn+1) n=0.. ZN = .e....s. In particular.... .w. Z2 = YN_1 i .1.4... 0 Corollary 4. WN be the Lindley process generated by Z1 = YN. ...2) satisfies the same recursion as in (4. Xo = 0. is defined by assigning Wo some arbitrary value > 0 and letting Wn+1 = (Wn + Zn+1)+• (4.1..... generated by Z1. I. there is an analogue of Theorem 3.1) Thus {Wn}n=o... n=0.. evolves as a random walk with increments Z1i Z2.h. i. as long as the random walk only takes nonnegative values. {Wn}n=0... Theorem 4. For a given i. R valued sequence Z1. Z2...N From this the result immediately follows.Y1 according to Wo = 0... .. 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. if Wo = 0 then (Z1+•••+Zn) WN = Zl+•••+ZN....1.d. with common distribution F (say). For discrete time random walks .. of (4.1 in terms of Lindley processes ....N (4.. Z2 . W1.. Most often. . (c) The Lindley process {WN} generated by Zl = Y1. (b) 1/i(u) = P(•r(u) < oo) > 0 as u * oo. Let further N be fixed and let Wo. Proof By (4. Then the events {r(u) < N} and {WN > u} coincide. 0).1.i.. Z2 = Y2... .. .d..2). ..1)).. hits (oo... just verify that the r .. W2. can be viewed as the reflected version of the random walk with increments Z1.. = Xo + Y1 + • • • + Y.1.min n=0. Z2.1 Let r(u) = inf In: u + Y1 + • • • + Yn < 0}.2) (for a rigorous proof. . . . N min (Y1 + + Yn).
=o. ±2. SOME GENERAL TOOLS AND RESULTS (d) m = inf. equivalently. the condition 00 F(YI+•••+ Yn<0)<00 n=1 is known to be necessary and sufficient ((APQ] p.1......2.. .ooa. .. e. there is a more general version of Corollary 4.2 and Theorem 4..the result is a sample path relation as is Theorem 3.5 does not necessarily lead to a random walk: if. doubly u infinite (n'= 0..1 is actually not necessary .1.s . 176) but appears to be rather intractable..o. YN). a sufficient condition for (e) is that EY is welldefined and > 0. the Lindley processes in Corollary 4. .) sup n=0. ±1.34 CHAPTER II. either M = oo a.. Remark 4 . Thus the assertion of Theorem 4. W v m and P(W > u) = P (m > u) = 0(u)..N so that WN _P4 M = supra=0. (Z1 + • • • + Zn) = m and P(W > u) = P(M > u) = i (u ). The converse follows from general random walk theory since it is standard that lim sup (Y1 + • • + Yn) = oo when Y1 + • • • + Yn 74 oo. w..l. 0 By the law of large numbers.. In that case .d.3 The i.. then the restrictions of Fx. One then assumes Yn to be a stationary sequence.s. .s. .1. Similarly. Next consider change of measure via likelihood ratios. . Proof Since (YN. . + Z. .s. assumption on the Z1. (Yi + • • • + Yn) > oo a. . ..1 have the same distribution for n = 0. In general..1. (e)Yi+•••+Yn 74 . Px to Fn are equivalent (have the same null sets) so that the likelihood ratio Ln exists.. (d) #. Combining these facts gives easily the equivalence of (a)(d).g. The following result gives the necessary and sufficient condition for {Ln} to define a new random walk: .i.. For a random walk. . By Kolmogorov's 01 law.. Y1) has the same distribution as (Y1. N.. 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.) and defines Zn = Yn. YN in Theorem 4..1 is equivalent to WN D MN = (Z1 + .. Clearly. a Markovian change of measure as in Theorem 2. ZN or. (e).l... the Y1..g. or M < oo a.
.5 corresponds to a new random walk with changed increment distribution F(x) = e'(a) Jr e"'F(dy) .4). Then the change of measure in Theorem 2. h(Yn) (4. In particular. (4. of F).4. Proof If (4.. Y) f (Y)] for all f and x. (a) = log F(a] is the c. for some function h with Eh(Y) = 1. Then the change of measure in Theorem 2. Y1).g. implying g (x. Breiman [78] p. cf.3) holds for n = 1. this means E(g(x. where h (y) = g(0. Conversely. Y) = h(Y ) a. y ).5 corresponds to a new random walk if and only if Ln = h(Y1) . . Y < x]. e.4 Let {Ln} be a multiplicative functional of a random walk with E_L. the changed increment distribution is F(x ) = E[h(Y). we get L2 = L1 (L1 o91 ) = h(Y1)g(X1.4) ({Ln} is the familiar Wald martingale . and so onforn =3. u A particular important example is exponential change of measure (h(y) = e°y'(") where r. We get: Corollary 4.. Since L1 has the form g (Xo.nrc(a )} (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.3) 1Pxa.. Y ) f (Y)] = E[g(O. RANDOM WALKS INDISCRETE OR CONTINUOUS TIME 35 Proposition 4.. The corresponding likelihood ratio is Ln = exp {a (Y1 + • • • + Yn) ..4.s.s.5 Consider a random walk and an a such that c(a) = log F[a] = log Ee° ' is finite.3) holds. and define Ln by (4. = 1 for all n and x..Y2) = h(1'i)h(I'a). In that case. For n = 2.g.f. 100 ). the random walk property implies Ex f (Y1) = Eo f (Y1 ).
The traditional formal definition is that {Xt} is Rvalued with the increments Xt(1)_t(o)... SOME GENERAL TOOLS AND RESULTS Discrete time random walks have classical applications in queueing theory via the Lindley process representation of the waiting time . v2 = 0 and v = 3B). {Xt} is a random walk. the claim surplus process for the compound Poisson risk model .Xt)I. see Chapter V. a positive measure on R with the properties e J x2v(dx) < oo. the pure jump process is given by its Levy measure v(dx). Xt =Xo+pt+oBt+Mt. 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. . however. but we omit the details ). which corresponds to the compound Poisson case: here jumps of {Mt} occur at rate 0 and have distribution B = v/0 (in particular . However. Roughly. a Brownian component {Bt} (scaled by a variance constant) and a pure jump process {Mt}. Xt(2)_t(l). or imbedded into continuous time processes .. < t(n) and with Xt( i)_t(i_l) having distribution depending only on t(i) . the interpretation is that the rate of a jump of size x is v(dx) (if f of Ixlv (dx) = oo. they arise as models for the reserve or claim surplus at a discrete sequence of instants.e.5) Note that the structure of such a process admits a complete description. say by recording the reserve or claim surplus just before or just after claims (see Chapter V for some fundamental examples). Xt (n)t(n1) being independent whenever t(O) < t(1 ) < .Ft] = Eof (X. {Xt} can be written as the independent sum of a pure drift {pt}. In discrete time. The appropriate generalization of random walks to continuous time is processes with stationary independent increments (Levy processes). corresponds to a process with stationary independent increments and u = p.36 CHAPTER II. say the beginning of each month or year .Xn < x). (4. with premium rate p.1). In continuous time. the tradition in the area is to use continuous time models. . this description needs some amendments. i. In risk theory.t(i .. we are .7) for all e > 0. (4.). An equivalent characterisation is {Xt} being Markov with state space R and E [f (Xt+e .6) More precisely. given by a the increment distribution F(x) = P(Xn+l . e f x:IxJ>e} v(dx) < oo (4. The simplest case is 3 = JhvMM < oo.
Q and distribution B of the service times of the arriving customers. is easily seen to be f3pB < 1. 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.6). Proposition 4. defined as a system with a single server working at a unit rate. Then the storage process {Vt} has constant release rate 1.1. cf.6 In the compound Poisson risk model with constant premium rate p(r) .v.4. WT = XT . Corollary 4.2).10) . Furthermore. then Ee'(xtxo) = Eoeaxt = etx(a).3 and decreases linearly at rate 1 in between jumps.7 If {Xt} has stationary independent increments as in (4. 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. T) = P(VT > u). VT + V for some r.e. 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. and the reflected version is then defined by means of the abstract reflection operator as in (4. [The condition for V < oo a. (ex . Now consider reflected versions of processes with stationary independent increments. 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). has upwards jumps governed by B at the epochs of a Poisson process with rate . O(u. where c(a) = ap + a2a2/2 + f 00 provided the Levy measure of the jump part {Mt} satisfies f". 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. V E [0.8) O<t<T (assuming Wo = Xo = 0 for simplicity). oo]. jxJ v(dx) < oo.] Processes with a more complicated path structure like Brownian motion or jump processes with unbounded Levy measure are not covered by Section 3. A different interpretation is as the workload or virtual waiting time process in an M/G/1 queue. Chapter III.min Xt (4.1)v(dx) (4. and b(u) = P(V > u). i. having Poisson arrivals with rate .
38 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Proof By standard formulas for the normal distribution.g. Chung [86]). if Lt = e9(xt . Xt +B . then the changed parameters in the representation (4. .xo)tk ( e).10) is the LevyKhinchine representation of the c.f.Xo is necessarily infinitely divisible when {Xt} has stationary independent increments. t. In particular. This is of course no coincidence since the distribution of Xl . (4. v(dx) = e9xv (dx). Xt Xo) with E2Lt = 1 for all x. Then l e" (a) = Eo [ Li ea "] = eK (9)Eo {e ( a+9)x1 J I = er(a+o )K(B) R(a) = K(a + 0) . of an infinitely divisible distribution (see. we use the characterization (4. Theorem 4 . e. let e" (a ) = Eoeaxl.5) and get E [f(Xt+B . 5 has stationary independent increments as well. 8 Assume that {Xt} has stationary independent increments and that {Lt} is a nonnegative multiplicative functional of the form Lt = g(t.Xt)I'Ftl = E [f(Xt+B . X8) = Eof (X8)L8 = Eof (X8)• For the second. Then the Markov process given by Theorem 2.g.6) are µ = µ + Oo2 .Xt)g(s.11) (eax .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 . Q2 = v2.Xt)I Ftl = Eof (X8)g(s. u Note that (4.Xt)L8 o 0tIFt] = E [f (Xt+s .1 ) v(dx) . Proof For the first statement . By explicit calculation .1.. use the representation as limit of compound Poisson processes.1)eexv(dx).1 that E eaMt = exp fmoo In the general case . we show in the compound Poisson case ( IlvIl < oo) in Proposition III. Eea(µt + QBt) = et{aµ +a2OZ/2}.
(3B(dx). Ei.9 If X0 = 0.3 and claim size distribution B # B. b = a = 0) the changed process is the claim surplus process of another compound Poisson risk process with Poisson rate .l3 and claim u size distribution B.. a = 0. abbreviated as MAP in this section2. As for processes with stationary independent increments .8. 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. where .(3 = . it is then easily seen that Lt = H dB(Ui) i:o.10 Let Xt be the claim surplus process of a compound Poisson risk process with Poisson rate . (5. one reason is that in parts of the applied probability literature. Ei instead of P2. 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 . Then we can write v(dx) _ /3eOxB(dx) = / (dx). Recalling that U1.. B(dx) = B[9] B(dx). then the martingale {eex(t)tk(e)} is the continuous u time analogue of the Wald martingale (4. the structure of MAP's is completely understood when E is finite: 2and only there .St)g(Jt+s)I. b with b(x) > 0 for all x such that b(x) > 0).3 and claim size distribution B. . the corresponding claim sizes .0 in the following.g.4).Ft] = Ejt. Example 4 .1) For shorthand . are the arrival times and U1. and let the Px refer to the claim surplus process of another compound Poisson risk process with Poisson rate.2. corresponding to p = 1.. 0. MARKOV ADDITIVE PROCESSES 39 Remark 4.3B[B]. U2. dB/dB = b/b when B. MAP stands for the Markovian arrival process discussed below.5. we write Pi. Example 4 . is defined as a bivariate Markov process {Xt} = {(Jt. <t whenever the RadonNikodym derivative dB/dB exists (e. u 5 Markov additive processes A Markov additive processes. B have densities b.o[f (S8)g(J8)]. .. Thus (since µ = p = 1..3 =.11 For an example of a likelihood ratio not covered by Theorem 4. v(dx) _ .0.
the converse requires a proof.it = A. Y1 E dx) where Y„ = S„ .6) depending on i. a MAP is the same as a semiMarkov or Markov renewal process.jEE (here pij = Pi(J1 = j)) and the probability measures Hij(dx)=P(Y1 EdxlJo=i. a jump of {Jt} from i to j # i has probability qij of giving rise to a jump of {St} at the same time.Sr_1. .. SOME GENERAL TOOLS AND RESULTS In discrete time.9 EE = (iii&ij[a])i j EE .. {Jt} is specified by its intensity matrix A = (Aij)i. (That a process with this description is a MAP is obvious. t+s) where Jt .jEE• On an interval [t. v. which we omit and refer to Neveu [272] or cinlar [87]. As a generalization of the m.[a) = (Ei[easl.1 For a MAP in discrete time and with E finite.g. An alternative description is in terms of the transition matrix P = (piA. Jn = j. a MAP is specified by the measurevalued matrix (kernel) F(dx) whose ijth element is the defective probability distribution Fij(dx) = Pi.40 CHAPTER H.. with the Y„ being interpreted as interarrival times. Proposition 5. 1 J1 ='^])iJEE = (Fij[a])i .f. this means that the MAP can be simulated by first simulating the Markov chain {J„} and next the Y1. {St} evolves like a process with stationary independent increments and the parameters pi.) If E is infinite a MAP may be much more complicated. If all Fij are concentrated on (0. Y2. by generating Yn according to Hij when J„_1 = i.o(Ji = j.. As an example.J1=j)= Fij (dx) Pij In simulation language. the distribution of which has some distribution Bij. In continuous time (assuming Dpaths). oo). Fn[a] = F[a]n where P[a] = P .i. Then a Markov additive process can be defined by letting t St = lim 1 I(IJB1 < e)ds E1o 2d o be the local time at 0 up to time t. consider the matrix Ft [a] with ijth element least Ei . let {Jt} be standard Brownian motion on the line. vi(dx) in (4. In addition..
MARKOV ADDITIVE PROCESSES Proof Conditioning upon (Jn. kEE Jn = k]Ek[e"Y" . a= . Jt = j] is given by etK[a]. \ diag Ft[a] = Ft[a]K.ijgij(Bij[a] . Bij (i. 00 r(i) (a) = api + a2ot /2 + f (e° . qij. this means that F't+h [a] = Ft[a] II+h(rc(i)(a)) +hA+h(Aijgij(Bij[a]1)) I. vi(dx) (i E E). we infer that in the discrete time case the .2 Let E be finite and consider a continuous time Markov additive process with parameters A. which in conjunction with Fo[a] = I implies Ft[a] = etK[a) according to the standard solution formula for systems of linear differential equations. Then the matrix Pt[a] with ijth element Ei [east.qkj + k?^j qkj Bkj [a] } = Ei [east. pi. In matrix formulation . Jn+1 = A] = 41 Ei[ e 5„.1)) . Jt = j] Ejesh'^ + E Ak j hEi [ease .(')(a)) diag + (). up to o( h) terms. Jt = k] { xk kEE j la] .1) } (recall that qjj = 0). u In the following. 013 . Proof Let {Stt) } be a process with stationary independent increments and pa rameters pi . Then. Jt = j] (1 + htc (j) (a)) j + Ak j qk j (Bk +h E Ei [east . u Proposition 5.4c). Sn ) yields Ei[easn+ '. aSt h = (1 + Ajjh) Ei [east . where K[a] = A+ (r. Jt = k] { 1 . assume that the Markov chain/process {Jt} is ergodic. j E E) and So = 0.1 )v(dx). vi(dx).5. By PerronFrobenius theory (see A. J1 = A which in matrix formulation is the same as Fn+1 [a] = Fn[a]F[a].
u Let k(a) denote the derivative of h() w. and appropriate generalizations of the Wald martingale (and the associated change of measure) can be defined in terms of .4 Eie"sth(a) = h=a)et?("). as will be seen from the following results. The corresponding left and right eigenvectors v("). Corollary 5.5 EiSt = tK'(0) + ki . Corollary 5. of a random walk.2) where 7r = v(°) is the stationary distribution.Jt+v = easttK( a)E [ee (st+vst)vK(a)h(a) jt+v I ^tJ = easttt(a)EJt (easesvK(a )h^a)1 = easttK(a)h^a).42 CHAPTER II. Furthermore. h(") are only given up to a constants. Jt = j] .t.h(a)vva)etw(a). Then h(°) = e. and we shall take V(a)h(a) = 1. Eie"sth^a) = e'Pt[a]h( a) = e.etx It then follows that E feast+^(t+v)K(a)h(a) I ^tl l . Corollary 5.f. h(") may be chosen with strictly positive components.3 Ei [east. its derivatives are 'asymptotic cumulants'. cf. . Since v(").e=e°tk. Proof For the first assertion.r. cf. Proposition 5.c(a) (and h(")). just note that [a]h(a) = eietK (a)h(a) = etK(a)h(a).Eikjt = ttc'(0) + ki .tK(a)h(a) J jj it L o is a martingale. Yrh(a ) = 1. 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. The function ic(a) plays in many respects the same role as the cumulant g. Proof By PerronFrobenius theory (see A. and write k = k(°). Jeast. In particular.4. (5.4c).7. We also get an analogue of the Wald martingale for random walks: Proposition 5. we are free to impose two normalizations. a.
u The argument is slightly heuristic (e. For the second .5 yields + W (a)k.g. [E. (5.4) .St]2 = t2/c'(0 ) 2 + 2ttc (0)vk . . summing and letting a = 0 yields E„ [St + 2Stkj.7 No matter the initial distribution v of Jo. we differentiate (5. Since it is easily seen by an asymptotic independence argument that E„ [Stkjt] u = trc'(0) E„kjt + O(1).6 For any stopping time T with finite mean.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. one obtains a generalization of Wald's identity EST = Er • ES.5. t a oo. 8 Also for E being infinite (possibly uncountable ). tam E tSt a (0).") }) .4.Eikjr . Ee"st typically grows asymptotically exponential with a rate ic(a) independent of the initial condition (i. In the same way. for a random walk: Corollary 5.5. More precisely.. E=ST = tc'(0)E7. 43 Ei [Steast h(a) + east k^a)1 = et"(a) (kia) + tic (a)hia)) . there is typically a function h = h(") on E and a ^c(a) such that Ey a"st t" (") * h(x).3) Let a = 0 and recall that h(°) = e so that 0=°) = h(o) = 1. Squaring in Corollary 5. (5. the existence of exponential moments is assumed ) but can be made rigorous by passing to characteristic functions.e.. Corollary 5. Remark 5 . ] = t2tc (0)2 + 2tK'(0)vk + ttc"(0) + O(1) . MARKOV ADDITIVE PROCESSES Proof By differentiation in Proposition 5. t im v^"St = '(0) Proof The first assertion is immediate by dividing by tin Corollary 5.a) + ttc (a)2hia ) Multiplying by v=. the distribution of Jo). subtraction yields Vary St = tic"(0) + O(1).2ttc (0)Evkjt + 0(i).+ k.
0) = h(i )( 1 + ttc(a)).9 The condition that (5. St)} be a MAP and let 0 be such that h(Jt) OStt. see. Usually.(9) {Lt}t>o = .44 CHAPTER IL SOME GENERAL TOOLS AND RESULTS for all x E E. this leads to h(i) + tcha( i. inconvenient due to the unboundedness of ea8 so we shall not aim for complete rigour but interpret C in a broader sense.4 that { h(Jt) easttK(a) L o (5. in particular that f is bounded. however. let ha(i. and the family {f LEE given by Theorem 2. St) } as follows. we take the martingale property as our basic condition below (though this is automatic in the finite case). First. In view of this discussion . Remark 5. From (5.s.e. where {Jt} is deterministic period motion on E = [0. We then want to determine h and x(a) such that Ejeasth (Jt) = etK(a)h(i).. gha(i.for the present purposes. For t small .f (x) tyo t provided the limit exists.5 defines a new MAP. (5.5) is a martingale can be expressed via the infinitesimal generator G of {Xt } = { (Jt. An example beyond the finite case occurs for periodic risk processes in VI. V. 1) (i. u forsEE).5) is a martingale . G is defined as Gf (x) = lim Exf (Xt) . this is. 0) = n(a) h(i).6) We shall not exploit this approach systematically.10 Let {(Jt.1) one then ( at least heuristically) obtains lim Ex eaSv v a) K( v+oo nEx easttK(a)EJt east(vt)K(a) u[J = Ex easttk(a)h(Jt) It then follows as in the proof of Proposition 5. s) = ea8h(i). Jt = (s+t) mod 1 P8a.6. h(Jo) Lo is a Px martingale for each x E E. however. Given a function h on E.6. xEE . 0 Proposition 5..5. Then {Lt } is a multiplicative functional. some extra conditions are imposed.3b and Remark VI.
1) holds for the P. and by A = Oh(°) K [0]Oh(e ) vi(dx) = e"xvi (dx). 0 < qij < 1 and Bij [0] > 0. 0<b<oo. Bi(dx) = Bi(dx). qij = r. this gives a direct verification that A is an intensity matrix: the offdiagonal elements are nonnegative because Aij > 0. In the infinite case .11 Consider the irreducible case with E finite. That 0 < qij < 1 follows from the inequality qb <1.(0)j. That the rows sum to 1 follows from Ae = Oh(e) K[O]h(B) .10 is given by P = eK(e) Oh e) F[e]Oh('). 1 + q(b .5.c(0)e = tc(0)e .1) .11 below in the finite case.Qi < oo and Bi a probability measure. u Theorem 5. In particular. Then the MAP in Proposition 5. MARKOV ADDITIVE PROCESSES Proof That { Lt} is a multiplicative functional follows from L8 ogt = h(Jt+s) es(St+ . Ai = µi + 0Q. then also vi (dx) is compound Poisson with e Ox ^i = /3iBi[0].1) eft ea' f ij (dx) = Hij (dx) Hij [0] . (5.. Here Oh(e) is the diagonal matrix with the h=e) on the diagonal. . vi (dx) = f3 Bi(dx) with .St)sl(e) h(Jt) 45 The proof that we have a MAP is contained in the proof of Theorem 5.7) In particular.. Bi. 0<q<1. ^i = of qij Bij [0] 1 + qij ( Bij [0] . in the discrete time case.tc(0)e = 0 . We omit the details. one can directly verify that (5. Bi [0] Remark 5.7(dx) Bij [0] Bij(dx) in the continuous time case .ic(0)e = ic(0)Oh e) h(e) .12 The expression for A means h(e) Aij = hie) Aij [1 + gij(Bij[0] i 0 j. if vi(dx) is compound Poisson.
F:j with a density proportional to eei .tc (') (0) corresponds to the stated parameters µ. In matrix notation . (dx) of a process with stationary independent increments follows from Theorem 4. SOME GENERAL TOOLS AND RESULTS Proof of Theorem 5. Letting a = 0 yields the stated expression for A.11. Hence the same is true for H=j and H.. this implies k[a] = A 1 ) (K[a + 0] . are probability measures .8) yields et'[a] = Ohie )et (K[a +e]K(e)I)Oh(°) By a general formula (A. it follows that indeed the normalizing constant is H1 [0]. Jl = j] :(Yi E dx. Now we can write K[a] =A+A ) ( K[a + 0] . .tc(0)' )Ah() = Oh(o) K[a + 0]Oh() . . this means that Ft[a] = etw ( e)Ohc) Ft[a + 9]oh (e) (5. since Hij. Here the stated formula for P follows immediately by letting t = 1. v.8). Jl = j) = Ei[Lt. (dx).t. v= . Yi E dx.tc(') (8)/ d)ag h 7 Aiiii (Bii[a + 0] . a = 0 in (5. H1.8) h(.46 CHAPTER II. Ji = j) h(e) eeyK(B)p h(8) h(e) eexK ( h=e) e)Fi.13) for matrixexponentials .r.K [O])Oh(e) (0) l + ( A + (tc(') (a + 0) . Similarly. Jt = j] = hie) .Bay [0]) That k(') (a + 0) . This shows that F. First note that the ijth element of Ft[a] is etK(e)Ej [e(a+B)st E:[east Jt = j] = Ej[Lteas' . in continuous time ( 5..tc(0)I. is absolutely continuous w.e) Consider first the discrete time case ..8. Further Fib (dx ) = P=(YI E dx. Jt = A.
which. 0]. 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). an extensive bibliography on aspects of the theory can be found in Asmussen [16].3 for an infinite E are given by Ney & Nummelin [266]. has no mass on (oo. [261].)Ajjgij(Bij[a+0] .6. is slightly less general than the present setting. Write r+ = T(0) and define the associated ladder height ST+ and ladder height distribution by G+(x) = 11 (S.1) = Aij4ij(Bij[a] .e. the literature on the continuous time case tends more to deal with special cases.+ < x.. however.. oo). Note that G+ is concentrated on (0. see also Fuh & Lai [149] and Moustakides [264]. 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].1). Though the literature on MAP's is extensive. For the Wald identity in Corollary 5. Notes and references The earliest paper on treatment of MAP's in the present spirit we know of is Nagaev [265]. there is. 7+ < oo). Conditions for analogues of Corollary 5. i. < x) = 11 (S. [225].7).6. h. hardly a single comprehensive treatment.(u) = inf {t > 0 : St > u} to ruin in the particular case u = 0 . however. [226] and Miller [260].Bij[0]) = hjel)ijgijBij[0](Bij[a] . Much of the pioneering was done in the sixties in papers like Keilson & Wishart [224]. 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. . and is typically defective. [262] in discrete time.
it follows that for g > 0 measurable.i. we shall first consider the compound Poisson model in the notation of Example 1. The first ladder step is precisely ST+. 0]. o 00 (6.T+ > t)dt = E f 0T+I(St E A) dt. Recall that B(x) = 1 . = ST+(1) Figure 6. see Fig. 0 f T+ (6.1 The term ladder height is motivated from the shape of the process {Mt} of relative maxima. In any case. define the prer+occupation measure R+ by R+(A) = E f o "o I(St E A. i. the dependence structure seems too complicated to be useful). oo). To illustrate the ideas. In simple cases like the compound Poisson model.00 ).2. 6. R+ is concentrated on (oo. Also. they have a semiMarkov structure (but in complete generality. a fact which turns out to be extremely useful.5 below.2) . For the proof of Theorem 6.. at present we concentrate on the first ladder height.1) The interpretation of R+(A ) is as the expected time {St} spends in the set A before T+.. 1 For the compound Poisson model with p = 01LB < 1. the second ladder height (step) is ST+(2) . where basically only stationarity is assumed.ST+(1) and so on. 6.1. the sum of all the ladder steps (if rl > 0. The main result of this section is Theorem 6.1.e. In other cases like the Markovian environment model. and the maximum M is the total height of the ladder. which gives an explicit expression for G+ in a very general setting. by approximation with step functions . i. has no mass on ( 0. SOME GENERAL TOOLS AND RESULTS M ST+(2) Sr. G+ is given by the defective density g + (x) =. On Fig. the ladder heights are i. g(y)R+(dy) = E f g(St)dt.48 CHAPTER K. Theorem 6 .e. Thus.d. the second ladder point is ST+(2) where r+(2) is the time of the next relative maximum after r+(1) = r+. Here bo(x) _ B(x)/µB.(3B(x ) = pbo(x) on (0.1.B(x) denotes the tail of B. there are only finitely many).
0]. 49 Proof Let T be fixed and define St = ST .O<t<T) = P(STEA. 0 < t < T) P(STEA. {St }o<t<T is constructed from {St}o<t<T by timereversion and hence. has the same distribution as {St}o<t<T. see Fig.2(b): r+ < t Thus. 6.ST<St. St S* t a Figure 6. That is.6. P(STEA.0<t<T) = F(ST E A. since the distribution of the Poisson process is invariant under time reversion.St<0. THE LADDER HEIGHT DISTRIBUTION Lemma 6 .2. ST < ST_t.ST_t. .T+>T) = P(STEA.2(a): T+ > t Figure 6.O<t<T). 0 < t < T.2 R+ is the restriction of Lebesgue measure to (00.ST<St.
and since the jump rate is /3. this is just the Lebesgue measure of A. That is. But since St 4 oo a. oo).t dT. Fig. U + St_ E A..3 G+ is the restriction of /3R+*B to (0. 6. .T+ > t] dt 0 T+ _ /3E f g( St) dt = 0 f g(y) R+(dy) 0 00 where g(y) = B(A .3 where the bold lines correspond to minimal values.St). s. SOME GENERAL TOOLS AND RESULTS Integrating w.y)R+(dy) 00 Proof A jump of {St} at time t and of size U contributes to the event IS. Figure 6.50 CHAPTER II. T+ > t] 0 _ /3 f E[B( A . oo).3 Lemma 6 . G+(A) = Q f 0 B(A .r.St _).. E A} precisely when r+ > t. for A C (0.2) in the last).St _)I(r+ > t). cf. we get G+ (A) = f 00 /3 dt E[B(A . it follows that R+ (A) is the expected time when ST is in A and at a minimum at the same time .y) (here we used the fact that the probability of a jump at u t is zero in the second step. The probability of this given { Su}u<t is B(A . and (6.
Uk) (k = 1. Lemma 6. . obviously. assuming basically stationarity in time and space. i.e. The points in the plane (marked by x on Fig.Q f r+(x .S8 )t> o = {St }t>o for all s > 0. we consider the claim surplus process {St }t>o of a risk reserve process in a very general setup. 0 Generalizing the setup.1 With r+(y) = I(y < 0) the density of R+. The traditional representation of the input sequence {(TT. this is equivalent to the risk process {St*} being stationary in the sense of (6.) where ak = Ti + • • • + Tk . 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.. oo). In the stationary case.4) are (ak.* ) and the second the mark (the claim size Uk ). .s > 0). this does not depend on h). the first component representing time (the arrival time o.4). 6 .3 yields g+ (x) = .6. {St+8 . We call M * stationary if M* o B8 has the same distribution as M* for all s > 0. Nt St =>Uk k=1 t where Nt = max{k = O. Fig.:T1 +•••+Tk <t}. The first ladder epoch r is defined as inf It > 0 : St > 0} and the corresponding ladder height distribution is * G+ (A) = P(S** E A) = P(ST+ E A.1.. 6 . Uk) for those k for which ak . h]} /h (by stationarity..T+ < oo). 2. U k)} k=1 a is as a marked point process M *. cf.. i. oo) x (0.. we define the arrival rate as E# { k : ak E [0 .s. 4 (the points in the plane are (ak .M o 08 shifted by s is defined the obvious way. The marked point process . THE LADDER HEIGHT DISTRIBUTION 51 Proof of Theorem 6.e. as a point process on [0.z)B(dz) _ f I(x < z)B(dz) _ f (x).
o.e. where TI = 0. See. where T is the first arrival time > 0 of M and h > 0 an arbitrary constant (in the literature. Oh becomes the approximate probability F(ri < h) of an arrival in [0..4 Given a stationary marked point process M*. SOME GENERAL TOOLS AND RESULTS M* U. Uk) k=1. h] Eco(M*) = 1 E f co(M o Bt)dt.4 Consider a finite Markov additive process (cf. = 0 .QiBi(dx). V(M* o eak ). The two fundamental formulas connecting M* and M are Eco(M) = aE E.52 CHAPTER II. Section 5) which has pure jump structure corresponding to pi = a = 0.g. we define its Palm version M as a marked point process having the conditional distribution of M* given an arrival at time 0 . i 1 U2 Us 1_ 0 or Q2 $ U3 *1 L 0 7 X I 11 1 Figure 6. This more or less gives a proof that indeed (6.s.5) does not depend on h. of (6. As above .. Note also that (again by stationarity) the Palm distribution also represents the conditional distribution of M* o Ot given an arrival at time t. most often one takes h = 1). We represent M by the sequence (Tk. vi(dx) = .. 0. and let T = T2 denote the first proper interarrival time . Sigman [348] for these and further aspects of Palm theory. h] and the sum approximately ^o(M*)I(ul < h). letting h J.2. k: vk E [0.5) represents the conditional distribution of M* given vi = 0. . Example 6 . i. Assume {Jt} irreducible so that a stationary distribution 7r = (1i)iGE exists. h.. e. the r.
we get a marked point process generated by Poisson arrivals at rate /3i and mark distribution Bi when Jt = i.p. THE LADDER HEIGHT DISTRIBUTION 53 Interpreting jump times as arrival times and jump sizes as marks. the ruin probability ./.5.p. 5 Consider a general stationary claim surplus process {St }t>o.oo a.*(0) with initial reserve u = 0 is p = /3EU0. let U0 be a r. Jt = j is iri(3i /. Then the ladder height distribution G+ is given by the (defective) density g+(x) = . After that.6.e. 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 .p.OEU0.OF(x). qij when {Jt} jumps from i to j and have mark distribution Bij.s. we note: Corollary 6.O for i # j.6iBi + Aijgij Bij j#i iEE iEE 0 Theorem 6 . and by some additional arrivals which occur w. the distribution of Ul) is the mixture B = E aii Bi + aij Bij J = j#i !i J.6 Under the assumptions of Theorem 6.= i. v. j) and let the initial mark Ul have distribution Bi when i = j and Bij otherwise. . Note in particular that the Palm distribution of the mark size (i. First choose (Jo_. Assume that St * . an arrival for M* occurs before time t + dt w. A stationary marked point process M* is obtained by assigning Jo distribution Tr.. having the Palm distribution of the claim size and F (x) = F(Uo < x) its distribution . It follows that we can describe the Palm version M as follows . aij for (i.O for i = j and iriAijgij/. If Jt_ = i. and that p = 0EU0 < 1. This follows by noting that iP*(0) = IIG+JI = J0 "o g+(x)dx = . the probability aij of Jt . Before giving the proof. Jo) w.O fo "o F(x)dx = . let the arrivals and their marks be generated by {Jt} starting from Jo = j.
..$St_ u. that M* and M have doubly infinite time (i.Su< 0. The result is notable by giving an explicit expression for a ruin in great generality and by only depending on the parameters of the model through the arrival rate 0 and the average ( in the Palm sense) claim size EU0.5. Then clearly * G+ (A) = P(ST+ E A) = Consider a process { f p(t)f3dt. in (oo. . Proof of Theorem 6.0<u<tIAt) = P(St EA. CHAPTER H. .Su_ <0. and the kth preceding claim arrives at time t .o<u<t where a claim arrives at time t and has size Uo. 6.0<u<t) = P(St EA. 0).0<u<tIAt) = P(St EA.s. SOME GENERAL TOOLS AND RESULTS V` (0) = E E Uk k: ak E [0.e. The sample path relation between { Su } and { Su } amounts to S„ = St .Mt). oo) x (0 . moves down linearly at a unit rate in between jumps and starts from S0 = U. which makes an upwards jump at time . The last property is referred to as insensitivity in the applied probability literature.g.. A standard argument for stationary processes ([78] p. 105) shows that one can assume w. T+ = t given the event At that an arrival at t occurs .St <.0<u<t) = P(StEA. Now conditionally upon At . oo) p(t) = P(St EA.. in (0. Let p(t) be the conditional probability that ST+ E A.Q_k and has size U_ k.. the mark at time Qk is denoted by Uk. oo)). (k = St}t>o 1. are point processes on (oo .).. We then represent M by the mark (claim size ) Uo of the arrival at time 0.l. the arrival times 0 < 0'1 < Q2 < .5).1] here the r .St*_ u. has a very simple interpretation as the average amount of claims received per unit time . 0<u<t) = P(St EA.5.o. h.o.St<Su. oo ) and the arrival times 0 > 0_1 > a_2 > .(left limit) when 0 < it < t and is illustrated on Fig .54 By (6. It follows that for A C (0.... { Su}0<u<t is distributed as a process {Su} .A. 2.Su<0.
5 where it = { St < Su. THE LADDER HEIGHT DISTRIBUTION 55 { A Su}0<u<t U0 U0 \t tt u>0 N U_1 Figure 6. and since by assumption St * oo a. and we let L(dy) be the random measure L(A) = fo°° I(St E A. 0 < u < t } is the event that { Su } has a relative minimum at t .6. the support of L has right endpoint U0. 2 therefore immediately shows that L(dy) is Lebesgue measure on (oo. A sample path inspection just as in the proof of Lemma 6 .s. Since So = U0. t a oo. Mt)dt = i3EL(A) o"o . 6. NIt)dt .5. the left endpoint of the support is oo. In Fig. Uo]. G' (A) = 3 f P(St E A. Thus. time instants corresponding to such minimal values have been marked with bold lines in the path of { St}. Fig.5 where the boxes on the time axis correspond to time intervals where {St } is at a minimum belonging to A and split A into pieces corresponding to segments where {Su} is at a relative minimum. 6. cf..
[263] (a special case of the result appear in Proposition VI.56 CHAPTER II. . SOME GENERAL TOOLS AND RESULTS = OE f 0 I(Uo>y)I (yEA)dy = Q f IP (Uo>y)dy A 0o a fA P(y) dy• 0 Notes and references Theorem 6. A further relevant reference related to Corollary 6.5 is due to Schmidt & coworkers [48].6 is Bjork & Grandell [67]. [147].2.1).
• the premium rate is p = 1. A common view of the literature is to consider such processes as perturbed compound Poisson risk processes .d. St = uRt = EUi t. i.e. being of the form Rt = Rt+Bt + Jt where {Rt } is a compound 57 . Panjer's recursion ( Corollary XI. {Rt} and the associated claims surplus process {St} are given by Nt Nt Rt = u+t EUi.6) and simulation methods ( Chapter X). 3).. U2. and assume that • { Nt}t>o is a Poisson process with rate j3.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. see Chapter IV. 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 . Some possibilities are numerical Laplace transform inversion via Corollary 3. say.4 below . Thus . For finite horizon ruin probabilities . exact matrixexponential solutions under the assumption that B is phasetype (see further VIII. i=1 i=1 An important omission of the discussion in this chapter is the numerical evaluation of the ruin probability. . 4. and independent of {Nt}. i. are i.. • the claim sizes U1. with common distribution B.
t = fltpB .'s etc. P = PAB = 1/(1 + rl) Proposition 1. where K(k) (0) is the kth derivative of is at 0.1). For (c). 0 . Dufresne & Gerber [126]. we get Ee8st = 00 e8t c` Ee8 (U1+. we shall start by giving the basic formulas for moments..1 is the expected claim surplus per unit time.t = E E [ U k k=1 k=1 Nt . 1 Introduction For later reference. Write pB^) = EUn' YB = Pali = EU. and this immediately yields (a).1 (a) ESt = t(13µ$ .g.t = E[Ntµs] .Rt. Furrer [150]. e .6pBa). (d) The kth cumulant of St is tf3p(k) for k > 2.1). We do not spell out in detail such generalizations.58 CHAPTER III. Proof It was noted in Chapter I that p . say stable Levy motion.g.1) . (b) Var St = t. and Schlegel [316].s. m. cumulants . A more formal proof goes as follows: Nt r Nt ESt = E > U k .u . THE COMPOUND POISSON MODEL Poisson risk process. and that B(k)[0] = Pak). See e.1) = t(p .)3t (fit' k t} = etk(8) exp {st '3t + B[s]f Finally. Schmidli [319].t = t(p . 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). of the claim surplus St .+Uk)P(Nt = k) k=O e8t k=O B[s]k .. for (d) just note that the kth cumulant of St is tic(k) (0).f. {Bt} a Brownian motion and {Rt} a pure jump process. [324]. (c) Ee8St = et" (8) where c(s) = f3(B[s] .
For example.Tk are i. cf. Obviously. The right hand inequality in (1. then lien inft. we get a discrete time random walk imbedded in the claim surplus process {St}.S„ attains its minimal value when there are no arrivals in (u. INTRODUCTION 59 The linear way the index t enters in the formulas in Proposition 1..V. (d) If 17 = 0.1 is the same as if {St} was a random walk indexed by t = 0.i. In this way. Indeed. where Tk is the time between the kth and the (k . S„+V > S„ .Tk.3) is proved similarly. u + v].h < St < S(n+1)h + h. meaning that the increments are stationary and independent. The connections to random walks are in fact fundamental. Sn+0 . then St 4 co. rather to view {St} directly as a random walk in continuous time.d. the Uk . v > 0. 1. then St 00. we need the following lemma: Lemma 1. lim supt. obviously 0(u) = F(maxk Sok > u). In particular. Here is one immediate application: Proposition 1.2 (DRIFT AND OSCILLATION) St/ta3'p1 ast >oo. For the proof. 2. and the value is then precisely v. then Snh .1. The point of view in the present chapter is. St = oo. . (a) No matter the value of 77. (b) If 77 < 0.1 = . then St> SnhV>Snhh.4.. which is often used in the literature for obtaining information about {St} and the ruin probabilities.. however. and there are at least two ways to exploit this: Recalling that ok is the time of the kth claim. II. St = oo.3 If nh < t < (n + 1)h. we have Sok . so that {Sok } is a random walk with mean EUET = EU..Sok_l = Uk .3EU01 = 1µs where rt is the safety loading. Proof We first note that for u. (c) If 77 > 0. if t = nh + v with 0 < v < h.1)th claim. We return to this approach in Chapter V.
{Snh}n=o. . it is seen that upcrossing occurs at least twice.3. h A similar argument for lim sup proves (a).2: Proposition 1. and hence it folz lows from standard central limit theory and the expression Var(St) = tf3pB (Proposition 1. if P(M > 0) = 1. or by a general result on discrete skeletons ([APQ] p. 169) stating that lim infra..1. and hence by the strong law of large numbers. p.1. u 307). Assuming that each risk generates claims at Poisson intensity /3 and pays premium 1 per unit time. Snh/n a4' ESh = h(p . This contradicts u St400. For any fixed h. where the size of the portfolio at time t is M(t). hence by induction i. it suffices to prove 4'(0) = F(M > 0) < 1.3.6 Often it is of interest to consider size fluctuations. is a discrete time random walk. h. then {St} upcrosses level 0 a.1). 0 Snh = 00. and (b)..o.p. However. {Snh}n=o.1.s. Notes and references All material of the present section is standard. Thus using Lemma 1... we get lim inf St t>oo t nroo nh<t<(n+1)h t = lim inf inf St h l++m of Sn 7t h = ESh = p . Remark 1 . THE COMPOUND POISSON MODEL Proof of Proposition 1.60 CHAPTER III. The general case now follows either by another easy application of Lemma 1. lim supn_.. If rl > 0. is a discretetime random walk for any h > 0.4 The ruin probability 0(u) is 1 for all u when 77 < 0. at least once.... 1 since St 4 oo) and repeating the argument. 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. Snh u = 00 (the lemma is not needed for (d)). There is also a central limit version of Proposition 1. Considering the next downcrossing (which occurs w.. Part (d) follows by a (slightly more intricate) general random walk result ([APQ]. and < 1 for all u when 77 > 0.1(b)) that the assertion holds as t 4 oo through values of the form t = 0.. Proof The case of 17 < 0 is immediate since then M = oo by Proposition 1.t .2. (c) are immediate consequences of (a). Corollary 1.1) 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)._.5 The limiting distribution of St .2. 2h.
The expression for g+ was proved in Theorem 11. Summing over n. Note that this . The decomposition of M as a sum of ladder heights now yields: 00 Theorem 2 . cf. [APQ] Ch. which we henceforth refer to as the PollaczeckKhinchine formula. As a vehicle for computing tIi(u).1. d. Fig. p < 1.34 or A.. equivalently. We assume throughout rl > 0 or.IIG +II)EG+ . The following results generalizes the fact that the conditional distribution of the deficit ST(o) just after ruin given that ruin occurs (i. 0 Alternatively. n=0 (2. we can rewrite the PollaczeckKhinchine formula as 00 (u) = P (M > u) = (1 .1 provides a representation formula for 0(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 . THE POLLACZECKKHINCHINE FORMULA 61 2 The PollaczeckKhinchine formula The time to ruin r(u) is defined as in Chapter I as inf It > 0: St > u}..1) is not entirely satisfying because of the infinite sum of convolution powers. i.1) representing the distribution of M as a geometric compound. Note that the distribution B0 with density bo is familiar from renewal theory as the limiting stationary distribution of the overshoot (forwards recurrence time ). the formula for the distribution of M follows .2. and we shall here exploit the decomposition of the maximum M as sum of ladder heights.P) E PnBon(u) .6. Thus . Combined with i/i(u) = P ( M > u).6. IV. the ladder heights are i.just before ruin is again B0. oo ). nevertheless.IIG+II) (the parenthesis gives the probability that there are no further ladder steps after the nth ). we may view the ladder heights as a terminating renewal process and M becomes then the lifetime. (2. 11. This follows simply by noting that the process repeats itself after reaching a relative maximum. B(x)/aB. but we shall be able to extract substantial information from the formula. It is crucial to note that for the compound Poisson model.e. 1e. where G+ is given n=0 by the defective density g+ (x) = 3B (x) = pbo(x) on (0.1. 1 The distribution of M is (1. that r(0) < oo) is Bo: taking y = 0 shows that the conditional distribution of (minus) the surplus ST(o). Theorem 2. and we further get information about the joint conditional distribution of the surplus and the deficit.
ladder heights so that the results do not appear not too useful for estimating 0(u) for u>0. THE COMPOUND POISSON MODEL distribution is the same as the limiting joint distribution of the age and excess life in a renewal process governed by B. Feller [143] or Wolff [384]. Again. Theorem A1. For the study of the joint distribution of the surplus ST(u)_ just before ruin and the deficit ST(„). As shown in Theorem 11. the PollaczeckKhinchine formula is often referred to as Beekman 's convolution formula. ST(o)) is given by the following four equivalent statements: B(z) dz.i. ST(o) > y. in this setting there is no decomposition of M as a sum of i. . In the risk theory literature.1 is traditionally carried out for the imbedded discrete time random walk.d. 1) and W has distribution Fw given by dFyy/ dB(x) = x/µB. Theorem 2. (1 .5. there is a general marked point process version. [62]. and the conditional distribution of ST(o)_ given ST(o)_ = z is Bo z) The proof is given in IV.V)W) where V. see Schmidli [323] and references there. where it requires slightly more calculation. 2 The joint distribution of (ST(o )_. 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). However.6.2 and it gives an alternative derivation of the distribution of the deficit ST(o) Notes and references The PollaczeckKhinchine formula is standard in queueing theory. Asmussen & Schmidt [49]. ST(o )) given r (0) < oo is the same as the distribution of (VW. Beekman [61]. (c) the marginal distribution of ST(o)_ is Bo .62 CHAPTER III. W are independent.6. V is uniform on (0. 7r(0 ) < oo) = Q 3 Special cases of the PollaczeckKhinchine formula The model and notation is the same as in the preceding sections.5. the form of G+ is surprisingly insensitive to the form of {St} and holds in a certain general marked point process setup. We assume rt > 0 throughout. cf. (d) the marginal distribution of ST(o)_ is B0. Theorem 2 . (a) 11 (ST(o)_ > x. f +b (b) the joint distribution of (ST( o). see for example [APQ]. The proof of Theorem 11. cf.2(a) is from Dufresne & Gerber [125].just after ruin. cf.
B0 is exponential with rate S and the result can now be proved from the Pollaczeck Khinchine formula by elementary calculations . As shown in 11.3 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.p. Alternatively. Thus . then. I. Let r ( x) be the failure rate of M at x > 0.3.p) E pn S n x n.1)1 00 ( 1 . But claims are exponential . Thus r(x) = S(1 . the formula for P(O) holds in a more general setting.0(u) = pe(aA)" Proof The distribution Bo of the ascending ladder height ( given that it is defined ) is the distribution of the overshoot of {St} at time r+ over level 0. however . Integrating from u to oo.p) = S . 3b Exponential claims Corollary 3.1 0(0) = p = Nl2B = 1 1 +71 Proof Just note that (recall that T+ = r(0)) 00 z/^(0) = I' (r+ < oo) = IIG+II = )3 f(x)dx =l3LB• Notes and references The fact that tp(u) only depends on B through µB is often referred to as an insensitivity property.1 e ax = n1 (n .e. the result follows . The result can. hence without memory.2 If B is exponential with rate S.p)pSe a ( l v)x = p( S . the current ladder step must terminate which occurs at rate S and there must be no further ones which occurs w.p. 0 . 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.O)e(b0)x.. use Laplace transforms. Bon is the Erlang distribution with n phases and thus the density of M at x > 0 is (1 . also be seen probabilistically without summing infinite series . and hence this overshoot has the same distribution as the claims themselves . a further relevant reference is Bjork & Grandell [67]. 1 . For a failure at x.6.
u + oo. (3. (3.3) Equivalently.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 . (b) use stopped martingales . and weights 1/2 for each.p + f u Z(u . the survival probability Z(u) = 1 . (Example VIII. if 3 = 3 and B is a mixture of two exponential distributions with rates 3 and 7. (u) 35eu + 35e6u. we show that expression for /'(u) which are explicit (up to matrix exponentials) come out in a similar way also when B is phasetype. A variety of proofs are available .y)G+(dy ) = f U V(u . 3c Some classical analytical results Recall the notation G+(u) = f^°° G+(dx).3. E. We mention in particular the following: (a) check that ip (u) = pe (60)u is solution of the renewal equation (3. (3.S. then 24 1 V.p + G+ * Z(u) = 1 .s.h.2).4) can be derived by elementary algebra from (3.3) below.3).T+ <oo).y)/3B (y) dy.S.64 CHAPTER III.4) is similar (equivalently.+ >u.g.+ <u.3)). and conditioning upon S.3. u . Corollary 3.y)G+(dy) For the last identity in (3.1 p pBo(u).i(u) satisfies the defective renewal equation Z(u) = 1 .T+ <oo)+P(M> u. (3. 0 Proof Write o (u) as P(M>u) = P(S. II. 2 is one of the main classical early results in the area. just insert the explicit form of G+.1) For a heavytailed B.2) Notes and references Corollary 3. is ?7+ ( u).4) zu P(M > u . The case of (3.+ <U.+ = y yields P(M>u. T+ <00) (3. Then the first term on the r. cf.1. THE COMPOUND POISSON MODEL In VIII. we use the PollaczeckKhinchine formula in Chapter IX to show that b(u) .y)f3 (y) dy.
Corollary 3.pBo[s] no (1 . SPECIAL CASES OF POLLACZECKKHINCHINE Corollary 3.4 The Laplace transform of the ruin probability is 65 fo Hence Ee8M 00 e8uiP(u)du . (3.3 .p)s s /3 .Ps s(.p = (1 .. 111112 or Feller [143]. Some relevant references are Abate & Whitt [2]. it is not surprising that such arguments are more cumbersome since the ladder height representation is not used.3. [APQ] pp.g.(3B[s] 1 . 191).f. 0 Notes and references Corollary 3. In fact./3B[s] . Griibel & Pitts [132].P)pB' (3. either of these sets of formulas are what many authors call the PollaczeckKhinchine formula.P)PB 2(1 . which yields the survival probability as 00 f u }t Z(u) = f f3eRtdt 0 from which (3.(3 . . numerical inversion of the Laplace transform is one of the classical approaches for computing ruin probabilities.PPB2) EM2 = PPB) + QZPBl 2(1 .g. We omit the details (see. Of course. Bo of B0 as m e8u B(du) = B[s] . [APQ] pp.5).p)2 3(1 .3 is standard . see e. In view of (3.4) can be derived by elementary but tedious manipulations. by analytical manipulations (L'Hospital's rule) from (3. eau B(u) du = f PB 3PB SPB 0 o (3.)3B[s]) (3. Embrechts.6) 00 = (I . for example.3 .7). g.p)s .s .s .p) E p"Bo[s]" = 1 . The approach there is to condition upon the first claim occuring at time t and having size x .. Griibel [179] and Thorin & Wikstad [370] (see also the Bibliographical Notes in [307] p./3B[s] which is the same as (3. e.7) and Corollary 3. Also (3. 206207).1 Bo[s] = f oc.8) Proof This can be shown.Ee8M) f ao e8' ( u)du = a8uP (M > u)du = 0 o 1 ( 1+ (1 .5 The first two moments of M are 2 EM .5 can be found in virtually any queueing book.7) s +.5) Proof We first find the m.5).
u) [N(k .1)! k=1 u1 .u + 1 )]k = QZ(u) .4) for Z( u) means f lhu Z(u) = 1.1). Z^ =eR(k.u)]k k! k0 The renewal equation (3.9).s. differentiation yields Z(u) _ /3Z(u) . then p) 1: ep(k u/. . THE COMPOUND POISSON MODEL 3d Deterministic claims Corollary 3.z/'(u) takes the form Z(u) L^J L.1). of (3.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. Assume (3.3+ 18+ J0 Z(uy).Q (k 1 k= n  [O(k .u)]k k! (1 L3) 1: e_O(ku) NIN (k (k .9) follows for 0 < u < 1./32(u .u)]k d 1 u) _ a) n ( du ( k! (1  .u/p)]k ko k! Proof By replacing {St} by {Stu/p} if necessary.66 CHAPTER III.u) [p(k .Q) k=0 k! E e0( = /32(u) .h.u)]k1 ku+1) [/3( k .9) shown for n .3I( 0<y<1)dy Z(y)/3I(0<u./3Z(u .1 < u < n and let Z(u ) denote the r.u) a)Qea" + (1 . For n < u < n + 1. eO('u) [)3(k . differentiation yields Z'(u) _ /3Z(u) which together with the boundary condition Z(0) = 1/3 yields Z(u) _ (1/3) eAu so that (3. we may assume p = 1 so that the stated formula in terms of the survival probability Z(u) = 1 .)3(1 .6 If B is degenerate at p.Q) 3e.
The answer is yes: inserting in (4. and define rce by (4. 00 the standard definition of the exponential family {F9} generated by F is FB(dx) = e°xK(e)F(dx). of F9.3) by t.4) works as well. 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) . it follows that Z(u) = 2(u) for n<u<n+1.3B[9].4) .r.g. B9(dx) = B[9] B(dx). we set up . K(a) = logEe'X = 109f 00 eaxF(dx) = logF[a]. We could first tentatively consider the claim surplus X = St for a single t. say t = 1: recall from Proposition 1. (4.) The adaptation of this construction to stochastic processes with stationary independent increment as {St} has been carried out in 11.a.f.f.4.1 that c(a) = /3(B[a] . and thus (4. 4 Change of measure via exponential families If X is a random variable with c.2). (4. Formalizing this for the purpose of studying the whole process {St}.1) . but will now be repeated for the sake of selfcontainedness.a. 0 Notes and references Corollary 3. we just have to multiply (4. F and c. or equivalently BB[a] = B[^+ Repeating for t 54 1.4.Qe(Bo[a] .3B = .2) shows that the solution is Ox [O]0].f.g. in terms of the c.1) .rc(9) = .g.1) or equivalently. co(a) = rc(a + 9) . (4. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 67 Since Z(n) = 2(n) by the induction hypothesis.6 is identical to the formula for the M/D/1 waiting time distribution derived by Erlang [139].f.2) (Here 9 is any such number such that r.d. See also Iversen & Staalhagen [208] for a discussion of computational aspects and further references.(9).(9) is welldefined. The question then naturally arises whether ie is the c.
3 and claim size distribution B.7) now follows by taking Z = eBST+TK(e)I(G) u Theorem 4 .d. oo) governing a given compound Poisson risk process with arrival intensity.7) Proof We must prove that if Z is FTmeasurable. Let FT = o(St : t < T) denote the o•algebra spanned by the St.d. But let Xk = SkT/n .. Then FB denotes the probability measure governing the compound Poisson risk process with arrival intensity.5) for the density. n) for a given n. for G E FT.i. (4.t.1. the PBT) are mutually equivalent on. t < T.3 Let T be any stopping time and let G E FT.6) F(G) = Po (G) = EB [exp {BST + Ttc(0)} .0e and claim size distribution Be.t. Then the Xk are i. The identity (4. The following result (Proposition 4. .2.S(k_1)Tln.. G C {T < oo}.8) follows by discrete exponential family theory. EeeBSt + tk(B) = 1.9) Proof We first note that for any fixed t. Z is measurable w. .Tic (0)} . the corresponding expectation operator is E9. and PBT) the restriction of PB to FT. . then EBZ = E [Ze9ST _T"(9)I .g.1 Let P be the probability measure on D[0. in particular the expression (4. = exp {BST . (4.r. Xn).2 For any fixed T.nr.FT. . and thus (4. with T taking the role of n) is the analogue of the expression exp{8(x1 + • • • + xn) . v(Xi. replications from Fe (replace x by xi in (4. G]. Ti(a)/n.68 CHAPTER III. BB by (4.4). and dP(T) dP^T) That is. THE COMPOUND POISSON MODEL Definition 4.r. it suffices to consider the case where Z is measurable w. Then P(G) = Fo(G) = EB [exp {BST + TK(O)} .5) for the density of n i. . .FTn) = Q(SkTIn : k = 0.10) . with common c. (4.(9)} (4.i..1) and multiply from 1 to n). and define 09. (4.f. G]. Proposition 4. (4.8) By standard measure theory.
5. t = T . (4. Given FT. Thus. c(a) is a convex function of a.9) holds with G replaced by GT. LUNDBERG CONJUGATION 69 Now assume first that G C Jr < T} for some deterministic T..9) holds for G as well. GT C_ Jr < T}.1(a).g. 5 Lundberg conjugation Being a c. Then GT = G n Jr < T} satisfies GT E FT. (a) rc (a) (b) KL(a) 'Y 'Y Figure 5. subject to the basic assumption ij > 0 of a positive safety loading.1) . Ee [exp { BST +Trc(9)} I(G) FT)] = 1.FT]] = EB [exp { BST + Trc(9)} I(G)] . Letting T t oo and using monotone convergence then shows u that (4.7) holds.10). and hence (4.5. so that PG = EeE0 [exp { 9ST+Trc(9)}I(G)I FT)] = Ee [exp { BST + rrrc(O)} I(G)EB [ exp {9 (ST .1 It is seen that typically) a ry > 0 satisfying 0 = r. 77 Thus.7 1 Some discussion further supporting this statement is given in the next section. The behaviour at zero is given by the first order Taylor expansion c(a) r.ST) + (T .r is deterministic. the typical shape of rc is as in Fig. Thus by (4. Now consider a general G.r)rc(9)}I . according to what has just been proved.1) _ 1 + a. (0) + rc'(0)a = 0 + ES1 a = a (p .(Y) = 13(B['Y] .f. Then G E FT. .
3. G = {T(u) < oo} in Theorem 4. An established terminology is to call y the adjustment coefficient but there are various alternatives around.3. we further note that ( 5. 5. Fig. Example 5 .s).1) .g.4) ELS1 = #L(0) cf. (5. we write FL instead of F7. u It is a crucial fact that when governed by FL. 5.1 Consider the case of exponential claims. (5. Lundberg conjugation corresponds to interchanging the rates of the interarrival times and the claim sizes.1(b).4) yields /3L = b and that BL is again exponential with rate bL =. .a = i(a + 7).3) cf.2 s As support for memory. Fig.1) is precisely what is needed for one of the terms in the exponent .2) 7 Figure 5.1(b). Thus. b[s] = 5/(b . Equation (5. (5. It is then readily seen that the nonzero solution of (5.2 is B(7) = 1 + ^. Note that KL (a) = /L (BL [a] .70 CHAPTER III.QL instead of /37 and so on in the following . Thus B[7] = 6/. Taking T = r(u).1) (or (5.3.1) is known as the Lundberg equation and plays a fundamental role in risk theory . an equivalent version illustrated in Fig. e. THE COMPOUND POISSON MODEL exists . the claim surplus process has positive drift > 0.2)) is 7 = 5/3. 5. and (4. the Lundberg exponent.
r.(u)} . V)(u) < e7u.1p .6 ).2 (LUNDBERG'S INEQUALITY) For all u > 0.3 takes a particular simple form. e(u) has a limit i. To this end.3 (THE CRAMERLUNDBERG APPROXIMATION) i'(u) .Ce7u as u 4 oo. V) (u) = P(T(u) < oo) = EL [exp {ryS.P Y j o' xeryxOB (x) dx /3k [Y] .G+L)(x)) dx ry^+L) J 00 f 0 (1 .5). (5. where C .u be the overshoot and noting that PL(T(u) < oo) = 1 by (5. Since a7' is continuous and bounded.1 (5.e7x)G+(dx).+ E A} in Theorem 4.7) 0 and all that is needed to check is that ( 5. 0 Theorem 5 . which shows that G(L) (dx) = e7xG +(dx) = e7x /3 (x) dx.3. we therefore have ELe7t(u) + C where C ELe7 (00) = µ+) f e7(1 . see A . PL ) with density 1 . ST+ E A] . Then P(ST+ E A) = EL [exp { 7S?+} .(oo) (in the sense of weak convergence w.ascending ladder height distribution and µ+ its mean.5) Theorem 5 . Proof Just note that e(u) > 0 in (5.1. T = T+.7) is the same as (5.1e. we can rewrite this as 0(u) = e"ELe7^(u). (5.6) Proof By renewal theory.G+ L) (x) G+L) (x) IL(+) µ+L) L) where G+L) is the FL. LUNDBERG CONJUGATION 71 to vanish so that Theorem 4. (5.8) .5. G = {S. Letting e(u) = ST(u) .4).t. T(u) < oo] . take first 0 = ry.
that 7 = S .72 CHAPTER III.10) VW = JI c* e° (x) dx = a (B[a] .1 .4). but some tedious (though elementary) calculations remain to bring the expressions on a final form.12) Example 5 .8) yields +L) J0 xel'B ( x) dx (5.3 (this was found already in Example 5. A direct proof of C = p is of course easy: B ['y] d S S (S7 )2 d7S y S 02' C 1p 1p _ 1p /3B' [7] 2 1 P1 p.")G + (dx ) = 1  J0 00 3B(x) dx = 1p. THE COMPOUND POISSON MODEL In principle.1) (5. Then 0(u) = pe(a_Q)u where p = /3/S.11) so that I 7B ['Y](B[7]1) BI [7]Q VP (7) 72 7 (using (5. Using (5. this solves the problem of evaluating (5. or equivalently of how close the safety loading 77 is to zero. of course. we get L where 00 (1 . (5.4 Consider first the exponential case b(x) = Seax.e. Noting that SIG(L)II = 1 because of (5.7).1)) and 7µ+L) = 'y/3 [7] 7 1/0 = /3B ['y] .1 above) and that C = p. . From this it follows. u .1 = ^7 The accuracy of Lundberg's inequality in the exponential case thus depends on how close p is to one.
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)
.
cf.(bA*)". we have according to the PollaczeckKhinchine formula in the form (3.Ps(/3max . THE COMPOUND POISSON MODEL These three equations have solutions 9/3µB2)3 30µa2)2 3µa2) (3) P+ (3) ' 0 .3 )1 } _ 1p 1 .s(/3max . Letting Bo be the stationary excess life distribution. the approximating risk process has ruin probability z.80 CHAPTER III.g./3)M converges in distribution to the 2a exponential distribution with rate S = B' Proof Note first that 1 . (/3max .PBo [s (/3max .3 and Corollary 3. Notes and references The approximation (7. or equivalently that /3 is only slightly smaller than /3max = 1/µ8.1./3)] 1 . Mathematically.8µBo  Ss' u where 6 = µB/µBo = 2µa/µB 2) .1. numerical evidence (e. Grandell [171] pp.p + p { 1 1p ti 1 . [174]) shows that it may produce surprisingly good results.p = (/3max 0)µB.b(u) = p*e. but has an obvious interpretation also in risk theory: on the average. p* _ .p .7) that Ee$(Amex /j)M _ 1p _ 1p Eo [s (0max 1 . 7c The heavy traffic approximation The term heavy traffic comes from queueing theory. That is.P . Proposition 1. 1924.2) was suggested by De Vylder [109]./3)PBo PB . we shall represent this situation with a limit where /3 T fl but B is fixed. the premiums exceed only slightly the expected claims. and hence the ruin probability approximation is b(u) e(bAln)u.(3)2 P PB 2µB 2µB Letting /3* = /3/P. heavy traffic conditions mean that the safety loading q is positive but small. Proposition 7.1 As /3 f Nmax.3* /S.)3 )PBo µB  . Though of course it is based upon purely empirical grounds.
p.7. We return to heavy traffic from a different point of view (diffusion approximations) in Chapter IV and give further references there .l3)M > (/3max . but has an obvious interpretation also in risk theory: on the average .4) suggested by the Cramer Lundberg approximation and Proposition 6. That is .0)u. Mathematically.ze a2unµB laB (7./3)u). and hence 2µ2B 1 .2. 2 provides the better mathematical foundation. in risk theory heavy traffic is most often argued to be the typical case rather than light traffic . then P(u) 4 e6„ Proof Write z'(u) as P((/3max . However . Notes and references Heavy traffic limit theory for queues goes back to Kingman [230]. [APQ] Ch. or equivalently that 0 is small compared to µB . light traffic conditions mean that the safety loading rl is positive and large . It is worth noting that this is essentially the same as the approximation (2) z/i(u) Ce. 7d The light traffic approximation As for heavy traffic .Q T /3max. u * oo in such a way that (3max . This follows since rl = 1/p . Numerical evidence shows that the fit of (7. obviously Corollary 7. light traffic is of some interest as a complement to heavy traffic .p _ 2rl11B PB p. In the setting of risk theory.1 1 . while the approximation may be far off for large u.ryu .2 If . . These results suggest the approximation Vi(u) e6(0_. 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 . the premiums are much larger than the expected claims . we shall represent this situation with a limit where 3 10 but B is fixed. VARIOUS APPROXIMATIONS FOR THE RUIN PROBABILITY 81 Corollary 7. the term light traffic comes from queueing theory./3)u * v.g. VIII).3) is reasonable for g being say 1020% and u being small or moderate. the first results of heavy traffic type seem to be due to Hadwiger [184].B AB ) 6()3max _'3) = However . Of course.
Q limIP ( u) + Q lim z/'(u) Amax &0 amax ATAm. For a more comprehensive treatment.5) follow by integration by parts.e. and hence 00 (U) /3pBBo (u) = 0 / B(x)dx.u. z/' (u) convergence P(U . ( 3 J O B dx.5) u Proof According to the PollaczeckKhinchine formula. 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 . 7e Interpolating between light and heavy traffic We shall now outline an idea of how the heavy and light traffic approximations can be combined. The crude idea of interpolating between light and heavy traffic leads to 0 (u) C1 .T > u). the Poisson case is much easier than the renewal case. 0 u Notes and references Light traffic limit theory for queues was initiated by Bloomfield & Cox [69]. 10 ( u The alternative expressions in (7.u)+. Light traffic does not appear to have been studied in risk theory. Indeed.= 1 aJ 1 a 0+ 1 = = p. Asmussen [19] and references there. i. THE COMPOUND POISSON MODEL Proposition 7. by monotone time T of the first claim . Omax max m. 0(u) /3 J B(x)dx = /3E[U .82 CHAPTER III. Again.(3 10.3 is the same which comes out by saying that basically ruin can only occur at the F(U . Sigman [347].T > u) = J o" B(x + u)/3eax dx . En'=2 • • • = O(/32) so that only the first terms matters.3 As . cf. ao n=1 00 n=1 (u) P) anllBBon(U) onPaBon(u) • Asymptotically. (7. . u Note that heuristically the light traffic approximation in Proposition 7. U > u] = /3iE(U . see Daley & Rolski [96]. [97].
we see that the following limits HT) (u'). to get nondegenerate limits . one may hope that some correction of the heavy traffic approximation has been obtained. The adaptation to risk theory is new. even if the safety loading is not very small. 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 ]. Let OLT) (u) denote the light traffic approximation given by Proposition 7. Thus . "/Qmex Cu) CLT(u ( /3max 0) + O16 CHT( U(Qmaz .8. Al . however.Wmax f(x ) dx + pee6mQ. available. 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. . Another main queueing paper is Whitt [380]. (7. Notes and references In the queueing setting . f / Qmax B(x)dx 00 eQmaxxdx 4/ Qmax 00 QmaxQ amaze" and the approximation we suggest is J B(x) dx = cLT(v) (say). ) M.3 and use similar notation for %(B) (u) = (u). that is. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS 83 which is clearly useless . [84]. B(2). with rate 1/µB = /3max. 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.VHT) ( ax QmQ ) h (B) ( .O(E)(u) 1 (1 ./3)) . no empirical study of the fit of (7. Instead. .O0 M. where further references can be found .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. Substituting v = u(.3).x . z/i(E) (u) = pe(QmaxQ)u.3n. ^IE) exist: 1 (B) HT QmsxQ hm J e e6" 2µE/µE2)'" = e(1 6)" =  Q1Qm. ^ LT Q maxQ m"^ Qlo V LT) ( CHT(v) (say).6) is . (U). _(E) (u).
. Bill is said to be convexly smaller than B(2) (in symbols. B(') <i. whereas (consider x2) B(2) has the larger variance. Taking probabilities. B(' <. this implies St T(l)(u) > r(2)(u) for all u so that 17(I) (U) < oo} C_ {T(2)(u) < oo}. In terms of the time to ruin. or the existence of random variables U(l). U(2) such that U(l) has distribution B('). then Bill = B(2). A weaker concept is increasing convex ordering: B(1) is said to be smaller than B(2) (in symbols. Recall that B(') is said to be stochastically smaller than B(2) (in symbols. most often the term stoploss ordering is used instead of increasing convex ordering because for a given distribution B.s. Rather than measuring difference in size.' 1)(u) < V)(2) (U) for all u. Here convex ordering is useful: Proposition 8. In the literature on risk theory. an equivalent characterization is f f dB(') < f f dB (2) for any nondecreasing convex function f. we shall need various ordering properties of distributions. this ordering measures difference in variability. Proposition 8. B(2)) in the increasing convex order if f BM (y) dy < f 00 Bi2i (y) dy x x for all x. B(') <d B(2)) if B(1)(x) < B(2)(x) for all x.ill(u) < V)(2) (U) for all u. u Of course. one can interpret f x°° B(y) dy as the net stoploss premium in a stoploss or excessofloss reinsurance arrangement with retention limit x. U(2) distribution B(2) and U(1) < U(2) a. THE COMPOUND POISSON MODEL To this end. we can assume that 1) < St 2l for all t. B(2)) if f fdB(1) < f fdB(2) for any convex function f. XI.1 If B(') <d B(2). Proposition 8.84 CHAPTER III. B(2) and PB(1) = µB(2).6. Proof According to the above characterization of stochastical ordering. we have the convex ordering. cf. then . the proof is complete. In particular (consider the convex functions x and x) the definition implies that B(1) and B(2) must have the same mean. for more detail and background on which we refer to Stoyan [352] or Shaked & Shantikumar [337]. then i.2 If B(') <j. Finally. 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). equivalent characterizations are f f dB(') < f f dB (2) for any nondecreasing function f.1 is quite weak.
8. we have Bol) (x) f ' B(1) (y) dy < ' f' B(2) (y) dy = Bo2) (x)• µ 85 I. A general picture that emerges from these results and numerical studies like in Example 8.5 If '0(1)(u) < p(2) (U) for all u and a. then /'(')(u) < 0(2)(u) for all u. (D) (u) < O(B) (U ) for all u.. Bo1) <_d Bo2) which implies the same order relation for all convolution powers. Then V. and here is one more result of the same flavor: Corollary 8.3 If B(1) <. This u implies that D <.1.p ) E /3"µ"Bo2)* n(u) _ V(2) (u) n=1 = Corollary 8.e. B.4 Let D refer to the distribution degenerate at 'LB .2 is the following: Proposition 8. with fixed mean.6 below is that (in a rough formulation) increased variation in B increases the risk (assuming that we fix the mean). it is seen that asymptotically in heavy traffic larger claim size variance leads to larger ruin probabilities. A first attempt would of course be to identify 'variation' with variance. Proof If f is convex. Hence by the PollaczeckKhinchine formula . COMPARISONS OF CLAIM SIZE DISTRIBUTIONS Proof Since the means are equal. we have by Jensen 's inequality that E f (U) > f ( EU).3 provides another instance of this. Corollary 8.1 and µB at 1 so that the safety loading 11 is 10%.4) certainly supports this view: noting that. then B(1) <. u We finally give a numerical example illustrating how differences in the claim size distribution B may lead to very different ruin probabilities even if we fix the mean p = PB. The heavy traffic approximation (7. Proof Consider the light traffic approximation in Proposition 7. A partial converse to Proposition 8. .(1) (.u) = (1 _ P) E /3npnBo( 1):n(u) n=1 00 < (1. larger variance is paramount to larger second moment.6 Fix /3 at 1/1. Example 8. from which the result immediately follows. B(2). B(2). The problem is to specify what 'variation' means.. and consider the following claim size distributions: B1: the standard exponential distribution with density ay. say to p.
lA.9A2e'2r where A.01%. the behaviour of which is governed by a parameter 9.e. 1/)(u. Pellerey [287] and (for the convex ordering) Makowski [ 252].01%. i. all distributions have mean 1. 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. 11 Notes and references Further relevant references are Goovaerts et al.e'\1x + 0. van Heerwarden [189]. we have 0r3 = 2 < or2 = 1 < 02 = 10 < 04 = 00 so that in this sense B4 is the most variable. A2 = 3.0' U0. 1%. In terms of variances o2. 0. A standard example from queueing theory is . However. with the hyperexponential distribution being more variable than the exponential distribution and the Erlang distribution less. Kluppelberg [234].86 CHAPTER III..1358. [166]. We return to ordering of ruin probabilities in a special problem in VI. B.000. Note to make the figures comparable. 0. = 0. B4: the Pareto distribution with density 3/(1 + 2x)5/2. 9 Sensitivity estimates In a broad setting. Let ua denote the a fractile of the ruin function.001 u0. in comparison to B2 the effect on the ua does not show before a = 0. 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).4142. THE COMPOUND POISSON MODEL B2: the hyperexponential distribution with density 0. B3: the Erlang distribution with density 4xe2x. and consider a = 5%.4.) = a. For B1i B2. B3 the comparison is as expected from the intutition concerning the variability of these distributions. One then obtains the following table: U005 U0. which appears to be smaller than the range of interest in insurance risk (certainly not in queueing applications!). and this is presumably a consequence of a heavier tail rather than larger variance.1%.
while /3 = j3 is an estimate.9.3. with 0 the vector of service rates at different nodes and routing probabilities. where the partial derivatives are evaluated at p = 1. Then a p ao = 00 Qa/.2 Consider a risk process { Rt} with a general premium rate p. Thus. In particular . Then the arrival rate /3(P) for { R(P) } is )31p. and hence a _ e(60)u + u e(60)u = ( i + which is of the order of magnitude uV. we may be interested in a'/ap for assesing the effects of a small change in the premium. Then ib = Pe(613)u. it follows that ' is approximatively normal N(0.(u) for large u. the distribution of %3 0 is approximatively normal N(0„ Q/t). 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. and hence the effect of changing p from 1 to 1 + Ap corresponds to changing /3 to /3/(1 + Op) /3(1 .e. increasing in u. and s(9) the expected sojourn time of a customer in the network. if = a e(6A)u. Similar conclusions will be found below. For example. Proof This is an easy time transformation argument in a similar way as in Proposition 1. Then if t is large . a/3 0 .19P a/ .. the standard deviation on the normalized estimate ^/1' (the relative error ) is approximatively . u Proposition 9. i.Ap).1. SENSITIVITY ESTIMATES 87 a queueing network. a2/t).01/2u. say estimated from data. where Q2 = fl ( l2 1113 / _ Ou2v)2. or we may be interested in aV)/0/3 as a measure of the uncertainty on '0 if 0 is only approximatively known. Example 9. the premium rate p and the claim size distribution B. t]. Let R(P) = Rtli. a0 as ao 80 19P . obtained say in the natural way as the empirical arrival rate Nt/t in [0. Thus at p = 1. In the present setting.1 Consider the case of claims which are exponential with rate 8 (the premium rate is one). Assume for example that 8 is known.
namely that of a twoparameter exponential family of the form Bo.()(0 +'0) ' (9 . /3 yields w e(e + Y.((dx ) = exp {Ox + (t(x) .Owe (9. 4) (9 .w(6. However .t. 9. Consider first the adjustment coefficient y as function of 3. but we shall concentrate on a special structure covering a number of important cases. we can rewrite the Lundberg equation as w(9+ y. for the ruin probabilities . but must look for approximations for the sensitivities 0. Similar notation for partial derivatived are used below. x > 0 (9. so that heuristically we obtain '00 50ryu = Coe"u . () Proof According to (9. (. The most intuitive approach is to rely on the accuracy of the CramerLundberg approximation . THE COMPOUND POISSON MODEL As a consequence.5) are similar.3. ^) . Differentiating w. it suffices to fix the premium at p = 1 and consider only the effects of changing .3 70 = 'Ye = = 7 /3(1 we(e +'y.(/3 + y)we(9 + 7.()YC = 1 +y/ /3 \ Q2 From this (9.^)] 1(/3+y)we(9+'y.6 below for some discussion of this assumption).3) follows by straightforward algebra. (9.3. In the case of the claim size distribution B. u Now consider the ruin probability 0 = 0 (u) itself. and write yp = 8y/8/3 and so on .r.1 or Proposition 9.2) (see Remark 9. we cannot expect in general to find explicit expressions like in Example 9.uypCe7u urypO. e. 5) (Q+'Y)[we(0+7.()^ 1 . Viei '0(. this intuition is indeed correct. (3+'y)PC (0+7. Consider first the case of 8/8/3: . ()} p(dx) . various parametric families of claim size distributions could be considered.g. Of course.6) As will be seen below. and the proofs of (9. () = log(1 + y//3).3 or/and B.()wC(e.0 = t/'(u) and the CramerLundberg constant C.w(O.10) below.4). Proposition 9.88 CHAPTER III. mathematically a proof is needed basically to show that two limits (u * oo and the differentiation as limit of finite differences) are interchangeable. (9. 3) ( 9 .
x). O} (9. ()} .4 As u oo. By dominated convergence.10) (9.C). 0(u) = /3 Ju"O B(x) dx + f 0 0(u .x) F(dx ) f u J C F(dx) = C as u 4 oo.3 (see in particular (5.8) (Section 5).x)B(x)dx.12)). we note the formulas Ee. PF = (1 .x).(e"U = = wS(O. ()} . u 0 Proceeding in a similar way as in the proof of the CramerLundberg approximation based upon (9.([a] = exp {w(9 + a.QB(x) dx. () exp {w(9 + a.3(x) dx. we multiply by e7" and let Z(u) = elt" cp(u). u 0 Then Z = z + F * Z and F is a proper probability distribution .8) Letting cp = e0/e/3 and differentiating (9.w(9. () .p)/C'y. the proof is complete. Z(u)/u a C//3PF where PF is the mean of F. z2(U) = e7" J u b(u . () . F(dx) = e'yy/3B(x)dx. But from the proof of Theorem 5.2 of the Appendix ).g. SENSITIVITY ESTIMATES Proposition 9.3) for z/'(u).8).w(O. Z= zl + z2 where zl (u) = e7u J m B(x)dx. ()] exp {w (O + y. we get p(u) = J "O B(x) dx + J U O(u . Hence by a variant of the key renewal theorem (Proposition A1. z2(u) _ 1 ^ e'ri`i7i( u .11) Ee. 11 For the following. BarndorffNielsen [58]). Be.x)B(x) dx + J U W(u . Further write de = [we (9 +'y.w(9.St (U) Ee.9. Combining these estimates . (9.we(9 .9) (9. it holds that 89 a ue ryu a/3 Q(1 P) 7C2 Proof We shall use the renewal equation (3. () . and alsoo zl(u) + 0 because of B['y ] < oo. w((9 + a. () .4t (U)e°`U = which are wellknown and easy to show (see e.
6C do 89 1p 8( 1p Proof By straightforward differentiation. this implies Z = z + F * Z. ()]B(dy) dx x 0 0C T ON O . oo z2 (u) f C . C) .wc(O.w( (0. ()]e7vB(dy) 'fCd 7 c .1) B(dy) 'f '[t(y) . Then as u > oo. THE COMPOUND POISSON MODEL [we(e+7. ()](e7v . 0 x Multiplying by e7" and letting Z(u) = e"uV(u). ^)} [wc (0 + 7.x). 2 z 07P N ue7u (3C de .12) f exp {O y + (t(y) .w (9. )}B(dy)• Letting cp it thus follows from (9.6e7u f "o f[t(y) .90 CHAPTER III. C)] (1 + 7 ) Proposition 9.11).e7x/3 f 00 [t(y) .lB(x) dx = e7uzl(u) + e7°zz(u) + V(u T where zl (u) = . z = zl + z2.we (0.8) that cp(u) . 8^ ue7u. By dominated convergence and (9. u Z2(U) = e7° f u ^/i(u .9)(9.w(e. F(dx) = e7x. ())B(dy) dx.x)f3 f ^[t(y) . 01 (i+) do = +'Y.QB(x)dx.5 Assume that (9. 8 8() 8( (9. ()]B(dy) dx. ()} 1z(dy) = f [t(y) .w( (0.2) holds.wc(9. ^) .w(0.wc (O.
6 Consider the gamma density b (x ) = Sa xa. . w(e. and the proof of the first one u is similar. () = C/9 = a/S.14) de = d( 7!3 76 = 7e = log ( \ ( \5a_ / \SSry ) 72 . < = a. U 7µF from which the second assertion of (9. () ='I'(t./35' a/i'y + aryl 625ry.yu/3C2do u86 89 1p' az/) = 8z/.1edz = 1 exp {Sx + a log x .) log(9) = %F(a) logs where %1 = F'/]F is the Digamma function. ue_Yu 'C2d( 8a 8( 1 p .QS 1 .a/35a&y' ' (9.rye) S 5rya. a /(S . Z(u) /3C 91 o c'o e11(t (y) ..18) (05 + 57 _'3_y .(log r(a) a log S)} • r(a) 1.17) (9. ( 9.9.3ary tog('Finally. t(x) = logx. () = log r(a) .16) (9. by inserting in the above formulas.Sry a/32 + a/37 + /37 . 9 = S.C log(9).13) (9. Example 9.15) (9. and also zj (u) 4 0 because of f Hence. We get w( (0. It follows after some elementary calculus that p = a)3/5 and.a log S = log r(c) . Here (9.pa+1 .2) holds with p(dx) = xldx. that C = a.ry) 5a1 cry (5 . SENSITIVITY ESTIMATES as u 3 oo..Y)a+1 ' (9.w((9. ())B(dy) < oo. we (9..1 .12) takes the form y) alp a .12) follows.
Straightforward but tedious calculations .9) (() .2) with µ(dx) = 2x3zrdx.S[a] = exp {w (9 + a. Be.21og 2. C) = B = Yc = de = do = . further yield . t(x) _ .92 CHAPTER III.CZ try)} 1 C C2 try . C) . ()} = exp {c (C . which we omit in part .l3 of Section 6a needed for the existence of ry becomes e^Q > 1+62 / 2.1 = eXP {c(C .7 Consider the inverse Gaussian density ( b(x) Zx37 exp This has the form (9."62 . w(e. 9 = . for a < a* = z (.2a) } Thus the condition B[a*] > 1 + a* /.log c = 2 In particular. C = .2 .1 16 +ry c C22ry 2( = + 70 We (e.22.2 log (0. THE COMPOUND POISSON MODEL Example 9.w(9.3. () = Cc .3Ee.([Y] .
then BT and hence ryT is undefined. Note that if NT = 0. in which case we can just let t(x) = 0. Thus.cue_7u)3C P Remark 9. To this end. sj=1 and let YT be defined by IKT('ryT) = 0. B are assumed to be completely unknown. by the LLN both F (NT = 0) and F (PT > 1) converge to 0 as T . Finally if k = 1.8 The specific form of (9. thus.2) is motivated as follows. 10 Estimation of the adjustment coefficient We consider a nonparametric setup where /3. the models there (e. In general.oo.2 of the parameters.1) . in u which case the extension just described applies. let NT 16T = ^T . Thus.. BT [a]= NT ^` e"U. (9. we have assumed k = 2 and ti (x) = x. or Ct(x).g..+UNT) > 1. the exponent is either Ox.7 and references there. Notes and references The general area of sensitivity analysis (gradient estimation) is currently receiving considerable interest in queueing theory. However. and hence explicit or asymptotic estimates are in general not possible. That it is no restriction to assume k < 2 follows since if k > 2. and we estimate y by means of the empirical solution ryT to the Lundberg equation. the main tool is simulation. queueing networks) are typically much more complicated than the one considered here.. we can just fix k . the exponent of the density in an exponential family has the form 01 tl (x) + • • • + 9ktk (x). ae t 1lEY u S _ .a. Also. to our knowledge. ESTIMATION OF THE ADJUSTMENT COEFFICIENT Finally. [379] consider a special problem related to reinsurance. the results presented here are new. kT (a) = /T (BT [a] .12) takes the form a = a 93 ar. . Van Wouve et al. if 1 PT = /3TNT(U1+. Comparatively less work seems to have been done in risk theory. for which we refer to X. However .10. then ryT < 0.3C2de 1p' z a = c . 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).
(27)/K'(7)2. then (10./^ B[27] .: N 0.v. 7T a4' 7. More generally. we need a lemma. it is easy to see that we can write \ V1 1 l _ .B[7]2 V2 . THE COMPOUND POISSON MODEL Theorem 10 . since NT /T .1) . B[2'Y]  /3T ) .b[Yp'V21 T { (E[7] . N ( n[7]. 1) r.2) follows from NT/T a4' .i3)(B[7] 1) + (3(BT[7]  .B[7]) + B [7] . B[27] .Q and Anscombe 's theorem.3T .1 As T 4 oo..1) .'Y .a BT[7] I B[7] I + .1) 'YT . a2 where a2 = /3r.'s.T y . B [7]2 (10. (10.B[7]) 0+ Iv/o(b[y].)vl+ N CO. For the proof.1)2 + E[27] . .94 CHAPTER III.: N ()3. If furthermore B[27] < oo.B[7]2 n Hence ( 10.3/ T). vfoVFB[2y]. Lemma 10 . 16T where V1. V2 are independent N (0.B[7]2 }) ( T 0 . Hence KT(7) = (F' + (OT a(B[7l 0))((BT [7] .7 + (.2) rT(7) N N (0.3) Proof Since Var(eryU) = we have B[7].If .2 As T * oo.
OT a 95 u 4 /3. 0 are estimated from data .KT(7) kT(7) K'(7) .1 By the law of large numbers.3). Combining ( 10.E) < 4T(7T) < 4T(7 + E).E ) < 4T(7T) < (7 +0' which implies 'T(ry4) a$' r. If ryT E (7  we have KT(7 . where ryT is some point between ryT and ry.1 can be used to obtain error bounds on the ruin probabilities when the parameters . NT BT [a] Hence r. Let 0 < E < ry.'(y). first note that e7TU N (e7U u2e27Uo'2/T) 7 . Theorem 10.2. it follows that 7T7 KT(7T) .4) + E).c'(7) N (0' T (2(7) / N (0.(ry + e) and hence KT(7 . To this end . Now write KT(7T)  kT(7) = 4T(7T)( 7T 7). BT[a] 3 B[a].e) < 0 < r. Proof of Theorem 10. By the law of large numbers.'T(a) = 1 E Uie°U' a$' EUe "u = B'[a]. Then r.Q.4) and Lemma 10. lcT(a ) 4 /c(a).e. and the truth of this for all e > 0 implies ryT at 'y. 7T E (y . 6"Y (10. I. °7IT) .E) < 0 < kT(7 + E) for all sufficiently large T . NT i =1 n'(a) for all a so that for all sufficiently large T K7 ..10. ESTIMATION OF THE ADJUSTMENT COEFFICIENT which is the same as (10.(ry .e. y + E) eventually.
i .info< „< t S.5%).. Herkenrath [192].1 : Vt = 0. Asmussen [23]) can then be used to produce an estimate of ry. Griibel & Embrechts [292]... various alternatives have been developed.g. [197].. = 1. Deheuvels & Steinebach [102].T = 3TKT ( 21T)IKT (^T)2 is the empirical estimate of vy and fc. Further work on estimation of y with different methods can be found in Csorgo & Steinebach [94]. Csorgo & Teugels [95]. Vt = St . For this reason .d.ueryuU ". if B is exponential with rate 8 so that ry = 8 . A major restriction of the approach is the condition B[2ry] < oo which may be quite restrictive.. This approach in fact applies also for many models more general than the compound Poisson one.C1e"a ( see e.3 or equivalently p > 1/2 or 11 < 100%. wn = inf{t > W.i. Notes and references Theorem 10.T VIT where r7ry..1 is from Grandell [170]. ft.Wn) are i .Q.. t]}. i. and the known fact that the Y„ = max Vt tE[W„1.96 CHAPTER III.g. satisfies b(.e.0) < 5.e.) = a (e. Letting Wo = 0. with a tail of the form P(Y > y) . . V.96 if a = 2. For example . the nth busy cycle is then [Wn1. it means 2 (8 . One (see Schmidli [321]) is to let {Vt} be the workload process of an M /G/1 queue with the same arrival epochs as the risk process and service times U1. Mammitzsch [253] and Pitts. > 0 for some t E [Wn_ 1.. U2. Hipp [196]. 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. Frees [146].f. Wn). 6 < 2. Embrechts & Mikosch [133].
the Poisson intensity is 0 and the claim size distribution is B with m.1 where p = 13µB. Further let 'Yo be the unique point in (0. 97 . 'y) where c(a) attains it minimum value. generalizations to other models are either discussed in the Notes and References or in relevant chapters.f.1) . See Fig.1 (the role of ryy will be explained in Section 4b).g.Chapter IV The probability of ruin within finite time This chapter is concerned with the finite time ruin probabilities 0(u. 0. The notation is essentially as in Chapter III. it is assumed that i > 0 and that the adjustment coefficient (Lundberg exponent) y. The safety loading is q = 1/p . In particular. defined as solution of c(ry) = 0 where ic(s) _ /3(B[s] . the premium rate is 1.s. B[•] and mean AB. 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. Unless otherwise stated. exists.
PL = 6/0 = 1/p > 1). 2 that E [T(u)k. By the likelihood identity III. 1 FL. using that the overshoot l. E[T(u) I T(u) < 00 ] = ELT (U). 1 Exponential claims Proposition 1.5 .1 The claims surplus is {St}.) = e7u ELe'Y^(u) ELT(U)k = e'Yu b ELT(u)k = O(u)ELT(u)k. 7.1) (1.. the time of ruin is T(u) and ^(u) = ST(t&) .1 In the compound Poisson model with exponential claims with rate S and safety loading 77 > 0. FL and independent of T(u). (u) is exponential with rate 0 w. In particular. EL refer to the exponentially tilted process 3 with arrival intensity S and exponential claims with rate / (thus .9).98 CHAPTER IV.(. Var[T(u) I T(u) < 00] = VarL T( U) . PROBABILITY OF RUIN IN FINITE TIME Figure 0. .2) Proof Let as in Example 111. we have for k = 1.u is the overshoot.(U) < 00] = ELT(u)ke'YS.t.(4. the conditional mean and variance of the time to ruin are given by E[r(u) I T (u) < oo] Var [T ( u) I T( u) < oo] /3u+1 J )3 _ 2/3Su+/3+S (S)3)3 (1.r.
1).h./3) .B = a. where = eBu I 1 ."(ry) = 26//32.s.0) . This means that /3(6/(6 .V/ is as asserted.I (1.1)2VarLT(u) + 2 Ca 1I VarLT(u).1.12 Thus the l. the Laplace transform of the time to ruin is given by Eea7( u) = E [eaT (u).1 (6)3)2 which is the same as the r.s.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 . T(u) < oo] fora > r.(PL .1)ELT(u).6a = 0 with solution 0 (the . the 1.6.2). 0 Proposition 1. Since Sr (u) and (PL . u + ELe(u) _ PL . EXPONENTIAL CLAIMS For (1 . is V1rLSr( u) +VarL ((PL .1 /3u + 1 u + 1 //3 = 6/3 6/01 For (1.1) .h.6 + a)0 .h./3 .(yo) = 2V . .1)T(u))2 = UL where = s.1//32 (6/)3 1)2 26(/3u + 1)/(6 .s.1)) ELST(u) ELT(u) (PL .1)T(u) are independent with QL the same mean .1)T(u)) = VarLe(u) + (PL . we have by Wald's identity that (note that ELSt = t(pL .2 In the compound Poisson model with exponential claims with rate 6 and safety loading rl > 0. Let 0 > yo be determined by ^c(0 ) = a. of (1. Wald's second moment identity yields 2 EL (Sr(u) . which leads to the quadratic 02 + (/3 .2) is aLELT( u) .
M(u) T(u) = T + E Tk k=1 where T = T(0) is the length of the first ladder segment .100 CHAPTER IV. T(u) < oo] = EB [exp {aT(u ) . Ti. But by the fundamental likelihood ratio identity ( Theorem 111. and M(u)+1 is the index of the ladder segment corresponding to T(u). are the lengths of of the ladder segments 2. . the result follows.Y1 Y2 Figure 1.0. Cf...'s with rate 5. Y(u) belonging to a convolution semigroup . Note that it follows from Proposition 1.3 that we can write EeaT( u) = eeuEe 017(o).3...4. T2 . PROBABILITY OF RUIN IN FINITE TIME sign of the square root is + because 0 > 0).v. 1.T+ Ti a t U T I 1 a i F.1 where Y1. Fig. T(u) < oo] = e. St Ti F..4) The interpretation of this that T(u) can be written as the independent sum of T(0) plus a r. (1.. Using 5 = 6 .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.. More precisely.. .3) we have E [e«T(u ).v. Y2. are the ladder heights which form a terminating sequence of exponential r.9ST(u) +T(u)!c(0)} .1 .
EN has an Erlang distribution with parameters (N.1.3 Assume that claims are exponential with rate b = 1..d. cf. U2. where U1.T the service times of the customers awaiting service .cos (u/.T) 1 I fl(O)h(0) fdO where (1. T.4. Let {Qt} be the queue length process of the queue (number in system. . Corollary 11. Since U1 .T. the following formula is convenient by allowing t.i (u. If QT = N > 0..v. .T are conditionally i. EXPONENTIAL CLAIMS 101 For numerical purposes . For j = 0.. 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 . including the customer being currently served).T + • • • + UN.6(u) = Vfl/j l(Su.6) fl(9) f2(0) = = fexp {2iTcos9(1+/3)T+u(/cos91) cos (uisin9) . 1). density xN lex/(N ..T.6.i. then VT = U1.1(u. the conditional distribution of VT given QT = N is that of EN where the r. .ST). Note that the case 6 # 1 is easily reduced to the case S = 1 via the formula V.I ex cos B cos j O dB fo " .T. .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.e. 2. let (cf.T is the residual service time of the customer being currently served and U2 . T) to be evaluated by numerical integration: Proposition 1. Then V(u.1 )!. i. UN. UN.0.3 sin0 + 29) f3(0) = 1+/32/cos9. and exponential with rate S = 1. Proof We use the formula ... [4]) 00 (x/2)2n+3 Ij (x) OnI(n+j)! .1. ..k + 1).
k + 1) = 1 k +1 + bj j=00 j=00 00 j=kk+1 j=k1 By Euler 's formulas. 00 E '3j/2 cos(je) j=k+1 00 _ j=k+1 ^j/ zeij = . in particular equations (1.(31/2 cos (( k + 2)9) .i(k +1)e R [/3( klal/2e:0 (01 /2 e .(31/2eie . Then (see Prabhu [294] pp.102 CHAPTER IV.ie . f3(0) .112 l 1( k +1)/2 [ 31/ 2 cos(kO) .3(k +1)/ 2ei(k + l)6 (.38).)3k+1 = e(1+0)T e201/2Tcos 7r 0 e )3(k +l)/2 [31/ 2 cos ( kO) . and define tj = e(1+R)Taj/2Ij(2vT T).44).1 00 ok+lR 00 j=k1 +1)/2e . 8789) 00 E aj j= 00 = 1. PROBABILITY OF RUIN IN FINITE TIME denote the modified Bessel function of order j.)3k +1 tj g'(QT >.31 /2eie L 1)] 1 I/31/2eie . (1.cos (( k + 1)0)] f3(9) Hence the integral expression in (1.cos((k + 2)9)] d9.8 ) yields F(QT > k + 1) . k k2 + $k+1 E bj 00 t j .1)] L _112 /(k+1)/2 [. let I _ j (x) = Ij (x). similar formulas are in [APQ] pp. 912.13(k +l)/2ei(k +1)9 R E .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 .1 R [.
.7) that _ [^ au ak+l (30 k L. expresses V)(0. or. oo (u)31/2e^e)k = )3k z cos(k9) = R k. F(x. u Notes and references Proposition 1.T) which. k! k=O k0 i/z Co Uk ate" o'/z e .3 was given in Asmussen [12] (as pointed out by BarndorffNielsen & Schmidli [59]. Seal [327] gives a different numerical integration fomula for 1 .0(u. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Since P(QOO > k + 1) = flk+1. however. we are concerned with describing the distribution of the ruin time T(0) in the case where the initial reserve is u = 0. equivalently. going back to Cramer. The first formula. We first prove two classical formulas which are remarkable by showing that the ruin probabilities can be reconstructed from the distributions of the St. 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. 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). there are several misprints in the formula there. t) = P . is numerically unstable for large T. from the accumulated claim distribution N.2. Related formulas are in Takacs [359]. 2 The ruin probability with no initial reserve In this section .e = e' COS a cos(uf31/2 sin 0). We allow a general claim size distribution B and recall that we have the explicit formula z/i(0) _ P(7(0) Goo) = p. t )). it follows as in (1. however. E Fk.. T) in terms of F(. The rest of the proof is easy algebra. Ui < x I / (note that P(St < x ) = F(x + t. T). the numerical examples in [12] are correct).
v]. T T o where the second equality follows from II.T].T))dv.T)) does not {Stv)} depend on v.1 In formulas. T]. . f T lStv)} 0<t<T by a 'cyclic translation'. 2. T) = P(Tr(0) > T) = P(M(0.104 CHAPTER IV. co ). Proof For any v E [0.T) T F(x.(6.T)) 1 fT P(M(v. resp. See Fig. we define a new claim surplus process St StM NJ Figure 2. [v.i.t)= {Stv) < SM.(.1.3) with A = (0.0<w<t} St+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. Then 1 .T)dx.T))dv E^T I(M(v. Stv^ _ Define M(v.b (0. meaning that we interchange the two segments of the arrival process of {St}o<t<_T corresponding to the intervals [0. ") } is at a minimum at time t.S„ 0 <t<Tv STS„+St_T+v Tv<t<T as the event that IS. PROBABILITY OF RUIN IN FINITE TIME Theorem 2 . 1 1 .(0.
ST on M (0. v). 0<t <Tv}n{ST<ST Sv+St T+v. For example.Sv.2. If ST < 0..2. Indeed. T) as {ST<St+ vS. Obviously. 0<t<v} = {ST < St . there exist v such that M(v. 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 . in which case there is a last time o where St downcrosses level u.v<t<T} = {ST<StSv. T Theorem 2 .T) occurs. We claim that if M(0. Fig 2.Tt))f(u+t. T)). It follows that if M(0 . w) for some small E. then M(v.T) occurs or not as long as ST < 0. letting w = inf It > 0 : St_ = mino<w<T Sw}. v). this integral is 0 if STv) . then i fT I(M(v. . T) occurs. t) denote the density of F(•. T].T) = F(u+T.T)) dv f T I(M(0. ST > 0. v < t < T} n M(0.2 10(u.T)f(I z /)(0. or it occurs. v). we can take v E (w E.t)dt. v<t<T}n{ST<STSv+St. T) occurs. It is then clear from the cyclical nature of the problem that this holds irrespective of whether M(0. T. T)) dv. Proof The event {ST < u} = { Ei T Ui < u + T j can occur in two ways: either ruin does not occur in [0. we can write M(v. Hence T TE f I( M(v.. where the last equality follows from ST < St on M(0. T) = M(0. v) = M(0. T T o i =1 Let f (•. T)) dv = TEST = T fP(ST < x) dx T T NT 1 f P(ST < x) dx = 1 f P Ui T .xdx.T) and Sv < 0 on M(0. cf. THE RUIN PROBABILITY WITH NO INITIAL RESERVE 105 Now consider the evaluation of fo I(M(v.
T) = C(z. O(T .t). The proof is combined with the proof of Theorem 111.p. z > 0.3 Define r_ (z) = inf It > 0 : St = z}. 0 < t < T.z. For a fixed T > 0. Hence P(ST<u) = 1 .(0.T) = . 2. Proposition 2.2.v. Proof of Theorem 111.2. C*(z. 0 < t <T . t + dt] occurs if and only if St E [u.ST_ t_ and let A(z. define St = ST .T) = {St < 0. ST_ _ z} . which is independent of St and has the stationary excess distribution B0.u+dt]).T)+ J0 T (1V.Tt))P(StE[u. The following representation of T(0) will be used in the next section.106 CHAPTER IV. u + dt] and there is no upcrossing of level u after time t. Let Z be a r. u which is the same as the assertion of the theorem. 0 < t <T.2 Here o. E [t.b(u. which occurs w. ST_ _ z}. {St > .2 . Then P(T(0) E • I T(0) < oo) = P(T_ (Z) E •). ST_ _ z}. PROBABILITY OF RUIN IN FINITE TIME u Q II T Figure 2. {S t > z.
Proof of Proposition 2.2. It follows by division by P(ST(o)_ E [z.3).3. A(z.1) z T . THE RUIN PROBABILITY WITH NO INITIAL RESERVE Then 107 P(r(0) E [T. (2.T) = C*(z. Thus P(Sr(o)_>x. r(0) < oo) = 3R(z) dz JP(C(z. we therefore have P(A(z.T))dT = Off(z) dz P(T_ (z ) < oo) = 3B(z) dz.1) that P(T(0) E [T.T(0)<oc) = f x F(U > y + z U > z) P(Sr(o)_ E [z. Hence integrating (2. u which is the assertion of Theorem 111. T(0) < oo) = OR(z) dz in (2. ST(o)_ E [z. T(0) < oo) B(y B(z) + z) f3B(z) dz = 3 f °^ B(y + z) dz = f3 + x v f B(z) dz. Fig.2.2.1) yields P(ST(o)_ E [z. z + dz].T)) = P(Cx. and since {St}o<t<T. z + dz].T + dT].ST(o) >y. z + dz]. Figure 2.T))f3B(z) dz dT. z + dz]) = P(A(z.T)). {St }o<t<T have the same distribution . z + dz].3 But by sample path inspection (cf. 2. . T + dT] I S7(o)_ E [z. 7( 0) < oo) = P (C(z)) dT.T).
1) where a > r.1 Eear( y) = eyr(a).T + dT] T(0) < oo) dT f ' P(C(z))P(Sr( o)_ E [z. Proposition 2.c(r(a)) l = l er( a)se+at } u yields 1 = eyr(a)Eear(y). who instead of the present direct proof gave two arguments. [329]. r(0) < oo. cf.3 was noted by Asmussen & Kl(ippelberg [36]. because of77>0. Lemma 3.(yo). I L Let ga(x) be the density of the measure E[ear(°). Notes and references For Theorems 2. Tak'ecs [359].2. z + dz].1 and the present proof is in the spirit of Ballot theorems. T(0) < oo) 0 = dT f 0 P(C(z))P(Z E [z.(3(B[r( a)] . In the setting of general Levy processes.s. Let T_ (y) be defined as Proposition 2.108 Hence CHAPTER IV. Note that T_ (y) < oo a.6. PROBABILITY OF RUIN IN FINITE TIME ]P(7(0) E [T. r(a) denotes the solution < 'Yo of the equation a = ic(r (a)) = .2 ga(x) = Qexr(a) f "o eyr(a)B(dy) x .5a). some relevant references are Shtatland [338] and Gusak & Korolyuk [181]. (3. one based upon a result of Asmussen & Schmidt [49] generalizing Theorem 11.1. Theorem 2. Proof Optional stopping of the martingale I er (a) 9 t. ^(0) E dx] (recall that ^(0) = Sr(o)) and write ga[b] = f OD ebxga(x) dx. 3 Laplace transforms Throughout in this section.5 and one upon excursion theory for Markov processes (see IX.3.1) .T+dT]). 2. a martingale proof is in Delbaen & Haezendonck [103].r(a). T(0) < oo) = dTP(T_(Z) E [T. z + dz]. Lemma 3 . see in addition to Prabhu [293] also Seal [326].
r(a) b . u .4 E[eaT (o).ic(b)/b x(b) + a eb"E[eaT(" ). (u . E[ear (o) I T(0) < oo . b . rr(0) < oo) = 1_ r(a) Proof Let b = 0.ST(o)_ just before ruin . £(0) E dx) = /3B(x + dy) dx and hence ga(x) = f e r)/3B(x + dy) _ /3 f x e(v.2 P(Z E [y. Further by Theorem 111. Z = y] = EeaT.5 f 00 o a/r(a) .r(a) = a [B[b] B[r(a)]] . the result follows after simple algebra. (Laplace transform) of the ruin Corollary 3.3. Corollary 3.g.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)) .3 ga[b] = c(b) Proof + b + a . Hence eb"du E[eaT(").f.°° ga(x)dx.2.1] evr(a)B(dy)[ b . LAPLACE TRANSFORMS 109 Proof Let Z be the surplus .x)ga (x) dx where za(u) = f.(v) = ev''(a). It is then easily seen that Za(u) is the solution of the renewal equation Za (u) = za (u) + fo Z. r(u) < oo).3. Then by Proposition 2.x)(a) B(dy)• Lemma 3 .ga [b] 0 TO Using Lemma 3. T(u) < oo] du = Proof Define Za(u) = E [eaT(" ).r(a) The result now follows by inserting /3B[s] = ic( s) +/3+ s and ic(r(a)) =a.T(0) < oo] = 20[b] = za[b] (9a[b] 9a[0])/b 1 . time T(u): u u Here is a classical result : the double m. y + dy].ga [b] 1 .3.
e. P = /3µB > 1. mu ) ( 0 m < ML '(u) 1 m > rL.1) i. T(u) a. the known results are even less explicit than for the exponential claims case.s.00 St = lim .3).1)Er(u) . St/t 1 1/m. = (p . For the second .s.(u ) = o(u) a. for any c > 0 P( Further. That is. The first main result of the present section is that the value umL.3LELU 1 1p' is in some appropriate sense critical as the most 'likely' time of ruin (here C is the CramerLundberg constant). for any m T(u) u . and take basically the form of approximations and inequalities. cf. For the proof.1.. Then given r(u) < 00. t T(u) T(u) T(u) t m = lim = lim = lim Utioo u + Sr(u) u+oo S. uoo u using e.r(u) = Er(u) • ES.w ) v/.2. T(u)/u mL as u + oo. we need the following auxiliary result: Proposition 4. Proposition A1. Theorem 4 .1 Assume 77 > 0.mu D 2 4 N(0. PROBABILITY OF RUIN IN FINITE TIME 4 When does ruin occur? For the general compound Poisson model.mL > E T(u) < 00 ) 40.. Later results then deal with more precise and refined versions of this statement. u 1 ET(u) 1 p1 u where Pw2 = 311B)m3• 7(u) .h(u.6. Then as u * oo. This proves the first assertion of (4. note that by Wald's identity u + EC(u) = ES. i. (4. (u) t.UProof The assumption 11 < 0 ensures that P(T(u) < oo) = 1 and r(u) a4' oo. and hence a. where _ 1 _ 1 1 C ML w(ry) 6B'[7J 1 .110 CHAPTER IV. By Proposition 111.2 Assume ri < 0.
apB ) . and (4. this can be rewritten as u + 1(u) . T(u) < oo f / 00) e7uE L [e_7 (t1).mL U > E. the result comes out not only by the present direct proof but also from any of the results in the following subsections.1).6. 4).1). For (4 . again Proposition A1.mu (2) '• m3/2 µB 7 . PL (•)+ 0. Thus.7 6  11 Proof of Theorem 4.mL >E By Proposition 4. WHEN DOES RUIN OCCUR? and that Ee(u)/u a 0. T) for T which are close to the critical value umL).g. If Z .1. proving (4.N(0. cf.^ N (o.4.5) St . which may be viewed both as a refinement of Theorem 4.6µB2) Z v m (3µB2) Z. of (4.r(u)/m T(u) ti µB2) Z.s.2) follows immediately from u (4.1) is T (u)  U mL P( T (u) < I > E.1).1 is standard.3). 1'r(U) . Notes and references Theorem 4.mu m . implying T(u) . Tu) T( u) . .3. 4a Segerdahl's normal approximation We shall now prove a classical result due to Segerdahl. Theorem 7.1 The l. According to Anscombe' s theorem (e. the same conclusion holds with t replaced by r(u). though it is not easy to attribute priority to any particular author.h. note first that ( Proposition 111.t/m D (2) 111 . T (u) < 00 J 0(u) e7'PL U \ I T u) .2.2 of [86]) and (4. and as a timedependent version of the CramerLundberg approximation.1 (by considering 0(u.
oo ). we need the following auxiliary result: Proposition 4. e'°'/b (u. we get E[ T (u) .f ( (oo)) .(u.) is readily seen to be degenerate at zero if ST(u•) > u and otherwise that of T(v) with v = u .mul h(oo)Eg(Z).w2) r.4 (SIAM'S LEMMA) If 71 < 0.t.6).VU T. oo). PROBABILITY OF RUIN IN FINITE TIME Corollary 4. Proof Define u' = u . and similarly as above we get E[f(^(u)) I Fr(u. with w2 as in (4.))I h(ul /4  ^(u)) I(6 (u') C ) f < ul /4 + f(e(u') .3). letting Z be a N(0.a C4'(y )• ( 4.l:(oo) (recall that rt < 0).L+YWLV'U) . Then for any y. and thus in (4.6) whenever f.v.T(u') given F. Then the distribution of T(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).112 CHAPTER IV.r.5) For the proof.^(T(u')).u1/4)I(S(u') > u1 /4) h(oo) + 0. g are continuous and bounded on [0. resp . Let h(u) = E f (^(u)).4). Using ( 4.ST( u') = u1/4 . Hence Ef (Vu )) 9 (T(u.um.)mu \ h(oo)Eg (r(ul) . E9(Z) (4. O . using that ul/4 .ul/4.e(u') oo w . we can replace T(u) by r(u'). one has 9 (r(u)_rnu) Ef (^(u)) * E. S( u ) < ul/4] < ET(ul / 4) = O(ul/4).T ( u')] = E[ T ( ul /4 . P because of ^(u') . then e(u) and r(u) are asymptotically independent in the sense that. (oo. Then h(u) 4 h(oo) = E f (6(oo)).
WHEN DOES RUIN OCCUR? Proof of Corollary 4. Segerdahl 's result suggests the approximation b(u. where we used Stain's lemma in the third step and (4. CL Fig.yK(ay)• (4.3 in terms of Edgeworth expansions . u needs to be very large). Theorem 4. For practical purposes . in practice one would trust (4. also Hoglund [204]. Cf. y u) < e 7v" . 3 is due to Segerdahl [333].z/)(u . define ay.8) Note that ay > 7o and that 7y > •y (unless for the critical value y = 1/ML). umL + ywL f) = e"P(T (u) < umL + ywL) = EL [e7V "). Thus .7) To arrive at this . .3 ery"z/i(u . T(u) < umL + ywL f.7) to be valid is that T varies with u in such a way that y(T) has a limit in (. ELe7E (") .(ay) = 17 7y = ay .7) to be good. The precise condition for (4. PL(T(u ) < umL + ywL) 113 4 C4(y). 10) '5(u) . just substitute T = umL + ywL in (4.7) whenever u is large and ly(T)l moderate or small (numerical evidence presented in [12 ] indicates .4.4) in the last.5) and solve for y = y(T). 0. oo ) as u * oo. e7v" y < ^'(7) (4 .5 '(u . Notes and references Corollary 4 . For refinements of Corollary 4.dependent version of Lundberg's inequality For y > 0.oo. y > k'(7) . however .T) Ce7"4 (T .9) ( 4 .umL wI V"U u (4. y u) < . that for the fit of (4. yy by 1 K. 4b Gerber's time. see Asmussen [12] and Malinovskii [254].1. The present proof is basically that of Siegmund [342]. see also von Bahr [55 ] and Gut [182].
and generalizations to more general models are given in Chapter VI. In view of Theorem 4. the point is that we want to select an a which produces the largest possible exponent in the inequalities. An easy combination with the proof of Theorem 111.7 i.9): Proposition 4. if y > 1/ic'(y).8). a.r.ay4(u)+ T(u)K(ay ). we have rc(ay) < 0 and get (u) . who used a martingale argument. yu < T (u) < oo 1 l e ayuEav [eT ( u)K(ay). For a different proof. T(u) < yu] < eayu + yUr(ay) Y < eayuEav [ eT(u)K(av )L T(u) < yu} Similarly. .3 yields easily the following sharpening of (4.h(u.5. see MartinLM [257] .v"U. yu ) = < eayuEay [eay^ ( u)+T(U)K ( ay). Numerical comparisons are in Grandell [172 ].Y' (u.114 CHAPTER IV.yu ) = eayuEav [e .t. which may be understood from Theorem 4. 0. yy is sometimes called the timedependent Lundberg exponent. dy) Notes and references Theorem 4 . which shows that the correct rate of decay of tp(u. u Differentiating w. From the proof it is seen that this amounts to that a should maximize ayic(a). Then ic(ay) > 0 (see Fig .2.6.6 It may appear that the proof uses considerably less information on ay than is inherent in the definition (4. Hoglund [203] treats the renewal case. 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( .8).1). 5 is due to Gerber [156 ]. yu) is e 'Yyu/ . f Some urther discussion is given in XI. However.8 below . and hence t. yu 11 < T(u) < oo j < eayu +Y UK(ay) Remark 4. we arrive at the expression in (4.b (u. the bound a7y° turns out to be rather crude .
(4.'(y ). yu ) ayay e ryyu ayay 27ry/3B"[ay] u Proof In view of Stam 's lemma.i(u.4. T(u) suggests heuristically that l t/. not inequalities. This idea is precisely what characterizes the saddlepoint method. if we want EaT(u) .2 yields EaT(u) u u r..c(&) = ic(ay) is < 0. The traditional application of the saddlepoint method is to derive approximations. WHEN DOES RUIN OCCUR? 115 4c Arfwedson's saddlepoint approximation Our next objective is to strengthen the timedependent Lundberg inequalities to approximations.5. then ay > 0. For any a > yo. then the relevant choice is precisely a = ay where y = T/u.ayuEay f eay^ ( u)+T(u)K(ay)..e.e. T(u) < yu] .yyu y l ay I 21ry/3B" [ay] V fU_ u + 00. Proposition 4.^3 ]1/ Bay [lay . We thereby obtain that T is 'in the center' of the Padistribution of T(u). u 4 oo.ayC() . it is instructive to reinspect the choice of the change of measure in the proof. yu) = e. and ii(u) .ay and get Ea e ayf (00) y _ 'Ya( ayKal lay C 1 . Using Lemma 111.13) .ay y 'Yay  ay . (4. (4. Ea . [eT(u )K( ay). the formula 0(u.yu) c ay ..8 If y < 1/ic'(ry).z. (0) r1 (a) ' I. the choice of ay. i.6 with P replaced by Pay and FL by Pay. and in case of ruin probabilities the approach leads to the following result: Theorem 4 .11) ' If y > 1/ r . As a motivation. yu ) eaauEaye . then the solution &y < ay of . we have ryas = ay .: T.12) < yu] Here the first expectation can be estimated similarly as in the proof of the CramerLundberg ' s approximation in Chapter III.(u. and b(u.ay a.
B[ay] /ay &y y(ay . (4. and the equation ic'(a) = 1/y is easily seen to have .1B[ay]1 ) y(ay . The proof of (4.a) .ay + ayl /BLay] .ay)K(ay) ay ayI&YI For the second term in (4. Example 4.I ay &y a ^c'(ay) a (1 +. T(U) < yu] = eyuk (ay)E''ay (ek(ay )"1/2WV.9 Assume that B(x) = eay.11) follows.1)3 = (jB"[ay]l (Pay .c(ay)ul/2W p 2ir = eyu(ay) dz 1 rc(ay ) 2. V < 01 Ir 00 er(ay)"1'2"'x eyur.3(5/(S .(j (1 .13) rigorously. Then ic(a) = . Writing r(u) and W2 = I3ay{.12) is 0 entirely similar. (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 .1)3 = y3/3B"[ay].116 CHAPTER IV.ay) ay +.(ay) _ y(ay .ay ) r. we get heuristically that Eay Ler (u)r(ay).a)2 .4). it seems tempting to apply the normal approxiyu + ul/2wV. PROBABILITY OF RUIN IN FINITE TIME ry I i .l'B)y /(Pay .13).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). a nr=. .13).1) under Pay mation (4. where V is normal(0.7ruw2 Inserting these estimates in (4. The difficulties in making the proof precise is in part to show (4.c'(a) _ /3a/(8 . i B[7ay .1) .1.
ay)3 0 3/2 and (4.tcp) Lo {Wo ( t)}t>0 . 2 = Var(Si ) the variance. A related result appears in BarndorffNielsen & Schmidli [59].= (s. is undefined for a > 5).5. DIFFUSION APPROXIMATIONS solution ay=5 117 V 1 (the sign of the square root is .i )( v s vc ('3 + s _2 / . (5. then { __ . yu) when y < 1/ic'('y) = p/1 . is the drift and o.. 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.8 is from Arfwedson [9]..p.because the c. The mathematical result behind is Donsker's theorem for a simple random walk {Sn}n=o. 0 Notes and references Theorem 4.g./4 ^y for 1/i (u. y) a''y" L '3 _ fl ) 51 /4(1 +1IY)3/4 \.1.1) .f.11) gives the expression '31/4 ( . c a 00.. and next to note that such an approximation in particular implies that the first passage probabilities are close.3+52 1+/351/y' sy 7 B ii[ay] 25 _ 251/2(1 + y)3/2 (5 . . It follows that 5^y =5ay = /«y =f3+ay=l3+d 1+1/y' V 1+^1/y /35 1+1/y /3' ay ay =Qay say =.
(5.3..2) t>o where p = pp = p . oo)). 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.a = Snp) and the inequalities Sn )C . this is an easy consequence of (5. and consider the limit p j p.tp). PROBABILITY OF RUIN IN FINITE TIME where {W( (t)} is Brownian motion with drift S and variance (diffusion constant) 1 (here 2 refers to weak convergence in D = D[0.1) with S. for the purpose of approximating ruin probabilities the centering around the mean (the tcp term in (5. we have o {i!t s: . p.e.p. of which a particular case is the claim surplus process (see the proof of Theorem 5. It is fairly straightforward to translate Donsker's theorem into a parallel statement for continuous time random walks (Levy processes).1)) is inconvenient. where p is the critical premium rate APBTheorem 5 . We want an approximation of the claim surplus process itself.3) takes the form LI S(P) { a2 to2/µ2 + t LI S (P) { a2 ta2/µ2 {W0(t)}. Letting c = a2/pp.t} _ {W_1(t)} . 0 .p/c < St(p) < S((n+l)/ c + Pp/c.7c).118 CHAPTER IV. a2 =/3µB2) Proof The first step is to note that { WC (St P) . such that the claim size distribution B and the Poisson rate a are the same for all p (i. Indeed . However.1 below). St = EN` U= . Lemma 111. cf.1. we shall represent this assumption on 77 by a family {StP) L of claim surplus processes indexed by the premium rate p. Mathematically.3) whenever c = cp f oo as p 1 p..z } {W_1(t )}t>o (5.tcpp) y = { WC (Sct) pct) } {Wo( t)}t>o (5. n/c < t < (n + 1)/c.1 As p J. + {Wo(t ) . and this can be obtained under the assumption that the safety loading rt is small and positive.
(u) ti IG(oo. ('.h. and the r. the continuity argument above does not generalize immediately. has a continuous distribution. this implies P sup 0<t<T a 12 Stu2 /µ2 > u 4 P ( sup W_1( t) > u O<t<T But the l.8 or [APQ] p.. (.4) Note that IG(.6) from Theorem 5. 263) that the distribution IG(•. However.5. (5.h. C. 196. u) =PIT( (u) < x) = 1 .T) IG(Tp2/ a2). [169] or [APQ] pp. is 1/ip (ua2 /IpI. see Grandell [ 168].1.s.1 . 1. u) of r( (u) (often referred to as the inverse Gaussian distribution) is given by IG(x. the continuous mapping theorem yields sup W Sz2 to lP 4 sup Wi(t)• O<t<T O<t<T a2 Since the r. 199. Because of the direct argument in Chapter III.Ta2 /p2). is IG(T.u). we omit the details .( ^ I + e2( \ I . u) is defective when < 0. Corollary 5.2 suggests the approximation u 0(u. u). (ua2 To2 op \ IPI > IG ( T .e. any probability measure concentrated on the continuous functions.s. w..s. and in fact some additional arguments are needed to justify (5.6) This is the same as the heavy traffic approximation derived in III.h.2 As p j p. For practical purposes . (5.1. DIFFUSION APPROXIMATIONS Now let Tp(u) = inf{t>0: S?)>u}.7c.r.5) Note that letting T * oo in ( 5.f I \\\ J \ (5.1 I 7= . ulpI/a2). TS(u)=inf{t>0: WW(t)>u}. we obtain formally the approximation V. ^ p2 Proof Since f 4 SUP0<t<T f (t) is continuous on D a. 119 It is wellknown (Corollary XI.t.. Corollary 5 . ulpl /a2) = e2"1µl / or2.5). since ti(u) has infinite horizon . .
See for example Billingsley [64].3 Consider a family {Ste) } oc claim surplus processes indexed by a parameter 9. PROBABILITY OF RUIN IN FINITE TIME Checks of the numerical fits of (5. For claims with infinite variance. [169]. Furrer.Po = 09µB6 . and which is much more precise. e. All material of this section can be found in these references. such that the Poisson rate Oe. 0) { 2 StQ2 /µ2 D { W_ i(t)}t>o t>o D 2 where p = pe = pe . . Further relevant references in this direction are Furrer [151] and Boxma & Cohen [75]. the B9.1. B0 * Boo. However. the claim size distribution B9 and the premium rate p9 depends on 0. (5. in Asmussen [12].g.5) combined with the fact that finite horizon ruin probabilities are so hard to deal with even for the compound Poisson model makes this approximation more appealing. in the next subsection we shall derive a refinement of (5. that 00 4090.6 of [APQ]. However. we have ^A. and two further standard references in the area are Grandell [168]. We conclude this section by giving a more general triangular array version of Theorem 5.00µB6 + 0.6) are presented.5) and (5.1 and Section VIII. The proof is a straightforward combination of the proof of Theorem 5.Pe.t.5) for the compound Poisson model which does not require much more computation. a2 = ae = 00µa6 Notes and references Diffusion approximations of random walks via Donsker's theorem is a classical topic of probability theory. as an example of such a generalization we mention the paper [129] by Emanuel et al. on the premium rule involving interest. In contrast. the simplicity of (5.120 CHAPTER IV. pe . Theorem 5. The picture which emerges is that the approximations are not terribly precise. Michna & Weron [152] suggested an approximation by a stable Levy process rather than a Brownian motion. in particular for large u.. for more general models it may be easier to generalize the diffusion approximation than the CramerLundberg approximation. The first application in risk theory is Iglehart [207]. as 0 * 00 and that the U2 are uniformly integrable w.r. Assume further that 039µB6 < pe. In view of the excellent fit of the CramerLundberg approximation. Then as 0 _+ 90. pt? 4 peo.6) therefore does not appear to of much practical relevance for the compound Poisson model.
ao (0) _ /c(s + 9 . let P9 refer to the risk process with parameters Q9 = QoB0[9] = QB[9 9o]. 0(0) = 0. and we want to consider the limit 77 10 corresponding to Oo f 0.90) and the given risk process corresponds to Poo where 90 = 'yo. claim size distribution B . 2. P9(r (u) < oo) = 1 for 9 > 0.c(s) = . 77 = 1/p . The setup is the exponential family of compound risk processes with parameters ( B9 constructed in III. this is because in the regime of the diffusion approximation . whereas there we let the given 3B.6. this means the following: 1.90) . For each 9. 3. which we have seen to play an important role for example for the CramerLundberg approximation .90] B(dx). B9(dx) =Bale] Bo(dx) e9z keo)z = B[9 .s and p = /3µB < 1. PB('r(u ) < oo) < 1 for 9 < 0. In this setup.6. Determine yo > 0 by r.9(s) = Ico ( s + 9) .1) . However . 77 is close to zero.4. risk process with safety loading 77 > 0 correspond to 9 = 0 .T) = Peo(r(u) < T) for 90 < 0. CORRECTED DIFFUSION APPROXIMATIONS 121 6 Corrected diffusion approximations The idea behind the simple diffusion approximation is to replace the risk process by a Brownian motion (by fitting the two first moments ) and use the Brownian first passage probabilities as approximation for the ruin probabilities. The objective of the corrected diffusion approximation is to take this and other deficits into consideration.1 > 0. it is more convenient here to use some value 9o < 0 and let 9 = 0 correspond to n = 0 (zero drift)./c(9 . Since Brownian motion is skip free. Bo(dx) = B[eo]B(dx). Then EOU' = Boki[0] = Biki[eo]/E[9o] and "(s) = k(sBo)k(9o). 9o T 0. . In terms of the given risk process with Poisson intensity .Q (B[s] . Let PO refer to the risk process with parameters e9oz Qo = QB[90]. . and we are studying b(u.'(yo) = 0 and let 90 = 'Yo. Then r. this idea ignores (among other things) the presence of the overshoot e(u).
1) . (U. 1) • Since L eatIG (dt.(y) = 0. tu2 ) i IG (t. 9otc0" (0) = 0061 = ul. 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.C. Vargo S.2) . (6. IGu+u2.7(u)/u2} eh(A.e.u. The first step in the derivation is to note that µ = k (0) = r0 (00) .2' where as ususal ry > 0 is the adjustment coefficient for the given risk process. and Si = QoEoU2 = Q B"'['Yo Eo U3 ]. (.S.122 CHAPTER IV. write r = T(u). () where h (A.1) IG(x.. means up to o(u1) terms): . u) = IG(x/u2. bl IG(t81. . _ ^(u) = ST . u) = euh(a ..3) this implies (take u = 1) Ego exp { . One has (6. The corrected diffusion approximation to be derived is (u. .. i.() The idea of the proof is to improve upon this by an O (u1) term (in the following. C . the solution of r. 0o to. (. for brevity. PROBABILITY OF RUIN IN FINITE TIME Recall that IG(x.. Theorem 5. (01.3 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. C.Varo S1 = f30Eo U2 = S1. 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.T) 1+u2 (6. C) = 2A + (2 .
2) is indeed o(u1).v. Note. calculated using numerical integration and Proposition 1. u is Eeazead2/++ Eeaz[1 + ab2/u] where the last expression coincides with the r. just replace t by Tb1/u2. 1. we have p =. distributed as Z . in (3) and (4).52/u where Z has distribution IG (•. The initial reserve u has been selected such that the infinite horizon ruin probability b(u) is 10% in (1) and (3). But the Laplace transform of such a r. it holds for any fixed A > 0 that Ego exp { Ab1rr(u)/u2} . bl I IG I t +2 . is the c. p = 0.z .7. that the saddlepoint approximation of BarndorffNielsen & Schmidli [59] is a serious competitor and is in fact preferable if 77 is large] .6. which is based upon exponential claims with mean µB = 1.1 + 629.f.h. . . 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. however.s. 9o T 0 in such as way that C = Sou is fixed. A numerical illustration is given in Fig.1 below is exact.5) according to (6. the r.ry2 .h.'yu /2)(1 + b2/u)} + Aug 1I J .3).5) Once this is established . the formal Laplace transform inversion is heuristic: an additional argument would be required to infer that the remainder term in (6.2 ).4.3 = 0. .d. In ( 1) and (2). and the dotted line the corrected diffusion approximation (6. .3. (6. however .1 + u2 I Indeed. CORRECTED DIFFUSION APPROXIMATIONS 123 Proposition 6.2). To arrive at (6. we get by formal Laplace transform inversion that C 2 u.v. The solid line represents the exact value . 6 . of (6. that whereas the proof of Proposition 6.exp { h(A.s. 1% in (2) and (4). of a (defective) r.1 As u + oo.
Similarly. Note that the ordinary diffusion approximation requires p to be close to 1 and '0 (u) to be not too small.^) .124 0.() Lemma 6.19)2 11 20 20 i0 T 1n0 Figure 6.02 I 90 120 160 2W A0 Z WT 40 80 120 160 100 240 280 T 111 WI. see Asmussen [12]..T) 0.08 a. the fit at p = 0.01 0.0 0.114 0. A51 7(SAT 3 3 h(X.2 e. and all of the numerical studies the author knows of indicate that its fit at p = 0.u2 2u3 (e . it gives the right order of magnitude and the ordinary diffusion approximation hopelessly fails for this value of p. PROBABILITY OF RUIN IN FINITE TIME 0. OM 0.07 0. BarndorffNielsen & Schmidli [59] and Asmussen & Hojgaard [34]. .OOIi O.7.08 0. (Inc 0s 0.111 W(U.1 proceeds in several steps.4 may not be outstanding but nevertheless. For further numerical illustrations.1 W IU.T1 00.(061 0. The proof of Proposition 6.aa1 .1 It is seen that the numerical fit is extraordinary for p = 0.W21 0.011 L1 60 T IM 11.05{ 0.TI CHAPTER IV.00 0.T) 111 0..7 or at values of Vi(u) like 1% is unsatisfying.EB 0 p ex p ( 7 S h ^)u .199 0.
. 1 = PB(T < oo) = Eo0 exp 125 {(B .6) u U3 Lemma 6 . the formulas Po(C(0) > x) Po(C(co) > x) imply 1 °° Po(ST(o) > x) = EIU fIP0 (U>y)dy . 3 lim Eof (u) = EoC(oo) = a2 Ep = 3EoU2 uroo Proof By partial integration .T (co (8/u) .s. exp ue } al 1J 3 exP I [2). () + C and note that 2 KO (0) = 102.C)C/u .h. 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 4.2u (B3 . 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 ..(3)Eea LauT exp i 3J .. CORRECTED DIFFUSION APPROXIMATIONS Proof For a>0.7) 2 2 .2 behaves like C l Eeo eXp r _ ^81T 1 Sl u2 1 u 2u3 [1+h(AC) S .. + a1b2 + .1) h(A.61a2T (B3 . () 62 Eeo exp u u2 J .+ h (A.00)(u +C)  'r (. (6. the result follows.4) that the r.(3) J t _ aa1T l + eh(A.r0 (00)) } Replacing B by 8/u and Bo by C/u yields e(B() = E eo exp { (e .C2 = 2).co ((/u)) } Let 8 = (2a + (2)1/2 = h().co (e) . in Lemma 6.6. (6.3 EoU2 + 103OoEoU3 + " 2 6 Using d2 .4 Ea.
4.h (A. C) ( 1+ u2 The result follows by combining Lemma 6 .S) d e 62 .\+ (2 (3 e 2u [ (2. Thus a2 y = 290 + O (u2). Thus by Taylor expansion around ( = 90u. [2+ (2 .2) for O(u) (indeed .2u [2A+ (2 3 . letting formally T * oo yields 7/)(u) C'e7u where C' = e7a2). PROBABILITY OF RUIN IN FINITE TIME The last term is approximately (e 3 (3) 27.e h(aS)h (^^ 262 exp {_h(.2 (^/2 + 3y9o + 390) + O(u3). () by h(\.s. we get the correct asymptotic exponential decay parameter ^/ in the approximation ( 6.() I 1 + u2 ) y .126 CHAPTER IV. There are two reasons for this : in this way.(2A + ()1/21 exp S h(A. 0 The last step is to replace h(A. yu/2) h(A.\ + () 1 2 / .() . and inserting this and 9o 2 = S/u on the r. yu/2) 11+ 62 I} S 1 \\\ u/11 l 62 (3 2u 2A Proof Use first (6. yu/2). and the correction terms which need to be added cancels conveniently with some of the more complicated expressions in Lemma 6. () .2. l Lemma 6 . 2 and (6. yields +90 62 0 + O(u 3) 2u2 +O(u 3)..7) and using eh(a. 5 exp { _h(A) (1 + / y u J)) exp 1. we get h(A.x.1 (y/2 + Oo)u .h.6  d h(A. 2 + 00 = .6) and 7co (Oo) = ico('y + Bo) to get 0 = 21 (^/2 + 2y90) + 1112 (_Y3 + 3_Y200 + 3y9o) + O(u4).
Fuh [148] considers the closely related case of discrete time Markov additive processes. The adaptation to risk theory has not been carried out. Hogan [200] considered a variant of the corrected diffusion approximation which does not require exponential moments. 7 How does ruin occur? We saw in Section 4 that given that ruin occurs. i.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. ()} 3 h (A. 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. () (i+a ) 2A + (2 . that is. with the translation to risk processes being carried out by the author [12].7.(i+ 62 exP{ h(A. the approach to the finite horizon case is in part different and uses local central limit theorems. 'yu/2) 127 ( i+ M pz^ exP { h (A. HOW DOES RUIN OCCUR? exp { h (x.e. ()} .5 in Lemma 6.4. the same as for the unconditional Lundberg process. . () I 1 + u2 )I 2u L 2A+C2_(2 exp { _h.1: Just insert Lemma 6. 0 1 Proof of Proposition 6. The answer is similar: the process behaved as if it changed its whole distribution to FL.T) has not been carried out and seems nontrivial. the analogous analysis of finite horizon ruin probabilities O(u. His ideas were adapted by Asmussen & Binswanger [27] to derive approximations for the infinite horizon ruin probability 'i(u) when claims are heavytailed. and to the Markovmodulated model of Chapter VI in Asmussen [16]. The corrected diffusion approximation was extended to the renewal model in Asmussen & Hojgaard [34]. We shall now generalize this question by asking what a sample path of the risk process looks like given it leads to ruin. In Siegmund's book [346]. the 'typical' value (say in sense of the conditional mean) was umL. () I 1 + u 2 ) } S 1 .
Theorem 7 .1. F(u)c] P(r(u) < oo) ?P(U) < EL[e7u.(u)_ is that i. stating roughly that under F(u).3L and the claim size distribution from B to BL. Recall that .ST(u)}t> o is just an independent copy of {St}t>o). P(u) and rate = aL w. Then also P(u)(F(u)) + 1. . PROBABILITY OF RUIN IN FINITE TIME changed its arrival rate from 0 to /3L and its claim size distribution from B to BL.2 If B is exponential.(u). so in the in the proof. Proof P(u) (F(u)c) = F(flu)c.FT(u) = o' (T(u ). and let M(u) be the index of the claim leading to ruin (thus T(u) = Ti + T2 + .TT(u) _measurable. {ST(u)+t . we give a typical application of Theorem 7. F(u)c] ti e' ru]PL (F(u)`) > 0.vi(u) Ce'Yu Corollary 7. the numerator becomes e'ruELe7^ (u)PL(F( u)t) = e7uCFL (F(u)°) when F(u) E .128 CHAPTER IV.. FL As example. Note that basically the difference between FT(u) and .F.T. In fact.1 Let {F(u)}u>0 be any family of events with F(u) E F. . {St}0< t<T(u)) Proof Write e'rsr(u ) = e'rue'r£(u). then P(u) and FL coincide on . the Poisson rate changes from .r. (u) and satifying PL(F(u)) * 1..F. Recall that 13L = (3B[ry] and BL(dx) = e'rxB(dx)/B[7]. In the exponential case.EL[e7S.t.(u) is not . . r(u) < oo) . 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.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. ^(u) is exponential with rate 8 w.3 to .(u)_ and similarly the denominator is exactly equal to Ce7u. + TMOO ). u * oo.t.(u)_ and ^(u) are independent .r.
see further XI. From a mathematical point of view. HOW DOES RUIN OCCUR? Corollary 7. This is currently a very active area of research.eaLx.(1 . Notes and references The results of the present section are part of a more general study carried out by the author [11]. Proof For the first assertion. take I(Tk < x) .7. 129 M(u) >2 I(Tk < x) M(tu) p(u) M(u) >2 I(Uk < x) BL(x). .3. A somewhat similar study was carried out in the queueing setting by Anantharam [6].eALx) M(u) k=1 u The proof of the second is similar.3 M(u) pcu) 1 . the subject treated in this section leads into the area of large deviations theory. who also treated the heavytailed case. however. the queueing results are of a somewhat different type because of the presence of reflection at 0.
This page is intentionally left blank .
with Nt = # {n: Un <t} the number of arrivals before t. with common distribution B.(1.1 (T1 = a1). the distribution Al of T1 is A as well.. and M is the maximum of {St}.Then no matter the distribution Al of T1i B. D'2. Then the arrival process is stationary which could be a reasonable assumption in many cases (for these and further basic facts from renewal theory. the one corresponding to the stationary case by 00)(u). Var(St) = 11Ba2A + I4AaB lim t goo t PA 131 . AA t*oo lim St = lim ESt t tioo t = p . . We use much of the same notation as in Chapter I. T3.. of the risk process form a renewal process: letting Tn = Qn . are i. the Tn are independent. {St} is the claim surplus process given by I... U2.1 Define p = !µ...7).d. The ruin probability corresponding to the zerodelayed case is denoted by 1/'(u). the claim sizes U1.1). and the one corresponding to T1 = s by 1/i8 (u). with the same distribution A (say) for T2.Q. In the socalled zerodelayed case.. A different important possibility is Al to be the stationary delay distribution A° with density A(x)/µA. see A.i.1. .Chapter V Renewal arrivals 1 Introduction The basic assumption of this chapter states that the arrival epochs O'1. Thus the premium rate is 1. . Proposition 1. . r(u) the time to ruin.
we get similarly by using known facts about ENt and Var Nt that Nt Var(St) = Var E Nt U. Here are two special cases of the renewal model with a similar direct interpretation: Example 1. Nt + EVar U. one could imagine that the claims are recorded only at discrete epochs (say each week or month) and thus each U. Nt = Var(PBNt) + E(4Nt) Q2 2 0`2 A tpB B + o(t). after E. the .1) follows . The simplest case is of course the Poisson case where A and Al are both exponential with rate 0. Proposition 1.1 gives the desired interpretation of the constant p as the expected claims per unit time. by the elementary renewal theorem (cf. This has a direct physical interpretation (a large portfolio with claims arising with small rates and independently). For (1 . and (1 .132 Furthermore for any a > 0.3 (SWITCHED POISSON ARRIVALS) Assume that the process has a random environment with two states ON. CHAPTER V. A.1 of the safety loading appears reasonable here as well.Nt] * a/PA. The renewal model is often referedd to as the Sparre Andersen process. stating that E[Nt+a . 2).t. s + t µA PA 0 Of course. the definition 77 = 1/p .St] = a(p . say at a. 3) follows similarly by Blackwell 's renewal theorem. Nt ESt = E E UI Nt t = ENt•pB .1). but the arrival rate in the ON state is .1) ENt/t + 1/µA. OFF. However .2 (DETERMINISTIC ARRIVALS) If A is degenerate.a is really the accumulated claims over a period u of length a.0 > 0. Thus. Example 1 . such that no arrivals occur in the off state. From this ( 1. RENEWAL ARRIVALS lim E [St+a . If the environment is Markovian with transition rate A from on to off and u from OFF to ON. t 4oo Proof Obviously. Sparre Andersen whose 1959 paper [7] was the first to treat renewal assumptions in risk theory in more depth.
INTRODUCTION 133 interarrival times become i.s < u). and for historical reasons. .y)B(dy).} with {S(d)} a discrete time random walk with increments distributed as the independent difference U . and then the whole process repeats itself).T between a claim U and an interarrival time T.1. the relevance of the model has been questioned repeatedly.. in general the mechanism generating a renewal arrival process appears much harder to understand. S o<t<oo n=0..t.oFF}..s. Ao. Therefore.2. (1. u For later use. A is phasetype (Example 1.. integrate (1. and the present author agrees to a large extent to this criticism. 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 . the first term represents the probability F(U1 .i. (an arrival occurs necessarily in the ON state. if for nothing else then for the mathematical elegance of the subject.1. initial vector (1 0) and phase generator 11 However. we have From this the result immediately follows.4) w. However.4) with phase space {oN. For the stationary case. More precisely. 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^.4) fo Indeed..1. U1 .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.r. arrival times.. Proof The essence of the argument is that ruin can only occur at claim times. The following representation of the ruin probability will be a basic vehicle for studying the ruin probabilities: Proposition 1. we feel it reasonable to present at least some basic features of the model.d. as follows easily by noting that the evolution of the risk process after time s is that of a renewal risk model with initial reserve U1 . the fundamental connections to the theory of queues and random walks.s > u) of ruin at the time s of the first claim whereas the second is P(r(u) < oo.
i. we shall be able to compute the ruin probabilities i(i* (u) for this model very quickly (. t.3* pB.1. Theorem 2 .134 CHAPTER V.a*PB• > 1.1 r* (u) One situation where this model is argued to be relevant is life annuities.1) . the remaining part of the prepayment (if any ) is made available to the company. 2. Using Lundberg conjugation . < 1.0* (u) = P (rr* (u) < oo) where rr* (u) = inf It > 0: Rt < 0} ) .Ut. A typical sample path of {Rt } is illustrated in Fig. are independent of {Nt} and i. A simple sample path comparison will then provide us with the ruin probabilities for the renewal model with exponential claim size distribution. U Figure 2. then 0*(u) = e 'r" where ry > 0 is the unique solution of 0 = k*(ry) = *(B*[ry] . If . each of which receive a payment at constant rate during the lifetime .1 If.d. St = t . 00). The initial reserve is obtained by prepayments from the policy holders. with common distribution B* (say) concentrated on (0. The compound Poisson model with negative claims We first consider a variant of the compound Poisson model obtained essentially by signreversion . That is . resp . then 0 * (u) = 1 for all u > 0. (2. b=1 !=1 where {Nt } is a Poisson process with rate .1) +ry. RENEWAL ARRIVALS 2 Exponential claims.3* (say ) and the U. At the time of death . the claims and the premium rate are negative so that the risk reserve process . the claim surplus process are given by Nt Nt Rt = u+^U.
Then the function k* is defined on the whole of (oo.2 sup St = inf St = 00 t>o t>o and hence 0* (u) = 1 follows. .2 Assume now . T_ (u) = inf { t > 0 : St = u 'r* (u). EXPONENTIAL OR NEGATIVE CLAIMS [Note that r. Proof Define 135 St =u . (a) is*(a) (b) .2(a)..1. B. 0) and has typically the shape on Fig.UB. and thus 1 = P(T. and the Lundberg conjugate of {St} is { St } and vice versa.(u ) < oo) = E {e7sr_ (u). Let B(dx) = ^e7x B*(dx). T_ ( u) < 001 e7"P(T_ (u) < oo) = e"V)* (u).* (a) = log Ee'st I. > 1 .g. Hence y exists and is unique. Then the c. Fig. 2.3*.Rt. If I3*pB* < 1. of {St} is c(a) = is*(a7). B* [7] and let {St} be a compound Poisson risk process with parameters .s. 0 Now return to the renewal model. Then { St } is the claim surplus process of a standard compound Poisson risk process with parameters 0 *. 2. cf. the safety loading of { St} is > 0.2.0. St=Rtu=St. Hence T_(u) < oo a.(a) 7 Figure 2. Define T_ (u) = inf It > 0 : St = u} . then by Proposition 111.2(b). B*. Since ic'(0) < 0.f.
which has the advantage of avoiding transforms and leading up to the basic ideas of the study of the phasetype case in VIII. Taking m. 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 . alternatively termination occurs at a jump time (having rate 8). and from Fig . However. Hence M* max {To + Ti + • • • + Tn .+Tn U1 Un. we get Ee'M(d) = Ee°M* _ Y/(.1) + ry = 0 which is easily seen to be the same as (2. A variant of the last part of the proof.7r+ 7r EeTo b/(Sa) + +. 3* = 6.•.Un } = max St = t>0 n=0. RENEWAL ARRIVALS Theorem 2 . Then B* = A.. and 5PA > 1. and hence the failure rate . To + max {Ul+•••+UnTI.. with the probability that a particular jump time is not followed by any later maximum values being 1 . u Hence P(M(d) > u) _ 1r+e'r".'s and noting that V)*(u) = P(M* > u) so that Theorem 2.1 it is seen that ruin is equivalent to one of these values being > u.2)....g..1.• • • .a) = 1 . 1) means that 8(A[ry] .e.f. To + M(d) in the notation of Proposition 1.2 If B is exponential.1.Ui . Now the value of {St*} just before the nth claim is To +T1* +.Y] (2. with rate S (say)..Ti = U1.. According to Theorem 2.1 means that M* is exponentially distributed with rate ry..Tn} n=0. respectively..Y a I.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 . then .4 goes as follows: define 7r+ = P(M(d) > 0) and consider {St*} only when the process is at a maximum value..Tr+.136 CHAPTER V. T2 = U2.1. and (2 ..4.)(u) _ 1r+e7" where ry > 0 is the unique solution of 1 = Ee'Y(uT ) = S 8 A[..u+ and lr+.. the failure rate of this process is y.
Bads (dx) = . 111.4.5.T to F(d)(x) = eK^d^(«) ^x e"vFidi(dy) 00 K(d) (a) = log F(d) [a] = log B[a] + log A[a] . a ladder step is the overshoot of a claim size. letting P(d) refer to the renewal risk model with these changed parameters . Putting this equal to y. However.3. 0 3 Change of measure via exponential families We shall discuss two points of view. 3a The imbedded random walk The key steps have already been carried out in Corollary 11.y/b.2. : S(d) > u} . CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 137 is b(1. The probability that the first ladder step is finite is 7r+. Thus a ladder step terminates at rate b and is followed by one more with probability 7r+. Hence the failure rate of M(') is 6(1 . we have ] A[a )3] E«d'efl' = Bad> [a] A ad> [Q] = B[a +. resp. Furthermore.. hence exponential with rate b.7r+) = ry and hence P(M(d) > u) = P(M(d) > 0)e7u = 7r+e'r".7r+) and hence r+ = 1.2. the relevant exponential change of measure corresponds to changing the distribution F(d) of Y = U ... 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^.. consider instead the failure rate of M(d) and decompose M(d) into ladder steps as in II.B(dx). the imbedded discrete time random walk and Markov additive processes.3 A[a] B[a] F( d) [a +)3] F(d) [a] = Fad) [^] Letting M(u) = inf in = 1. we see that ry = 6(1. B^d) where Aad> (dt) = ^[ a] A(dt).6. This follow since.7r+). which states that for a given a.
in the easiest nonexponential case where B is phasetype. For the stationary case.138 CHAPTER V. (a) '(u) < eryu. and claim (a) follows immediately from this and e (u) > 0. Consider now the Lundberg case. 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 .. we get: Proposition 3. 7µA .1 implies Cu) = e«uE ( 7d)e«^(u) . Corollary 3.(u) .1).T is nonlattice. the evaluation of C is at the same level of difficulty as the evaluation of i/i(u) in matrixexponential form. ik. This is known to be sufficient for ^(O) ]p (d) ([APQ] Proposition 3. In fact. E(d)e 1' (u).u the overshoot . VIII.Ce"u where C = limu. Proof Proposition 3. let 7 > 0 be the solution of r.1 For any a such that k(d)' (a) > 0. to converge in distribution since p(yd) (r(0) < oo) = 1 because of r (d)' (y) > 0.2 p. 187) and thereby for ^(u) to be nonlattice w. For claim (b). just note that F7(d) is nonlattice when F is so . O(u) = eauE (d)ea{ (u)+M(u)K (d)(a) . We have the following versions of Lundberg' s inequality and the CramerLundberg approximation: Theorem 3 .2 In the zerodelayed case. cf. 00)(u) .C(°)eryu where C(O) = C0[7] .3 For the delayed case Tl = s.C8e7u where Cs = Ce78B[7].e.. (d) (7) _ 0.t. i .p)/($B'[7] . (b) V)(u) .1) is explicit given 7.. provided the distribution F of U .r. RENEWAL ARRIVALS be the number of claims leading to ruin and ^(u) u = SM(u) .4.
(°) ( Ce8B[7] Ao(ds) similar manner. Equating this to tch(s) and dividing by h(s) yields h'(s)/h(s) _ . According to Remark 11. (s. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES Proof Using (1.. where G is the infinitesimal generator of {Xt} = {(Jt.3. 3b Markov additive processes We take the Markov additive point of view of II.y) B(dy) 0 For the stationary case.5.5. The expressions are slightly more complicated and we omit the details. delayed version of Lundberg's inequality can be obtained in a e7u.1) = C(O).9. B(x) = o(e7x) and dominated convergence.(s.h'(s).dt ) eadt = h ( s) .. (u + s . 0) = ah (s) . Let P8f E8 refer to the case Jo = s. 0) = tc(a)h(s).a . St)} and h. we get r u +8 e"8(u) 139 e7uB(u + s) + 4 0 + L 00 J e7(v8)e7(u+8v). IPA 0 Of course.4).0 ) = Eo[ha ( Jdt.St)} can be defined by taking Jt as the residual time until the next arrival. E8h0 (Jdt. y) = e°yh(s). Here K. For s > 0. we look for a function h(s) and a k (both depending on a) such that Gh. Sdt) = h(s .1) (normalizing by h(0) = 1)./c. The underlying Markov process {Jt} for the Markov additive process {Xt} = {(Jt.Sdt] = Ee'uh(T) means 1 = f ' e°^B(dy) f ' h( s)A(ds). h(s) = e(a +x( a))8 (3. To determine boundary 0. we invoke the behavior at the 1 = h«(0. another use of dominated convergence combined with Ao[s] = (A[s] 1)/SPA yields 00 u) e7u iP8(u) Ao(ds) + f 0 = CB['Y](A[y] . 0 0 .dt(ah ( s) + h'(s)) so that Gha ( s.
.140 CHAPTER V.2). Proposition 3. Further. however..s and P(d).5.e. i. Ba where Aa (dt) . rc(a) = 0 (B[a] . Aa as asserted .s governing {(Jt. .c(a)] B[a] Proof Pa. . T2.13]A[b .a ..s is the probability measure governing a renewal risk process with Jo = s and the interarrival distribution A and the service time distribution B changed to Aa. U.rc(a)] = B = Ba[13]Aa[5].. Note that the changed distributions of A and B are in general not the same for Pa. = J8 = T2. (3. (3. [e1U1 + 6T2ea ( U1s)stc ( a)e(a+K(a ))( T2s)I B[a +. . resp.s(Jo = s) = 1 follows trivially from Lo = 1. .. we can now for each a define a new probability measure Pa. RENEWAL ARRIVALS B[a]A[a . ] = E.tK( a)e.e(«+k(a))t esy A(dt).2) means 1 = B[a]/3/(/3+a+rc (a)).1)a in agreement u with Chapter III.rc(a)] B[a] A[a . since JT.( a+r' (a))(Jt s) h(s) where c(a) is the solution of (3. Remark 3 . the determination of the adjustment coefficient ry where the defining equations rc(d) (ry) = 0 and rc(ry) = 0 are the same.rc(a)] = 1. Ba(dx) = B(dx). [a + /3] A[b . An easy extension of the argument shows that U1.c(a)] which shows that U1. . .2) As in 11. Ease AU1+6T2 [ AU1+6T2 = Ea a LT. An important exception is.8. J n+1 u are independent with distribution Aa for the Tk and Ba for the Uk. A[a . T2 are independent with distributions Ba. St)}too by letting the likelihood ratio Lt restricted to Yt = a((J. 5 For the compound Poisson case where A is exponential with rate .S„):0<v< )be Lt = eaSt tK(a) h(Jt) = east . resp.4 The probability measure Pa..a .
which is the same as the asserted inequality for 0. Let M(u) be the number of claims leading to ruin .4..t.4. see e.1 and n.(u.r.)+1 e J j e(ay+w(ay))8 e . Proof As in the proof of Theorem IV. The virtual waiting time Vt at time t is the residual amount of work at time t. not after time T. and U„ the service time of customer n. u The approach via the embedded random walk is standard. see in particular Dassios & Embrechts [98] and Asmussen & Rubinstein [45]. and define yy = ay . XII.rc( ay)] = e(aa+(r ))sb[a ]e7yu L y1 In particular. T(u) < yu h(JT(u)) < eayu+yuk(ay ) ( Eia y Le(a(+k(ay))s v.s eaysr(")+r(u ) K(ay) h (s) . . or FCFS = first come first served) queueing discipline and renewal interarrival times.6 Let y < let ay > 0 be the solution of ic'(ay) = 1/y. Then J(rr(u)) = TM(u)+1 and hence Ws(u. 2.yx(ay). .5 Proposition 3. the amount of .ay+ray))TM(.4.y yuAa y [ay + K(ay) . defined as the single server queue with first in first out (FIFO. The actual waiting time Wn of customer n is defined as his time spent in queue (excluding the service time). Notes and references 4 The duality with queueing theory We first review some basic facts about the GI/G/1 queue. yu ) < e7yu. The claim for the zerodelayed case follows by integration w. yu).g. the time from he arrives to the queue till he starts service.yu ) e7vu A[ay . and assume that T„ is the time between the arrivals of customers n . yu) = F'ay. A(ds). [APQ] Ch. . Using the Markov additive approach yields for example the following analogue of Theorem IV. for the zerodelayed case zp8(u.5. that is. 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.. Label the customers 1. For the approach via Markov additive processes. say finite horizon ruin probabilities where the approach via the imbedded random walk yields results on the probability of ruin after N claims. it is easily seen that ic(ay ) > 0. that is. Then "^ e(ay+w(aY))8 Ys(u.
4: Proposition 4. and we have P(V > u) = ?/iiol(u).2) (b) as t * oo. then Wn v M.1. Then: (a) as n + oo.2 Let Mnd) = maxk=o. (u).. we have Wn = Van(left limit). and obviously z/'(u) = limN.1).4. (4.n1 (U1 +• • •+Uk Tl . Let the T there be the random time UN.Tn)+ Proof The amount of residual work just before customer n arrives is VQ„ . (4. .142 CHAPTER V.3. equivalently.1 and Corollary 11.1) 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. Then P(r(u) < T) is the probability z/iiNi (u) of ruin after at most N claims.• • • Tk ). Wn converges i n distribution to a random variable W. p < 1. but we shall present a slightly different proof via the duality result given in Theorem II.4. the waiting time a customer would have if he arrived at time t.1 Wn+1 = (Wn + U. since customer n arrives at time on. Vt converges in distribution to a random variable V. the proposition follows. 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. whereas in [On ... 0 Applying Theorem 11.1.2. and we have P(W > u) = V. and combining with (4.3) Proof Part (a) is contained in Theorem 11. If W1 = 0.3 Assume rl > 0 or. but interchanging the set . 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.Tn)+. Also {Zt}o<t<T evolves like the leftcontinuous version of the virtual waiting time process up to just before the Nth arrival. we get: Corollary 4.4): Proposition 4."^ Vi(N) (u).4. It then jumps to VQ„ . The traffic intensity of the queue is p = EU/ET.+ Un. Thus. (4. in which case {V} remains at zero until time on+1. equivalently. Thus Vos}1 _ = (Wn + Un .. an+1) = [on.
5 (LINDLEY'S INTEGRAL EQUATION) Let F(x) = P(U1T1 < x). but this requires some additional arguments (involving regeneration at 0 but not difficult) that we omit.. where U*.. For part (b). as WN.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.. (4.T*)+ < x) = P(W + U* .2). hence (since the residual lifetime at 0 and the age at T have the same distribution . Hence for x > 0. 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).T* = y yields K(x) = P ((W + U* . x > 0.T* < x) fK(x_y)F(dy) (x > 0 is crucial for the second equality!). Letting n oo in Corollary 4. by an obvious reversibility argument this does not affect the distribution . THE DUALITY WITH QUEUEING THEORY 143 (T1. and hence in particular ZT is distributed as the virtual waiting time just before the Nth arrival. T1. Then K(x) = J x00K(x .. cf. conditioning upon U* .1. i. u Now return to the Poisson case .. Corollary 4. K(x) = P(W < x).4... Ti) and similarly for the U. Then the arrivals of {Rt} in [0. we let T be deterministic . convergence in distribution hold for arbitrary initial conditions .. and we get: . T] form a stationary renewal process with interarrival distribution A. In fact .Ao in (b).oo in Proposition 4. (4. TN) with (TN.2. A. which implies the convergence in distribution and (4. T1 .4 The steadystate actual waiting time W has the same distribution as M(d). resp .le) the same is true for the timereversed point process which is the interarrival process for { Zt}o<t < T• Thus as before .T*)+.5) Proof Letting n .y)F(dy).T* are independent and distributed as U1. However. namely W1 = 0 in (a) and Vo = 0.e. Then the corresponding queue is M/G/1... we get W = (W + U* . we obtain: Corollary 4.(N)(u) has the limit tp(u) for all u.
Some early classical papers are Smith [350] and Lindley [246]. Cohen [88] or [APQ] Ch. implying P(W > u) = P(V > u) for all u. see e. Asmussen [24] and references there. Note that (4. The equation (4. W v V. VIII). . Hence '(u) = Ali(°)(u).6 For the M/G/1 queue with p < 1.g. despite the fact that the extension from M/G/1 is of equally doubtful relevance as we argued in Section 1 to be the case in risk theory. RENEWAL ARRIVALS Corollary 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 . the actual and the virtual waiting time have the same distribution in the steady state.5) to hold for all x E R and not just x > 0).g.5) is in fact a homogeneous WienerHopf equation. Proof For the Poisson case. That is.5) looks like the convolution equation K = F * K but is not the same (one would need (4.144 CHAPTER V.
N. As in Chapter I. 145 Oj( u. i=1 0 and r(u) = inf It > 0: St > u}. and can be computed as the positive solution of WA = 0. • The premium rate when Jt = i is pi. {Jt} describes the environmental conditions for the risk process. dv.T) = Pi (T(u) < T). M = supt>o St. {St} denotes the claim surplus process. The ruin probabilities with initial environment i are '+ki(u) = pi(T(u ) < oo) = Pi (M > u). here it exists whenever A is irreducible which is assumed throughout. • Claims arriving when Jt = i have distribution Bi. Ire = 1. The intensity matrix governing {Jt} is denoted by A = (A. Thus..(3i when Jt = i.Chapter VI Risk theory in a Markovian environment 1 Model and examples We assume that arrivals are not homogeneous in time but determined by a Markov process {Jt}0<t<oo with a finite state space E as follows: • The arrival intensity is . .f pi. t St = E Ui .)iJEE and its stationary limiting distribution by lr.
this is no restriction when studying infinite horizon ruin probabilities. respectively. MARKOVIAN ENVIRONMENT where as usual Pi refers to the case Jo = i. We let p Pi = /ji/AB. with rates Ani and Ain. An example of how such a mechanism could be relevant in risk theory follows.5 below. and assume that weather conditions play a major role for the occurence of accidents. i and corresponding arrival intensities Qn.1 Consider car insurance. f3i and claim size distributions Bn. r^ = P (1. we can approximate A(i) with a phasetype distribution (cf. Cl The versatility of the model in terms of incorporating (or at least approximating) many phenomena which look very different or more complicated at a first sight goes in fact much further: Example 1. According to Theorem A5. the intensity matrix is A OW) T(i) T(n) t(n)a(i) where t(n) = T(n)e. T(=)). P = E 7riPi. = iii when j E E(i). cf.T(n)). Proposition 1.4) with representation (E(i).146 CHAPTER VI. say. leading to E having two states n. Example 1 . in blockpartitioned form. we shall assume that pi = 1. t(i) = T(')e are the exit rates. /3 = Nn when j E E(n). cf. u . meaning that accidents occuring during icy road conditions lead to claim amounts which are different from the normal ones. which is clearly unrealistic. Then the state space for the environment is the disjoint union of E(n) and E(i). one expects that 3i > on and presumably also that Bn # Bi. Example 1.1 implicitly assumes that the sojourn times of the environment in the normal and the icy states are exponential.14. assume that the sojourn time in the icy state has a more general distribution A(i).. the operational time argument given in Example 1. Thus. For example.11 below. we could distinguish between normal and icy road conditions.2 (ALTERNATING RENEWAL ENVIRONMENT) The model of Example 1.1) iEE Then pi is the average amount of claims received per unit time when the environment is in state i. Unless otherwise stated. a(i).2. and we have f3. 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('). and p is the overall average amount of claims per unit time. say.a(').
where W = (w.T(n)). and 1/ii(u) = t/ii(u).. MODEL AND EXAMPLES 147 Example 1 . but assume now that the arrival intensity changes during the icy period. we assume again pi = 1 so that the claim surplus is Nt St = ?Ui_t.n.3i/pi. t(n) = T("i)e..p. the parameters are ^ij = aid/pi. iq (visited in that order) and letfOil >. n8}. St = SB=(t). u Example 1. such that a sojourn time of type rt is followed by one of type c w. In the car insurance example. i E E(n) }. resp...4) with states i1. T(9) +wggt(9)0.j = . u Example 1 .3i. Approximating each A('?) by a phasetype distribution with representation (E('l)..3. and . 0 Then (by standard operational time arguments) {St } is a risk process in a Markovian environment with unit premium rate. (9) where q = CHI. say it is larger initially. say..3 Consider again the alternating renewal model for car insurance in Example 1. Example VIII..tEH is a transition matrix. A('^).. depending only on 77. w.J017.2. one could for example have H = {i1.1. it = Jel(t). the state space E for the environment is { ('q. The simplest model for the arrival intensity amounts to . i8f n1. ..Q.a(n). i ) : n E H. . Then for example wi.5 (MARKOVMODULATED PREMIUMS) Returning for a short while to the case of general premium rates pi depending on the environment i. is the probability that a long icy period is followed by a short normal one. u From now on. such that the icy period is of two types (long and short) each with their sojourn time distribution A('L).. and similarly for the normal period. 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. One way to model this would be to take A(') to be Coxian (cf.. dt. Indeed.>. let T 9(T) = f pi. Qi = . This amounts to a family (A(")) ?CH Of sojourn time distributions.. 1 .1. T(1) +w11t(1)a(1) w12t (1)a(2) w21t(2)a(1) w1gt(1)a(9) w2gt ( 2)a(q) T(2) +w22t( 2)a(2) A = wg1t(9)a(1) wg2t(9)a(2) .
the empirical distribution of the claims is B*.(3iBi(dx). iEE iEE )3 These parameters are the ones which the statistician would estimate if he ignored the presence of Markovmodulation: Proposition 1. Pi (Ti E dx. the Markov additive structure will be used for exponential change of measure and thereby versions of Lundberg's inequality and the CramerLundberg approximation.(Qi)diag)• More precisely.148 CHAPTER VI.A . A remark which is fundamental for much of the intuition on the model consists in noting that to each risk process in a Markovian environment. t l=1 Note that the last statement of the proposition just means that in the limit. In particular. The key property for much of the analysis presented below is the following immediate observation: Proposition 1. vi(dx) = . Note also that (as the proof shows) 7ri/3i//3* gives the proportion of the claims which are of type i (arrive in state i). B* = 1 /^* Bi.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. )3*. one can associate in a natural way a standard Poisson one by averaging over the environment. qij = 0 in the notation of Chapter 11. Nt Nt a . Next we note a semiMarkov structure of the arrival process: Proposition 1. More precisely. N > 1(Ul < x) a4 B*(x). o = 0. dx. JT1 = j) = Qj • e.7 The Pidistribution of T1 is phasetype with representation (ei. MARKOVIAN ENVIRONMENT We now turn to some more mathematically oriented basic discussion. .8 As t oo. 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. . we put )3* = E 7fi/3i.5.e(A(Oi)d'sg)xe.
cf. y Ni) i Nti) t a.. 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. this converges to the exponential distribution with rate 0* as a * oo. iEE 13 A different interpretation of B* is as the Palm distribution of the claim size.7. By Proposition A5.4. In particular. Proposition 1.1. given {Jt}0<t<0. 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$. Then it is standard that ti lt '4' iri as t > oo. Example 11. However .aA . Bi(x). 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 . has distribution (7ri)3i //3*)iEE and is independent of Ti. Bi.. we may view Nt`i as the number of events in a Poisson process where the accumulated intensity at time t is Niti. aA. Bi. the limiting distribution of the first claim size U1 is B*.(/3i)aiag). A. Then {St)} + {St*} in D[0.2...6. Proof According to Proposition 1. N + oo Hence 1 Nt 1 N`+) Nits Nt E I ( Ut <. oo) as a 4 oo. Nt a' t t iEE Also.9 Consider a Markovmodulated risk process {St} with param eters Ni. and let {St °i} refer to the one with parameters Pi. and furthermore in the limit JT. zli( (u) . e. the Fidistribution of T1 in {St(a ) } is phase type with representation (E.x) = Nt E > I (Uk) X) Nt Bi(x) 1=1 iEE k=1 iEE 1: t5 Bi( x) = B*(x).. i.* (u) for all u. B*. In particular. {St} to the compound Poisson model with parameters 0 *. Hence Nt'> a . Conditioning . ^j 7riNi.
s = o in state 1 and with intensity 1 .1. there are p = 2 background states of {Jt}. from Theorem 3. the company even suffers an average loss.. 1.=1. shows similarly that in the limit (T2. state 1 appears as more dangerous than state 2. since E3 is a more dangerous claim size distribution than E7 (the mean is larger and the tail is heavier). 0 Example 1. e. resp. and (at least when a is small such that state changes of the environment are infrequent)..FT.2 +2 2 = 3. Computing the parameters of the averaged compound Poisson model.s = 1o in state 2. MARKOVIAN ENVIRONMENT upon FT.. T) + ?P* (u. U.. On Fig. Continuing in this manner shows that the limiting distribution of (T. Claims of type E3 arrive with intensity 2 . Thus. T) for all u and T. U2) are independent of .1 with periods with positive drift alternating with periods with negative drift.2.. B1=3E3+2E7..). oo). the overall drift is negative since it = (2 2) so that p = 71P1 + 112P2 = 7. lines in the path of {St}.1 of [145].g. From this the convergence in distribution follows by general facts on weak convergence in D[0. 1. thick.. 132=2. B2=1E3+4E7.10 Let E_{1. we may imagine that we have two types of claims such that the claim size distributions are E3 and E7.l3* and U2 having distribution B*. The fact that indeed 0(a) (U) 3 0* (u) follows. we first get that 3 (3* = 2. is as in {St }. That is. those of type E7 with intensity z s = 5 in state 1 and with intensity z . A= (  a a ) \ a a 5 5 J 9 3 2 a1=2. > 1.. marked by thin. s 5 in state 2. 5 5 where E5 denotes the exponential distribution with intensity parameter 5 and a > 0 is arbitrary. 9 .2}. with T2 being exponential with rate .150 CHAPTER VI.31µB 2 = 2 5 3 7 70 Thus in state 1 where p..2. which also yield O(a) (u. 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 = . the paths of the surplus process will exhibit the type of behaviour in Fig.
1.8.1) 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. (b) St/t * p .11 (a ) ESt/t * p .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). t + oo. the averaged compound Poisson model is the same as in III. That is. note first that EN Uk')/N a4' µgi. Proof In the notation of Proposition 1..1. 0 The definition (1.s. t * oo. For (b). a fraction r. = P. iEE .1 Thus.3* = 3/4 of the claims occur in state 1 and the remaining fraction 1/4 in state 2.(3.1 a. Hence (i) Nti) 1 U(i) k' N(i)k=1 E t 4 St + t = iEE Nt t 1: 7ri Qi µs. 01 /. Hence B* = 415E3+5E7/ +4 ( 51E3 +5 E7) = 1E 3 +2E7. MODEL AND EXAMPLES 151 Figure 1.
The case 77 > 0 is similarly easy.1)Eiw = 0. n n Thus {SWn l is a discrete time random walk with mean zero.2(a) p. Statistical aspects are not treated here.. The mainstream of the present chapter follows [16].. and hence 1/ii(u ) = 1 for all i and u.1 and the Corollary are standard. some early studies are in Janssen & Reinhard [211]. X3. Since the X„ are independent .a form a renewal process . In risk theory. having the Pidistribution of X. X 1 =Sty. 38) Eiw1 = 1/ir. Proposition 1. 136 or A. MARKOVIAN ENVIRONMENT Corollary 1.s. limit p . If 77 > 0. X2 =SW2 So. [212]. There seems still to be more to be done in this area. with some important improvements being obtained in Asmussen [17] in the queueing setting and being implemented numerically in Asmussen & Rolski [43].ld.Jt=i}.4. also + . PB.1 of St / t is > 0. with X2. see the Notes to Section 7. See Meier [258] and Ryden [314]..Eiw o'o Eiw • E ^ifjµs. see [APQ] p. Eiw. s. 2 The ladder height distribution Our mathematical treatment of the ruin problem follows the model of Chapter III for the simple compound Poisson model. and a more comprehensive treatment in Asmussen [16]. and so on. w2=inf {t>w1:Jt_#i. Theorem II. .\ i and EiX1 Ei f 13 J.. and hence M = 00. [APQ]. and hence oscillates between 0o and oo so that also here M = oo.12 If 77 < 0.1(b) is essentially the same as the proof of the strong law of large numbers for cumulative processes.152 CHAPTER VI. let some state i be fixed and define w=wl=inf{t >0:Jt_#i. [302]. EiX = 0.g.0i(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. then M < oo a.Jt=i}. Now let r) = 0. then M = 00 a... Then by standard Markov process formulas (e. and involves a version of the .. Now obviously the w.. dt . [315]. Proof The case 77 < 0 is trivial since then the a..s. The proof of Proposition 1. . + Xn SWn ](1 a . and hence wn /n a4. and .1 jEE = (p ..
j. cf. Thus. j. . Proposition 2.2) R(dx)S((y . Let further R denote the preT+ occupation kernel.g..3*B *(y)dy. B* in Section 1. which represents a nice simplified form of the ladder height distribution G+ when taking certain averages : starting {Jt} stationary.a/i. •) * G +(k. oo)) = f R(i.2 (a) The distribution of M is given by 00 1 . only with the product of real numbers replaced by convolution of measures.j. THE LADDER HEIGHT DISTRIBUTION 153 PollaczeckKhinchine formula (see Proposition 2. let G+(i.A) =ZI(St E. for i.A) = Pt(ST+ E A. (y. the definition of . •)• kEE Also. oo)). dx)/jBj(y .j E E.x). •).2(a) below ) where the ladder height distribution is evaluated by a time reversion argument.Jr+ =j. 6. by specializing results for general stationary risk processes (Theorem II .(u) = Pi(M < u) = e' E G+ (u)(I .IIG +II)e.1 irG+(dy)e =. we define the convolution operation by the same rule as for multiplication of realvalued matrices. The form of G+ turns out to be explicit (or at least computable). That is.4) we obtain the following result . j. IIG+ II denotes the matrix with ijth element IIG+(i.x. oo) = J ao 0 G+(i. k.i.2. j. T R(i.5.6*. we get the same ladder height distribution as for the averaged compound Poisson model.6. However .dx). •) II = JG+(i. For measurevalued matrices. e. Proposition 2. T+ < oo) and let G+ be the measurevalued matrix with ijth element G+(i. j.1) 0 (b) G+ (y. 00 (2. G+ is the matrix whose ijth element is E G +(i.j. n=0 (2. Define the ladder epoch T+ by T+ = inf It : St > 0} = r(0). see also Example II. and S (dx) the measure valued diagonal matrix with /3 Bj(dx) as ith diagonal element.EA.Jt=j)dt. but is substantially more involved than for the compound Poisson case .
{mx} is a non terminating Markov process on E. III to bring R and G+ on a more explicit form .3) We let {St*} be defined as {St}. lines in the path of {St}.3.1 for an illustration in the case of p = 2 environmental states of {Jt}. and let further {my} be the Evalued process obtained by observing {Jt } only when {St*} is at a minimum value. The u proof of (2. .3 When q > 0. JJ = j. MARKOVIAN ENVIRONMENT Proof The probability that there are n proper ladder steps not exceeding x and (x)ej. marked by thin.1) follows by summing over n and j. thick. only with {Jt} replaced by {Jt } (the /3i and Bi are the same ). 0  x Figure 2. and that the environment is j at the nth when we start from i is e . we need to invoke the timereversed version {Jt } of {Jt} .154 CHAPTER VI. G+ the probability that there are no further ladder steps starting from environment j is e^ ( I . St < S* for u < t. hence uniquely specified by its intensity matrix Q (say). resp.IIG+II)e. To make Proposition 2.1 The following observation is immediate: Proposition 2. we need as in Chapters II. To this end . From this (2.2 useful . mx = j when for some (necessarily unique) t we have St = x. That is.2) is just the same as the proof of Lemma 11.6. the intensity matrix A* has ijth element * 7r ^i3 7ri and we have Pi(JT = j) = 7rj P2(JT = i)7ri (2. see Figure 2.
} is a minimum value at v = t. the sequence {Q(n)} A* defined by Q(O) = . The definitions are illustrated on Fig.2.4 Q satisfies the nonlinear matrix equation Q = W(Q) where 0 co(Q) = n* .2. and the excursion ends at time s = inf {v > t : S.*. If there are no jumps in (t.(/3i)diag + T S(dx) eQx. we say that the excursion has depth 0. corresponding to two subexcursions of depth 0. {S. Furthermore. we recursively define the depth of an excursion as 1 plus the maximal depth of a subexcursion.2 where there are three excursions of depth 1.. For example the excursion of depth 2 has one subexcursion which is of depth 1. ( Q( n)) converges monotonically to Q. Figure 2. Otherwise each jump at a minimum level during the excursion starts a subexcursion. and the excursion is said to have depth 1 if each of these subexcursions have depth 0. THE LADDER HEIGHT DISTRIBUTION 155 Proposition 2. An excursion of {St*} above level x starts at time t if St = x.(/3i) diag. Proof The argument relies on an interpretation in terms of excursions. Q( n+l) _ ^.and a jump (claim arrival) occurs at time t. s]. 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. In general. 0 mms1   ^O \ T. = x}. 2.2 . and S(dx) is the diagonal matrix with the f3iBi(dx) on the diagonal.
4). mx+dx = j) occurs in two ways .5 R(i.St EA. The proof of Q = W(Q) then immediately carries over to show that the subintensity matrix of {mil) } is cp (Q(o)) = Q(l). Suppose mx = i. 0)).j.. i. h. Similarly by induction .156 CHAPTER VI. A). Similarly.St <S*.. MARKOVIAN ENVIRONMENT Let p=7) be the probability that an excursion starting from Jt = i has depth at most n and terminates at J8 = j and pij the probability that an excursion starting from Jt = i terminates at J8 = j. It follows that qij = A. p1^) Define a further kernel U by f U(i.5) A (note that we use A = {x : x E Al on the r.j +/3ipij. It is clear that { mini } is a terminating Markov process and that { mio) } has subintensity matrix A* . Q = W(Q) follows. A) = L' U(j. j. Fi(mh =i ) = 1 + =hflh+Qihpii+o(h) implies qii = 'iii /i +)3ipii. Then a jump to j (i. Theorem 2 . Writing out in matrix notation . A) = f Pi(mx = j) dx eie4xej dx A u (2. StEA . 7rE Proof We shall show that Fi(Jt=j.s. Now let {m ( n) } be {mx } killed at the first time i7n (say) a subexcursion of depth at least n occurs . the subintensity matrix of {min+i ) } is cp (Q(n)) = Q(n +l) which implies that qgj +1) = \!.6) .Qi + )%pij) Now just note that t pij and insert (2. either due to a jump of {Jt } which occurs with intensity A= j. or through an arrival starting an excursion terminating with J. e.T+>t) _ ^iF 7ri (JJ =i. = j. we first compute qij for i $ j.u< t).4) To show Q = cp(Q). it becomes clear that pij = r [eQh] 0 ij Bi (dy) • (2. By considering minimum values within the excursion.(01)diag = Q. of the definition to make U be concentrated on (co. (2.
0<u<t) = P. (Jo = j.StEA. oo)) = f o' eIXS((x + z. K( n (d) the sequence converges monotonically to K.t. the CramerLundberg approximation (Section 3).=StSt.1 can be rederived using the more detailed form of G+ in Corollary 2. THE LADDER HEIGHT DISTRIBUTION 157 from which the result immediately follows by integrating from 0 to oo w. x < 0.St EA. oo))dx.7 It is instructive to see how Proposition 2.(.g.z+>t) = P. we shall see that nevertheless we have enough information to derive.Jo=i.2.0<u<t.6 is hardly all that explicit in general.6 (a) R(dx) = eKxdx. St < St U.6). We may then assume Ju=Jtu. e. where A is the diagonal matrix with 7r on the diagonal: Corollary 2.r. To this end.S„<0. Remark 2. St EA. consider stationary versions of {Jt}. and we let k be the corresponding right eigenvector normalized by Irk = 1. 0<u<t).. dt. Jt = i. (b) for z > 0.. St E A. {Jt }. and to obtain a simple solution in the . u It is convenient at this stage to rewrite the above results in terms of the matrix K = 0'Q'A. and this immediately yields (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). and get irPi(Jt =j. G+((z.(Jt=j..St <Su. From Qe = 0. S.4].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. `` {K(n)} [the W(•) here is of course not the same as in Proposition 2. it is readily checked that 7r is a left eigenvector of K corresponding to the eigenvalue 0 (when p < 1). 0 < u < t) = 7rjPj(Jt =i. 0 +1) = cp (K( n)) defined by K(o) = A ..Qi)diag.
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
j) * coj)(u) _ f u ^Pj(u .9) For the proof. yu) f h(av)e v avuE«v. we have ic(ay) < 0 and get 'i(u) . 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. 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.(ay)}. 1 Similarly.. for a vector <p(u) = (cpi (u))iEE of functions .y)G+(z. dy)• o iEE jEE .3. r(u) < yu] hiay)C+ h=av)C+ o) (y)eayu+yuw(av).i [e*(u)K(av). yu < r(u) < 00 h 4(u) < h(av)C+o)(y)eavuEav . However. r(u) yu o)(y)eavuEav. We further write G(u) for the vector with ith component Gi(u) = EiEE G+(i. hj P .00 su e7( ( 3.i [eT(u)K(av ). (3. Chie ryu < Vi(u ) < C+hie 7u.11 Let Bj (x) C_ = min 1 • inf jEE hj x>o f2° e'r( vx)Bj(dy) ' C+ _ mE 1 Bj(x) J Y x)Bj (dy).5) will produce the maximal ryy for which the argument works. exp {e() + r(u))} .7. (u.5). oo)) and. if y > 1lk'(ry). we let G+ * W(u) be the vector with ith component E(G+(i.j.8 ) Then for all i E E and all u > 0.i I (a) exp {aye(u) + r(u)r.V)i(u. we shall need the matrices G+ and R of Section 2. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 165 hiav)e _avuE. as in the classical case (3. av 'i [h.
x ) R(i.13 For all i and u.x) jEE 00 u 0 //^^ C+E.12 Assume sup1. if r+ (n) is the nth ladder epoch.(0) ] (u) < sup Jt t. dy) : 1(u) < C+ > hj u e(1tL)G+(i. 00 Thus C+ > hj f"o e7(Yu)G +(i. dx). Then cpin)(u) sit (u) as n + oo. 0 G+(i. jEE u 0 j. Proof Write UN = EN G+ .7. and define W(n+1) (u) = G(u) + (G+ * tp(n))(u). = Eo G+ G. j.u IMP:°) (u) I < oo. dy) = aj f Bj(dy .j. Then iterating the defining equation ip(n+1) = G + G+ * V(n) we get W(N+1) = UN * G + G+N+1) * ^(o) However. dx) f e7( vu)Bj (dy .u Iv 2°)(u)I Pi(rr+(N + 1) < oo) + 0. 00 f C_ hj f e(Y)G+(i. Lemma 3 .x ) = Gi(u). n > oo.x) x) jEE 0 E Qj f jEE R(i. dx ) Bj (u . j.166 CHAPTER VI. dy) 00 C+ ijhj f R(i. °O .& (u). j.dy). MARKOVIAN ENVIRONMENT Lemma 3 . Hence lim cp(n) exists and equals U * G.ery(&u+x)Bj (dy) Bj(u Bj (u .j. j. U = U". dx) 100 C . _ To see that the ith component of U * G(u) equals ?Pi (u). we have G *(N +1) * ^. 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.3jhj // f 00 R(i.
MT < u.11).Pi(MT > u) = Pi(MT < u. We claim by induction that then cpin) (u) > C_ hie7u for all n. from which the lower inequality follows by letting n * oo. this is obvious if n = 0. T) = Pi (7. j.10 ) by Lemma 3 .10: Theorem 3 . ST < u] < C+(yo)e7ouEi [h^7o)e70ST1 l T J = C h(7o)e7ou8T .(u) < T ) to 0i (u) which is different from Theorem 3. dy) jEE o (3.y)G+(i. dy) (3. +i .T) = Pi(M > u) . and assuming it shown for n. we get Wo n +1) (u) = ? 7 i ( U ) + E J u gyp.tpi(u. y]hj = C_ e7uhi. 13 and the second by the induction hypothesis .13 Let first cp=°)(u) = C_ hie"u in Lemma 3. and let 8 = e'(70).3.M>u) = Ei [VGJT (u . letting MT = maxo<t<T St.ST).11. 9 for the last equality in (3.13. (3. jEE estimating the first term in (3. 14 Let yo > 0 be the solution of 'c'(yo ) = 0. let C+(yo) be as in (3.M > u) = Pi(ST<u. T) < C+(')' o)hi7u)e7ou8T . Then 0< Vi (u )  0i(u. (3. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES proving the upper inequality.n) ( u . u The proof of the upper inequality is similar .11) C_e7u 57 O+[i. we have Vii (u) . taking cps°) (u) = 0.u)G+(i. Here is an estimate of the rate of convergence of the finite horizon ruin probabilities 'i (u. and using Lemma 3 .13) Hence.12) Proof We first note that just as in the proof of Theorem 3. 167 u Proof of Theorem 3. j. j. dy) jEE u U +C_ hje7( yu)G jEE"" +(i. j.MT<u. and the proof of the lower one is similar.8) with y replaced by yo and hi by h=7o ). Indeed.10) C_ 1 f hje7(y. it follows that Vi(u) < C_(yo) h=70)e7ou.
2) (4. where it has been observed repeatedly that Markovmodulation increases waiting times and in fact some partial results had been obtained.. It was long conjectured that 0* Vi..3) Bl <_s. B2 <_s. Occasionally we strengthen (4. u > 0. For the notion of monotone Markov processes. V)" if z/i'(u) <'c"" (u)..1) Obviously.4) To avoid trivialities. < . The Markov process {Jt} is stochastically monotone (4..0.3) to B = Bi does not depend on i. The conditions which play a role in the following are: .5) Note that whereas (4. we define the stochastic ordering by 0' < s.. we also assume that there exist i # j such that either /3i <. this correponds to the usual stochastic ordering of the maxima M'. where o*(u) is the ruin probability for the averaged compound Poisson model defined in Section 1 and . 4 Comparisons with the compound Poisson model 4a Ordering of the ruin functions For two risk functions 0'.33 or Bi 0 Bj. ". [177].2) alone just amounts to an ordering of the states. Further related discussion is given in Grigelionis [176].3p. this is not the case for (4.. M" of the corresponding two claim surplus proceses (note that 0'(u) _ P(M' > u).168 CHAPTER VI. but that in general the picture is more diverse. (4.. (4.. The motivation that such a result should be true came in part from numerical studies. MARKOVIAN ENVIRONMENT Notes and references The results and proofs are from Asmussen and Rolski [44].o. in part from the folklore principle that any added stochastic variation increases the risk.31:5)32 .3). is the one for the Markovmodulated one in the stationary case (the distribution of J0 is 7r). Bp. 0"(u) = P(M" > u)) Now consider the risk process in a Markovian environment and define i' (u) _ >iEE irioi(u). we refer to . and finally in part from queueing theory. <s. The results to be presented show that quite often this is so.o. (4.o.o.
. E 7r i Wi(u . Conditioning upon the first ladder epoch. then j is the more risky state . we obtain (cf.4. the second follows from an extension of Theorem I1.4) is automatic in some simple examples like birthdeath processes or p = 2 . 7(0) < oo) = pirf+).1.9 ) below)...1) which with basically the same proof can be found in Asmussen & Schmidt [49].. ..2.9) (4. = aP or b1 = . 0 Here (4. T(0) < oo) = Bi(x) dx/tcai .10) Q*B*(u)+. 1:7riaibi > E 7riai i=1 i=1 j=1 The equality holds if and only if a1 = .6).1 for the first term in (4..8) ^j Tri/iBd(x) . Lemma 4 . = b. 2 If al < ..13* J0 u 0*(u .r (JT(o) = i.. COMPARISONS WITH THE COMPOUND POISSON MODEL 169 Stoyan [352].10) and (4.5 (cf. Then V. (b) P. dx (4. then P P P 7rjbj.x) dx u o i =1 i=1 (4. Section 4. . where 7r2+) = QiµBilri/p. (4.4) say basically that if i < j .7) 7ri.3iBi(x)YPi(u . Proschan & Walkup [140].1 Assume that conditions (4.r(u x)dx.3 for the second) *(u) _ /3 *B* (u) +.. Conditions (4.6..2)(4.. Proposition 2.2)(4. and it is in fact easy to show that Vii(u ) < t/j(u) (this is used in the derivation of (4. Theorem 4 .x)dx _ /3*B*(u) + f u / ^ t=1 > 3 * B* ( ) + f (4. we need two lemmas...4) hold. Comparing (4.x)B*(x) dx. The first is a standard result going back to Chebycheff and appearing in a more general form in Esary.7) and Lemma 4.r (Sr(o) E dx Jr(o) = i. 3 (a) P. p). also Proposition 2. note that (4. < bp and 7ri > 0 (i = 1.3* f uB(x) z/^.6) 7r= fl*B*(u) + p> s=1 +) fu 0 b (u  x)Bt (x) /pB. ^i 7ri = 1. Proof of Theorem 4.* For the proof.x) of i and using Lemma 4. b1 < .2. Lemma 4 . it follows by a standard ..9) follows by considering the increasing functions 3iBi (x) and Oi (u . < a.
that P P /^ 1r1NiµBi /^2 /^ ^i/ji pBi < 1il3i i=1 i=1 (4. i=1 i=1 7'r(0) _ EFioiwi(0) .h. and from this the claim follows./3*.s. except possibly for a very special situation . (4.11) is of order 104 and the r.4 is not vacuous.* (0).0. 0. it will hold for all sufficiently large u.6). Using (4. of (4. Frey. Q2 = 1.2. As is seen.4 Assume that . 01 = 103.s.3µi < 1 for all i. Recall that .4) is essential (the present author conjectures it is). (u) may fail for some u.. Then i/i*(u) < ..8) we get P P '*' (0) = 3* + /3*1* (0) _ > lri'3qqi • E 7i/ipBi . it is sufficient to show that 0'. u To see that Proposition 4. Bi as e J.4 is the understanding of whether the stochastic monotonicity condition (4. Proof Since 0. 4b Ordering of adjustment coefficients Despite the fact that V)* (u) < *.1 of [145] for a formal proof) that z/ii(u) converges to the ruin probability for the compound Poisson model with parameters .3i.170 CHAPTER VI. let = ( 1/2 1/2 ) .(0) < b *'(0) for e small enough.h.1 and Proposition 4. (u) is not in general true: Proposition 4.(0) = V. Rolski & Schmidt [32]. Notes and references The results are from Asmussen.0*• i=1 But it is intuitively clear (see Theorem 3.r (u ) fails for all sufficiently small e > 0. this ruin probability is /3iPBi.6). u Here is a counterexample showing that the inequality tp* (u) < V). dominates the solution 0* to the renewal equation (4. Then the l. µB2 = 104. of order 101.. µB. they are at present not quite complete. = 102.11) i=1` and that A has the form eAo for some fixed intensity matrix A0. For u = 0. MARKOVIAN ENVIRONMENT argument from renewal theory that tk... What is missing in relation to Theorem 4.
(4. Asmussen [20]) as discussed in 11.12) iEE Theorem 4. in particular . Now we can view {Xt} as a cumulative process (see A. Then {(Jt.1) .13) (4. is nondegenerate unless bi does not depend on i E E. cf.(a) > 0 for all a 0 0. with strict inequality unless rci (y*) does not depend on iEE.1) .5. e. (4. Jt = i])' EE = vA+n(6..a = E irirci(a).14) A„(O) iioo varXt t t By convexity.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) > 0.13) implies A(a) > 0 for all a. Xt)} is a Markov additive process (a socalled Markovian fluid model. with strict inequality unless a = 0 or bi = 0 for all i E E. Hence if 5i 54 0 for some i E E. COMPARISONS WITH THE COMPOUND POISSON MODEL 171 the adjustment coefficient for the Markovmodulated model is defined as the solution y > 0 of ic(y) = 0 where c(a) is the eigenvalue with maximal real part of the matrix A + (rci(a))diag where rci(a) = ai(Bi[a] .5.14) is nonzero so that A"(0) > 0.Jt=kI A (the return time of k) where k E E is some arbitrary but fixed state.4.5 y < ry*. It is clear that the distribution of X. it follows by Proposition A1.a.. (4. and by Proposition II. Proof Define X= f &ids.4(b) that the limit in (4. Further (see Corollary 11.2 we have (Ei[e"X'. which in view of EiEE 1ibi = 0 is only possible if Si = 0 for all i E E.)a.ld) with generic cycle w = inf{t>0: Jt_54 k. 0 . This implies that A is strictly convex.g.g. Lemma 4. The adjustment coefficient y* for the averaged compound Poisson model is the solution > 0 of rc*(ry*) = 0 where rc*(a) _ 13*(B*[a] .5.7) )i is convex with A'(0) = lim EXt tioo t = iEE 70i = 0.
. Let bi = rci(y*).p yi and compute 8y 8a a=0 In both cases. Rolski & Schmidt [32]. (4. y. h'(0) = (Ao .15) once more and letting e = 0 we get . h(0) = e.eir)h'(0). (4.16) Differentiating (4. a = 1 in Lemma 4.15) Normalizing h by 7rh = 0. Further a(1) = rc(y*) by definition of A(.172 CHAPTER VI. Hence letting e = 0 in (4.5 is from Asmussen & O'Cinneide [40]. Since ic is convex with rc'(0) < 0 . whereas the .Qi and Bi are fixed . In the case of e. Then > risi = 0 because of (4. Frey.5.) and rc (•). we have 7rh' = 0.6.15) yields 0 = (Ii(y*)) diage + Aoh'(0) = (rci('Y*)) diage + (Ao .e7r)1 (Ici(Y*))diage. note that y(a) + mins=1. If rci(y* ) is not a constant function of i E E. h depend on the parameter (e or a).12) and rc*(y*) = 0.. and our aim is to compute the sensitivity ay e a E=O A dual result deals with the limit a 4 oo. 4c Sensitivity estimates for the adjustment coefficient Now assume that the intensity matrix for the environment is Ae = Ao/ e. where A. multiply the basic equation by a to obtain 0 = (A0 + e(r£i(y))diag)h. MARKOVIAN ENVIRONMENT Proof of Theorem 4. this implies that the solution y > 0 of K(y) = 0 must satisfy y < y*. the basic equation is (A + (rci(y))diag)h = 0. we get rc (y*) > 0 which in a similar manner implies that u y < y*. 0 = ((ri(Y))diag + ery (4{('Y))diag)h + (A0 + e(?i'Y))diag)h'. Thus y(e) * y* as e 10.. improving upon more incomplete results from Asmussen. Notes and references Theorem 4. Hence rc (y*) > 0. The corresponding adjustment coefficient is denoted by ry(e).. Here we put a = 1/e.
i(7' *))diagh'(0). (4. (4. then 8a a=o All rci (0) Notes and references The results are from Asmussen. i = 2. . p.19) holds. .18). 0 = (Ao + ry'(ii(Y)) diag )h + (aAo + (Ki(7'))diag)h'.5. Inserting (4. and we have proved: Proposition 4. multiplying (4. which has recently received much attention in the queueing literature. The additional feature of the model is the following: • Certain transitions of {Jt} from state i to state j are accompanied by a claim with distribution Bid.20) and multiplying by el to the left we get 0 = All + 7'(0)rci (0) + 0 (here we used icl (ry(0)) = 0 to infer that the first component of K[7(0)]h'( 0) is 0). The analogue of Proposition 4.19) Then 'y ^ ryl as a ^ 0 and we may take h(0) = el (the first unit vector). the intensity for such a transition (referred to as marked in the following) is denoted by Aii l and the remaining intensity ..8 when ryi < 0 for some i is open.17) (4.16) yields Proposition 4. and may have some relevance in risk theory as well (though this still remains to be implemented). (4. Frey. We get 0 = (aAo + ( lc&Y))diag)h. Rolski & Schmidt [32].17) by 7r to the left to get (4. THE MARKOVIAN ARRIVAL PROCESS 173 0 = 27'(0)(ri(`Y *)) diage + 2(ci('Y* )) diag h' (0) + Aoh" (0) . We assume that 0 < y < 7i.18) 0 = 27'(0)p+27r(rs.7 8ry aE = 1 7r(ci ('Y*))diag ( Ao e7r)1(Xi(Y*))diage *=0 P Now turn to the case of a.20) Letting a = 0 in (4..8 If (4. 5 The Markovian arrival process We shall here briefly survey an extension of the model.
Again . with common distribution B.4). then {Nt} is a Markov additive process if and only if it corresponds to an arrival mechanism of the type just considered. In the above setting. A(1'k) A(2 k1). that Bii = Bi . let { Jt 1) }.d. Here are some main examples: Example 5 . A(1) = A . j(2) } be two independent environmental processes and let E(k). The extension of the model can also be motivated via Markov additive processes: if {Nt} is the counting process of a point process. the definition of Bi is redundant because of f3i = 0.(13i )diag. and thus 1i = 0. MARKOVIAN ENVIRONMENT f o r a transition i + j by A .174 CHAPTER VI. Note that the case that 0 < qij < 1. and the marked transitions are then the ones corresponding to arrivals.2 (SUPERPOSITIONS) A nice feature of the setup is that it is closed under superposition of independent arrival streams . T).^) etc. A(l) = tv.i. Jt2)) (2. II.2). the Markovmodulated compound Poisson model considered sofar corresponds to A(l) = (. where qij is the probability that a transition i * j is accompanied by a claim with distribution.2 for details).6i ) diag. For i = j.1 (PHASETYPE RENEWAL ARRIVALS) Consider a risk process where the claim sizes are i. B. the claim surplus is a Markov additive process (cf. and that are determined by A = A(l ) +A(2) where A is the intensity matrix the governing {Jt}. Bii = Bi . This is the only way in which arrivals can occur. Bij = B. we use the convention that a1i = f3i where 3i is the Poisson rate in state i. the definition of Bij is redundant for i i4 j. but the point process of arrivals is not Poisson but renewal with interclaim times having common distribution A of phasetype with representation (v. . we may let {Jt} represent the phase processes of the individual interarrival times glued together (see further VIII. Indeed.2) A(1) = A(' 1) ® A(1. Jt = (Jtl). u Example 5 . A ( 2) = A (2`1 ) ® A. Thus . We then let (see the Appendix for the Kronecker E = E(1) x E(2). is neither 0 or 1 is covered by letting Bij have an atom of size qij at 0. A(l) = T. refer to notation) { Jt k) }.
say... where ik = 0 means that the kth policy has not yet expired and ik = 1 that it has expired. 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 ..kj = Bik) B13 4k = Bak) 175  (the definition of the remaining Bij. e. iN. E 10. The versatility of the setup is even greater than for the Markovmodulated model. iN = all BOi2.g. iN. E = { WORKING. • upon a claim.iN = C27 All other offdiagonal elements of A are zero so that all other Bii are redundant. This means that the environmental states are of the form i1i2 • • • iN with il.5. u Example 5 . after which it starts afresh. or. INVALIDIZED. more recently.. 11.. RETIRED..iN C17 AilO. Thus. In this way we can model.1i2 .. claims occur only at state transitions for the environment so that AN2. WIDOWED. Example 5 . In fact .4 (A SINGLE LIFE INSURANCE POLICY ) Consider the life insurance of a single policy holder which can be in one of several states. Assume further that the ith policy leads to a claim having distribution Ci after a time which is exponential. assume that there is a finite number N of policies....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). possibly having a general phasetype sojourn time.. and that the policy then expires.. the idea of arrivals at transition epochs can be found in Hermann [193] and Rudemo [313]. Bilo. DEAD etc.iN = a2.. as the Markovian arrival process ( MAP).iil.. superpositions of renewal processes. THE MARKOVIAN ARRIVAL PROCESS Bij.iN..iN. Hermann [193 ] and Asmussen & Koole [37] showed that in some appropriate .}.kl is redundant)... DIVORCED. with rate ai.... Similarly. Easy modifications apply to allow for • the time until expiration of the kth policy is general phasetype rather than exponential.iN are zero and all Bi are redundant.iil. However . i2i .1i2. the kth policy enters a recovering state.. 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. MARRIED. all Al i2.
.1) Then the average arrival rate is /3* and the safety loading rt is 77 = (p* . For the Markovmodulated model. one needs to assume also (as a minimum) that they are measurable in t. [248]. p * = 0 p(t) dt. Sengupta [336].p)/p. but now exhibiting (deterministic) periodic fluctuations rather than (random ) Markovian ones. The basic assumptions are as follows: • The arrival intensity at time t of the year is 3(t) for a certain function /3(t). B* = J f B(t) ((*) dt. one limitation for approximation purposes is the inequality Var Nt > ENt which needs not hold for all arrival streams. 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. 1). Let 1 1 /3* _ f /3(t) dt. Lucantoni et at. )3 t 1 J (6. let the period be 1. 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 . We denote throughout the initial season by s and by P(8) the corresponding governing probability measure for the risk process. 1).3*µs • p = f /3(v) dv 0 0 (6. where i f00 xB(°) (dx) _ . Obviously. we may assume that the functions /3(t). for s E E = [0. p(t) and B(t) are defined also for t t [0. a claim arrives with rate /3(s + t) and is distributed according to B(8+0 . • The premium rate at time t of the year is p(t).176 CHAPTER VI. Thus at time t the premium rate is p(s + t). Lucantoni [248]. • Claims arriving at time t of the year have distribution B(t). 0 < t < 1. we talk of s as the 'time of the year'. Without loss of generality.2) Note that p is the average net claim amount per unit time and µ* = p//3* the average mean claim size. By periodic extension. Some main queueing references using the MAP are Ramaswami [298]. Neuts [271] and Asmussen & Perry [42]. from an application point of view.
t. not random. The arrival process {Nt}t>0 is a timeinhomogeneous Poisson process with intensity function {/3(s + t)}t>0 .6. We u assume in the rest of this section that p(t) .w(t)) dt). Thus. In contrast. The claim surplus process {St } two is defined in the obvious way as St = ^N° Ui . and thus the averaged standard compound Poisson models have the same risk for all A. RISK THEORY IN A PERIODIC ENVIRONMENT 177 In a similar manner as in Proposition 1.1 As an example to be used for numerical illustration throughout this section. 0 Then (by standard operational time arguments ) {St} is a periodic risk process with unit premium rate and the same infinite horizon ruin probabilities. Example 6 .3(t) = 3A(1 + sin 27rt).3) Note that A enters just as a scaling factor of the time axis. 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 . for Markovmodulated model typically the adjustment coefficient is larger than for the averaged model (cf.10. the conditional distribution . u Remark 6 . Thus . one may think of the standard compound Poisson model with parameters 3*. let . we shall see that for the periodic model increasing A increases the effect of the periodic fluctuations. or. equivalently.1) and Example 1. it turns out that they have the same adjustment coefficient. Section 4b).8. In particular. Many of the results given below indicate that the averaged and the periodic model share a number of main features.w(t). respectively. the discussion in 111. p* as an averaged version of the periodic model. since the added variation is deterministic. In contrast. the average compound Poisson model is the same as in III.1. It is easily seen that . 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. in agreement with the general principle of added variation increasing the risk (cf. (6.3* = 3A. of the periodic model as arising from the compound Poisson model by adding some extra variability.9). B*. St = SeI(t).2 Define T 6(T) = p(t ) dt. and we recall from there that the ruin probability is 24 1 *(u) _ 3 5eu + 35e6u.(3. The behaviour of the periodic model needs not to be seen as a violation of this principle.
with the underlying Markov process {Jt} being deterministic period motion on E = [0.4) At a first sight this point of view may appear quite artificial.3(s + t)dt[B(8+t)[a] . and define h(s. of the claim surplus process..8). J Theorem 6 .east B(8+t) [a] east .(8) [eaSt+dt I7t] = = (1 . 6b Lundberg conjugation Motivated by the discussion in Chapter II. Jt = (s + t) mod 1 P(8) . of the averaged compound Poisson model (the last expression is independent of s by periodicity).al.^8 [. let f 8+1 tc *(a) _ (B* [a] .g.tc* (a)] dv then h (. Dassios & Embrechts [98] and Asmussen & Rolski [43].(1 .e.g.5. i.adt +. The exposition of the present chapter is basically an extract from [44]. we obtain E. As usual. we start by deriving formulas giving the m.1) a = J8 .T) = P(8)(r(u) <T).. r(u) _ inf It > 0 : St > u} is the time to ruin .a. To this end.f. given that the ith claim occurs at time t is B(8+t).1) . [101] . a) is periodic on R. 1).a . Notes and references The model has been studied in risk theory by.5 (see in particular Remark 11. 3 E(8)eaSt = h(s. The claim surplus process {St} may be seen as a Markov additive process. (6.g. with some variants in the proofs. a) etw*(a) h(s+t.s . t + dt] or not. see the Notes to Section 7).f. .a be the c. 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.Q(v) (B(„) [a] .1) dv . Daykin et. 0 (5)(u. e. [44] (the literature in the mathematical equivalent setting of queueing theory is somewhat more extensive.178 CHAPTER VL MARKOVIAN ENVIRONMENT of U. and the ruin probabilities are 0(8) (U) = P(s )(r(u) < 00)..(3(s + t)dt)e«St adt + /3(s + t)dt .a) Proof Conditioning upon whether a claim occurs in [t.1]) .a) = exp { .3(v)(B(vl [a] .
1]) .5 The formula for h(s) = h(s.1)dv l og E(8) et where atetk•(a) h(t. B) eoSt t. h(s + t. RISK THEORY IN A PERIODIC ENVIRONMENT E(8)east+ dt d Et. 0) P(8)a.1]) .s. a) h(s + t. a) . 9) is a P ( 8)martingale with mean one. a) as well as the fact that rc = k` (a) is the correct exponential growth rate of Eeast can be derived via Remark 11.(e) Let = h( h(Jo.c* (e) {Le.t. it then suffices to note that E(8)Le.9 as follows. E (8)east (a +.5. Proof In the Markov additive sense of (6. a) Corollary 6.* (a) h(s.adt +. 0) exist and are finite.1)dv  o h(t.3. u Remark 6. St)} and .1]. a).6.3(s + t)[D(8 +t)[a] .6 .(8)east 179 = = = = = E(8)east (1 .0(s + t)dt[B(8+t)[a] .t} is a multiplicative functional for the Markov process { (Jt. dt log E(8)east a + f3(s + t) [B(8+t) [a] . a) = h(s.4 For each 0 such that the integrals in the definition of h(t . at + f log h(s + t.4). so that obviously {Lo. St)} . According to Remark 11. a) = exp I f t3(v)(kv)[a) . we can write Lo Jt.log h(s.9) eastt.. + v)(B([a] .2.t = 1 by Theorem 6. a) Thus E(8)east = h(s + t. a) et.t}t>o = h(s. With g the infinitesimal generator of {Xt} = {(Jt.
That is.y) = eayh(s). see [44] for 11 a formal proof. of St is as for the asserted periodic risk model.'y).3(s)dt • B(s)[a]h(s) = gha(s. J s [. the requirement is cha(i.1) .3(s)ks)[a]h(s)} ah(s) 13(s)h(s) + h'(s) +. .3.3(v)( Bi"i [a] .180 CHAPTER VI. Proposition 6. For each 0 satisfying the conditions of Corollary 6. correspond to a new periodic risk model with parameters ex . Lemma 6 .60(t) = a(t)B(t)[0].3(s)h(s) + h'(s) +. we put for short h(s) = h(s.1. [70] .3(s)B(s) [a]h(s). yo is determined by 0 = k* (70) = QB*.0) = Kh(s). St)} with governing probability measures Fes). Equating this to rch (s) and dividing by h(s) yields h(s ) = h(s) = a + .a .g. Proof (i) Check that m.6 The P(s). B(s).tc] dv} (normalizing by h(0) = 1).2. Sdt) = h(s + dt) eadt (1 . Now define 'y as the positive solution of the Lundberg equation for the averaged model. When a = y. Proposition 6.7 When a > yo.f. y solves n* (y) = 0. such that for any s and T < oo. ry)) at which n* (a) attains its minimum. P(s) (T(u) < oo) = 1 for all u > 0.T. MARKOVIAN ENVIRONMENT ha(s. A further important constant is the value yo (located in (0. Bet)(dx) = ^ B(t ) (dx). ( iii) use approximations with piecewiese constant /3(s). (iv) finally. the restrictions of Plsi and Pest to Ft are equivalent with likelihood ratio Le. 0) = h(s) + dt {ah(s) . That is. 0 < s < 1.4. (ii) use Markovmodulated approximations (Section 6c). cf. it follows by Theorem II. say. as above E (s) ha(Jdt. However.5 that we can define a new Markov process {(Jt. That rc = is*(a) then follows by noting that h(1) _ u h(0) by periodicity.6 ( s ) exp { 0( s )&s) [a] + tc .(3(s)dt) +.
The relevant likelihood ratio representation of the ruin probabilities now follows immediately from Corollary 11.1.1) the distribution of (l: (oo). Here and in the following. 1). T) = h(s. a) e«uE(8 ) e «^ . 9(u)) for any bounded continuous function (e. 0(u)) * (b(oo). u which is > 1 by convexity.4. xEJ 0 (s)b(8)(x) > 0. Wu).u is the overshoot and 9(u) = (T(u) + s) mod 1 the season at the time of ruin.9(u))} u>0. a) TI h(9(u). The proof involves machinery from the ergodic theory of Markov chains on a general state space. considered with governing probability measures { E(8) }E[ . has a unique stationary distribution. s E I. T(u) < (6. a) a > ry0 (6.7) h(B(u). we get: . a)e«uE (a iP(s) (u) = h( s)e7uE(` ) h(O(u)) To obtain the CramerLundberg approximation from Corollary 3.6(v) dv Jo ' xe«xB (°) (dx) r^ xe«xB'(dx) = Q'B' [ a] = ^' J 0 = ^c"'(a) + 1. and we refer to [44].10) Then for each a. the Markov process {(^(u).2. J C R+ such that the B(8). RISK THEORY IN A PERIODIC ENVIRONMENT Proof According to (6. q) = eryx/h(q)). we need the following auxiliary result . and no matter what is the initial season s. say s0. the mean number of claims per unit time is p« 181 = Jo 1.6. e(cc)) Letting u > oo in (6.9) 0(')(u) = h(s. which is not used elsewhere in the book.8) (6.9) and noting that weak convergence entails convergence of E f (^(u). B(oo)). (6. Lemma 6 .2). ^(u) = ST(u) . Corollary 6. f (x.8 The ruin probabilities can be computed as (u)+T(u)k'(a) ^/i(8) (u. have components with densities b(8)(x) satisfying inf sEI.9 Assume that there exist open intervals I C [0.g.
10 Under the condition (6. 11 7/'O (u) < C+°)h(s) ery".182 CHAPTER VI. elementary calculus yields h(s) = exp { A C 2^ cos 2irs  4^ sin 21rs + 11 cos 41rs . A=1/4 A=1 A=4 0 Figure 6.) C = E1 h(B(oo)) u + oo.W.6 for the Markovmodulated model: Theorem 6 .11) Note that ( 6. Vi(8) (u) . where e.11) gives an interpretation of h(s ) as a measure of how the risks of different initial seasons s vary.16. which may provide one among many motivations for the Markovmodulated approximation procedure to be considered in Section 6c. For our basic Example 6 . illustrating that the effect of seasonality increases with A.1. At this stage . this provides an algorithm for computing C as a limit. Theorem 6 .1 In contrast to h.ir) } Plots of h for different values of A are given in Fig. Among other things. we obtain immediately the following version of Lundberg ' s inequality which is a direct parallel of the result given in Corollary 3. Noting that ^(u) > 0 in ( 6. 10 shows that certainly ry is the correct Lundberg exponent.10) of Lemma 3. MARKOVIAN ENVIRONMENT Theorem 6. it does not seem within the range of our methods to compute C explicitly. 1.Ch(s)ery".9). 6. (6. where C(o) = 1 + info < t<i h(t) .1.
we substitute T = yu in 0(u.11 as well as it supplements with a lower bound. e7 ( yx)B(t)(dy) > Then for all $ E [0.167r I Cu. #c( ay) < 0 when y > 1/tc'('y).12 Let 00)(y) 1 Then info < t<i h(t.(8) (u.42 • exp {J_ cos 27rs . the proofs are basically the same as in Section 3 and we refer to [44] for details.13 to our basic example.16 In order to apply Theorem 6. where ay is the unique solution of W(ay) =y• (6.15) The next result improves upon the constant C+) in front of eryu in Theorem 6. Theorem 6 .0(8) (u+ yu) (6. we obtain Co) = 1. Just as in IV.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. (6. whereas ay < y. 1 ) and all u > 0.7e .(ay) > 0 when y < 1/ic' (7). we first note that the function fu° exu {w • 3e . e. Theorem 6.g. (ay).42 so that 183 tp(8) (u) < 1. We state the results below..12) As for the Markovian environment model. in our basic example with A = 1.47r sin 27rs + 167r cos 47rs . C_h(s)e7u < V.w ) • 7e u{w • 3e3x + ( 1 .6.7x j dx _7x } _ 6w + 6(1 . ay) • (6.14) < C+)(y)h(s) e7yu. . RISK THEORY IN A PERIODIC ENVIRONMENT Thus. Consider first the timedependent version of Lundberg's inequality.3x + (1 .13) Elementary convexity arguments show that we always have ryy > Y and ay > ry. r. Lundberg's inequality can be con siderably sharpened and extended.w) .17) (6.w)e4u dx 9w + 7(1 .yr. yu) 000 (u) .w)e4u .4. 1 (6.(s)(u) < C+h(s)e7". T) and replace the Lundberg exponent ry by ryy = ay .
where the environment at time t is (s + t) mod 1 E [0.4^ sin 2irs + 16^ cos 41rs . The idea is basically to approximate the (deterministic) continuous clock by a discrete (random) Markovian one with n 'months'.181 s(u) < 1. 6c Markovmodulated approximations A periodic risk model may be seen as a varying environment model.. Thus C_ = 2 inf ex cos 2irs . we have the following result: Theorem 6 .(8)(u.013. and in fact. 1) for the environment).20). 0 <'p(8)(u ) .1 sin 27rs + 1 cos 47rs . . 14 Let C+('yo) be as in (6. 1). MARKOVIAN ENVIRONMENT attains its minimum 2 /3 for u = oo and its maximum 6 /(7 + 2w) for u = 0. Some of the present proofs are more elementary by avoiding the general point process machinery of [44].0.184 CHAPTER VI. .I eu.cos 27rs . This observation motivates to look for a more formal connection between the periodic model and the one evolving in a finite Markovian environment.16) with 'y replaced by yo and h(t) by h(t. for A = 1 (where 3 e0.66.g. yo). much of the analysis of the preceding section is modelled after the techniques developed in the preceding sections for the case of a finite E. n}.19 } 0 <8<1 8 + cos 21rs Thus e. exp 2^ cos 21rs .\ 3 C+ = sup 6 exp { A (.20 •exp { 2n cos 27rs . with s the initial season.'Yo)e (6. C+ = 1. Thus.16.9 3 0<8<1 p 27r 47r 167r 161r 2 _ _e. such a deterministic periodic environment may be seen as a special case of a Markovian one (allowing a continuous state space E = [0.19 I eu. 1/i18 1 s (u) > 0.\ = 0 .T) < C+('Yo)h( s.18) Notes and references The material is from Asmussen & Rolski [44]. and let 8 = er' (Y0).1 sin 2irs + 16_ cos 47rs .L sin 27rs + 1 I cos 47rs . Of course. . completing a cycle . the nth Markovian environmental process {Jt} moves cyclically on {1. Then 7oudT .. Finally.013.66. but thereby also slightly longer.
T) can be expressed in a simple way in terms of the waiting time probabilities of a queueing system with the input being the timereversed input of the risk process. z/'i (u. one simple choice is Oni = 0( i . and the ruin probability corresponding to the initial state i of the environment is then Y'yn)(t) = F (M(n) > t). but others are also possible. Notes and references See Rolski [306].7. DUAL QUEUEING MODELS 185 within one unit of time on the average . This queue is commonly denoted as the Markovmodulated M/G/1 queue and has received considerable attention in the last decade. Thus.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.20) be the claim surplus process of t>o the nth approximating Markovmodulated model. Bi.jEE of the risk process. . 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. AE= Aii'r?/7ri• The arrival intensity is /3i when Jt = i. (6. We let {Stn)} (6.21) which serves as an approximation to 0(1)(u) whenever n is large and i/n s. We want to choose the /3ni and Bni in order to achieve good convergence to the periodic model. Let 0j. M(n) = Supt>o Stn). To this end. it is desirable to have formulas permitting freely to translate from one setting into the other.1 ((i 1)/n) ) and Bni = B . since the settings are equivalent from a mathematical point of view. so that the intensity matrix is A(n) given by n n 0 ••• 0 0 n n ••• 0 A(n) _ (6. 7 Dual queueing models The essence of the results of the present section is that the ruin probabilities i/ (u).
P(V > u. JJ = i). 0 < t < T and that the risk process {Rt}o<t<T is coupled to the virtual waiting process {Vt}o<t<T as in the basic dualitylemma (Theorem 11. JT = j} and {VT > u. . Jt ).4) where 0* = >jEE 7rj/3j.3. Then Pi(T(u) < T.2) Oi(u) = 1. JT = i} coincide. Proposition 7. JT = Z). just sum (7. (7.0i (u . {Jt }o<t<T• Then we may assume that Jt = JTt. and the virtual waiting time (workload) process {Vt}too are defined exactly as for the renewal model in Chapter V. Jo = i. For (7.n(VT > u. Proposition 7.. MARKOVIAN ENVIRONMENT • Customers arriving when Jt = i have service time distribution Bi. Jo = j.. (7.oo in u (7. J* = i). JT = j) = LjPj (VT > u.2).=1 .T(V > u I J* = i). The first conclusion of that result then states that the events {T(u) < T. and for (7. Proof Consider stationary versions of {Jt}o<t<T.186 CHAPTER VI. I* = i).1). T) = 7ri 1 P. In particular.2) and use that limF (VT > u. (7. let T . JT = i) = 'P. J* = i) for all j. (7. ii (u) = it /3 P(W > u.1 Assume V0 = 0. and (7. .. Now let In denote the environment when customer n arrives and I* the steadystate limit.3).2 The relation between the steadystate distributions of the actual and the virtual waiting time distribution is given by F(W > u. • The queueing discipline is FIFO. (VT > u I JT = 2). J*) is the steadystate limit of (Vt. The actual waiting time process 1W1.1) 7ri In particular. J* = i) = P. I* )3i P(V > u.1) over j. JT = j) = 7rjPj(VT > u.3) 7ri where (V. Taking probabilities and using the stationarity yields 7riPi(T(u) < T.1) follows. JT = i) = P(V > u. 2 .
that /3(t). on average /32TP(V > u.o. . DUAL QUEUEING MODELS 187 Proof Identifying the distribution of (W. The relation (7. P(W >u. the dual queueing model is a periodic M/G/1 queue with arrival rate 0(t) and service time distribution B(') at time t of the year (assuming w. P(. with (7.7. n=1 N However. Taking the ratio yields (7.=i) a4. [243]. if T is large. a paper relying heavily on classical complex plane methods. and of these.4) and (7.1 is from Asmussen [16]. and one has PI'>(rr(u) < T) = P('_T)(VT > u). a general formalism allowing this type of conclusion is 'conditional PASTA'.8) For treatments of periodic M/G/1 queue.l. on average 0*T customers arrive in [0. and Rolski [306]. In the setting of the periodic model of Section 6. I*) with the timeaverage .3).5) follows from (7.I *=i). T]. p < 1 then ensures that V(*) = limNloo VN+9 exists in distribution. With {Vt} denoting the workload process of the periodic queue. Proposition 7. see Regterschot & van Doorn [123]. see in particular Harrison & Lemoine [186].4) can be found in Regterschot & de Smit [301]. P(1')(r(u) < oo) = P(')(00) > u). A more probabilistic treatment was given by Asmussen [17]. J* = i) see W > u.g.6) (7. and further references (to which we add Prabhu & Zhu [296]) can be found there. Lemoine [242].7) of that paper. N * oo. (7.7) (7. u Notes and references One of the earliest papers drawing attention to the Markovmodulated M/G/1 queue is Burman & Smith [84].T)(T(u) <T) = P(8)(VT > u). >u.I.3) improving somewhat upon (2. we have 1: I(W. I* = i..4). The first comprehensive solution of the waiting time problem is Regterschot & de Smit [301]. B(t) have been periodically extended to negative t). and (7.
This page is intentionally left blank .
z/i(u) = F IinffRt< 0IRo=u 1 (1.2) tk(u.At + p(R8) ds.. the aggregate claims in [0. Thus in between jumps.T) = F(T(u) < T). t] are Nt At = Ui (1. with common distribution B and independent of {Nt}. i&(u. and that the claim sizes U1.d. Thus. Zt As earlier. and the evolution of the reserve may be described by the equation Rt = u . and T(u) = inf {t > 0 : Rt < u} is the time to ruin starting from Ro = u so that '(u) = F(T(u) < oo).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 . are i. However . U2.T) = FloinfTRt< OIRo=u1 denote the ruin probabilities with/initial reserve u and infinite.1) (other terms are accumulated claims or total claims).i. finite horizon. 189 .6. resp . {Rt} moves according to the differential equation R = p(R). the premium charged is assumed to depend upon the current reserve Rt so that the premium rate is p(r) when Rt = r.. .
However. dividends are paid out at rate pi . that {Rt} will reach level u before the first claim arrives. i. That is. If Ro = v < u.3 (ABSOLUTE RUIN) Consider the same situation as in Example 1. say e. Now return to the general model. P(r) _ p + e(r . say at interest rate b. there is positive probability. Assume 0(u) < 1 for some u. but assume now that the company borrows the deficit in the bank when the reserve goes negative. 1 . and the probability of absolute ruin with initial reserve u E [p/S.2. pi > p2 and p(r) = One reason could be competition.4 Either i.p/S) r > p/S p5(p/5r) 0<r<p/5 Then the ruin problem for {Rt } is of the type defined above.'(u)) > 0 so that V'(v) < 1.e.1 Assume that the company reduces the premium rate from pi to p2 when the reserve comes above some critical value v.Vi(v) u > e(1 . No tractable necessary and sufficient condition is known in complete generality of the model. Example 1.p2.2 (INTEREST) If the company charges a constant premium rate p u but invests its money at interest rate e.190 CHAPTER VII. Proposition 1. A basic question is thus which premium rules p(r) ensure that 'O(u) < 1. we get p(r) = p + er. but when the reserve comes above v. Proof Obviously '(u) < ilb(v) when u > v. it seems reasonable to assume monotonicity (p(r) is u . where one would try to attract new customers as soon as the business has become reasonably safe. Another could be the payout of dividends: here the premium paid by the policy holders is the same for all r. we can put Rt = Rt + p/S. In this situation. or o(u) < 1 for all u. Thus at deficit x > 0 (meaning Rt = x).i(u) = 1 for all u. Example 1. the payout rate of interest is Sx and absolute ruin occurs when this exceeds the premium inflow p. when x > p/S. RESERVEDEPENDENT PREMIUMS The following examples provide some main motivation for studying the model: Example 1 . rather than when the reserve itself becomes negative. Hence in terms of survival probabilities. oo) is given by i (u + p/S).
5 (a) If p(r) < /.uo) < 1.f p(Vs) ds. obviously infu<uo z/'(u) > 0. which was proved in 11. T] i n such a way that the events {r(u) <T} and {VT > u} coincide. Let Op(u) refer to the compound Poisson model with the same 0. say V. cf. {Vt} decreases at rate p(v) when Vt = v (i. let uo be chosen such that p(r) > p = 0ILB + e for r > uo. That is.+ p(r) exists.6 For any T < oo. [APQ] pp. In case (a).2(d). Starting from Ro = uo. and P(Rt + oo) > 0. (b) If p(r) > /3µB + e for all sufficiently large r and some e > 0. Here {Vt}two is a storage process which has reflection at zero and initial condition Vo = 0. Theorem 1. and hence by a geometric trials argument. In particular. one can couple the risk process and the storage process on [0.b(u. (1. However. if and only if V)(u) < 1 for all u. We next recall the following results.2) we have t Vt = At .I3IB requires a more detailed analysis and that µB < oo is not always necessary for O(u) < 1 when p(r) 4 oo. Proof This follows by a simple comparison with the compound Poisson model. instead of (1.5) and the process {Vt} has a proper limit in distribution . 296297): Theorem 1.1. (1. INTRODUCTION 191 decreasing in Example 1. appealing to Proposition 111. Proposition I1I. Then if u > no. Then 0(u) = P(V > u). (1.2 once more. Hence ik(u) < 1 for all u by Proposition III.3. let uo be chosen such that p(r) < p = /3µB for r > uo. In case (b).1. then ?(u) = 1 for all u. the probability that Rt < uo for some t is at least tp(0) = 1 (cf.4. In between jumps.o(uo) = 1 so that t/'(u) = 1 for all u by Proposition 1. { Vt} remains at 0 until the next arrival). V = p(V)).2(d)). B and (constant) premium rate p. we have z/i(u) <p(u .T) = P(VT > u). then l/i(u) < 1 for all u. .uo) and.1 and increasing in Example 1.1.3µB for all sufficiently large r.e.6) .1.. that u zPp(u . This is basically covered by the following result (but note that the case p(r) . hence Rt < uo also for a whole sequence of is converging to oo.2) for r sufficiently large so that p(oo) = limr.4) 0 and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0.
t + dt] if and only if Vt E [x. the flow of mass from [0. and is succesful if the jump size is larger than x . u Define ^x 1 w(x) Jo p(t) dt. t + dt] can be neglected so that a path of {Vt} corresponds to a downcrossing in [t. and the other being given by a density g(x) on (0. Corollary 1. say.6x and that w(x) < oo for all x > 0. Note that it may happen that w (x) = oo for all x > 0. x] to (x. Jo AX) (1.Sx} dx. we arrive at the desired interpretation of the r. say if p(r) goes to 0 at rate 1 /r or faster as r j 0. An attempt of an upcrossing occurs as result of an arrival.8 Assume that B is exponential with rate b.9) Proof We may rewrite (1.s.8) Proof In stationarity.8) as g(x) = p 1 {yo13e_6x +. B(x) = e. Oeax f x e'Yg (y) dy } = p) eaxa(x) . where g(x) = p( ^ exp {.7 p(x)g(x) = tofB (x) + a f (x . we thus need to look more into the stationary distribution G. It is intuitively obvious and not too hard to prove that G is a mixture of two components. Now obviously. yo ^ 1 + oo Q exp {. x + p(x)dt]). the l.8) is the rate of downcrossings (the event of an arrival in [t. In view of the path structure of {V t }. Then the ruin probability is tp (u) = f' g(y)dy.y)g(y) dy.192 CHAPTER VII.s. RESERVEDEPENDENT PREMIUMS In order to make Theorem 1. say. oo) must be the same as the flow the other way. Considering the cases y = 0 and 0 < y < x separately.6 applicable. one having an atom at 0 of size 'yo. (1. It follows in particular that 0(u) = fg(Y)dy. for the storage process {Vt}. of (1.y.7) Proposition 1.Qw(x) .h.Sx}. (1. oo).6w(x) . this means that the rate of upcrossings of level x must be the same as the rate of downcrossings.h. say when {Vt} is in state y.8) as the rate of upcrossings. of (1. Then w(x) is the time it takes for the reserve to reach level x provided it starts with Ro = 0 and no claims arrive.
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
2y +µ ) dy .V.pa.01'E) + (p/E)al aO l febP/E } IF (0 /0 jF From this (2.g.13 /E) r (.3/E) By the memoryless property of the exponential distribution. ./3 log(b + a)] = log ePa/f (a + a ) e which shows that Z is distributed as p/E . /^ u Example 2 .3. The process {St} corresponds to {Wt} so that c(a) or2a2/2 .b P/E dx /' P/ ' (p/  x)p/e 150/f I' (/3/E) (6P1'E. Proof 2 We use Theorem 2. From ic(a) = . 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 .200 CHAPTER VII.2) follows by elementary algebra. and the c.2) follows by elementary algebra.a) . of Z is IogH[a] = f ytc(y)dy = e fa (0. assume that {Wt} is Brownian motion with drift µ and variance v2. then {Rt} is the diffusion with drift function p+Ex and constant variance a2.3 is also valid if {Rt} is obtained by adding interest to a more general process {Wt} with stationary independent increments.pa. with density x(3/e1aQ/e fV (x) _ e 6X ' x > 0. 13/E). where V is Gamma(b.x) dx e.V < x)]0 + f P(V > p/E ) + eby fv (p/E .5 The analysis leading to Theorem 2. i.e. r (j3/E) In particular.f. H(u) = P(Z r < u) = P(V > u + p/E) = (8(p + Eu)/E. RT(u) has an exponential distribution with rate (S) and hence E [H(RT(u))I r(u) < oo] L Pe6'r (P/C . As an example.3a/ (5 . RESERVEDEPENDENT PREMIUMS u from which (2.
[282]. it is also used as basis for a diffusion approximation by these authors. [129]. Further studies of the model with interest can be found in Boogaert & Crijns [71].d. Paulsen [281]. A r.. write Vi* (u) for the ruin probability etc. Paulsen & Gjessing [286] found some remarkable explicit formulas for 0(u) beyond the exponential case in Corollary 1. Q2/2E).i.1) .3) was derived by Emanuel et at. see e.g. It must be noted. of the form Ei° p"X" with the X„ i. Some of these references also go into a stochastic interest rate. se e. however. 134 (the time scale there is discrete but the argument is easily adapted to the continuous case). it follows that the ruin probability is Cu) H(u) H(0) 11 Notes and references Theorem 2. Goldie & Griibel [167].e. Gerber [157] p. 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 The local adjustment coefficient.3 is from Harrison [185]. [283]. Gerber [155]. Paulsen & Gjessing [286] and Sundt & Teugels [356].4 is classical. Corollary 2. not even Erlang(3) or H3.p*. for a martingale proof.g. THE LOCAL ADJUSTMENT COEFFICIENT _ Q2a2 pa 4e E 201 I. Delbaen & Haezendonck [104].2 is a special case of a perpetuity. [357]. Z is normal (p/E.3. The formula (2. Emanuel et at..8. write y* for the solution of the Lundberg equation f3(B[ry *] . and recall Lundberg 's inequality . and since RT = 0 by the continuity of Brownian motion.v. that the analysis does not seem to carry over to general phasetype distributions. [129] and Harrison [185]. Logarithmic asymptotics For the classical risk model with constant premium rule p(x) . as in the proof of Proposition 2. or to nonlinear premium rules p(•).Y*p* W*(u) < ery*u = 0.
Then lim sup u>oo u and e E''p(r) + 0.E). log ?i(u) < < 00 JO ..2) such that p(x) < c(.5) which implies inf.>o 7(x) > 0. x>0 (3. obviously O(u) can be bounded with the probability that the Cramer Lundberg compound Poisson model with premium rate p* downcrosses level uE starting from u . RESERVEDEPENDENT PREMIUMS and the CramerLundberg approximation V. 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. choose uo such that p( x) > p* when x > u0E. as will hold under the steepness assumption of Theorem 3.1. B(x) > C(2)e(ao+f)x for all x.ap(x). e(1o+e)2 (x ) u > 00. Letting first E * 0 and next ry * T 5o yields the first statement of the theorem.i)eex. a first step is the following: Theorem 3 .1 ). oo for all E > 0. then log u (u) In the proof as well as in the remaining part of the section . and that p(x) * oo. as solution of the equation n(x. let p* be a in (3.202 CHAPTER VII. 1) and for a given E > 0. we will use the local adjustment coefficient 'y(x).4) we assume existence of y(x) for all x.1. which in turn by Lundberg's inequality can be bounded by ery*(1E)" Hence limsup„.*(u) . When u > uo. i. c(. Let y* < So. i. (3.e. The steepness assumption and p(x) + oo ensure 'y(x) * So.C*ef*".'y ( x)) = 0 where r.w (u) J dt > c(3)eeu v 1 p(u+ t) .1 Assume that for some 0 < 5o < oo. (3. it holds that f3[s] T oo. For the last asssertion . 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).log '(u)/u < ry*(1 . a) = f3(B[a] . and (for simplicity) that inf p(x) > (3µs . choose c(. Proof of Theorem 3.3) When trying to extend these results to the model of this chapter where p(x) depends on x. x * oo. Then we have the following lower bound for the time for the reserve to go from level u to level u + v without a claim: w(u + v) .e. (x.. If 60 s f 6o.1) .
Bucklew [81]). THE LOCAL ADJUSTMENT COEFFICIENT 203 where c. or (b) Condition 3. one can then assume that e = 1 is small enough for Theorem 3. Then limelog l/ie (u) = I(u).13 below holds. the limit is not u + oo but the slow Markov walk limit in large deviations theory (see e. However.v. Theorem 3. 2. The form of the result is superficially similar to the CramerLundberg approximation. and hence '(u) > c(4)eeuc( 2)e(do+e)u The truth of this for all e > 0 implies lim inf log V.' (u) < eI("). the result is not very informative if bo = oo.e.7) CIO Remarks: 1.4)eE" Given such an arrival. the asymptotics u * oo and c . UJU > x cannot have a much heavier tail than the claim U itself. Theorem 3 . (u) = O(u/e).1 only presents a first step. Then . u Obviously. For e > 0. I. 3.3 to be reasonably precise and use e` (u) as approximation to 0 (u). 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.6) The second main result to be derived states that the bound in Theorem 3.0 are the same. (3. . noting that in many cases the constant C is close to 1.g.3 Assume that either (a) p(r) is a non decreasing function of r. (3.2). ruin will occur if the claim is at least u + v. {Rte)} defined as in (1. Condition 3. The rest of this section deals with tail estimates involving the local adjustment coefficient.3. by cU2.(u) > so. which essentially says that an overshoot r.13 is a technical condition on the claim size distribution B. The first main result in this direction is the following version of Lundberg's inequality: Theorem 3 . 2 Assume that p(x) is a nondecreasing function of x and let I(u) = fo ry(x)dx. 3) = (1 . If p(x) = pis constant . then Rte) = CRtie for all t so that V). let 0e (u) be evaluated for the process only with 3 replaced by /0/e and U. and in particular.ea°/(ecf1)).2 is also an approximation under appropriate conditions. Therefore the probability that a claim arrives ( before the reserve has reached level u + v is at least c(..
5. rather than eI(u)).1 + b f e. and r j 1 'Yo v(x)dx = bu  a J0 p(x)ldx = Integrating by parts.4 Consider again the exponential case B(x) = eax as in Corollary 1. J0 ^oo g(x ) dx f AX) lexp IOW (X ) bx + b u 1 exp low(x) .bx} dx oo exp low(x) bx dx 70 Ju r oo = b J exp low (x) . First.bx}]o + b /' oo exp low (x) .6). 3.(iw(x) . One would expect the behaviour in 2) to be important for the quantitative performance of the Lundberg inequality (3. we show how to rewrite the explicit solution for ti(u) in Corollary 1. u .bx} dx .8. the logaritmic form of (3.bx} dx = 1 + J0 dodx(x) exp {. we get = 1+ J" AX) exp {(3w (x) .8 in terms of I(u) when the claims are exponential: Example 3 . Then y(x) = b .exp {/33w(u) . 3a Examples Before giving the proofs of Theorems 3. RESERVEDEPENDENT PREMIUMS 4. As typical in large deviations theory.bx} dx fo 00 1 + [exp {/(3w(x) .2. say.204 CHAPTER VII. it is formally needed only for Theorem 3.3. we consider some simple examples.3.bu}.(3/p(x).(x) dx.bx} dx 1+0.7) is only captures 'the main term in the exponent' but is not precise to describe the asymptotic form of O(u) in terms of ratio limit theorems (the precise asymptotics could be logI(u)e1(U) or I(u)"e_I(u). However.
and (3. It is well known that (see Theorem XI.10) where AE = e log 000 e.ev 0 O /E) J0 70 70 Yo This implies lim inf A. Be = e log U000 e. (X) = µ(x).BE. ry(x /b I u o e f0 °° e  e.3. applying the inequality 7(x + u) > 7(x) yields immediately the conclusion of Theorem 3. 0.3 in the particularly simple case of diffusions: Example 3.0. u .2. > lime log e = 0 and AE * 0. (3.I ( v )dy fo +u) dxdy . 3. Choosing yo. we get r 00 e.1. Similarly.f y(x)dxd y If 7(x) is increasing .fa 7(x+u)dx/Edy o The analogue of (3. and (3.fo 7(x) dx /E dy > av 'yo /Edy = E (1 ..2(X) = ev2(x) so that 7e(x) = 7(x)/e. In particular.9) yields e log .2. THE LOCAL ADJUSTMENT COEFFICIENT and hence 205 f°° eI(v )dy . IE(u) = I(u)/e. 70 > 0 such that 7(x) < 7o for y < yo. (3.8) 7(x)dxdy 11 We next give a direct derivations of Theorems 3. (u) = I(u) + AE . in the definition of AE converges to 0.1/8 .fo 7(x)dx/Edy f . For Theorem 3.10 or Karlin & Taylor [222] pp.3.9 ) 11000 eI(v)dy f000 e.fory(x+u)dxdy ( 3.5) is infx>o 7(x) > 0 which implies that f °O .7) follows. oo) with drift µ(x) and variance a2 (x) > 0 at x..5 Assume that {Rt} is a diffusion on [0. BE * 0. note first that the appropriate slow Markov walk assumption amounts to u. 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. 191195) that 1P (U) = fu0 eI(v)dy = eI(u) follo e.e.I ( u) fool. the integral is bounded by 1 eventually and hence lim sup AE < lim sup a log 1 = 0.
5 for risk processes with exponential claims is as follows: Example 3 . > .7o C 15 I I. . . Thus 7e(x) _7(x)/e and (3. 7(x) is typically not explicit. .e. Ignoring 1/5 in the formula there.5) and 7* = 5 . RESERVEDEPENDENT PREMIUMS The analogue of Example 3.(u) oo. G. the results are suggestive in their form and much more explicit than anything else in the literature. Of course.6/p* so that u 1 I (U) = bu . however .5. + Gq(u) + o(G9(u))• Gi (u) It should be noted .. I(u ) = G1(u) + .10) holds if we redefine AE as AE = flog (j °° efo 7(x)dx/edy _ E/5 I and similarly for B. E+o e*O By (3.. . lim sup Af < lim sup c log(1 .6 Assume that B is exponential with rate S.. so our approach is to determine standard functions Gl (u). Further. this leads to (3.1 3. that the interchange of the slow Markov walk oo is not justified and in fact.0/e. . 0. Nevertheless .5.206 CHAPTER VII. G. the slow Markov walk assumption means 5E = b/c.0) = 0.. _ . (u) representing the first few terms in the asymptotic expansion of I(u) as u + oo. 0 Now (3. ) Note that this expression shows up also in the explicit formula for lk(u) in the form given in Example 3.+1 (u) = o( 1).0./3 1 AX dx. we have 5 > 7o and get lim inf AE > lime log e .5. Then the solution of the Lundberg equation is y* = b .Q/p*. I.4. G. the slow Markov limit a * 0 and the limit u walk approximation deteriorates as x becomes large.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. As in Example 3.6) exactly as in Example 3.
y/s)dy sn 1 1 f ' evy'7ldy = cse8r(T7) as s T oc.1 =$ f cse8 Sn f e"B(x)dx = e8 Jo s eIB ( 1 .8).:.c3 logu a= 1 J 0 a + bx 1/ ( c4ul 1/° a > 1 where c3 = c2 /b. more generally.Y .12) with y > 1.. e.11) that b[s] * co as s f S and hence 7* T S as p* + oo. ry* loge*+ g7loglogp*. u Example 3 .c2 Su a dx ) Su a<1 Su . Hence (3. . .1/a).clxale5x 207 (3. 1. 1) and 17 = k + 1 if B is the convolution of k uniforms on (0. u(logu + r7loglogu). if the phase generator is irreducible ( Proposition VIII.1) leads to . y = 2 if B is uniform on (0. phasetype distributions (Example 1. fu I(u) Su .ry*°p*. x T 1.cs(1 .1) leads to (S7T N Ocp a.4) or gamma distributions.8 Assume next that B has bounded support. 2.C2p* C2 = (3clr( a))11'. I(u) Pt.g. It follows from (3. (3.3. Here B[s] is defined for all s and B[s] . For example. 77 = 1 if B is degenerate at 1. B[s] = 1 + s exB(x)dx = 1 +c1SF(a) ('+o(')) (S . c4 = c2b 1/'/(1 .s)C' f "o o as s T S.11) with a > 0.1/k). in the phasetype case .x)n1. This covers mixtures or convolutions of exponentials or.7 Assume that B(x) . THE LOCAL ADJUSTMENT COEFFICIENT Example 3 . say 1 is the upper limit and B(x) . the typical case is a = 1 which holds .3cse7*I7(77) . More precisely. and hence (3.
1 0 3e. (3. I (u) c8u log u 0 where c8 = 2/c7.1) .Ul up to the first claim (here ru (•) denotes the solution of i = p (r) starting from ru(0) = u).10 Assume that p(x) is a nondecreasing function of x. The assumption implies that ru(t) . (3.c8 log .f. .B[7o (u)] .g. one could also have considered the increment ru (T1) .15) Proposition 3.e7o ( u)(ul+u r. this is only possible if 7o(v) 2 7o(u)• .14) for the m .•. h].u is a nondecreasing function of u.u .13) We get b[s] .2 We first remark that the definition (3. RESERVEDEPENDENT PREMIUMS Example 3 .1 Cgs o"O 0 esxex2/2c7 dx = cgsec782/2 f .sp(u). g.9 As a case intermediate between (3.12).ru(TI)) . ec78)2/2c7 dx C7 .. This leads to an alternative local adjustment coefficient 7o(u) defined as solution of 1 = Ee''o(u)(vi+u . 1 = E. 7 * . (3.4) of the local adjustment coefficient is not the only possible one: whereas the motivation for (3. x f oo . Hence for u<V. of U1 + v .4).Ote7o( u)(u.11) and (3.f. (b) 'y(x) <'Yo(x)• Proof That 7(x) is nondecreasing follows easily by inspection of (3. Then: (a) y(x) and 7o(x) are also nondecreasing functions of x. h 10. assume that B(x) CO x2/2c7.r„(Ti).(t))dt.3 (B[s] .css 2%rc7eC782/2.(T1)) > Ee7o(u)(ul+vr»(Ti)). 3b Proof of Theorem 3.208 CHAPTER VII.r^. By convexity of the m . of the increment in a small time interval [0.log p*.4) is the formula h logEues ( Rhu) .
11 Assume that p(x) is a nondecreasing function of x. THE LOCAL ADJUSTMENT COEFFICIENT For (b). Then (u) < efo Yo(x)dx. u We prove Theorem 3. The case n = 0 is clear since here To = 0 so that ik(°)(u) = 0.u > tp(u).2 in terms of 7o. Also.10(b): Theorem 3. note that the assumption implies that ru(t) . Hence 1 = EeYo(u)(U1+uru(T1)) < E.16) Proof Define 411(n)(u) = P('r(u) < on) as the ruin probability after at most n claims (on = TI + • • • + Tn).(n+l) (u) 1 . we obtain „I.(n+1) (u) efo Yo(x)dxI^"Q exyo( I u u)Fu(dx )+ J . it is easily seen that fu x7o(y)dy < xyo (u).e70(u)(U1P(u)T1) 209 0 + 7o(u)p(u)' 0 <_ 00['Yo( u)] . fa 7o(y)dy < u7o(u) < xyo (u) for x > u.ru(T1) < x). this is only possible if yo(u) > 7(u).17) from which the theorem follows by letting n + oo.17) shown for n and let Fu(x) = P(U1 + u .u[70(u)] fo eyo(x)dx .x)Fu(dx) 00 U efo J = o (y) dYF (dx) )+f I 11 /' / 00 e f oFu fu dx) + of u :7o(Y)dYFu(dx) 00 J u 1 l` Considering the cases x > 0 and x < 0 separately.3. We shall show by induction that (' Y'(n) (u) < e fo 'Yo(x)dx (3.1) . the case of 7 then follows immediately by Proposition 3.es'Yo(u)Fu(dx)} o0 e fo yo( x)dx j. Separating after whether ruin occurs at the first claim or not. Hence „/. Assume (3.7o (u)p(u)• Since (3.Fu(u ) + J ^(n)(u . (3.4) considered as function of 7 is convex and 0 for y = 0.
3 is required.. 3c Proof of Theorem 3. Lemma 3. ryk. 0.n AX). The probability of this is at least n n. and. Proof For ruin to occur. 3.15). the probability that ruin occurs in the CramerLundberg model with p* = pn. Y*u /E.E (u/ n) Y'E (un .x/n..10(b ) that the bound provided by Theorem 3.n.11 is sharper than the one given by Theorem 3.12 lim sup4^o f log O.3/e and U. x + x/n] by two classical risk processes with a constant p and appeal to the classical results (3. 0 It follows from Proposition 3. given downcrossing occurs. resp. Further.nbe C*. in accordance with the notation i/iE (u). (u) < I(u).n.I. we have chosen to work with y(u) as the fundamental local adjustment coefficient. P k.n) must first downcross unl. {RtE)} (starting from u = un. (3.n. RESERVEDEPENDENT PREMIUMS where the last identity immediately follows from (3.: y(u). C*e where the first equality follows by an easy scaling argument and the approximation by (3.2). (u).3 The idea of the proof is to bound { R( f) } above and below in a small interval [x . we used also Proposition 3. To this end. pk n = uk_l.n) pn niE (u /n) n n_1 n.11 be reasonably tight something like the slow Markov walk conditions in Theorem 3.n (starting from u/n) without that 2u/n is upcrossed before ruin. For these reasons. define uk.E (u/n)..n inf n uk1..E (u) denote the ruin probability for the classical model with 0 replaced by .210 CHAPTER VII. (un2. Also. and here it is easily seen that yo(u) ..n <Z auk}l.n) > k =1 II v ^k n. for either of Theorems 3.E (u/n) Now as e .3).n u k}1.e (u) = v'. Let Ck.n. in .n = sup p(x). 0.2.2.3). y* evaluated for p* = Pk. by €U=.. let Op*. yo(u) appears more difficult to evaluate than y(u). However. W O . op*.10(a) for some of the inequalities.E ( u/n) ^•e.n so that n.n = ku. the value of {R(E)} at the time of downcrossing is < unl.
(u) CIO < Letting n 4 oo and using a Riemann sum approximation completes the proof.nu /En) o(1)). *p•. in obvious notation one has tC (x) = y(x)/e. 0 with n and u fixed.i.e. It follows that n log V'C (u) k =1 log Ypk.3.nk=1 limsupelogv). uk_1. (3.n + 0(1).E (u/n) Op•. 211 Clearly. /' (u/n) 'T nk. V < oo such that (i) for any u < oo there exist Cu < oo and a (u) > supy <„ 7(x) such that P(V > x) < Cue. . so that Theorem 3. since ry' is an increasing function of p'. 3 now follows easily in case (a).7k. i.19) .n cE (2u/n) Ck ne7k.nu/en(1 + where o(1) refers to the limit e . B(x) (3. Indeed .2 gives 7PE (u) < eIi"i/f = lim inf Clog 0E (u) > I (u). 40 Combining with the upper bound of Lemma 3. v. THE LOCAL ADJUSTMENT COEFFICIENT particular.. y > 0 it holds that F(U>x +yIU>x) B(x + y) < F (V > y). also ryk. for all x . we need the following condition: Condition 3.n. k=1 k=1 n u _ nE7 k.nu /fn( 1 Ck  e. 11 Theorem 3.12 completes the proof.E (u/n) OP +^p•.n = sup ?'(x).E (urn) < \ *I.18) (ii) the family of claim overshoot distributions is stochastically dominated by V. In case (b).F (2u/n).log Ck.a( u)z.n . ne7k.n <X<Uk..13 There exists a r..! (u/n) n n m 7k..
(u/n . V < u/En] + P(V > u/En) (u/En .EV) = El + E2.eV) • P (T(E) (u. Then Y'E (u) ^(E) (u.^'' = E [ . T() (u. u/n) < oo] E [OE (u/n .R<) (u v). (R.2y 1 ' . P (T(E) (u.of:>2 in n(x).( . Then the standard Lundberg inequality yields El < E?. The probability of ruin in between two downcrossings is bounded by Epp .E (2u/n . RESERVEDEPENDENT PREMIUMS To complete the proof.E (0) cf.. u/n) < oo] . u/n)) .nu/En0(1) .E(E) (u.EV) = e.EV) = EiI 1 . u/n) < oo) . Ei + E2 < e71.^(E) (u. infx>2u /n P(x) .nu /EnE [e71. u /en 0(i) _n so that E2 < e2ryl nu/En0(1).n < ery1. (u/n .V) = e71 nu/Eno(l) (using (3.. u /n) < oo] l = = < E [OE (u/n .18) for the last equality).. let v < u and define T(E) (u. (3.. For E2. v ) = v . N with EN < 1 = infx>2u/nA(x) = 0(1). v ) = inf { t > 0 : R(c ) < v R) = u } .E (u/n . ) (u u /n)) . u/n)) I T(E) (u. we first note that the number of downcrossings of 2u/n starting from RoE) = 2u/n is bounded by a geometric r. T(E) (u. .1 n. Write EO. u/n) < oo) EV).v.QEU 1 .n V.5) and the standard formula for b(0). where El is the contribution from the event that the process does not reach level 2u/n before ruin and E2 is the rest.212 CHAPTER VII.
where the key mathematical tool is the deep WentzellFreidlin theory of slow Markov walks (see e .g. the results are from Asmussen & Nielsen [39].4) and the prime meaning differentiation w.t.7) for ruin probabilities in the presence of an upper barrier b appears in Cottrell et al. Djehiche [122] gives an approximation for tp(u. (u) 40 213 lim inf e log(Ei +E2) + logP (r(`) (u. Comparing these references with the present work shows that in the slow Markov walk setup. 0 ) (= p(x) .T) = P „(info<t <T Rt < 0) via related large deviations techniques.1. Typically. whereas the most probable path leading to ruin is the solution of r(x) _ k (x.)Ui } .J y(Rs)dR.3EU) (3. u/n) < oo { 40 )I U nryl n+liminfelogP (T(')(u.21) to pass from u to 0.20) (with ic(x.)ds + Y(R2. Bucklew [81]). T) is maximized over T by taking T as the time for (3.13. the risk process itself is close to the solution of the differential equation r(x) _ r (x. [89]. 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 3.7) then comes out (at least heuristically) by analytical manipulations with the action integral. .r. u Notes and references With the exception of Theorem 3. l o JJJ o .=1 J An approximation similar to (3.21) (the initial condition is r(0) = u in both cases).7(x)) (3. s) as in (3. 0 and b T 00 are interchangeable in the setting of [89]. the approximation (3.J r(Rs)p(R.) = exp .u/n) < oo) CI  > u n n ryi n' i=1 Another Riemann sum approximation completes the proof.3. s). they also discuss simulation based upon 'local exponential change of measure' for which the likelihood ratio is ( /'t /'t Ns Lt = exp S . it might be possible to show that the limits e . Similarly. Whereas the result of [122] is given in terms of an action integral which does not look very explicit. THE LOCAL ADJUSTMENT COEFFICIENT Hence lim inf e log Ali. one can in fact arrive at the optimal path by showing that the approximation for 0(u.
3. . For different types of applications of large deviations to ruin probabilities .214 CHAPTER VII. the exponential distribution ).g. We should like. 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. e. see XI.. RESERVEDEPENDENT PREMIUMS the simplest being to require b[s] to be defined for all s > 0 (thus excluding .
that is. Typically. on Eo = E U {A} where A is some extra state which is absorbing. A distribution B on (0. F (Jt = A eventually) = 1 for all i E E 1 and where all states i E E are transient. 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. Note that since (1. if v = (vi)iEE is a probability distribution.1) is 'Here as usual . 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.Chapter VIII Matrixanalytic methods 1 Definition and basic properties of phasetype distributions Phasetype distributions are the computational vehicle of much of modern applied probability. refers to the case Jo = i. P. More precisely. we write Pv for the case where Jo has distribution v so that Pv = KER viPi• 215 . We often write p for the number of elements of E. a terminating Markov process {Jt} with state space E and intensity matrix T is defined as the restriction to E of a Markov process {Jt}o<t<. if a problem can be solved explicitly when the relevant distributions are exponentials. A proper knowledge of phasetype distributions seems therefore a must for anyone working in an applied probability area like risk theory. This implies in particular that the intensity matrix for { it } can be written in blockpartitioned form as T 0 0 . and not in other cases.
C is the lifetime sup It > 0 : Jt E E} of {Jt}.e. Then a = a1 = 1. i. MATRIXANALYTIC METHODS the intensity matrix of a nonterminating Markov process.0 = t11. the exit rates ti and the transition rates (intensities) tij: tj 3 aj ai i ti tk tjk FkJ ak Figure 1. We now say that B is of phasetype with representation (E. E = {i. that is. 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. 0 2this means that tii < 0. Equivalently. and we have t = Te.3.1 The phase diagram of a phasetype distribution with 3 phases. and the phasetype distribution is the lifetime of a particle with constant failure rate /3. In particular. j. k}. t1 = /3.e.2) The interpretation of the column vector t is as the exit rate vector.1 Suppose that p = 1 and write . a. Here are some important special cases: Example 1 . i. B(t) = Fa(^ < t ). Thus the phasetype distributions with p = 1 is exactly the class of exponential distributions. T) (or sometimes just (a. A convenient graphical representation is the phase diagram in terms of the entrance probabilities ai. (1.216 CHAPTER VIII.T)) if B is the Padistribution of the absorption time C = inf{t > 0 : it = A}. an exponential distribution with rate parameter . The initial vector a is written as a row vector. tij > 0 for i 54 j and EjEE tij < 0 . T is a subintensity matrix2. the ith component ti gives the intensity in state i for leaving E and going to the absorbing state A.
. ... 0 0 0 T= t= 0 ••• S S 0 0 0 0 0 0 ..1)!e Since this corresponds to a convolution of p exponential densities with the same rate S. 0 ••• 0 0 Sp1 0 0 t= 0 0 00 •. .3 The hyperexponential distribution HP with p parallel channels is defined as a mixture of p exponential distributions with rates 51.. so that the density is P E ai6ie6. . 00)) S s o .1. 0 SP 0 and the phase diagram is (p = 2) . PHASETYPE DISTRIBUTIONS 217 Example 1. 6.. the EP distribution may be represented by the phase diagram (p = 3) Figure 1. . ..x i=1 Thus E _ Si 0 T 0 S2 0 0 . 0 0 0 0 S 6 . 0 •.. 0 S 6 Example 1..2 corresponding to E = {1... a = (1 0 0 . .. p}...2 The Erlang distribution EP with p phases is defined Gamma distribution with integer parameter p and density bp XP1 6x (p.
36) yields s d. j E E. p:. B[s] = f0°O esxB (dx) is a(sI T)lt (d) the nth moment f0°O xnB(dx) is (.f.aeTxe. Then: (a) the c. Proof Let P8 = (p ^) be the sstep EA x EA transition matrix for {Jt } and P8 the sstep E x Etransition matrix for {Jt} . see A. . Then for i . (c) the m. dp.g. E t ikp kj = kEE kEE 3For a number of additional important properties of matrixexponentials and discussion of computational aspects . ds^ = ds' = ttlaj + tikpkj. MATRIXANALYTIC METHODS Figure 1. i.1)"n! aT"e. a. and is defined as the class of phasetype distributions with a phase diagram of the following form: 1 617 ti t2 2 b2.1 tP1 1 Figure 1.t2 yt bP.f is B (x) = 1 .4 (COXIAN DISTRIBUTIONS) This class of distributions is popular in much of the applied literature. the backwards equation for {Jt} (e. 5 Let B be phasetype with representation (E. (b) the density is b(x ) = B'(x) = aeTxt.g. Recall that the matrixexponential eK is defined by the standard series expansion Eo K"/n! 3.3 . the Erlang distribution is a special case of a Coxian distribution.d. the restriction of P8 to E.4 For example. Theorem 1 . The basic analytical properties of phase type distributions are given by the following result . [APQ ] p.e.3 0 Example 1 . T).218 CHAPTER VIII.
and (b) then follows from 1: aipF. tij / .jEE B'(x) _ cx Pxe = aeTxTe = aeTxt (since T and eTx commute).1.5) Indeed .5) as hi(tii + s) = ti  t ij hj.g.f. 1.tii we go to A... the solution is P8 = eT8.f. (1) n+1n!aT .s) is the m . Then h tit ti + ti3 h j . ti/ . of the initial sojourn in state i.n lt .n1t = (1)nn!aTn1Te (1)nn! aTne. For (c).p.tii and have an additional time to absorption either go to state j which has m . this means in vector notation that (T + sI)h = t. (Jx E E) = this proves (a). and since b[s] = ah. Part (d) follows by differentiating the m. d" dsn a (.tii tii .s I .f.T) 't = (. Since 1 . h = (T + sI)1t. define hi = Eie8S. we i w. . we arrive once more at the stated expression for B[s].e.1 ) n +l n ! a (s I + T ) . and since obviously P° = I.12) for integrating matrixexponentials yields B[s] = J esxaeTxt dx = a ( f°°e(81+T)dx ) t a(sI . in which case the time to absorption is 0 with m . j#i jEE tijhj + his = ti.p. B(n)[0] = _ Alternatively.f.g.5) ki = 1 + tii L jj:Ai tii (1. Rewriting ( 1. or w. the rule (A. d8 P8 = TP8.g.s j# tii i (1. i. Alternatively.T) 1t.tii is the rate of the exponential holding time of state i and hence (tii)/(tii .B(x) = 1'a (( > x) = P. i.6) . PHASETYPE DISTRIBUTIONS 219 That is. for n = 1 we may put ki = Ei( and get as in (1. hj .g. After that. = aPxe.
MATRIXANALYTIC METHODS which is solved as above to get k = aTle. One obvious instance is the hyperexponential distribution. another the case p = 2 where explicit diagonalization formulas are always available. 0 Example 1. we get the density as 9 9 6 (1 1) 10 7 1 0 10 2 aeTyt = e x .s. are idempotent. T= 2 111 so that 2 2 Then (cf. making the problem trivial.220 CHAPTER VIII. 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."n! ( ( l 2 2 ) 17 9 0 \ 1 / 10 10 32 n! 35 6" +n!353 Similarly. see the Appendix. Consider for example 3 9 a= (2 2). there are some examples where it is appealing to write T on diagonal form. Example A3.7) the diagonal form of T is 9 9 1 9 T 10 7 10 70 1 10 6 10 7 0 70 9 1 10 where the two matrices on the r.6 Though typically the evaluation of matrixexponentials is most conveniently carried out on a computer.h.
. i. hail = E=EE a.7) Proof According to (A.e 11BIJ = 1laDD < 1. and in fact one also most often there allows a to have a component ao at A. where the initial vector a is substochastic.f. i. < 1.hall. a random variable U having a defective phasetype distribution with representation (a. 0 Sometimes it is relevant also to consider phasetype distributions.1.4. B[Q] of B is f3[Q] = J e'1zB(dx) _ (v (9 I)(T ® Q)1(t ® I).29) and Proposition A4.11aDD. 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).e a mixture of a phasetype distribution with representation (a/llall. then the matrix m. or one just lets U be undefined on this additional set. 5 and serves at this stage to introduce Kronecker notation and calculus (see A.g. • The phasetype distribution B is zeromodified. T) is then defined to be oo on a set of probability 1. (1. This is the traditional choice in the literature.T) with weight hall and an atom at zero with weight 1 . There are two ways to interpret this: • The phasetype distribution B is defective.4b for definitions and basic rules): Proposition 1.T).7 If B is phasetype with representation (v. 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.
q be the eigenvalue of largest real part of T. The Erlang distribution gives an example where k > 0 (in fact.F.8) Proof By PerronFrobenius theory (A.. 1.8 are far from necessary ( a mixture of phasetype distributions with the respective T(') irreducible has obviously an asymptotically exponential tail. Using B(x) = aeTxe . 77 > 0 and k = 0. here k = p1). where C. In Proposition A5.. the text is essentially identical to Section 2 of Asmussen [26]. See in particular the notes to Section 6. it is easily seen that the asymptotic form of the tail of a general phasetype distribution has the form B(x) _ Cxkenx. but todays interest in the topic was largely initiated by M. and we have eTx . 0 Of course. Example A5. All material of the present section is standard.hve7x. Then the tail B(x) is asymptotically exponential. cf. h can be chosen with strictly positive component.222 CHAPTER VIII. v. MATRIXANALYTIC METHODS la Asymptotic exponentiality Writing T on the Jordan canonical form. Neuts. Schmidli. (1. Schmidt & Teugels [307] and Wolff [384].8). the conditions of Proposition 1. Notes and references The idea behind using phasetype distributions goes back to Erlang. cf.8 Let B be phasetype with representation (a. Here is a sufficient condition: Proposition 1. Other expositions of the basic theory of phasetype distributions can be found in [APQ]. one has k = 0. let .g. see his book [269] (a historical important intermediate step is Jensen [214]). In older literature.f. not only in the tail but in the whole distribution. T). O'Cinneide [276] gave a necessary and sufficient for a distribution B with a rational m. but the relevant T is not irreducible. 2. . h be the corresponding left and right eigenvectors normalized by vh = 1 and define C = ah • ve .1 of the Appendix. let v. No satisfying .f.4c). assume that T is irreducible . the result follows (with C = (ah)(ve)).Ce7'. we give a criterion for asymptotical exponentiality of a phasetype distribution B. distributions with a rational m. i is real and positive. B(x) . but in many practical cases. Rolski. Lipsky [247]. 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.g. x * oo. (or Laplace transform) are often used where one would now work instead with phasetype distributions. the Erlang case).
: U1 + .r. ... the renewals form a Poisson process and we have u(x) = 0.1) Proof Let {Jtk)} be the governing phase process for Uk and define {Jt} by piecing the { J(k) } together. oo) w. . U2.t. If B is exponential with rate 0. known. 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 +. or the density is available ) is. For this reason... and U(A) is then the expected number of replacements (renewals) in A. U1<t < U1+U2. Lebesgue measure. . JtJt1) Then { 0<t<U1 . with common distribution B and define4 U(A) = E# {n = 0.+UnEA).1 of the Appendix. (2. +UnEA} 00 = EEI(U1 +. + U0 is 0 .1 Consider a renewal process with interarrivals which are phasetype with representation (cr. Let U1... we refer to U as the renewal measure.2. we denote the density by u(x) and refer to u as the renewal density. but is in part repeated below.d. 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..i. The explicit calculation of the renewal density (or the renewal measure) is often thought of as infeasible for other distributions. Jt={Jt?ul}. is Markov and has two types of jumps .. RENEWAL THEORY 223 algorithm for finding a phase representation of a distribution B (which is known to be phasetype and for which the m.g. as the lifetimes of items (say electrical bulbs) which are replaced upon failure.1.. but nevertheless. the problem has an algorithmically tractable solution if B is phasetype: Theorem 2. be i...f.T). however. if U is absolutely continuous on (0. Then the renewal density exists and is given by u(x) = ae(T+ta)xt. n=O We may think of the U... A related important unsolved problem deals with minimal representations: given a phasetype distribution .
(b) £(t) has a limiting distribution as t * oo.e. see Fig.T) where vt = ae (T+ta)t . IIafl < 1.3 Consider a renewal process with interarrivals which are phasetype with representation (a. Hence the intensity matrix is T + ta. define the excess life e(t) at time t as the time until the next renewal following t. . which is ti in state i. u The argument goes through without change if the renewal process is terminating. the phasetype assumptions also yield the distribution of a further quantity of fundamental importance in later parts of this chapter . 2. i. The renewal density at x is now just the rate of jumps of the second type. i..1) follows by the law of total probability. Then: (a) the excess life t(t) at time t is phasetype with representation ( vt.1) remains valid for that case. the density is veTxt = B(x)/µB.U1 U3 U2 U3 U4 Figure 2. and let µB = aTle be the mean of B. T). Hence ( 2. MATRIXANALYTIC METHODS of the last type from i to j occurs at rate tiaj .1.. Corollary 2.2 Consider a terminating renewal process with interarrivals which are defective phasetype with representation (a.224 CHAPTER VIII.1 Corollary 2. + Uit_1 where s. and the jumps of the first type are governed by T. which is phase type with representation (v. Proof Just note that { it } is a governing phase process for the lifetime.T + ta). However. This is defined as U1 + . since Uk = oo with probability 1 . .IIBII which is > 0 in the defective case. and hence ( 2. that is. as the time of the last renewal. Equivalently. is the first k with Uk = 00. the lifetime of the renewal process. fi(t) U2 U1 . and the distribution of Jx is ae ( T+t«)x.T) where v = aT1 /µB. this is welldefined. B is defective .e. u Returning to nonterminating renewal processes .T). Then the lifetime is zeromodified phase type with representation (a.
The time of the next renewal after t is the time of the next jump of the second type. T1 and eTx commute. cf. Next appeal to the standard fact from renewal theory that the limiting distribution of e(x) has density B(x)/µB. = qz ql (x1 xz) = ql + qz ql + q ' and the nonzero eigenvalue A = ql . (2.2.2) v(T + ta) = 0. the unique positive solution of ve = 1.6. i. RENEWAL THEORY 225 Proof Consider again the process { Jt } in the proof of Theorem 2. Al. Here are two different arguments that this yields the asserted expression: (i) Just check that aT1/µB satisfies (2. we first compute the stationary distribution of Q.4 Consider a nonterminating renewal process with two phases.e.1. The renewal density is then aeQtt = (al a2) ( 7i 7"2.T) where vt is the distribution of it which is obviously given by the expression in (a). According to Example A3. (ii) First check the asserted identity for the density: since T. hence e(t) is phasetype with representation (vt.e. we get B(x) aeTxe aT1eTxTe µB µB PB = veTxt.) ( t2 ) . Hence in (b) it is immediate that v exists and is the stationary limiting distribution of it. 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). u Example 2 .2): aT1 e = AB = 1 µB µB a + aT'Tea aT1(T + ta) µB PB a + aea a + a µB µB =0.q2.
Hence 7r = (1/2 1/2). .4 yields the renewal density as u(t) = 2 (1 . )t (51 .t2) 1 + eat (a17r2 .5 Let B be Erlang(2).52a1.226 CHAPTER VIII.a27r1) (t1 . and Example 2. A = 25. t1B 0 Example 2 .(biaz + aza. Then _ Q Hence 51 0 0 52 + 51 52 _ 5152 51a2 ) (al a2) 52a1 62a1 Slat + 52a1 51a2 51a2+52a1 A = 51a2 .t2) .`t (al a2) + C 11 172 ir12 / \ t 2 ) r1 (7r1 7r2) ( t2 7rltl + J + eAt (al a2) ( 71(t2 .tl) 7r2t2 + eat (a17r2 .6 Let B be hyperexponential. The present treatment is somewhat more probabilistic.4 yields the renewal density as u(t) = 5152 e.a27rl) (tl . Then Q= 0 55 )+(1o)=( j ad ).e2bt) 13 Example 2 .52) 25152 51x2+5251 51a2+5251 Notes and references Renewal theory for phasetype distributions is treated in Neuts [268] and Kao [221]. and Example 2. MATRIXANALYTIC METHODS e.
T + to+). Corollary 3. Then: (a) G+ is defective phasetype with representation (a+. Considering the first. Within ladder steps. Here we have taken the terminating Markov process underlying B with two states.3. however. The stars represent the ladder points ST+(k). with 0 denoting the Poisson intensity. and if there is a subsequent ladder step starting in j whic occurs w.2. T).e. add a more selfcontained explanation of why of the phasetype structure is preserved. Now just observe that the initial vector of {mx} is a+ and that the lifelength is M.(u) = a+e(T+tQ+)u Note in particular that p = IIG+II = a+e. and rewriting in matrix form yields the phase generator of {my} as T + ta+. {St} the claim surplus process. cf. Proof The result follows immediately by combining the PollaczeckKhinchine formula by general results on phasetype distributions: for (a).3. r(u) the time of ruin with initial reserve u. T). 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. and M is zeromodified phasetype with representation (a+.1 on the next page.p. marked by thin and thick lines on the figure.i. 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 2. itself phasetype with the same phase generator T and the initial vector a+ being the distribution of the upcrossing Markov process at time ST+_. (b) V. a+j. the Markov processes representing ladder steps can be pieced together to one {my}. For (b). Thus the total rate is tip + tia+. Next. T) where a+ is given by a+ = . i. which occurs at rate ti. use the phasetype representation of Bo. . Then each claim (jump) corresponds to one (finite) sample path of the Markov process. Since the results is so basic.1 Assume that the claim size distribution B is phasetype with representation (a. 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. 3. The essence is contained in Fig.) = F(ST(o) E •. B the claim size distribution. we shall.f3aT1. We asssume that B is phasetype with representation (a. T(0) < oo) the ladder height distribution and M = supt>o St. G+(. represent the maximum M as the lifetime of a terminating renewal process and use Corollary 2.
Q = 3 and b(x) = . T = (3 . Figure 3.3.QaT1. This is in fact a simple consequence of the form of the excess distribution B0.2 Assume that .t t d kkt S..228 CHAPTER VIII. 7e7x 2 2 Thus b is hyperexponential (a mixture of exponential distributions) with a (2 2 ).1 . 0 Example 3...M {mx} ST+(2)  S .1 This derivation is a complete proof except for the identification of a+ with .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 . see Corollary 2. 3e3x + . MATRIXANALYTIC METHODS t .
For the compound Poisson model. Fig. 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.T).1 In the zerodelayed case. 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 .2 are taken from Gerber [157].6. T). cf.1 can be found in Neuts [269] (in the setting of M/G/1 queues. That is. 3. In the next sections.4. For further more or less explicit computations of ruin probabilities. The parameters of Example 3. The result carries over to B being matrixexponential. It is notable that the phasetype assumption does not seem to simplify the computation of finite horizon ruin probabilities substantially. T) for some vector a+ = (a+. this was obtained in Section 3. 3. For an attempt.^(u) = a+e( T+ta+)ue = 24eu + 1 e6u 35 35 0 Notes and references Corollary 3. the discussion around Fig. 4 The renewal model We consider the renewal model in the notation of Chapter V. if we define {mz} just as for the Poisson case (cf.6). We shall derive phasetype representations of the ruin probabilities V) (u). but there the vector a+ is not explicit but needs to be calculated (typically by an iteration). THE RENEWAL MODEL This is the same matrix as is Example 1.1 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+. We assume p = PB/µA < 1 and that B is phasetype with representation (a.1): Proposition 4. see Stanford & Stroinski [351] . where a+ is the (defective) .and Markovmodulated models. see Shin [340]. his derivation of +'(u) is different. but that such a simple and general solution exists does not appear to have been well known to the risk theoretic community. see Section 6. with A denoting the interarrival distribution and B the service time distribution. 0(8) (u) (recall that z/i(u) refers to the zerodelayed case and iY(8) (u) to the stationary case).j).4. the duality result given in Corollary 11. (a) G+ is of phasetype with representation (a+.
We have now almost collected all pieces of the main result of this section: Theorem 4 . where u w(a +) = aA[T + to+) = a J0 e(T+t+)1A(dy).3.1. B0 is phasetype with representation (aT1/µa. the calculation of the first ladder height is simple in the stationary case: Proposition 4.3 a+ satisfies a+ = V(a+). Then . (c) {mx } is a (terminating) Markov process on E. Then {m.6.2 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).230 distribution of mo. The key difference from the Poisson case is that it is more difficult to evaluate a+. But by Corollary 2. (4. which for numerical purposes can be solved by iteration. Fig. it follows by integrating y out that the distribution a+ u of mo is given by the final expression in (4. the form in which we derive a+ for the renewal model is as the unique solution of a fixpoint problem a+ = cp(a+).T).4 Consider the renewal model with interarrival distribution A and the claim size distribution B being of phasetype with representation (a. Proof Obviously. where B0 is the stationary excess life distribution corresponding to B. where a(8) = aT1/PA. Hence by Theorem 11. CHAPTER VIII. In fact. Also. with intensity matrix Q given by Q = T + to+. 4.1).*} from {St+y .Sy} in the same way as {mx} is defined from {St}.T)• Proposition 4. MATRIXANALYTIC METHODS (b) The maximum claim surplus M is the lifetime of {mx}. G(') = pBo.1) Proof We condition upon T1 = y and define {m. obviously mo = m. cf.*'} is Markov with the same transition intensities as {mx}. the Palm distribution of the claim size is just B. Since the conditional distribution of my given T1 = y is ae4y. but with initial distribution a rather than a+.5. Nevertheless.T).
•. It remains to prove convergence of the iteration scheme (4.2 ) follows from Proposition 4.3) (defined on the domain of subprobability vectors .1). a+) > 0 = a+o) implies a+) _ (a+) > W (a+)) = a+) .e.0.^(8)(u) = a ( 8)e(T+ta +) xe.1/pA. a+2) = ^p (a+l)) .1 . and that this is given by Proposition 4.M. .4. Hence ^p(.1(b). the maximum claim surplus for the stationary case has a similar representation as in Proposition 4. .2. I {mx} .^(u) = a+e ( T+ta+)xe.0) is an increasing function of /3. (4.1) and a(8) _ aT.. THE RENEWAL MODEL 231 . a+ can be computed by iteration of (4. In particular .3). by a+ = lim a +n) where a+°) . The second follows in a similar way by noting that only the first ladder step has a different distribution in the stationary case. (4. Furthermore .2) where a+ satisfies (4. i y ^ T1= y `•r Figure 4.3) Proof The first expression in (4...1 by noting that the distribution of mo is a+. thus .. only with initial distribution a(*) for mo. The term tf3 in cp(i3) represents feedback with rate vector t and feedback probability vector (3. a+l ) = cp (a+°)) . i.
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)'t • A[s] (4.ST. Theorem 4. F[s] being interpreted in the sense of the analytical continuation of the m. To prove the converse inequality. 0 0 We next give an alternative algorithm. limn4oo a ) < a+.} can contain at most n . the corresponding right eigenvector may be taken as (sI .4) whenever EeR(S)U < oo.4). Similarly. (4.g.T)It. Thus by (4. . Thus . both quantities are just 0 .T)1t.4) makes sense and provides an analytic continuation of F[•] as long as s ¢ sp(T). s ¢ sp(T). so to complete the proof it suffices to show that &+ < a+) for all n. (4. with B[s].5) Since s $ sp(T). which links together the phasetype setting and the classical complex plane approach to the renewal model (see further the notes). MATRIXANALYTIC METHODS and (by induction ) that { a+ n) } is an increasing sequence such that limn. Let Fn = {T1 + • • • + Tn+1 > r+}be the event that {my} has at most n arrivals in [T1. 7+ ]. n) &+n) T a+. For n = 0..232 CHAPTER VIII. To this end. Then s is an eigenvalue of Q = T + ta+ if and only if 1 =. Proof Suppose first Qh = sh. Obviously. a+ ) exists . In that case. let F be the distribution of U1 . Fn ). 0 = a+) < a+ yields a+) _ (a+0)) (a+) = a+ (n and by induction that a(n) < a+ for all n . Then e4'h = e82h and hence sh = Qh = (T + taA[Q])h = Th + A[s]tah. Then F[s] = a(sI .f.P[s] = A[s]B[s].) = P(mTl = i.1. and hence we may assume that h has been normalized such that ahA[s] = 1.1 arrivals (n arrivals are excluded because of the initial arrival at time T1 ). this implies that ahA[s] # 0. the normalization is equivalent to F(s) = 1. Assume the assertion shown for n .2.5) yields h = (sI . and let &+". Then (4. However. Then each subexcursion of {St+Tl .4.T1.5 Let s be some complex number with k(s) > 0. we use an argument similar to the proof of Proposition VI.
.. T) with a+ = a(QT)/at.. t(ry) > 0. Pd in the domain ER(s) > 0 . the classical algorithm starts by looking for roots in the complex plane of the equation f3[y]A[ry] = 1. 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 ). hd..6.T)lt = sh. Let d denote the number of phases.lt we get Qh = (T + to+)h = T(sI . the matrix Q in Theorem 2. and hence by the WienerHopf factorization identity (A.T)lt + t = s(sI . . and define hi = (piI .. Hence with h = (sI T). Q = CD1 where C is the matrix with columns hl. pdhd. ....p1i . . Pd with corresponding eigenvectors hl. W v M(d) in the notation of Chapter V)...6 Suppose u < 0. THE RENEWAL MODEL 233 Suppose next F(s) = 1. hd..T) = 1 ata+ = a+. This immediately implies that Q has the form CD1 and the last assertion on the diagonal form .type with representation (a+..6) i=1 i=1 Proof Appealing to Theorem 4. .. . and the topic is classic both in risk theory and queueing theory (recall that we can identify 0(u) with the tail P(W > u) of the GI/PH /1 waiting time W.. we get at a(Q . Notes and references Results like those of the present section have a long history. As in Corollary 4. Further.T)It. Given T has been computed. (4. letting vi be the left eigenvector of Q corresponding to pi and normalised by vihi = 1 ..5. .. 0).' that the equation F(s) = 1 has d distinct roots p1. Since R(s) > 0 and G _ is concentrated on (oo.. in turn. 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. In older literature .9) we have G+[s] = 1 which according to Theorem 1.1 has the d distinct eigenvalues . Then G+ is phase. Q has diagonal form d d Q = dpivi®hi = dpihivi. D that with columns p1 hl. we have IG_ [s] I < 1 . and the solution is . Corollary 4.5(c) means that a+(sI T)1t = 1.4.. yd satisfying R(ryi) > 0. This gives d roots 'y.
We assume that each B. see Neuts [269]. where R is an unknown matrix. Neuts and his students.4. 5 Markovmodulated input We consider a risk process {St } in a Markovian environment in the notation of Chapter VI.. The exposition here is based upon [18]. starting around in 1975.) d (see. The matrix. The solutions are based upon iterations schemes like in Theorem 4. whereas the approach was introduced in queueing theory by Smith [350].type assumption . [270] and Latouche & Ramaswami [241]. the intensity matrix is A and the stationary row vector is ir . The distribution of W comes out from the approach but in a rather complicated form . but the models solved are basically Markov chains and processes with countably many states ( for example queue length processes ). a pioneering paper in this direction is Tacklind [373]. and the distribution of an arrival claim is B. MATRIXANALYTIC METHODS d F 1 + a J e°" ip(u) du = Ee°w = 11(t. For further explicit computations of ruin probabilities in the phasetype renewal case .g. T('). which contains somewhat stronger results concerning the fixpoint problem and the iteration scheme.. the fixpoint problems look like R=Ao+RAI+R2A2+ . the background Markov process with p states is {Jt}. involving . That is . 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. The number of elements of El=> is denoted by q. e.234 then in transform terms CHAPTER VIII. Numerical examples appear in Asmussen & Rolski [43]. similar discussion appears in Kemperman [227] and much of the queueing literature like Cohen [88].type assumptions are basic. The arrival rate in background state i is a. an alternative approach (the matrixgeometric method ) has been developed largely by M. the ruin probability can be found in matrixexponential form just as for the renewal model.. is phasetype. [119].contained derivation). For surveys . with representation say (a(' ). In queueing theory. Here phase. see Dickson & Hipp [118].exponential form of the distribution was found by Sengupta [335] and the phasetype form by the author [18]. E(t)).F. It turns out that subject to the phase. In risk theory. Asmussen & O'Cinneide [ 41] for a short self. and appears already in some early work by Wallace [377].
The key unknown is the matrix K. Diagonalization Consider a process {(It. say with slope r(i) on intervals where It = i. 5.Vt)} obtained by time reversing the I component. p = ql = Q2 = 2.1 In Fig. •.1. We start in Section 5a with an algorithm involving roots in a similar manner as Corollary 4. However. 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 two environmental states are denoted o. O. 5. MARKOVMODULATED INPUT 235 some parameters like the ones T or a+ for the renewal model which need to be determined by similar algorithms. The stationary distribution is obtained by finding the maximum of the Vcomponent of the version of {(It. 5a Calculations via fluid models.6. The connection between the two models is a fluid representation of the Markovmodulated risk process given in Fig. and the one E(•) for B.1. Section 5b then gives a representation along the lines of Theorem 4.2. for which the relevant fixpoint problem and iteration scheme has already been studied in VI. has states o. This calculation in a special case gives also the ruin probabilities for the Markovmodulated risk process with phasetype claims. states . the phase space E(°) for B.5. (a) 0 0 ♦ o ° tl ♦ • 0 0 o } o o (b) 0 } ♦ • 0 o f o Figure 5. Vt)}t>o such that {It} is a Markov process with a finite state space F and {Vt} has piecewiese linear paths.4. 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.
t. corresponding to the partitioning + Epp). The fluid model on Fig . Thus F = {o.Vt)} is then obtained by changing the vertical jumps to segments with slope 1. MATRIXANALYTIC METHODS 4.31a(l) (/3i)diag . we have more martingales at our disposal. V. a) = 1. In the general formulation . If s is such a number. whereas Ee8s' = oo for all t and all s > so where so < oo. < oo for all s. j = 1. 5. a E E(i) } . consider the vector a satisfying (A + (13i(Bi[ s] . Second. of E into components indexed by E. in the fluid model Eel'. resp. This implies that in the fluid context.1))diag ) a = sa and the eigenvector b = . Eli) + Proposition 5. '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. 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. Bi[s] = a(i)(T(i) + sI)it('). 2. 4. 4.1(a).1(b) {(It .(Ni)diag r(i. First. F is the disjoint union of E and the Eli). Let E denote the matrix .1) if and only if s is an eigenvalue of E.1))diag + sII = 0 (5. a) of {It}. A claim in state i can then be represented by an E()valued Markov process as on Fig.A 0 Or 1A/ _ t(i) 0 t(2) 0 0 0 0 0 t(3) 0 T1 0 0 0 .92a(2) 0 0 T(2) 0 0 0 f33a(3) 0 0 T(3) with the four blocks denoted by Ei„ i. o. Recall that in the phasetype case.236 CHAPTER VIII. •. r(i) _ 1. 5. The intensity matrix for { It} is (taking p = 3 for simplicity) I A .1 A complex number s satisfies 'A+ (f3i(Bi[s] . i E E. a) : i E E. F = E U { (i. 4}.
0 . resp .sI ()3i)diag .E21a + sd = sd.sI 0 0 0 T(2) .T('))1t(i) .(sI . and let d = (sI .E12E22 E21 I .A . Then E21c+E22d = E21a .sI 0 0 0 T(3) .sI + E12 (sI .E22)1 E21) a = 0.32a(2) (/3i)diag .sI+ ((3ia(i)(T(i) .E22 . it follows that Ell E12 ( E 21 E22) (d) = s 1 d I .sI) (sI . with Eii replaced by Eii . E(1) + + E(P). MARKOVMODULATED INPUT 237 indexed by E.1). t(1) 0 0 then also 0 t(2) 0 . Noting that E11c + E12d = se by definition. it follows that if Qla(1) 0 0 . For the assertions on the eigenvectors. assume that a is chosen as asserted which means (Ell . where c. iEE (a> of 0* 1 AI.E22)1 E21a E21a . d = (sI E22)1E21a = E ai(sI .sI. d correspond to the partitioning of b into components Proof Using the wellknown determinant identity Ell E12 E21 E22 E22 I ' I Ell .sI)1t)) iag I = 0 which is the same as (5.5.E22)1 E21a.A .Nla(1) 0 0 T 1. Then (up to a constant) c = a. c = a.sI 0 0 t(3) 0 0 = 0.
4 Assume that E has two states and that B1.5..v) = v) I. pi(u. j. .. .. a) = Pi (Vw(u. We can take a = c = 1 and get d = (s + b)16 = 5/(3 = 1/p. e89uc(e)) (d(1) . < 0 and let b(v) = I d(„)) be the right eigenvector corresponding to s. a) and noting that i1 (u) = >I j.4 that {e"1b(v) is a martingale .v) = j).v) = (j. j.a)d^ ).. w(u. s2 are the negative eigenvalues of Al +01 A1 E _ A 2 b1 0 52 A2 +32 0 ..v) = Optional stopping at time w (u. Thus 0(u) = esu/d = pe7 ° as u should be.v)=inf{t >0:Vtu orVt=. j.v}..j)c v . a)). we first look for the negative eigenvalue s of E = I 0 I which is s = ry with yy = b . a )d(a + e8 °vpi (u . q. Letting v ^ oo and using Rsv < 0 yields e8'u = Epi(u.v. w(u)=inf{t >O:Vtu}.3 Consider the Poisson model with exponential claims with rate 5.v) = = p i( u .. Proof Writing Or'Alb( v) = svb( v) as (AI .sv)b(v) = 0. Iw(u.upi(u. d("))1 e. .2 Assume that E = Or 'Al has q = ql + + qp distinct eigenvalues si. j) pi( u . v > 0. .. it follows by Proposition II./' u = e' (esiuc ( 1) . a)).. Example 5 . B2 are both exponential with rates 51 i b2. To determine 0 (u). c j. Example 5 ...pi(u. v. define w(u. Here E has one state only. a). j.O. For u. the result u follows. v = 1.. v.( u.Q. u) Iw(u.. v. a) = (j.j.238 CHAPTER VIII. j. Then . Then we get V)i (u) as sum of two exponential terms where the rates s1. sq with $2s. MATRIXANALYTIC METHODS Theorem 5. . v) yields C{V) = e8 .a Solving for the pi(u. I' i( V P2 (w (u) < oo. ..
is 0 /3 f R(i . (y. •) is phasetype with representation (E(i).y = to B k7 j # k In particular. (') a T( However . i.xxej • a 00 oo el . oo)) j)ye. (5.x) 00 f ° (') (j) eT (yy)edx .6 For i E E.5 G+(i. 9('). 0 Theorem 5 . j.33(e = 0 a(j))(K ®T ( j))(ej (9 I). U) where t(j) + t(j)O(j j = k uja. 8^')IT(j)) where e 3^') =. j. In terms of K. dx)Bj(y .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. j.k.b (u) = Pi(M > u) = 9(i)euue.3j eye. MARKOVMODULATED INPUT 239 5b Computations via K Recall the definition of the matrix K from VI.3) .h.2. according to VI. 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.s. the Pidistribution of M is phasetype with representation (E(1) + + E(P). Proof We must show that G+ (i.2) the l.( 2.5.
the initial value of (i. An alternative characterization is that such a distribution is matrixexponential. a) is obviously chosen according to e(`).5).. and lifelength M. i. t) is the representation of the matrixexponential distribution/density): Proposition 6. Furthermore. t = (0 0 . some square matrix T and some column vector t (the triple (a.1 Let b(x) be an integrable function on [0. Starting from Jo = i. (6. in many cases where such expressions are available there are classical results from the prephasetypeera which give alternative solutions under the slightly more general assumption that B has a Laplace transform (or.. and a new ladder step of type k must start in phase y. T. +bn0i1 0n +a10n1 +. 6 Matrixexponential distributions When deriving explicit or algorithmically tractable expressions for the ruin probability... Piecing together these phase processes yields a terminating Markov process with state space EiEE E('). which occurs at rate t^^7. if b* [0] = b1 +b20+b302 +. a m. we have sofar concentrated on a claim size distribution B of phasetype. u Notes and references Section 5a is based upon Asmussen [21] and Section 5b upon Asmussen [17]. bn1 bn).2) .) which is rational. Bk7 .p. i. which occurs w. However.y) to occur when j # k.. For j = k. the current ladder step of type j must terminate. say.e. equivalently. a) to (k. the ratio between two polynomials (for the form of the density. intensity matrix U.e. MATRIXANALYTIC METHODS Proof We decompose M in the familiar way as sum of ladder steps . we have the additional possibility of a phase change from a to ry within the ladder step. Associated with each ladder step is a phase process. 0 1)'. This yields the asserted form of uja. with phase space EU> whenever the corresponding arrival occurs in environmental state j (the ladder step is of type j)... which occurs at rate t(i).240 CHAPTER VIII. oo) and b* [0] = f °O eBxb(x) dx the Laplace transform. and it just remains to check that U has the asserted form. see Example 1. +aii10+anI then a matrixexponential representation is given by b(x) = aeTxt where a = (b1 b2 . Then b*[0] is rational if and only b(x) is matrixexponential.2.. .f. Numerical illustrations are given in Asmussen & Rolski [43].g.k y. For a transition from (j. that the density b(x) can be written as aeTxt for some row vector a.
4) 0 0 1 c This representation is complex. For a proof. namely to asssume the roots 6l.47x2 3 1 0 .. T= 0 0 1 .1) as E 1 c. u giving b(x) = E 1 ciebiz/bY. shows that the distribution B with density b(x) = c(1 cos(21r x))ex. T. where c = 1 + 1/47r 2..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. since 1 + 4ir2 03 + 302 + (3 + 47x2)0 + 1 + 47r2 it follows by (6. (6.. but as follows from Proposition 6. . (6. cannot be phasetype.. .e(tai1)x/2 + e'T) it follows that a matrixexponential representation ()3. Namely. .an_3 an _ 4 .h.s.3) that we can take 0 1 0 0 a= (1 + 47r2 0 0). MATRIXEXPONENTIAL DISTRIBUTIONS 241 T = 0 1 0 0 0 . we can always obtain a real one (a.2 A remarkable feature of Proposition 6. bn of the denominator to be distinct and expand the r. S = f c/2 0 21ri .3 A set of necessary and sufficient conditions for a distribution to be phasetype are given in O'Cinneide [276]. S. 1 .3) was suggested by Colm O'Cinneide./(0 + bi).3) 0 0 0 0 0 .1. 0 0 . t= 0 . u Remark 6. Writing b(x) = c(e( 2ni1 ) y/2 . b(x) > 0 for x > 0. s) is given by 27r i . s = c/ 2 ..1 0 . t). Example 6 . then b*[0] = a(0I T)1t which is rational since each element of (01 .. 0 1 an an1 an _2 . matrixexponentiality implies a rational transform. ..(6.T)1 is so.1 0 0 )3 = (111).. Thus. see Asmussen & Bladt [29] (the representation (6.. personal communication). 0 0 0 0 1 0 0 .2).2). One of his elementary criteria. . of (6.6. The converse follows from the last statement of the theorem. a2 a1 Proof If b(x) = aeTxt.47r2 3 . (6.
For the first. 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. 0. then: Proposition 6. Consider the distribution with density = 15 ((2e2x . and that the minimal number of phases in a phasetype representation increases to 0o as 5 .1 to get i (u) = f3esus.3. recall that t = Te) that if B is phasetype and (a. then 5(u) = a+e(T+t+)uTle where a+ = /3aT1.4 This example shows why it is sometimes useful to work with matrixexponential distributions instead of phasetype distributions: for dimension reasons .1)2 + 6). . Then (cf. we take as starting point a representation of b* [0] as p( O)/q(9) where p. T. (6. (6.5) Thus. MATRIXANALYTIC METHODS Example 6 . For the second algorithm.6) in Section 3 seems to use the probabilistic interpretation of phasetype distribution in an essential way. q are polynomials without common roots. t) a phasetype representation with a the initial vector.5 (6. As for the role of matrixexponential distributions in ruin probability calculations. we have represented ti* [0] as ratio between polynomials (note that 0 must necessarily be a root of the numerator and cancels). we shall only consider the compound Poisson model with arrival rate 0 and a matrixexponential claim size distribution B.242 CHAPTER VIII. and present two algorithms for calculating '(u) in that setting. We recall (see Section 3. Corollary 111.6) holds true also in the matrixexponential case. T. and can use this to invert by the method of Proposition 6.4) the Laplace transform of the ruin probability is /g(e)PO 0*[e] _ /' eeu^G(u)dU = 0 9(/3a0p(9)ap (9)/q(9)) . we use a representation (a. leading to matrix calculus in high dimensions when b is small.6) The remarkable fact is. 7 + 155ex b(x) Then it is known from O'Cinneide [276] that b is phasetype when 6 > 0. T the phase generator and t = Te. t) of b(x). that despite that the proof of (6.1 shows that a matrixexponential representation can always be u obtained in dimension only 3 independently of J.
T)1t. Presumably.T)1 J0 00 b(x) dx = f aT1t.T .T)1 + (6I .1 = A1 .7) 9( cf. .1t = f3a (0I T)1T1t .T)1t)1a +(9I .to+)1T . with A = 91T.b* (6. (6. since (91T)1T .T . Now.T)1T 2 = and 1 = AB IT2 + 82T .to+)1T .T . we get (91.'t. Then in Laplace transform formulation . b+ = a +(BI .1UB(B + BVA1UB).T)1 (91.T)1T1t.T)1 so that b* b** b** a+(9I .A .1BVA1. but we shall give an algebraic proof.6.T)1t ( l . b+ = a+(9I . MATRIXEXPONENTIAL DISTRIBUTIONS 243 Proof Write b* = a(9I .T)1ta+(OI . 519) (A + UBV ).a+(9I . (6.t.to+)1 = (BI .1 = ^(T1 + ( 91T)1).B=land V=a+. the assertion is equivalent to a+(BI . this can be verified by analytic continuation from the phasetype domain to the matrixexponential domain . From the general matrix identity ([331] p.1 + 82 (9I .1 + b+ = b++ 1 . (91. U =.6b* .1t du = . we get b+ = 0aT1(9I T).T).1t = b* .T)1 + 1 ib* (91. xb(x) dx = aT2t.6).5 ).
3a (1 0 T 2 + 1 T 102 (9I + 02 1 T)1) t P + 7. 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. which is selfexplanatory given Fig. to piece together the phases at downcrossing times of {Rt} (upcrossing times of {St}) to a Markov process {mx} with state space E.1.la.h. 0 Notes and references As noted in the references to section 4.1. cf. In Corollary VII.82b*. A key tool is identifying poles and zeroes of transforms via WienerHopf factorization.s.3 is taken).T)1T2t . Much of the flavor of this classical approach and many examples are in Cohen [88].5 is similar to arguments used in [29] for formulas in renewal theory. (for some remarkable explicit formulas due to Paulsen & Gjessing [286]. T). VII.b*). but the argument of [286] does not apply in any reasonable generality). 7. Lipsky [247] and Asmussen & O'Cinneide [41].1t = /3a (9I .8. For expositions on the general theory of matrixexponential distributions. 7 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate 0. the ruin probability(u) was found in explicit form for the case of B being exponential.244 CHAPTER VIII. a key early paper is Cox [90] (from where the distribution in Example 6. of (6. premium rate p(r) at level r of the reserve {Rt} and claim size distribution B which we assume to be of phasetype with representation (E.T)1)t = 8 (1 .1) is the same as the r. see Asmussen & Bladt [29]. The proof of Proposition 6. 3. See Fig. a. From this it is straightforward to check that b**/(b+ .1. 7a Computing O(u) via differential equations The representation we use is essentially the same as the ones used in Sections 3 and 4. MATRIXANALYTIC METHODS . some key early references using distributions with a rational transform for applied probability calculations are Tacklind [373] (ruin probabilities) and Smith [350] (queueing theory). see the Notes to VII.7).8 a(T1 + (01.1. . /3aT1(0I .T)1T.
t) is the vector of state probabilities for mt.1 The difference from the case p(r) = p is that {m2}. Since v(u) = (vi(u))iEE is the (defective) initial probability vector for {m8}.1z I. P(tl. though still Markov. Figure 7.1 A(0) = v(u) and A'(t) = A(t)(T + tv(u . >iEE Vi (U) is the ruin probability for a risk process with initial reserve 0 and premium function p(u + •). Also. 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. i. t2) = exp where Q(t) = ds [P(t.e.u)e = A(u)e (7. O<.1) where A(t) = v(u)P(0.7.t2) be the matrix with ijth element P (mt2 =j I mtl = i). Given the v(t) have been computed. t + s) . Note that in general >iEE Vi (U) < 1. in contrast to Section 3. Let P(tl. RESERVEDEPENDENT PREMIUMS 245 Rt l0 u . Proof The first statement is clear by definition. Ai(t) = P(mt = i). In fact. is no longer timehomogeneous.I] I 80 { tq f Q(v) dvl t1 1 . 0 < t < u.tl < t2 < u.t)). By general results on timeinhomogeneous Markov processes. the A(t) and hence Vi(u) is available by solving differential equations: Proposition 7. we obtain V)(u) = P(m„ E E) = v(u)P(0. the definition of {m8} depends on the initial reserve u = Ro.
A'(t) = A(t)Q(t) = A(t)(T + tv(u . given A. we get vi(u) = aidt + (1 .4) jEE jEE Proof Consider the event A that there are no arrivals in the interval [0. Given A.t)).246 CHAPTER VIII. those corresponding to state changes in the underlying phase process and those corresponding to the present jump of {Rt} being terminated at level u . or it stops between level u + p(u)dt and u. (7. the probability of which is 1 . Hence Q(t) _ T + tv(u . jEE .(tai + vi(u) E vj(u)tjp (u)  Q + vj (u)tjip ( u). from a computational point of view the remaining problem is to evaluate the v(t). the probability that level u is downcrossed for the first time in phase i is ai. {mx} has jumps of two types. 0 Thus.(u) p ( u) = . the probability that level u + p(u)dt is downcrossed for the first time in phase j is vj (u + p(u)dt). In the first case. Given A'.t). Given this occurs. 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. MATRIXANALYTIC METHODS However.3dt. vi. 0 < t < u.t) for the second.Qdt) vi(u) + vi'(u)p(u)dt + p(u) dt E{tji+tjvi(u)}. two things can happen: either the current jump continues from u + p(u)dt to u. Proposition 7. 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.t and being followed by a downcrossing.Sj i)p(u)dt • tji = Sji + p(u)tji dt. Thus. whereas in the second case the probability is p(u)dt • tjvi(u). dt].2 For i E E. The intensity of a jump from i to j is tij for jumps of the first type and tivj(u . the probability of downcrossing level u in phase i is 8ji(1 + p(u)dt • tii) + (1 .
say.^ 0. Since the processes Rt and Rt coincide under level B. Rt . vi (U) = lim v= (u). (u) on both side and dividing by dt yields the asserted differential u equation. after a certain level v. we have p(r) = p = vi (u) 0aTe.5).) is the tail of a (defective) random variable so that P(Bv) + 0 as v 4 oo.i7rT1/p. Let p" (t)...3 For any fixed u > 0. supRt>v l t<7 I where o. then P(A n Bv) _ P"(A n BV'). When solving the differential equation in Proposition 7. we can first for a given v solve (7. Thus. . To deal with this..h. consider a modification of the original process {Rt} by linearizing the process with some rate p. (v) is given by the r. From Section 3. we face the difficulty that no boundary conditions is immediately available. Then pv(r) p(r) r < v p r>v ' and (no matter how p is chosen) we have: Lemma 7. denotes the time of downcrossing level u . This yields v. <oo.) + 0 as v + oo. refer to the modified process. F" etc. (u) for any values of u and v such that u < v.7. RESERVEDEPENDENT PREMIUMS 247 Subtracting v. P u which implies that 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. and similarly P"(Bv) .. say.4) backwards for {va (t)}v>t>o.2.)P"(AnB. of (7. Then P(B. V .) P"(AnBv) = P(AnB. starting from v"(v) = .s. Now since both P(A n Bv) 3 0 and P"(A n Bv) .0 as v + 00 we have P(A) P"(A) = P(AnBv)+P(AnBv) P"(AnB.
the evaluation of Vi(u) requires q(u) = 1 .e.zp1(u)/(1 . r<v r > v. i.1.1a. while the fourthorder RungeKutta method implemented in [30] gives 0(n5). 0 < u < v. (u)}. the probability of ruin between a and the next upcrossing of v.RQ (defined for or < oo only) is defective phasetype with representation (v(u)..7) f o (the integral is the contribution from {R. Recall that q(w) is the probability of upcrossing level v before ruin given the process starts at w < v... say. < 0}). such that Rt coincide with RI under level v and with Rt above level v. To evaluate p1(u). p(r) P. MATRIXANALYTIC METHODS Next consider a sequence of solutions obtained from a sequence of initial values {v.1. (7.. T). assuming u > v for the moment. Then v(u) is the initial distribution of the undershoot when downcrossing level v given that the process starts at u. > 0} and the last term the contribu tion from {R. as well p1(u). Therefore u pl(vvueTa t 1.x) dx f v(u)eT xt dx .7) equals 01 (v .z51(v)). Corollary 3. We recall from Propositon VII..v v(u)eTat 1 1 . The f iin in (7. cf. let v(u) = a+2ieiT +ta+>)(uv).248 CHAPTER VIII. 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. 2u. where v = inf It > 0 : Rt < v}.q(v dx +( ) ) = ( ) ( q( )) vueTva (7. for u > v the distribution of v .10 that in addition to the O'(•). 7b Twostep premium rules We now assume the premium function to be constant in two levels as in VII. 1/n. where. The algorithm based upon numerical solution of a Volterra integral equation (Remark VII. v = u. The precision depends on the particular quadrature rule being employed. typically the complexity in n is 0(n2) for integral equations but 0(n) for integral equations. However.V" M 0 . 3u etc. which is available since the z/i'(.1. p2. 2/n.9. The trapezoidal rule used in [288] gives a precision of 0(n 3). Let ii'( u) = a+'ie(T+ta +^)"e denote the ruin probability for R't where a+ = a+i) = laT1/pi.) are so.6) We may think of process Rt as pieced together of two standard risk processes RI and Rte with constant premiums p1. numerically implemented in Schock Petersen [288]) and the present one based upon differential equations require both discretization along a discrete grid 0.
v(u ) eTV e J v(u)eTxtz/)l (v .8) equals v v(u)eTxta+2) e(T+ta +))( vx)edx which using Kronecker calculus (see A. so we consider the nontrivial case example p2 = 4 and p1 = 1.to+))11 {e{T®(Ttoy+ ))}„ .(7 The arrival rate is (i = 3. Then one gets X20 20 21 f 1ea1(u v) + 1 3 3 ^ A 2(u e .e6u 35 .8) The integral in (7.v) 1eai(u v) + 7 7 1 e\2(u v) 1 3 ^') eA2 (u.21 = ? yields 0(u) = 1.jl (t ®e) Thus.4) can be written as (Y(u) ®a+)e(T+t°+>)°1 (T ® (T . RESERVEDEPENDENT PREMIUMS 1 .be the eigenvalues of T + to( 2 ). (7.2V"2.4 Let {Rt } be as in Example 3.7.v(u)eTVe .x) dx} V 1 1(v) f V v(u) eTxt.1 from which we see that pl (u) = 1 + 1 249  1 .e6v Let Al = 3 + 2V'2.01 (v . I.e. B is hyperexponential corresponding to 3 0 3 a(2 2)' T= ( 0 7 t.u .24ev . Example 7.and A2 = 3 .24e. p2 < 3. Since µB = 5/21.2.^1(v) 1 . 01(u) _ 24 u + 35 e6u 1 35 e 4(u) _ 35 . From Example 3.v) + (2^ + 3v2 ea'(u " .2.v(u)eTve).x) dx 1 ^(v) ( 1 . all quantities involved in the computation of b(u) have been found in matrix form..
MATRIXANALYTIC METHODS From (7.1)' ?.1 V2 = 4e5"+6 35e6v . and one gets 12e5" .1 Thus. pi (u) = p12(u)/p1 l(u) where p1i(u) p12(u) 35e6v .1. Notes and references [30]. 192esv + 8 P1 .2 35e6v . 21 3 In particular. The analysis and the example are from Asmussen & Bladt .21(35e6v .24es" . ) e sv + ( 2v/2./2 ea1(u") .+ it (3 4'I 1 ea2(uv e1\2(u") 7 + ( 32 +4.7) we see that we can write pi (u) = v(u)V2 where V2 depends only on v.24es" .24e5v .v)esv + 7 4_ 2.b(v) = 192esv +8 35e6v + 168esv + 7* Thus all terms involved in the formulae for the ruin probability have been exu plicitly derived.250 CHAPTER VIII..24e5v ./2) ea 1(u .
and instead we shall work within the class S of subexponential distributions . For example. B(x) = L(x)/x" where a > 0 and L(x) is slowly varying. L(tx)/L(x) 4 1. For further examples.4. (b) the lognormal distribution (the distribution of eu where U . x 4 oo. we require that B is concentrated on (0. 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.f.g. the exponential change of measure techniques discussed in II. B(x) = ex0 with 0<0<1.N(µ. x 2iror2 (c) the Weibull distribution with decreasing failure rate . III. a2)) with density 1 e(logyFh) 2/2az . see I.B(x). Some main cases where this lighttail criterion are violated are (a) distributions with a regularly varying tail. for all t > 0. 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. A rough distinction between light and heavy tails is that the m.46 and at numerous later occasions require a light tail. For the definition . oo ) and say then that B is subexponential (B E S) if 251 .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) = 1.2b.
X2 > x) = 2B(x) .'s X1. proving (a). X2 with distribution B. X2) > x}. We later show: Proposition 1. the behaviour in the lighttailed case is illustrated in the following example: Example 1. X2) > x} C {X1 + X2 > x}. and thus the lim inf in (b) is at least lim inf P(max(Xi. X2 but none of them exceeds x. As contrast to Proposition 1.1 Let B be any distribution on (0. To capture the intuition behind this definition. That is.B(x)2 .p. 1/2 'typical' (with distribution B) and w. X2) > u x)/B(x) = 2. P(Xi <yI Xi+X2>x) 1B(y). the r. 1/2 it has the distribution of X1I X1 > x.p. x 3 00.252 CHAPTER IX.2 If B E S. Then X1 +X2 has an Erlang(2) distribution with density yeY so that B*2(x) xex. note first the following fact: Proposition 1. Thus the liminf in Proposition 1. that is. (b) liminf BB(() ) > 2.v. then (with high probau bility) so are both of X1. In terms of r.F(X1 > x. B(x) ax.2B(x).2. oo). That is. Thus . then P(X1>xI X1+X2>x)* 2. oo). Then: (a) P(max(Xi. given X1 + X2 > x. X2) > x) ^' 2B(x). X1 is w. the distribution of independent r. P(max(Xi.v. 1). .v. one can check that x x where U is uniform on (0. Proof By the inclusionexclusion formula. X2) > x) is P(X1 > x) + P(X2 > x) . In contrast.'s. in the subexponential case the only way X1 + X2 can get large is by one of the Xi becoming large. we have {max(Xi. 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. B(x) Here B*2 is the convolution square. (1. Since B is concentrated on (0.3 Consider the standard exponential distribution. HEAVY TAILS B*2\ 2.1) then means P(X1 +X2 > x) 2P(Xi > x).1(b) is oo. if X1 + X2 is large .
Proposition 1.y)/B(x) > 1. Using the identity B*(n+1)(x) = 1+ + 1)(x) 1+ 2 1 .z B(x) .1.2) B(x) B(x ) B(x) Jo with n = 1 and splitting the integral into two corresponding to the intervals [0. B( 0 . SUBEXPONENTIAL DISTRIBUTIONS Here is the simplest example of subexponentiality: Proposition 1. We now turn to the mathematical theory of subexponential distributions.S)x.5)x)' + 0 _ 2 L(x)l xa (16) Letting S 10.yo]. Hence lim sup a+oo B*2(x) 2B((1 . yo] as X + 00. then the overshoot X .'s: if X .B*(n ) B(dz) (1. then B(B(x)y) * 1 uniformly in y E [0.B*n(x . [In terms of r. and combining with Proposition u 1. If X1 + X2 > x. we therefore get lim sup B*2(x)/B(x) > 1+B(y)+ 1 .6)x)/((1 .B(y) = 2.B(y)) .xIX > x converges in distribution tooo. a contradiction. The uniformity now follows from what has been shown for y = yo and the obvious inequality y E [0. 1 < B(x ) B( x) Y) < B( 0).] Proof Consider first a fixed y. This follows since the probability of the overshoot to exceed y is B (x + y)/B(x ) which has limit 1.B E S. then either one of the Xi exceeds (1 . If lim sup B(x . x].S)x + B(Sx)2 < lim sup B(x) xaoo B(x) lim sup 2L((1 x^oo . y] and (y.1(b) we get B*2(x)/B(x) * 2. 253 Proof Assume B(x) = L(x)/xa with L slowly varying and a > 0.v. we get limsupB*2(x)/B(x) < 2. Finally lim inf B(x .y)/B(x) > 1 since y > 0. or they both exceed Sx. we get BZ(x)) > 1 + B(y) + B(B()y) (B(x) .5 If B E S. Let 0 < 5 < 1/2.4 Any B with a regularly varying tail is subexponential.
Given e > 0.z ) (x ) = 1 + (^ B(x . Proposition 1. Then by (1.B*2 (x) B(x) (x .6 If B E 8.z) B(dz). B(x) \Jo _ B(x .z) B(dz) 2B(x) o rv 2 0 2 using Proposition 1.y) sup v>o B(v) B(x) which converges to 0 by Proposition 1.5 that B(n) > e6B(n . so assume the proposition has been shown for n.B(x . Proof For 0 < 5 < e. then e"R(x) * oo. 0 Proof of Proposition 1.5 and dominated convergence.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 .2). Proof We use induction. and this immediately yields the desired conclusions. its intuitive content is the same as discussed in the case n = 2 above. b[c] = oo for all e > 0. HEAVY TAILS Corollary 1..2. we have by Proposition 1.z) B(dz) _y B(x) 111 Lx B .1) for all large n so that B(n) > cle6n for all n.z) B(x) Here the second integral can be bounded by B*n(y) B(x) . This implies B(x) > c2e5x for all x.nI < e for x > y.z) B(dz) (n + O(e)) ^x JO B(x) (n + 0(0) I B (x) . The case n = 2 is just the definition.5 and the induction hypothesis. then for any n B*n(x)/B(x) * n. x oo. B*(n+1) (x I xy + Jxx y) W. The first integral is y B(x .7 If B E S. O The following result is extremely important and is often taken as definition of the class S. choose y such that IB*n(x)/B(x) .
ajB(x)Ai(v) = ajB( x)(1+o„(1)) (j = 3 .z) B(x) < 1 + A + an sup f x B(x . Combining these estimates and letting a 4.z) B(x . Proof Define 5 > 0 by (1+5)2 = 1+e.v. an+1 fX B*n( *n(x . an = supx>o B*n(x)/B(x). 0 Proposition 1.y))/B(x). choose T such that (B(x)B*2(x))/B(x) < 1 + b for x > T and let A = 1/B(T). e > 0. SUBEXPONENTIAL DISTRIBUTIONS 255 Here the first term in {•} converges to 1 (by the definition of B E S) and the second to 0 since it is bounded by (B(x) . then there exists a constant K = KE such that B*n(x) < K(1 + e)nB(x) for all n and x.X2 > xv) < A1(xv)A2(x v) .8 If B E S.B(x .'s such that Xi has distribution Ai.5 easily yields P(X1 + X2 > x.3) Using the necessity part in the case Al = A2 = B yields f xv B(x . Since P(X1+X2 > x.i).y)Ai(dy) v) f o .1.z) B(dz ) + sup < 1 + sup f x<T B ( x) x>T 0 B(x . Then Al * A2(x) = P(X1 + X2 > x). Then Al * A2 (x) . oo) such that Ai (x) _ aiB(x) for some B E S and some constants al. it follows that it is necessary and sufficient for the assertion to be true that JX_VA (x . A2 be distributions on (0.0 completes the proof.(al + a2)B(x). Then by (1. 0 Lemma 1. X2 be independent r.4) . 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. Proof Let X1.y)Ai(dy) = (x)o(1) (1.2). a2 with a1 + a2 > 0.z) B(dz) x . For any fixed v.9 Let A1. Proposition 1.z) B(dz) < 1 + A + an(1 + d) .y)B(dy) = B(x)ov (1)• v (1.ala2B(x)2 which can be neglected. Xi <= v Ai (x .X1 > xv.
y)Ai(dy) = B(x)o„(1). a1 = a2 = a yields A*2(x) . if q(x) aB(x) for some B E S and some constant a > 0.5) becomes x B(x . of (1.11 Let B E S and let A be any distribution with a ligther tail. the l. B2 E S. i = 1.v)B(v) + _'U Aq(x . That is. A2 = B so that a1 = 0. u Corollary 1.10 The class S is closed under tailequivalence.3) follows if CHAPTER LX. B1 * B2 E S does not hold in full generality (but once B1 * B2 E S has been shown. B1 * B2 (x) .256 Now (1. whereas the two first yield B(x)(Ai(v) . However.B(x) Proof Take Al = A. u Corollary 1.Bl (x) + B2 (x) follows precisely as in the proof of Proposition 1. with a > 0 and L1. In the regularly varying case.12 Assume that Bi(x) = Li(x)lxa. . L2 slowly varying.aiB(v)) = B(x)o„(1). then so is L = L1 + L2.4).5) Here approximately the last term is B(x)o„(1) by ( 1. That is.13 Let B have density b and failure rate A(x) such that . a2 = 1.s. Proof Taking Al = A2 = A.Bl (x) + B2 (x) when B1. Recall that the failure rate A(x) of a distribution B with density b is A(x) = b(x)/B(x) Proposition 1.h. A(x) = o(B(x)).2. it is easy to see that if L1. it should hold that B1 * B2 E S and B1 * B2 (x) . Then L = L1 + L2 is slowly varying and B1 * B2(x) sim L(x)/x«. Then A * B E S and A * B(x) .9). Hence Corollary 1.2A(x). HEAVY TAILS 'VV B(x . then A E S. u It is tempting to conjecture that S is closed under convolution.Ai(x .2aB(x) . L2 are slowly varying. V (1. f " By a change of variables. We next give a classical sufficient (and close to necessary) condition for subexponentiality due to Pitman [290].y)B(dy).(x) is decreasing for x > x0 with limit 0 at oo.v)Ai(v) . Then B E S provided fo "O exA(x) b(x) dx < oo.
1.y ) b(y)dy = B (x) o ox _ J = ox/2 eA( x)A(xy ).y) < yA(x . Since ) (x . Define A(x) = fo .U) /or) v 2x This yields easily that ex.A(xy)A ( y). we can use the same domination for the second integral but now the integrand has limit 0 . 0 A(x) . In the regularly varying case. the first integral has limit 1 .(y) dy. and exa(x)b(x) = (3x01e(10)x9 is integrable.(x . L(x) y° (a . Thus A(x) is everywhere decreasing.A(y)\(y) dy + fox/ 2 eA(x ).1 has limit 1 + 0. replace B by a tail equivalent distribution with a failure rate which is everywhere decreasing). Thus B*2(x )/ B(x) . a(x) = ax01.16 For L(x) slowly varying and a > 1.14 Consider the DFR Weibull case B(x) = ex0 with 0 <.12. Then B(x) = eA(x). the u lognormal distribution is subexponential. The middle bound shows that it converges to b(y) for any fixed y since \ (x .e009xv)2/2a2/(x 2irv2) logx ( ) 't ((logx .13 works in this setting. we first quote Karamata's theorem (Bingham. Then b(x) = Ox0lexp.. Thus by dominated convergence .A(y)a(y ) = ev'(y) b(y). Jo For y < x/2. SUBEXPONENTIAL DISTRIBUTIONS 257 Proof We may assume that A(x) is everywhere decreasing (otherwise.3 < 1. f ' L(y) dy . By (1. Goldie & Teugels [66]): Proposition 1. the DFR Weibull distriu bution is subexponential. x . Further. Thus.2)..`(x)b(x) is integrable.A(x .y) dy. B*2(x) .y) y\(y)• The rightmost bound shows that the integrand in the first integral is bounded by ey"(v).1)xcl1 .y) < A (y) for y < x/2. subexponentiality has alrady been proved in Corollary 1. Thus.15 In the lognormal distribution. To illustrate how Proposition 1. an integrable function by assumption. Example 1. proving B E S. Example 1.1 B(x) eA( x)A(xv )A(y)A(y) dy f B(x . elementary but tedious calculations (which we omit) show that A(x) is ultimately decreasing.y) * 0.
L(x)/x" and )t(x) .1)])a 1 1 . yo] .y(x)B(x).1)] I X > x) L(x[1 + y/(a .13 may present a problem in some cases so that the direct proof in Proposition 1.258 From this we get CHAPTER IX. then 7(x) x/(a . we get (1.1)]) xa L(x) (x[1 + y/(a . (c) Under the assumptions of either ( a) or (b).)/Y(x) > y) (1 + y/(a . Proof ( a): Using Karamata's theorem.4 is necessary in full generality. HEAVY TAILS Proposition 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) and P(X (.1/A(x) and P(X ixil'Y (x) > y) * e'.ea b(x) is integrable. Then: Proposition 1.1))^ ' (b) Assume that for any yo )t(x + y/A(x)) 1 A(x) uniformly for y E (0. Thus exa(x)b(x) . then B(x) .xjX > x. More precisely. .6) EX(x) . the monotonicity condition in Proposition 1. However.E(X . (1 + y/(a . f O B(y) dy .17 If B has a density of the form b(x) = aL(x)/x°+1 with L(x) slowly varying and a > 1.1)X(x)/x > y) = P(X > x[1 + y/(a . We conclude with a property of subexponential distributions which is often extremely important: under some mild smoothness assumptions. let X W = X . 'y(x) = EXix>.18 (a) If B has a density of the form b(x) = aL(x)/xa with L(x) slowly varying and a > 1.x)+ _ 1 °° P(X > x) P(X>x )J L PX >y)dy 1 x L(y)/ydy L(x)/((a1)x'1) x )l ° J °° ( ()l a x a1 Further P ((a . Then 7(x) .1))a .a/x.
2 Let Y1..A(x + y/A(x))} =a(x) a(x + x) dx = ex p ex P . We assume p = /3µB < 1 and are interested in the ruin probability V)(u) = P(M > u) = P(r(u) < oo). 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.7) is referred to as 1/A(x) being selfneglecting. P The proof is based upon the following lemma (stated slightly more generally than needed at present). St > u}. .nn.1 If Bo E S. 2 The compound Poisson model Consider the compound Poisson model with arrival intensity /3 and claim size distribution B.and lognormal distributions . Kliippelberg & Mikosch [134].. r(u) = inf it > 0. i.(x). u a oo. cf.. d.n0 1•P(K= n)•n = EK. Examples 1. be i. Notes and references A good general reference for subexponential distribution is Embrechts. Bo(x) = f0 B(y) dy / µB.2. nG(u). Then P(Y1 + • • • + YK > u) . Recall that B0 denotes the stationary excess distribution.. 1. with common distribution G E S and let K be an independent integervalued r.14.15.v.t be the claim surplus at time t and M = sups>0 St. 0 G(u) L G(u) . then Vi(u) P Bo(u).A(x) I X > x) = exp {A(x) .EK G(u).1/.f yl 0 0 = exp {y (1 + 0(1))} 0 fY A( x + u /A( x)) a(x) du } The property (1. THE COMPOUND POISSON MODEL 259 We omit the proof of (c) and that EX (x) . Y2. Let St = Ei ` Ui . with EzK < oo for some z > 1.+YK> u) = ^•P(K = n)G* n(u ) . We get p(yl+.8) in (b) then follows from P (A(x)X (x) > y) = F(X > x + y/. Lemma 2. . Theorem 2 . The remaining statement (1. It is trivially verified to hold for the Weibull.
The PollaczeckKhinchine formula states that (in the setup of Lemma 2. _ B(x^sx Bo(x) µ8 I aoB(y )dy = (^) . P(K = k) = (1.2) M = Yl + • • • +YK where the Yt have distribution Bo and K is geometric with parameter p. The approximation in Theorem 2. see Abate.. See also Embrechts & Veraverbeeke [136]. Bo is more heavytailed than B . Since EK = p/(1. and for the lognormal and Weibull cases it can be verified using Pitman 's criterion (Proposition 1. u x+a Notes and references Theorem 2. Bo ¢ S.x^ ) B(x) _ f or ( lox . we have fx B(y)dy = a B0 (x) > lim inf lim inf x+oo B(x) . the result follows immediately from Lemma 2. then Bo(x)/B(x) + 00. Bo E S. (2.p) and EzK < oo whenever pz < 1.1)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 ) .x400 PBB(x) PB Leta+oo. In general: Proposition 2. Bo E S is immediate in the regularly varying case. a]. Proof of Theorem 2. mathematically one must note that there exist (quite intricate) examples where B E S.1.1 is essentially due to von Bahr [56]. Note that in these examples .18. Borovkov [73] and Pakes [280].. .p)p'.2.1) In particular . x 4 00.1 is notoriously not very accurate. as well as examples where B ¢ S. Weibull) one has Bo(x ( B(x) .µ J ) . Proof Since B(x + y)/B(x) * 1 uniformly in y E [0. r(1/Q) xlQexp B(x) = ex' From this . u The condition Bo E S is for all practical purposes equivalent to B E S.µB(01 . The tail of Bo is easily expressed in terms of the tail of B and the function y(x) in Proposition 1. HEAVY TAILS u using dominated convergence with >2 P(K = n) Dz" as majorant. lognormal .260 CHAPTER IX. However. For some numerical studies.3 If B E S.?(xµ 8 (x). in our three main examples (regular variation . The problem is a very slow rate of convergence as u ^ oo.13).
let t9+ = i9(0) be the first ascending ladder epoch of {Snd> }. t9(u) = inf {n : Snd> > u} . i=1 .Ti. . Define further 0 = IIG+II = P(r9+ < oo).} Then ik(u) = F ( M > u) = P(i9 (u) < oo). E. in [219] p. Based upon ideas of Hogan [200]. M = sup s$ . also a second order term is introduced but unfortunately it does not present a great improvement.. Asmussen & Binswanger [27] suggested an approximation which is substantially better than Theorem 2.y)/B (x) > 1 uniformly on compact y internals. The main result is: Theorem 3 .1.9+ < oo) = P(S. Let U= be the ith claim . We assume positive safety loading.1 when u is small or moderately large.1 gives 1010.. Snd) = Xl +. 3 The renewal model We consider the renewal model with claim size distribution B and interarrival distribution A as in Chapter V. Kalashnikov [219] and Asmussen & Binswanger [27].1) this end . one may have to go out to values of 1/'(u) which are unrealistically small before the fit is reasonable. (3. This shows that even the approximation is asymptotically correct in the tail. T1 the ith interarrival time and Xi = U. G+ (A) = P(Sq+ E A. Thus G+ is the ascending ladder height distribution (which is defective because of PB < PA).y + as usual denotes the first ascending ladder epoch of the continuous time claim surplus process {St}. To Bo(u) u + 00.. {n= 0. there is a representation of M similar to the PollaczeckKhinchine formula. [279].3. 1 Assume that (a) the stationary excess distribution Bo of B is subexponential and that (b) B itself satisfies B(x .] The proof is based upon the observation that also in the renewal setting.. T+ < oo) where r+ = T1 + • • • + T.. In [1]. Then l/i(u) 1 P P [Note that (b) in particular holds if B E S... THE RENEWAL MODEL 261 Choudhury & Whitt [1]. 195 there are numerical examples where tp(u) is of order 105 but Theorem 2.+ E A.g.e. i. p = iB /µA < 1. Then K M=EY. + Xn. Somewhat related work is in Omey & Willekens [278].
1 IPG_ I / F(x .B(x)..(. B(x) _ J O° B(B(x)y) A(dy) f 1 . FI (x) _ fz ° F(y) dy.1) holds for a general random walk satisfying the analogues of (a). (b) and does not rely on the structure Xi = Ui . the contribution from the interval (.N. A. Lemma 3 ..3 G+ (x) .i. 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. As for the compound Poisson model. G_(A) = P(S. Write G+( x) = G+ ( x.FI(u).d)) E A) denote the pre19+ occupation measure and let and U_ = Eo G'_" be the renewal measure corresponding to G_. Let further 19_ _ inf {n > 0: S^d^ < 0} be the first descending ladder epoch. Proof By dominated convergence and (b). (3..2 F(x) . HEAVY TAILS where K is geometric with parameter 9. Proof Let R+(A) = E E'+ ' I(S. 0] normalized by IPG_ I so that we should have to G+(x) . cf. u a 00. and hence FI(x) . Lemma 3 .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_. we will use the fact that the proof of (3.y) dy = 1 Pi (X) oo IPG_ I . x > 0. Then 0 0 F( x .PBBo(x).y_ E A) the descending ladder height distribution (IIG II = 1 because of PB < P A) and let PG_ be the mean of G_.Y2.y+ given r+ < oo). U_ (dy) is close to Lebesgue measure on (. P(K = k) = (1 .d. Let F denote the distribution of the Xi and F1 the integrated tail. A(dy) = 1.262 CHAPTER IX. are independent of K and i.Ti).9)9'' and Y1. The heuristics is now that because of (b). oo) = F(S.oo. x * oo.g+ > x.2). 0] to the integral is O(F(x)) = o(FI(x)).3) and we will prove it in this form (in the next Section.1) is equivalent to P(M > u) " . with distribution G+/9 (the distribution of S. x + oo.FI(x) /IPG_I. 0 The lemma implies that (3. whereas for large y . d+ < oo).
Hence using dominated convergence precisely as for the compound Poisson model.1.2). u Proof of Theorem 3.=1 BIp G_ I (1.1 I .1.1)/F(n) < 1 + e for n > N (this is possible by (b) and Lemma 3. n] F1 ( n=N _1 1+e E F(x+n) 0 + limsup xr00 FI(x) FAG.3. we can assume that the span is 1 and then the same conclusion holds since then U(n .O[s])(1 . choose N such that F(n . the proof is complete. 0] x+00 FI(x) 00 + lim up 1 x) E F(x + n) U_ (n .1.9)IpG_ I Differentiating the WienerHopf factorization identity (A. Given e.3.F[s] = (1 . (3. We then get lim sup G+(x) xro0 Fj(x) < lim sup X)00 o F(x . n] < (1 + e)/1µc_ I for n > N.y) U. then by Blackwell 's renewal theorem U_ (n . n] is just the probability of a renewal at n. n] + 1/I µG_ I.y) U_ (dy) 00 FI (x) < lim sup F(x) U(N. By Lemma 3. and that U_(n . Similarly.UG_ I x. > (1 .2) yields 00 F F I (u) P(M > u) _ E(1 . In the lattice case.0)0k k I(u) A.e) z lim inf G+(x)  FI (x) Ip G_ I Letting a 10. If G_ is nonlattice.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 0.9) 1 .1.(dy) fN FI ( x) + lim sup ZY00 N F(x .G+[s]) .I n=N (1 E)2 r00 F(x + y) dy + e) lim sup . and in the last that FI is asymptotically proportional to Bo E S.1. THE RENEWAL MODEL 263 We now make this precise. F(Y= > x) FI(x)/(OIp _ 1).
Se(u)_1 < a) = o(Fj(u)).So(u)) are available. Mn < u}.1)6+[0] . with roots in von Bahr [56] and Pakes [280]. on the set {M > u. Note that substantially sharper statements than Lemma 3. see Asmussen & Kliippelberg [36].. P(M > u. allowing also for possible dependence between the arrival process and the claim sizes.1 is due to Embrechts & Veraverbeke [136].a) N P(M > u).(u)+n . u)). Therefore by Lemma 3.l.Sty(u)_I < a} we have w(u) < oo.2. Proof Let w(u) = inf {n : Sid) E (u .a. HEAVY TAILS µF = (1 .(1 . Notes and references Theorem 3.264 and letting s = 0 yields CHAPTER IX. ..0)ua_ . 10(0) But since P(M > u .a.a. 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. FJ(u) UBBO(U) PBo(u) N = (10)Ipc_I JUA .yiui_1. S+q(u) . Then P(M E (u . 4 Models with dependent input We now generalize one step further and consider risk processes with dependent interclaim times.a.AB iP We conclude by a lemma needed in the next section: Lemma 3 . S+9(u) . must attain a maximum > 0 so that P(M > u.SS(u)}n=o. we have P(M E (u . Sty(u) . and {Su.IIG+II)µc_ = (1 ..4 on the joint distribution of (S. u)) > P(w(u) < oo)(i lp (0))• On the other hand.4 For any a < oo.So( u)_1 < a) < P (w(u) < oo)j/i(0) < 0(0) P(M E (u .u)) = o(P (M > u)) = o(FI(u)).
. Theorem 4.4 below. M..d..1 based upon a regenerative assumption. M* = max S.1..1..i. t>0 S. We give here one of them. examples and counterexamples. and apply it to the Markovmodulated model of Chapter VI. see [47].. (corresponding to the filled circles on Fig. 4. Schmidli & Schmidt [47]. The zerodelayed case corresponds to Xo = Xl = 0 and we write then F0.1 except for the first one) is a random walk. = Sx. and the distribution of {Sxk+t .... We let F* denote the Podistribution of Si.. (viewed as random elements of the space of Dfunctions with finite lifelengths) are i..F*(X) = P0(Si > x) .Sxi}0<t<x2Xl . We return to this point in Example 4. < 0 and EoX < oo where X = X2 .X1 is the generic cycle.Sxk}o<t<xk+1xk is the same for all k = 1.. assume pp. See Fig. For further approaches.. 0o(u) etc. 4.1) . 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 <. . such that {SXo+t  SXo}0<t< X 1Xo . M = sup St.. The idea is now to observe that in the zerodelayed case.1) is assumed..1 = max k=0. Figure 4.1 Note that no specific sample path structure of {St} (like in Fig.. Thus the assumption ..1 where the filled circles symbolize a regeneration in the path.. {Sn}n=o..X2 < . E0.. +1.n n=0. 2.4. {SX1+t . MODELS WITH DEPENDENT INPUT 265 Various criteria for this to be true were recently given by Asmussen. Define S. G(x) (4.. 4.
3) hold. The one we focus on is Fo (Mix) > x) .2 Theorem 4.* i o<t<xn+1x..266 CHAPTER IX. Since clearly M(x) > Sl . it suffices by (4.2) Imposing suitable conditions on the behaviour of {St} within a cycle will then ensure that M and M* are sufficiently close to be tail equivalent. N N Xi=0 N Figure 4. 4. jF11 F* (U).1) and (4. Sxn +t .4) liminf u>oo F(M > u) .Sxn = sup Sxn+t .2) to show F(M* > u) > 1.. the assumption means that Mix) and Sl are not too far away. (4. See Fig.3) where Mnx) = sup o<t<xn +1 X. Then '00 (u) = Fo(M > u) .S.. (4.3) applicable so that F(M* > u) 141 F*(u). (4. u p 00.2. HEAVY TAILS for some G such that both G E S and Go E S makes (3.1 Assume that (4. Fo(Si > X).. Proof Since M > M*.
5) which follows since Po (M > u. Let a > 0 be fixed.e)Po (MMX> > x).4).: Sn > u} .1 = limti00 St/t.+Mn+1>u} 267 (note that {M> u} = {3(u) < oo}).Po (M* > u. We shall use the estimate Po(M > u) Miu^+ 1 < a) = o(Po (M > u)) (4. Under suitable conditions . MW O(u)+1 < a) IN ( U n=1 A1. Letting first u + oo and next e . Theorem 4.: S.( u)1 > a) 00 1: Po(Mn<u. choose a such that Po(Si > x ) > (1 .a. assume the path structure Nt St = EUit+Zt i=1 . 2..Sn 0<t<x„+j ( 1 .Sn+1Sn>aV(uSn*)) n=1 00 > (1E)EPo(Mn<u.. To this end.4.. /3(u) = inf{n=1. Given e > 0..1 can be rewritten as 00 (U) (4. M^xu)+l > a) . x > a. MODELS WITH DEPENDENT INPUT Define 79* (u) = inf {n = 1 .(1 .2. Po(M* > u) . Mn+l > a V (u .Mn +1 >aV(u n=1 00 S.e)Po (M > u. 0 yields (4.. u)} < P(M* E (u .. S. Then by Lemma 3.E) Po ( n max St u.4.. . E (u . u))/P(M* = 0) = o(Po(M* > u)).6) 1 p pBo(u) u where B is the Palm distribution of claims and p .e)Po (M > U).a.(u) .S. )) > (1 .
Corollary 4.4). (ii) EozNX < oo for some z > 1. the proof of Lemma 4.6 below.1 is in force. and the rest is just rewriting of constants: since p = 1+tlim St = 1+ .I u J Po(Sl > x) dx 1 EoNxB(x) dx EoX(1 . HEAVY TAILS N` U.7). oX (see Proposition A1.8) x Write .6) holds with p = . independent of {> CHAPTER IX. (iii) For some o field Y. The same is true for Sl. and also for Mix) since Nx FNX U. i=1 (4. Mix) < > UE + i=1 o<t<x Thus Theorem 4. Assume further that (i) both B and Bo are subexponential. since the tail of Zx is lighter than B(x) by (iv).Q = EoNx/EoX. Proof It is easily seen that the r.268 with {Zt} continuous. X and N. are Fmeasurable and NX Po J:U=>x i=1 (iv) Po sup Zt > x / (0:5t<x o(B(x)) Then (4.v. cf. Then the Palm distribution of claims is B(x) = E N Eo 0 I( U1 < x) .X both have tails of sup Zt.} and satisfying Zt/t N.p) Ju P Bo(u) 1p 0 .2 Assume that {St} is regenerative and satisfies (4.'s order EoNx • B(x). a4' 0. and ENX Ui .3PB. we get 00 (u) 1 IPF.
MODELS WITH DEPENDENT INPUT 269 Example 4 . < 1.3 that (4.9). Bo E S.t} is standard compound Poisson and {Zt} an independent Brownian motion with mean zero and variance constant a2. Again . and taking F = o. and for some constants ci < oo such that cl + • • • + c.0 (thus (iv) is trivial).6) u holds. Then (4... X3 = 2. Taking again Xo = Xi = 0.6 with arrival rate /3(t) at time t (periodic with period 1) and claims with distribution B (independent of the time at which they arrive). > 0.t + EN'I Ui where {>N`1 Ui . (iii) is obvious..4 Assume that St = Zt .6) holds. X2 = 1. consider the periodic model of VI. We consider the case where one or more of the claim size distributions Bi are heavytailed. 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 . in particular lighttailed. i. The key step of the proof is the following lemma. More precisely. . Theorem 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. . Zt . In particular.4. Assume that B E S. Example 4 . X2 = 1.e. we assume that B E S. 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.5 Consider the Markovmodulated risk model with claim size distributions satisfying (4.. Bo E S. Thus we conclude that (4.(NX).. of claims arriving in [0. The regenerative assumption is satisfied if we take Xo = Xi = 0. 3 The average arrival rate / and the Palm distribution B of the claim sizes are given by P P Q = ir i/i.6) holds. i=1 B = >2 7riaiBi i=1 and we assume p = 014 B = Ep ri/3ipB. . The arrival rate is /3i and the claim size distribution Bi when Jt = i. X3 = 2. . then (iv) holds since the distribution of supo<t<i Z(t) is the same as that of I Zl 1.3 As a first quick application. we conclude just as in Example 4. 1) is Poisson with rate /3 = fo /3(s) ds so that (ii) holds. (i) holds.
F) < CG(x)zn'1+. we can define the regenerations points as the times of returns to i..2. u Proof of Theorem 4.Fmeasurable. If Jo = i..v. Assume EzN1+"'+Np < oo for some z > 1 and all i. Then P P(Yx > x) .270 CHAPTER IX. 2 . and the rest of the argument is then just as the proof of Corollary 4... and F a aalgebra such that (N1.ciG(x).X i=1 j=1 where conditionally upon F the Xi. . For lighttailed distributions. . 6 Let (N1. It follows by a slight extension of results from Section 1 that P P(Yo > x I Y) G( x) ci Ni. Thus dominated convergence yields ( P(Yo>x P(Yo>x . are independent with distribution Fi for Xij. X > 0 a r. NP ) be a random vector in {0. HEAVY TAILS Lemma 4 .F) = P(Yo > X+x I •^) G (x +x)>2ciNi i=1 . i =1 P(Yo > x I ^ ) < CG(x)zN1+ +Np for some C = C(z) < oo. as x a oo. 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 ."+Np .. NP ) and X are .G( x ) > ciNi . Let {Fi}t=1 P be a family of distributions on [0.}P.^•) G(x) P ^ E ciNi = C.c'(x) where c = ciENi . . 1.. and that for some + cp distribution G on [0.5. i=1 Proof Consider first the case X = 0. i=1 P(Yx > x ^) < P(Y0 > x I. . An easy conditioning argument then yields the result when Jo is u random. cp with cl + > 0 it holds that Fi(x) . .. oo) and define p Ni Yx = EEX'i .. P P P(YX and > x I. i1 = E\ G(x) In the general case. The same dominated convergence argument completes the proof. . .
5 shows that basically only the tail dominant claim size distributions (those with c.p)Bo(u).d. in particular Proposition 2. Then O(u) . The main result of this section. For further studies of perturbations like in Corollary 4. IV. As usual. VI. I T(u) < oo). r(u) is the time of ruin and as in IV.4... It follows from Theorem 4. An improvement was given in Asmussen & Hojgaard [33].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. as well as a condition for (4. we let PN"N = P(. and the final reduction by Jelenkovic & Lazar [213]. 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. and independent of (T1. see Schlegel [316]. Combined with the approximation for O(u).5. The present approach via Theorem 4. 5 Finitehorizon ruin probabilities We consider the compound Poisson model with p = /3pB < 1 and the stationary excess distribution Bo subexponential. ). the discussion provides an alternative point of view to some results in Chapter IV. Within the class of risk processes in a Markovian environment.T2. Theorem 5. Schmidli & Schmidt [47].. Theorem 4. FINITEHORIZON RUIN PROBABILITIES 271 probabilities for large u.7.1. Theorem 2. this should be compared with the normal limit for the lighttailed case. m is a (orfinite) .6) to hold in a situation where the interclaim times (T1.T2. this then easily yields approximations for the finite horizon ruin probabilities (Corollary 5.. 5 was first proved by Asmussen. cf. this is applied for example to risk processes with Poisson cluster arrivals. ) form a general stationary sequence and the U.pl(1 .2 and Example 4. for lighttailed distributions the value of the adjustment coefficient y is given by a delicate interaction between all B..4. In contrast. Floe Henriksen & Kliippelberg [31] by a lengthy argument which did not provide the constant in front of Bo(u) in final form. cf. i. Notes and references Theorem 4.e.7).i..4. We start by reviewing some general facts which are fundamental for the analysis. That paper also contains further criteria for regenerative input (in particular also a treatment of the delayed case which we have omitted here). cf. > 0) matter for determining the order of magnitude of the ruin probabilities in the heavytailed case.4. Essentially.1 is from Asmussen. states that under mild additional conditions. i.3. 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).
and starting from So = y. to consider only the case Px(w(F`) = 0) 0. k on E. the whole of R and not as usual impose the restrictions x > 0. resp. {Rt}. Rt is distributed as x + t .00). w(Fc) < oo ) 'In general Markov process theory.s. and (5. k as indicator functions. Thus. 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} . a main difficulty is to make sense to such excursions also when Px(w(F°) = 0) = 1. Proof Starting from Ro = x. j.t. The equality of the l. {St} and {Rt} are in classical duality w. HEAVY TAILS measure on E such that L for all measurable A C E and all t > 0. The simplest example is a discrete time discrete state space chain. For the present purposes it suffices .r. (note that we allow x.= y.t. t. follows by the substitution y = x . to the r.). say. . Let G denote the distribution of ENt U.h.z) dx G(dz) = ffh(y + z) k(y)dy G(dz). m. We let QS be the corresponding distribution and Qx.272 CHAPTER IX.rij = mjsji where r13. St is distributed as y .2) with t = 1 means m.)k(x .2) means ffh(a.2) for all bounded measurable functions h.s=j are the transition probabilities for {St}. x = 0+ and F = (0. oo). Then there is a Markov process {Rt} on E such that fE m(dx)h(x)Exk(Rt) = Lm(dy)k(y)Eyh(St) (5. for states i.h.1 A compound Poisson risk process {Rt} and its associated claim surplus process {St} are in classical duality w . where we can take h. . but the example of relevance for us is the following: Proposition 5. r. Then (5.t + EI U. Say {St} is reflected Brownian motion on [0. y = 0).y = Qx (. Sw(F.s. however .>N` Ui. a familiar case is time reversion (here m is the stationary distribution). Lebesgue measure. u For F C E. y to vary in. in the terminology of general Markov process theory.z.
in E F.1 The sample path in (a) is the excursion of {St} conditioned to start in x = 0 and to end in y > 0..y = Qy Q. z > 0. That is. Qx y is the distribution of an excursion of {St} conditioned to start in x E F and terminate in y E F. i1..1 for the case F = (oo. In particular: Corollary 5. this simply means the distribution of the path of {Rt} starting from y and stopped when 0 is hit.. 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.= y) Theorem 5 . 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). . and we let Qy y refer to the time reversed excursion .. Thus.2 Qy. QR and QRy are defined similarly. oo) = r(0) x= St y (a) Figure 5. io = x. .(0)_ = y < 0 is the same as the distribution of w(y) where w(z) = inf It > 0 : Rt = z}..).SS(F.s. Sw(F)1 = y) . /^s x (S1 = Z1. Sn = in = y.y as a measure on all strings of the form i0i1 . Qx. 5. when p = .itt) = P Px(w(Fc) < 00. Sn+1 E Fc) nx. .3 The distribution of r(0) given r(0) < oo.y() = P ({SW(F`)t} 0<t<w(F °) E So = x. [note that w(z) < oo a.2.13AB < 1] Proof of Theorem 5. 0]. w(0.5. The theorem is illustrated in Fig .y(2p21 . in with i0. in = y. the one in (b) is the time reversed path. S. .. But in the risk theory example (corresponding to which the sample paths are drawn).. FINITEHORIZON RUIN PROBABILITIES 273 y E F (in discrete time. We can then view Qy. x = 0. Sw(Fo)_ should be interpreted as Sw(F^)_1)..
Sn+1 E Fc) n=1 i1.. in E F... S. in = x.(F<)1 = Y) S S and Qx y( ipil . .... ....gilt' k=1 ii ...rin_1in E Txj jEFC m21 s2120 m2252221 m in Ssn n1 mjSjx Mx m2p mil min1 jEF` 1 Sinin _ 1 . Si l io E mjSjx. .. HEAVY TAILS E E Px (Si = 21i .. Silt' E Sxik_1 . 2p). in with 20.. .. = in = y... S.. (Fc)1 = y) 00 CHAPTER IX..i„_iEF Similarly... 2p) when 20.....J (i. in)  Pt' (R1 = ii. .. Rn = in = x. MY Thus Qx(ioii ... Si1y k=1 i1 .. R Qy x(2p21 .. R .in E F.ii ...TI( 2n2n _1 .274 note that Fx(w(Fc) < 00. .in1 . . 2n) = Qx..y(inin _ E SYj jEF` 00 Sxik _1 . Rn = in = x.... in = x. 20 = y.... Si11 S 1 . ...ik1EF Similarly but easier Sxin_1 .. To show Q y x (i0 i 1 . note first that Pt' (R l = il. in) = oo jEF^ Sxin1 . Rn+1 E F`) F (w(Fc) < 00. 21 . i0 = y.. t' y and Qy x are measures on all strings of the form ipi l .....ik_1EF . Rn+1 E FC) TioilTili2....... Si1y 00 jEF° E E 5xik_ 1 .....ik_1EF Sxin_1 .. Silt' E SO k=1 i1. .. i0) Q x. in) = Qx.
v. P(") = P(. Let Y = Yl = Sr+( 1) be the value of the claim surplus process just after the first ladder epoch . Z follows the excess distribution B(Y) given by B(Y) (x) _ B(y + x)/B(y). Y > y} .p. that is.')distribution of Yu is Bo"). We are interested in the conditional distribution of T(u) = T(0) given {T(0) < oo.t. S.UBBo(u)].2.'s are defined w. The formulation relevant for the present purposes states that Y has distribution Bo and that conditionally upon Y = y. the distribution w. Now the P(u. Bo") is also the P(u.B(a) +a PBBo(u) . FINITEHORIZON RUIN PROBABILITIES 275 5b The time to ruin Our approach to the study of the asymptotic distribution of the ruin time is to decompose the path of { St} in ladder segments . Z) is described in Theorem 111. U T(O) = T (u) Y Figure 5.(o) > y} = {T(0) < oo.')distribution of Z since P(Z>aIY>u) = 1 °° B(y) B(y + a) dy FLBBo(u) B (y) J°° (z) dy . the P(u. That is. To clarify the ideas we first consider the case where ruin occurs already in the first ladder segment . Z = Zl = ST+( 1)_ the value just before the first ladder epoch (these r.r. the case r (O) < oo. Y > u).r.5.t. 1 w .2.2. P(o) ). y > u. ST(o) > y.')density of Y is B(y)/[. 5.2 The distribution of (Y. see Fig. 7(0) < oo. that is.
the random vectors (YI. HEAVY TAILS Let {w(z)}Z^. 4 Assume that Bo E S and that (5.i. r(u)/Z 4 1/(1 .: r+ (n) < oo. let r+(1) = T(0).3 implies that the P("'1)distribution of T(u) = r(0) is that of w(Z).. That is .e. the duration T+ (n) .. this in principle determines the asymptotic behaviour of r(u).1) of the last ladder segment can be estimated by the same approach as we used above when n = 1. We let K(u) = inf In = 1.. Then 7(u)/y(u) ^ W/(1 .o be defined by w(z) = inf It > 0: Rt = z} where {Rt} is is independent of {St}. Y1 + • • • + Yn > u} denote the number of ladder steps leading to ruin and P("'n) = P(• I r(u) < oo. Since w(z)/z a$. Recall the definition of the auxiliary function y(x) in Section 1.3) where the distribution of W is Pareto with mean one in case ( a) and exponential with mean one in case (b).p) then yields the final result T(u)/y(u) + W/(1 ..p) in Pi"'')distribution. Zn).p). Z). it therefore follows that T(u)/Z converges in Pi"'')probability to 1/(1 . i. conditionally upon r+ (n) < oo. Yn_1 'typical'. Zk be defined similarly as Y = Y1.e. In the proof. . cf. 5. K(u) = n). (Y.. Then Corollary 5. Since the conditional distribution of Z is known (viz. a slight rewriting may be more appealing. and distributed as (Y. denote the ladder epochs and let Yk. 2. However.T+(2).3. Fig. Z/'y(u) * W in Pi "' ')distribution ..p).1. . then by the subexponential property Yn must be large. It is straightforward that under the conditions of Proposition 1....276 CHAPTER IX.. more precisely. We now turn to the general case and will see that this conclusion also is true in P(")distribution: Theorem 5 . P(Z < a I Y > u) 3 0.. > u with high probability.. are i.. must be large and Z1.18(c) Bo")(yY (u)) + P(W > y) ( 5.r+ (n .3) holds. . i. Zn_1 'typical' which implies that the first n1 ladder segment must be short and the last long. Z = ZI but relative to the kth ladder segment. . . we get the same asymptotics as when n = 1..p) in F(u) distribution.d. 1/(1 . and since its dominates the first n . and YI. The idea is now to observe that if K(u) = n. z ^ oo. Now Bo E S implies that the Bo ")(a) + 0 for any fixed a. Z1). Hence Z. in particular of Z. .. Bo") ). Then. .
.n). Y„1..Yl+ +Yf1>u}. . > u}.n) (y1. I A'(u)) = P(u. A"(u) are events such that P(A'(u) AA"(u)) = o(F (A'(u)) (A = symmetrical difference of events). A"(u) _ {K(u)=n} = {Y1+ P(. Yn . .. then IIP( I A'(u)) Taking A'(u) = {Y.u) E •) .5 Ilp(u.Bo (ri1) ®B( . I A"(u ))II + 0.u) II 0. the condition on A'(u) A A"(u) follows from Bo being subexponential (Proposition 1. II ' II denotes the total variation norm between probability measures and ® product measure. P(. suitably adapted). P (Yj.. Proof We shall use the easily proved fact that if A'(u).u) E • I A'(u)) = Bo (n1) ®Bou) . +Yn1<u.. FINITEHORIZON RUIN PROBABILITIES 277 16 Z3 Z1 r+(1) T+(1) T+(1) Figure 5..3 In the following. Lemma 5. .5. . Further.2..Yn1iYn .
. The first step is to observe that K(u) has a proper limit distribution w..1. y > u.. Then according to Section 5a. Proof of Theorem 5.r.6 IIPIu'n ) CHAPTER IX.4). the F'distribution of r(u) is the same as the P'distribution of w1(Zl) + • • • + wn(Zn). the density of Yn is B(y)/[IBBO(u)]. Z. ..t. wk(Zk) has a proper limit distribution as u + oo for k < n.. Similarly (replace u by 0).. .. 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 •) ..u has distribution Bout That is. .. and clearly Zi. +wn(Z n))l7( u ) > 1y) ^' P(u'n)(wn (Zn)/7(u) > y) 4 NW/(1 . Zn) E •) .6. . then 11P(Z E •) . .. the discussion just before the statement of Theorem 5. The same calculation as given above when n = 1 shows then that the marginal distribution of Zn is Bou). and that Yk has marginal distribution B0 for k = 1..7 O (u. .). It therefore suffices to show that the P(u'")distribution of T(u) has the asserted limit. n_1 < u.. see Fitzsimmons [144])... P(u) since by Theorem 2.. the marginal distribution of Zk is Bo for k < n. Y1 +. . .. + Y" > u) Flul (K (u ) = n) _ Cu) P"F(1'i +.. Proof Let (Y11. Y'. n.... in particular his Proposition (2..i. whereas wn(Zn) has the same limit behaviour as when n = 1 (cf.. . n .d... k = 1. (Y. Zn are independent.P) Bo(u) for n = 1. Notes and references Excursion theory for general Markov processes is a fairly abstract and advanced topic.P(Y' E •)II * 0.Bo (n1) ®Bo' 0. By Lemma 5. Thus F(u'n)(T(u) /7(u) > y) = F(u'n)((wl (Z1) + .1 P PBo(u) • P(W/(1 .1 and Y„ .p) < y).4. {wn(z)} be i.. ..278 Lemma 5 . copies of {w(z)}.' = y is BM. For Theorem 5. 2. ..P) > y) Corollary 5.P(Z' E •)II > 0 (here Y.1). Zn). Let {wl(z)}.y(u)T) . Z' are arbitrary random vectors. Z11)..+y 1 p"F(Yn > u) P)Pn1 P/(1 .. in our example Y = (Y1. . Y") u etc. be independent random vectors such that the conditional distribution of Zk given Y.2. HEAVY TAILS ((Z1'.
B(u). = supo<t<0. 3.6. We will show that the stationary density f (x) of {Vt} satisfies f (x) /B(x) r(x) We then get V. RESERVEDEPENDENT PREMIUMS 279 The results of Section 5b are from Asmussen & Kluppelberg [36] who also treated the renewal model and gave a sharp total variation limit result .1. The heuristic motivation is the usual in the heavytailed area. the probability that is exceeds u is then B(u . x > oo. Then P(MT > u) . i. V. that fo p(x)1 dx < oo. 6 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate /3. p(Y) and the result follows. the results only cover the regularly varying case. T) when T + oo with u fixed. Asmussen & Teugels [53] studied approximations of i (u. Assume for simplicity that {Vt} regenerates in state 0 . Corollary II. Proof of Theorem 6. max VB>0I Vo=0^ o<s<t J11JJJ Lemma 6 . The rigorous proof is. that MQ becomes large as consequence of one big jump. however.e. . Then 0 (u) Qf "O ^) dy. and premium rate p(x) at level x of the reserve. Theorem 6 .y) .. cf.(3 u u J B(y) dy .1 Assume that B is subexponential and that p(x) > 00. 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). one expects the level y form which the big jump occurs to be 0(1). and define the cycle as a = inf{t>0: Vt=0.(u) = P(V > u) = f f (y) dy . Extensions to the Markovmodulated model of Chapter VI are in Asmussen & Hojgaard [33].2 Define M./3Ea B(u). nontrivial and we refer to Asmussen [22].1) The key step in the proof is the following lemma on the cycle maximum of the associated storage process {Vt}. u (6.2. claim size distribution B. More precisely.
It is also shown in that paper that typically.P(MT > u) $B(u) Ft µ(1 . . Then D(u) = f(u)p(u) and.280 CHAPTER IX. Further the conditional distribution of the number of downcrossings of u during a cycle given Mo > u is geometric with parameter q(u) = P(Mo > u I Vo = u). 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. u Notes and references The results are from Asmussen [22].q ( u)) 1 . Hence f (u)r(u) = D(u) = Do(u) .q(u) Now just use that p(x) * oo implies q (x) + 0. there exist constants c(u) 4 0 such that the limiting distribution of r(u)/c(u) given r(u) < oo is exponential. D(u) = DQ(u)/µ. where also the (easier) case of p(x) having a finite limit is treated . by regenerative process theory.
Ripley [304]. vrN(z .. replicates Zl.2) is an asymptotic 95% confidence interval .z) 4 N(0. and this is the form in which the result of the simulation experiment is commonly reported.i..Chapter X Simulation methodology 1 Generalities This section gives a summary of some basic issues in simulation and Monte Carlo methods . 281 . z) 2 = Zit NE ii ii According to standard central limit theory . ZN. Hence 1.96s z f (1. where a2 = Var(Z ). We shall be brief concerning general aspects and refer to standard textbooks like Bratley. 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. .d. The crude Monte Carlo ( CMC) method then amounts to simulating i. topics of direct relevance for the study of ruin probabilities are treated in more depth. Fox & Schrage [77]. Rubinstein [310] or Rubinstein & Melamed [311] for more detail . estimating z by the empirical mean (Z1 + • • + ZN)/N and the variance of Z by the empirical variance N s2 = E(Z{  N 2. 4Z)..
generated at the same time as Z. variance reduction is hardly worthwhile.b(u. We mention in particular ( regression adjusted) control variates and common random numbers. and a longer CPU time to produce one replication. Say that Var(Z') = Var(Z)/2. This is a classical area of the simulation literature. The situation is more intricate for the infinite horizon ruin probability 0(u). SIMULATION METHODOLOGY In the setting of ruin probabilities. it is straightforward to use the CMC method to simulate the finite horizon ruin probability z = i. Sections 24 deal with alternative representations of Vi(u) allowing to overcome this difficulty. Further. Z = I inf Rt < 0 (0<t<T = I('r(u) < T). writing Var(Z) = Var(E [Z I Y]) + E(Var[Z I Y]) . conditional Monte Carlo and importance sampling. However. Therefore. The difficulty in the naive choice Z = I(T(u) < oo) is that Z can not be simulated in finite time: no finite segment of {St} can tell whether ruin will ultimately occur or not. 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. we then have EZ = EZ = z. so that Z' is a candidate for a Monte Carlo estimator of z. 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'. an added programming effort. typically by modifying Z to an alternative estimator Z' with EZ' = EZ = z and (hopefully) Var(Z') < Var(Z).282 CHAPTER X. there are others which are widely used in other areas and potentially useful also for ruin probabilities. Conditional Monte Carlo Let Z be a CMC estimator and Y some other r . v. 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. Letting Z' = E[Z I Y]. and in most cases this modest increase of N is totally unproblematic. one can argue that unless Var(Z') is considerable smaller than Var(Z). Typically variance reduction involves both some theoretical idea (in some cases also a mathematical calculation). We survey two methods which are used below to study ruin probabilities.
Nevertheless. even if the optimal change of measure is not practical. . Thus. it may often be impossible to describe P in such a way that it is straightforward to simulate from P).1. it gives a guidance: choose P such that dP/dP is as proportional to Z as possible.zrs) = 2 1 N 2 2 2 i=1 i=1 N > Lt Zi . (1. the argument cheats because we are simulating since z is not avaliable analytically. Thus we cannot compute L = Z/z (further. . using the CMC method one generates (Z1. However. This may also be difficult to assess . Variance reduction may or may not be obtained: it depends on the choice of the alternative measure P.. i..e.[E(LZ)] = E Z2 Zz . Then z Var(LZ) = E(LZ)2 . L such that z = EZ = E[LZ]. In order to achieve (1. LN) from P and uses the estimator N zrs = N > L:Zj i=1 and the confidence interval zrs f 1.3) Thus.zrs. and the problem is to make an efficient choice.z2 = 0. To this end. L1). the obvious possibility is to take F and P mutually equivalent and L = dP/dP as the likelihood ratio. L = z/Z (the event {Z = 0} is not a concern because P(Z = 0) = 0).v. one would try to choose P to make large values of Z more likely. (ZN. .3).96 sis v^ N 2 1 where srs = N j(LiZi . but tentatively. GENERALITIES 283 and ignoring the last term shows that Var(Z') < Var(Z) so that conditional Monte Carlo always leads to variance reduction.E [Z Z]2 = z2 . it appears that we have produced an estimator with variance zero. 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. a crucial observation is that there is an optimal choice of P: define P by dP/dP = Z/EZ = Z/z.
A = {T(u) < T} or A = {r(u) < oo} and the rare events assumption amount to u being large. we may try to make P look as much like P(•IA) as possible. 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.1. say 10%. Two established efficiency criteria in rare events simulation are bounded relative error and logarithmic efficiency. say of the order 103 or less.100 .e.0. assume that the A(u) are rare in the sense that z(u) * 0. the optimal P is the conditional distribution given A. To introduce these. i.96 2Z ( 1 .96 2 z2 z increases like z1 as z . An example where this works out nicely is given in Section 3. as is the case of typical interest. it does not help telling whether z is of the magnitude 104.5 or even much smaller .284 CHAPTER X. Thus.z) which tends to zero as z ^ 0. but if the point estimate z is of the order 105.z) 1001.. assume that the rare event A = A(u) depends on a parameter u (say A = {r(u) < oo}). Another way to illustrate the problem is in terms of the sample size N needed to acquire a given relative precision . SIMULATION METHODOLOGY 1c Rare events simulation The problem is to estimate z = P(A) when z is small . However. let z(u) = P(A(u)).1. the issue is not so much that the precision is good as that relative precision is bad: oZ z(1 . and further it is usually not practicable to simulate from P(•IA). if z is small. z I.z) 1 > 00. The CMC method leads to a variance of oZ = z(1 . We then . in terms of the halfwidth of the confidence interval. However.. The optimal change of measure ( as discussed above) is given by P(B) = E [ Z] i. a confidence interval of width 10 4 may look small.e. I. Z z V5 In other words .B = iP(AB) = P(BIA).e. In ruin probability theory. u + oo. Z = I(A) and A is a rare event. large sample sizes are required. Again.96oz /(zV) = 0. 10 . For each u. We shall focuse on importance sampling as a potential (though not the only) way to overcome this problem. and let Z(u) be a Monte Carlo estimator of z(u). N . This leads to the equation 1.
XK from the density bo(x). O (u) = z = EZ.0. with common density bo(x) = B(x)/µB and K is geometric with parameter p. Generate X1. Generate K as geometric.d. We shall here present an algorithm developed by Asmussen & Binswanger [ 271. are i. Therefore .e. the PollaczeckKhinchine formula III. This allows Var(Z(u)) to decrease slightly slower than z(u)2.p)pk.4) for any e > 0. 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. According to the above discussion. where M = X1 + • • • + XK.log z(u) of (1. it is appealing to combine with some variance reduction method . so that NE (u) may go to infinity. let Z +. which gives a logarithmically efficient estimator . the mathematical definition puts certain restrictions on this growth rate. see Asmussen & Rubinstein [45] and Heidelberger [190]. it is not efficient for large u . X2. Notes and references For surveys on rare events simulation. where X1.4). where Z = I(M > u) may be generated as follows: 1. and in practice. P(K = k) = (1 . Otherwise. If M > u. F(K = k) = (1 . let Z +.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. . 3. 2 Simulation via the PollaczeckKhinchine formula For the compound Poisson model.. Let M . logarithmic efficiency is almost as good as bounded relative error. 2. However. this means that the sample size N = NE(u) required to obtain a given fixed relative precision (say a =10%) remains bounded. i.2..1) may be written as V) (u) = P(M > u).. . The algorithm gives a solution to the infinite horizon problem .i.log Var(Z(u)) lim inf > 2 u+oo .p)pk. The term logarithmic comes from the equivalent form .. . Thus.X1 + + XK. Var(Z(u)) hm sup U+00 z (u) 2E < oo (1. but as a CMC method .(2.
.. As a conditional Monte Carlo estimator . note first that To check the formula for the P(X(n) > x I X(1).XK_1 and let Z( 1)(u) = Bo (Y) (if K = 0. The idea of [27] is to avoid this problem by discarding the largest X. y < 0). A first obvious idea is to use conditional Monte Carlo: write i.X(n_1)) Bo(X(„_l) V X) Bo(X(n1)) .. Z(1) (u) has a smaller variance than Zl (x). XK...1) V)(u) .XK1] = EBo(uX1 .2.. and considering only the remaining ones.. .S( K_1)) V X(K1)) / Bo(X(K 1)) where S(K_l) = X(1) + X(2) + • • • + X(K_1)..p/(l . compute Y = u .. To see this.. form the order statistics X(1) < X(2) < . . just note that EZ(1)(u ) 2 > E[Bo (x .... . 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. we generate only X1. Thus. Z(1)(u) is defined as 0). 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.Xl . .X(2). Xl > x. < X(K) throw away the largest one X(K).X(K1)) . and let Z(2)(u) = _ P (SK B0((u > u I X(l).XK_1).SK1)2. XK1. However.. Then (cf. and that Bo(y) = 1. conditional probability...X(2).. we thus generate K and X1i . assume in the following that Bo(x) ... X1 + + XK_ 1 > x when X1 > x.p)Bo(x). K > 2] = P2p(Xl > x) = P2Bo(x) (here we used that by positivity of the X. asymptotically it presents no improvement : the variance is of the same order of magnitude F(x).. So..286 CHAPTER X.. SIMULATION METHODOLOGY when the claim size distribution B (and hence Bo) has a regularly varying tail...+XK > uIXl.... For the simulation. Theorem IX.L(x)/x`' with a > 0 and L(x) slowly varying..b(u) = P (Xl +•••+XK>u) = EF[Xl + .X1 .
BL(dx) = e7sB(dx)/B[y].. B. . using 13L. the algorithm for generating Z = Z(u) is: 1.S (n1)) V X (. .. that is. Binswanger and HOjgaard of [28] give a general survey of rare events simulation for heavy tailed distributions ..1 is elementary but lengty.. X(2). and we refer to [27]. However .y. in the renewal or Markov. 3 Importance sampling via Lundberg conjugation We consider again the compound Poisson model and assume the conditions of Ce7".1) .5). Notes and references The proof of Theorem 2. The algorithm is sofar the only efficient one which has been developed for the heavytailed case. 111. BL instead of 0. 1 Assume that Bo (x) = L(x)/x° with L(x) slowly varying. BL by I3L = /3B[y]. and define )3L.3. and that paper contains one more logarithmically efficient algorithm for the compound Poisson model using the Pollaczeck. . Asmussen .. Thus.. . 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. X .u is the representation 0(u) = e7sr(u) overshoot (cf. and simulate from FL. For practical purposes. X (. the continuoustime process {St} is simulated by considering it at the discrete epochs {Qk} corresponding to claim arrivals.S(n_1) I X(1). l)) . . use the the CramerLundberg approximation so that z(u) = '(u) = e7"ELe7E(") where ^(u) = ST(") . . for the purpose of recording Z(u) = erysr(u). IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 287 We then get P(S" > x I X( 1). X(2).modulated model P(r+ < oo) and G+ are not explicit ). Then the algorithm given by { Z (2) (u) } is logarithmically efficient. X(n1)) P(X(TZ) + S(. X(n1)) Bo((x .Khinchine formula and importance sampling . Compute y > 0 as solution of the Lundberg equation 0 = K(y) = )3(B[y] . X(2)._1) > P(X(n) > _ X X(1). 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. l)) BO(X(n1)) Theorem 2 .
to deal with the infinite horizon problem . Let Sf0 CHAPTER X. = X1 + . Proof Just note that EZ(u)2 < e . . If S > u. BL).r(u) < oo) = 1.QL. b different from . b) # (/3L.Q. The answer is no. with distribution F.. It resolves the infinite horizon problem since FL(. r(u) < oo) and FL (both measures restricted to.i. The proof is given below as a corollary to Theorem 3. SIMULATION METHODOLOGY 3. More precisely. one must restrict attention to the case 4µB > 1. 4. the discussion at the end of Section 1b. Let FL (dx) = 'For the renewal model. X2. so that changing the measure to FL is close to the optimal scheme for importance sampling .T. Otherwise.3. return to 3. let S. In detail . In fact: Theorem 3. We formulate this in a slightly more general random walk setting '.l3 and U from B. Let X1.1) (simulated with parameters ^3.. P'[y] < oo for some ry > 0. the results of IV. A > AL as in Chapter V. . + X. and we have: Theorem 3. The estimator is then M(u) /3eQT' dB Z(u) (Ui) j=1 )3 e $Ti dB where M(u) is the number of claims leading to ruin. Let S .. We may expect a small variance since we have used our knowledge of the form of 0(u) to isolate what is really unknown.S+U .7 tell that P(. There are various intuitive reasons that this should be a good algorithm. M(u) = inf {n : S„ > u}. BL could improve the variance of the estimator .2ryu _ z (u)2/C2. B) is not logarithmically efficient when (/3. be i..d. Ti. namely ELery£(").288 2.(u)) are asymptotically coincide on {r(u)} < oo. Generate T as being exponential with parameter . and avoid simulating the known part e7".1 The estimator Z(u) = e'rs* "u) (simulated from FL) has bounded relative error. cf. and assume that µF < 0 and that F[y] = 1.2 The estimator (3. let Z F e_'s. and the change of measure F r FL corresponds to B > BL.F. Xi = U... u It is tempting to ask whether choosing importance sampling parameters . The algorithm generalizes easily to the renewal model ..
it thus follows that for 0 < e < e'/ELXi. Proof The first statement is proved exactly as Theorem 3 .. are i. The importance sampling estimator is then Z( u) = e'rSM( ). More generally. = c'. Here ELK. 1.2ryELXi.yu+elu u +oo etry' 1 > lim up C2e2..+KM(u)}. F(XM(u)). For the second. where Kl og (X) (j) 2 ) = log dFL (Xi) . EFZ(u)2 EFZ(u)2 lim sup z(u)2eeU = lim cop C2e2.2'X1 ..d.yu = G. .3 The estimator (3.2) dF Theorem 3. + KM(u))} = exp {ELM(u)(E . write W(F IF) _ F(XI).2 > 0.. Since K1..P = FL. let F be an importance sampling distribution equivalent to F and M(u) dF Z(u) _ I (Xi) . By the chain rule for RadonNikodym derivatives. IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 289 e7yF(dx).. (3.. K2..2ryELXi)} .. When F # FL.2) (simulated with distribution F of the X3 has bounded relative error when . Since ELM(u)/u + 1/ELXi. Jensen's inequality and Wald's identity yield EpZ(u)2 > exp {EL(K1 + .i. where e' = EL Iog dFL (Xi) > 0 by the information inequality. EFZ(u)2 = EeW2(FIF) = Ep [W2(FIFL)W2(FLIF)] = EL [W2 ( FIFL)w(FLIF)] = ELexp {Kl+.3. is not logarithmically efficient. .
Consider compound Poisson risk process with intensities /3'. The queueing literature on related algorithms is extensive .4 indicate that we can expect a major difference according to whether y < 1/r.T' = U" .3"eQ x 0 J eQ zB (z) dz x (x > 0) and /3' = /3". then /3' B' = B". Further discussion is in Lehtonen & Nyrhinen [245]. T) with T < oo. The easy case is y > 1/k'(y) where O(u.3. optimality is discussed in a heavy traffic limit y 10 rather than when u + oo. then the estimator Z(u) = e7Sr(°)I(r(u) < yu) (simulated with parameters /3L. the references in Asmussen & Rubinstein [45] and Heidelberger [190].T" has a left exponential tail with rate /3'. 4 Importance sampling for the finite horizon case The problem is to produce efficient simulation estimators for '0 (u.2. U' . we conclude by differentiation that Bo(x)=B' (x)forallx > 0.1 is from Lehtonen & Nyrhinen [244].1 If y > 1/ic'('y). SIMULATION METHODOLOGY u Proof of Theorem 3. B" and generic claim sizes U'.T' D U" . from 3' P(U'T'>x) ^ = ^ eQ'zB (z) dz. all that needs to be shown is that if U' .i.T' has a left exponential tail with rate /3' and U" . The optimality result Theorem 3. U". This immediately yields / = 3". /3'eQ'YR'( x + y) dy = .3'eO'x f f P (U" . U' . As in IV.B'=B". generic interarrival times T' . The extension to the Markovian environment model is straightforward and was suggested in Asmussen [ 16].T".4. /3".T".'(y) or y > 1/r.T" > x) J /3"e0 yB (x + y) dy = . yu) is close to zk(u). CHAPTER X. claim size distributions B'. BL) has bounded relative error. In [13]. 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 . Next. The results of IV.e. we write T = yu. First by the memoryless distribution of the exponential distribution .290 which completes the proof. Then according to Theorem 3. so that one would expect the change of measure F 4 FL to produce close to optimal results. see e. T". In fact: Proposition 4.g. .'(y). with the present (shorter and more elementary) proof taken from Asmussen & Rubinstein [45].
3 Theorem IV.2 The estimator (4.log 4')u) 4 u (Theorem IV.log Var(Z(u)) _ .8 has a stronger conclusion than (4. We next consider the case y < 1/r.1. 7y (4.2) Since the definition of ay is equivalent to Eay r(u) . and that ryy > ry.'(7).1). Let Qy2 = .to g x ( u ) u u so that (1. The corresponding estimator is Z(u) = eavS' ( u)+T(u)K (ay)I(T( u) < yu). and in fact.4. Remark 4 . Bay) is logarithmically efficient. yu)/z. yu) is of order of magnitude a71.yu. We recall that ay is defined as the solution of a'(a) = 1/y.1) which is all that is needed here can be showed much easier . yu) in the sense that .5) follows.4.yk(ay) determines the order of magnitude of z'(u. .1). T( u) < yu] e2ryyuEay le 2ay^(u). Bounding u ELZ(u)2 above by a7u. (4. yu) = eayu Eay Leay^(u)+r(u)K(ay). lim inf uoo 27yu . the result follows as in the proof of Theorem 3.(u) * 1 (Theorem IV. (4. IMPORTANCE SAMPLING FOR THE FINITE HORIZON CASE 291 Proof The assumption y > 1/n'(y) ensures that 1fi(u. that ryy = ay .log Var(Z(u)) l im of .4.8). T(u) < yu] e Hence by (4. Further .(ay). T(u) < yu] . one would expect that the change of measure F Pay is in some sense optimal.yy> 2 . we have ic(ay ) > 0 and get Eay Z(u)2 = Eay [e  2aySr( u)+2r(u )r.3) (simulated with parameters /gay.O(u.4.1) so that z(u) = zP(u. Proof Since ryy > y.3) and we have: Theorem 4.
1) (see Proposition IV. related discussion is given in a heavy traffic limit q J. the algorithm in Section 3 produces simulation estimates for the tail P(W > u) of the GI/G/1 waiting time W).292 CHAPTER X. the object of interest is {Vt} rather than {Rt}. yu .1) is used to study Voo by simulating {Rt} (for example. Hence lira inf log ryyu + vyu 1/2 tc(ay) . Z (u)2 above.yu)/(uyu1/2) .b(u.4). 5 Regenerative simulation Our starting point is the duality representations in 11. N(0.a yu +l/ur' (av)Ei`av reav^(u)+(T(++)(U) yu . and (5. there exists a dual process { V t} such that i. In Asmussen [13].(av)Eav l e. 0 Notes and references The results of the present section are new. yu .2) .3: for many risk processes {Rt }.1): then by Proposition A1. 0 rather than when u 3 oo. '%(u) = P I info Rt < 0) = P(VV > u).yu1/2 <1 T(u) < yu l r > e7vu +avul/ 2r. > u) = E f I(VV > u) dt 0 (5.a vt(u).Qyu1/2 < T(u) C yu e.ryyu +oy u1/2K'(av)Eo l v 1/2) where the last step follows by Stam's lemma (Proposition IV.u1/2 < r(u) < yu Le ] l = e.o.7y x(u) > hm inf u+Oo U . (5. However.4. SIMULATION METHODOLOGY Vara„ (r(u))/u so that (T(u) .1) where the identity for Vi(u) requires that Vt has a limit in distribution V.1) may be useful. Then z(u) = Eay Z(u) > Eay avS'(u)+T( u)k(av 1 ). In most of the simulation literature (say in queueing applications).4.2).T) = P O<t<T inf Rt < 0 = P(VT > u)..3.o . One main example is {Vt} being regenerative (see A. we believe that there are examples also in risk theory where (5. zi(u) = INV.uaoo U That lim sup < follows similarly but easier as when estimating En.
Z2 = N (X21' + .h (zl.. Then (Z1z1i Z2z2 ) 4 N2(0.+Z(N) z 1. Vh = (8h/8z1 8h/ 8z2). record Zi'i = (Z1'). provides estimates for F ( V. letting J0 'o I (Vt > u) dt . Z2'> the time during the cycle where { Vt} exceeds u and zj = EZJ'). i (^(u) . and Z2'>) where Zi'i = w. Z2 a4* z2. EZ2'i = z2 = E Thus. which we survey below ... REGENERATIVE SIMULATION 293 where w is the generic cycle for {Vt}.. To derive confidence intervals . Thus the method provides one answer on to how to avoid to simulate { Rt} for an infinitely long time period. i.. For the ith cycle. oh) for h : R2 ^ R and Ch = VhEVh.. The method of regenerative simulation.....t(u)) 4 N(0. Zl the LLN yields Z1 a$' Z(1) +. > u) (and more general expectations Eg(V. Simulate a zerodelayed version of {V t } until a large number N of cycles have been completed. + Z1N>) . For details .E). a standard transformation technique (sometimes called the delta method) yields 1 V 2 (h (Zi. consider first the case of independent cycles . the regenerative estimator z%(u) is consistent.d. Taking h(zl. .5. (u) ?2 = E fo I(Vt > u) dt = 0( u ) zl Ew as N > oo.. Z2 . z2) z2/z1 yields Vh = (z2/z2 1/zl). is the cycle length. + Z2N)) . EZ1'i = z1 = Ew. j = 1. Thus. 2. z2)) > N(O. Z1 = (Zl1i +.3) . . Therefore . let E denote the 2 x 2 covariance matrix of Z('). Then Z(1). 02) (5. Z(N) are i . )).
Notes and references The literature on regenerative simulation is extensive.5) Z1 Z1 Z1 and the 95% confidence interval is z1 (u) ± 1. There is potential also for combining with some variance reduction method. The regenerative method is not likely to be efficient for large u but rather a brute force one.z^ i=1 so a2 can be estimated by 2 2 = 72 S11+ 12 S22 . say risk processes with a complicated structure of the point process of claim arrivals and heavy tailed claims . the expectation z = EZ of a single r. Z of the form Z = ^p(X) where X is a r . consider an extremely simple example . in some situations it may be the only one resolving the infinite horizon problem .294 where 01 2 CHAPTER X. asymptotic estimates were derived using the renewal equation for z /i(u).2. 6 Sensitivity analysis We return to the problem of 111 . Rubinstein [310] and Rubinstein & Melamed [311]. v. Then z(() = f cp(x) f (x. In 111. Before going into the complications of ruin probabilities . We here consider simulation algorithms which have the potential of applying to substantially more complex situations.0 . Let X have a density f (x. () dx so that differentiation yields zS d( fco(x)f(x.2 E1 2 z1 z1 Z2 The natural estimator for E is the empirical covariance matrix N S = N 1 12 (ZW . Here are the ideas of the two main appfoaches in today 's simulation literature: The score function ( SF) method . However . to evaluate the sensitivity z/i( (u ) = (d/d() 0(u) where ( is some parameter governing the risk process .g S12 (5. SIMULATION METHODOLOGY 2 Eli = Z2 z1 + 2 E22 .C)dx = f w(x) d( f ( x. see e. () dx f Ax) (dl d()f (x' () f ( z.g. 9. () depending on C.96s/v"N.Z) ^Z(=) .v.9. with distribution depending on a parameter (. () dx = E[SZ] f(X.
Thus. however . Infinitesimal perturbation analysis (IPA) uses sample path derivatives. ()) d( hc(U. The following example demonstrates how the SF method handles this situation.t. () is an unbiased Monte Carlo estimator of zS. IPA will estimate zS by 0 which is obviously not correct. C) f(X. ()) h((U. Thus . ()). Then . with density f (x. I(r(u) .log U/(. giving h( (U. one.t. /3 is 0. say W(x) = I(x > xo) and assume that h(U. In the setting of ruin probabilities . () = log U/(2. Let M(u) be the number of claims up to the time r(u) of ruin (thus. 11 /3oeOoT. So assume that a r. one can take h (U. Then z(() = Ecp(h(U.1 Consider the sensitivity tka(u) w. () _ (eSx. cp(h(U. To see this. For the SF method. () can be generated as h(U. () Thus.v. cp' (h(U. ( where h( (u.r. C). () = d log f (X. SENSITIVITY ANALYSIS where 295 S = (d/d()f (X.r. ()) is 0 for C < Co and 1 for C > Co so that the sample path derivative cp'(h(U. A related difficulty occurs in situations involving the Poisson number Nt of claims: also here the sample path derivative w. zc = E [d co(h(U. () is increasing in C. () where U is uniform(0. just take cp as an indicator function . The likelihood ratio up to r(u) for two Poisson processes with rates /3. The derivations of these two estimators is heuristic in that both use an interchange of expectation and differentiation that needs to be justified. () = . ()) is 0 w .1). C)). for some Co = (o(U). For example . /3o is M(u) Oe (3T: < oo) .() d( is the score function familiar from statistics . this is usually unproblematic and involves some application of dominated convergence . = E [`d (h(U. the Poisson rate /3 in the compound Poisson model. () = (8/8()h (u. For IPA there are. p. this phenomenon is particularly unpleasant since indicators occur widely in the CMC estimators . Example 6 . SZ is an unbiased Monte Carlo estimator of z(. nonpathological examples where sample path derivatives fail to produce estimators with the correct expectation. r(u) = Tl + • • • +TM(u)).6. if f (x.
9 . whereas for the SF method we refer to Rubinstein & Shapiro [312].T(u)) e7uerVu) for ?P. There have been much work on resolving the difficulties associated with IPA pointed out above. To resolve the infinite horizon problem .t. We then arrive at the estimator ZZ(u) = (M(u) . change the measure to FL as when simulating tp(u). different parameters. since ELZp(u)2 < (M(U) _T(u) \ 1 2 a2ryu = O(u2)e27u.3 (u) (to generate Zp (u). However. the risk process should be simulated with parameters . for different models and for the sensitivities w. Example 6.T(u)) I(T(u) < co) ] .296 CHAPTER X.t. the estimation of z(ip(u) is subject to the same problem concerning relative precision as in rare events simulation . a relevant reference is VazquezAbad [374].1 is from Asmussen & Rubinstein [46] who also work out a number of similar sensitivity estimators. j3 and letting flo = 0.0(1) so that in fact the estimator Zf(u) has bounded relative error. Thus. ) we have VarL(ZQ(u)) ZO(u)2 O(u2)e2 u2e2ryu yu . BL).3 (u) is of the order of magnitude ue7u. we get 1 M(u) 00(u) = E (_Ti)I(T(U)<) E [(M(u) . 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. differentiating w. 4) that V5.r. We recall (Proposition 111.r. SIMULATION METHODOLOGY Taking expectation.3L. In the setting of ruin probabilities. .
(u) is defined as the probability of being ruined (starting from u) before the reserve reaches level a > u. are i. .. and {1. either this makes no difference (P(R. a) = r(u)).i. with P(Xk = 1) = 9.d. X2. Besides its intrinsic interest . a) = r(u) A T+(a). Y'a(U) = P(T (u) = r+(a)) = 1 ..P(•r(u. That is.. The twobarrier ruin problem The twobarrier ruin probability 0..Chapter XI Miscellaneous topics 1 The ruin problem for Bernoulli random walk and Brownian motion..+• • •+X.. as e. wherel T(u) = inf {t > 0 : Rt < 0} .g.. 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 ). T+(a) = inf It > 0 : Rt > al. 297 . Consider first a Bernoulli random walk. Oa(U ) can also be a useful vehicle for computing t/i(u) by letting a * oo.1}valued . defined as Ro = u (with u E {0. }).1. where X1. in most cases . T(u....(u) = 0 ) = 0) or it is trivial to translate from one setup to the other. in the Bernoulli random walk example below.
C1_0\a. By optional stopping. The martingale is then {zuzXl+•••+X„ } = {zR° }.a) Y..(u) I\ e = 1 oa ' ()i a = u. We choose a = ry where ry is the Lundberg exponent.o» = z°P (RT ( u.. then 'Oa(u) _ au a We give two proofs .1) o If 0 = 1/ 2. i.1) = (19)4/'0(a3)+9ba(a1). the solution of F[. one elementary but difficult to generalize to other models.a(u))..2). 7/la(a .e.y] = 1. and in view of the discrete nature of a Bernoulli random walk we write z = e7. = (1 . The Lundberg equation becomes 1=F[ry]=(19)+9z.1 For a Bernoulli random walk with 0 0 1/2. = z°Va(u) + za(1  . and insertion shows that ( 1.. Wald's exponential martingale is defined as in 11.(1B)u oJ 0.4) by ea(u+Xl+.r(u. (1. In a general random walk setting . where a is any number such that Ee°X = F[a] <oo. MISCELLANEOUS TOPICS Proposition 1. and the other more advanced but applicable also in some other settings..o)'t/1a(a . Conditioning upon X1 yields immediately the recursion 'a(1) = 19+00a(2). u + 1. Proof 1.o)T/la (1) + 8z/'u(3).(4.. z and the solution is z = (1 .298 CHAPTER XI.2) Oa(a ..0)/0.1.a) = 0) + zap ( R. zu = EzRO = EzRT(u.1) is solution.+Xn) F[ a]n n=0. u Proof 2... tba(2) _ (1 .
(1. {Rt} is itself a martingale and just the same calculation as in the u proof of Proposition 1. pa( u) _ u Corollary 1. RANDOM WALK. then Vi(u) = 1. However.1).5) .zu)/(za .2) is trivial (z = 1). BROWNIAN MOTION. If p<0.} 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)). 1h (u) = a el u \1 If 9 < 1/ 2. i1(u) = e211 .ba(u) = e2µa . thenz1 (u)=1.. {R.u) = et(a2 /2 +aµ) the Lundberg equation is rye/2'yp = 0 with solution y = 2p. . } yields e7u = Ee7R° = e°Wa(u) + e7a(1 .3 Let {Rt} be Brownian motion starting from u and with drift p and unit variance . Corollary 1. TWO BARRIERS 299 and solving for 4/la(u) yields t/ia(u) = (za . and solving for 9/la(u) yields Z/)a(u) = (e 76 .u)/u. (1. Then for p 0 0. Proof Let a+ oo in (1.• a2µa e2µu . then Proof Since 'Oa (U)  au a Eea(R°.1) for p # 0. u Proposition 1. If 9 = 1/2. If p = 0.e7u)/(e7° .1.1). Applying optional stopping to the exponential martingale {e7R.2 For a Bernoulli random walk with 9 > 1/2.0a(u)).1 If p = 0.4 For a Brownian motion with drift u > 0.1 yields 't/la(u) = (a .
6 If the paths of {Rt} are upwards skipfree and 7//(a) < 1. the paths are upwards skipfree but not downwards.e7a Again . It may then be easier to first compute the onebarrier ruin probability O(u): Proposition 1.a) = a) = 5 y = P (R (u. 1 . however. Here the undershoot under 0 is exponential with rate 5. However. CHAPTER XI. this immediately yields (1. VIII. say.300 Proof Let a * oo in (1.7) . Ic 5ry 'pa(u) Using y = 6 . and hence e7u = Ee7Ro E [e7R(. valid if p < 1 (otherwise . and thus one encounters the problem of controlling the undershoot under level 0. 0.3. (u) _ O(u) .7/la(u)).a) = a on {r (u. 7/'(u) = 1). (1..a ) < 0) + e 7aF ( R (u. implying R(u.4).5a). For most standard risk processes . Here is one more case where this is feasible: Example 1.a) = a ) + e ' ° ( 1 . letting a * oo yields the standard expression pe7u for . 5). passing to even more general cases the method quickly becomes unfeasible (see.0 (u) (where u p =. MISCELLANEOUS TOPICS u The reason that the calculations work out so smoothly for Bernoulli random walks and Brownian motion is the skipfree nature of the paths.a) = r+ (a)} and similarly for the boundary 0.5 Consider the compound Poisson model with exponential claims (with rate. .+^a(u))^(a) If 7k(a) < 1.616).vi(a) Proof By the upwards skipfree property. a) I R(u a ) < 0] P (R(u .0(a) 0 < u < a. 7O(u) = 7/la(u) + (1 . we obtain 'Oa a7u .e7a (u) = 6 /0 .7).a) < 0) + e7°P (R(u.
Hence P(MT>u. in particular symmetric so that from time r(u) (where the level is level u) it is equally likely to go to levels < u and levels > u in time T .11) VIT ) Proof For p = 0. MT > u) = P (ST > u) + P (ST > u.8) Proof In terms of the claim surplus process { St} = {u .)_ _( u)µ2 /2. we have ili(u. TWO BARRIERS 301 Note thas this argument has already been used in VII.11 ) is the same as (1.10) Pµ (T(u) < T) !. T(u) E dT. (1.8 Let {Rt} be Brownian motion with drift .1.8 ).µ so that {St} is Brownian motion with drift µ .. = eµuTµ2/2Po (T( u) E dT) 2 eµuTµ2/2 u T3/2 ex p u 27r p 12 T .ST<u) = P(MT>u.ST>U).7 For Brownian motion with drift 0. MT > U) = P(ST > u) + P(ST > u) (1. + µ2T) } . Then the density and c.f. Here {St } is Brownian motion with drift 0 (starting from 0).4) I = . For µ # 0. the density dPµ / dP0 of St is eµsttµ2/2.9) = 2P(ST > u).1a for computing ruin probabilities for a twostep premium function.. 0(u.d. We now return to Bernoulli random walk and Brownian motion to consider finite horizon ruin probabilities. (1.. and (1 ..Rt}. (i). Corollary 1. P(MT > u) = P(ST > u) + P(ST < u. BROWNIAN MOTION. = 1 . 10) follows then by straightforward differentiation.. For the symmetric (drift 0) case these are easily computable by means of the reflection principle: Proposition 1.µ%T (1.r(u).2 . ( 1.T) P(MT > u) where MT = maxo<t<T St.µ T I + e2µ"4) ( . of r(u) are ( U2 Pµ (T(u ) E dT) = 2^T 3/2 exp µu . RANDOM WALK. T ) = P(T(u) < T ) = 241. and hence Pµ('r(u) E dT) = Eo [e µsr(.
is zero for all u > 0 but that nevertheless Rt ^4 0 (the problem leads into the complicated area of boundary classification of diffusions. S(oo) = f c s(y)dy. Vi(u. but we omit the details. the behaviour at the boundary 0 is more complicated and it may happen. whenever u. and in a similar spirit as in VII.10). oo) with drift µ(x) and variance a2 (x) at x.10) and that the value at 0 is 0.. We finally consider a general diffusion {Rt} on [0.9) goes through unchanged. close to x {Rt} behaves as Brownian motion with drift µ = u(x) and variance a2 = a2(x). Thus.13) The following results gives a complete solution of the ruin problem for the diffusion subject to the assumption that S(x)..10 Consider a diffusion process {Rt} on [0. 226). that 0(u).13) with 0 as lower limit of integration.h. 0 0 (1. The expression for F ( ST = v) is just a standard formula for the u binomial distribution. oo).11) then follows by checking that the derivative of the r.12) is the same as ( 1. Here {2T( (v}TT)/2) v=T. Breiman [78] or Karlin & Taylor [222] p. The same argument as used for Corollary 1. We assume that u(x) and a2 (x) are continuous with a2 (x) > 0 for x > 0. T are integervalued and nonnegative.302 CHAPTER XI.g.8 also applies to the case 9 54 1/2.. (1.T+2. MISCELLANEOUS TOPICS which is the same as (1.. e.T)dx.T (1. Let s(y) = ef0 ry(.g.9). see e.9 For Bernoulli random walk with 9 = 1/2. as defined in (1. u Small modifications also apply to Bernoulli random walks: Proposition 1. Theorem 1. S(x) = f x s(y)dy.s. and (1.T2.3 we can define the local adjustment coefficient y(x) as the one 2µ(x)/a2(x) for the locally approximating Brownian motion. such that the drift µ(x) and the variance a2(x) are continuous functions of x and that a2(x) > 0 . Proof The argument leading to ( 1. is (1. If this assumption fails.T) = P(ST = u) + 2P (ST > u).12) P(ST = v) = 0 otherwise. as defined above as the probability of actually hitting 0. is finite for all x > 0.
b = 0.b(u) be the probability that {Rt} hits b before a starting from u.0(u) = 1 for all u > 0. RANDOM WALK. . S(oo) < oo separately u completes the proof. TWO BARRIERS 303 for x > 0. For generalizations of Proposition 1. then 0 < 2l. E„ q(Rdt) = q(u)+Lq(u)dt. A good introduction to diffusions is in Karlin & Taylor [222].17) Hence L.14) fails. If b < u < a. (1 .b(u) = S(a) . Further references on twobarrier ruin problems include Dickson & Gray [116]. A classical reference for further aspects of Bernoulli random walks is Feller [142]. O. Lemma 1. where Lq(u) = 0'22u) q "(u) + p(u)q(u) is the differential operator associated with the diffusion. see in particular pp. elementary calculus shows that we can rewrite L as Lq(u) d 1a2 (u)s(u)d [ s (u) ? ] . 1'.e LVa. 0 Proof of Theorem 1. Assume further that S (x) as defined in (1.10. and we get Wo. we can ignore the possibility of ruin or hitting the upper barrier a before dt. i. 191195 for material related to Theorem 1.b = 0 implies that VQ b/s is constant.11 Let 0 < b < u < a and let t&0. The obvious boundary conditions '0a.b(a) = 0 then yield the result. Notes and references All material of the present section is standard. In view of (1.S(u)/S(a). 0 in (1.e. 15) i. Letting a T oo and considering the cases S(oo) = oo. Then YIa.b(b) = 1.b (Rdt) = Oa.10. If (1. if (1.6 to Markovmodulated models .ba.16) S(a) . Letting b J.b(u) + L.ba.16).b('u) = Eu .b(u)dt.S(b) Proof Recall that under mild conditions on q.16) yields 4b (u) = 1 .S(u) (1.b = a+/3S. Using s'/ s = 2p/a2. so that Y)n. [117]. then. Wa.1.13) is finite for all x > 0.(u) < 1 for all u > 0 and ^ S^ Conversely.b(Rdt). (1.14) S(oo) < 00. see Asmussen & Perry [42]. BROWNIAN MOTION.b('u) = Eu &0. the function S(x) is .
aRo ..1. 1 y < k (y).o•K(a) = Ee .3) < e 7yu. Remark 11. 2 Further applications of martingales Consider the compound Poisson model with adjustment coefficient ry and the following versions of Lundberg 's inequality (see Theorems 111. MISCELLANEOUS TOPICS referred to as the natural scale in the general theory of diffusions (in case of integrability problems at 0. but by duality. yielding eau = Ee. yu) '+/1(u) .aR.2. where C_ = B(x) _ B(x) sup 2no fy° e7(Y )B(dy)' f2e7(Y2)B(dy)' C+ i/i(u. defined by the density 1/va(u)s(u) showing up in (1.13)).3. much of the literature dels with the pure drift case. equivalently. Another basic quantity is the speed measure M .6.5): _ z/'(u) < e 7u. one works instead with a lower limit 5 > 0 of integration in (1.4. which is motivated from the study of modern ATM (asynchronous transfer mode ) technology in telecommunications.6) .4) I. information on ruin probabilities can be obtained . variance 0. The emphasis is often on stationary distributions .4. 111 .1) (2.t&(u. IV. See Asmussen [20] and Rogers [305] for some recent treatments and references to the vast literature. (2. (2. They all use the fact that ( tx(a) l ( eaRt = eau + aSttx(a) < e7yu.ytc (ay).5. Markovmodulated Brownian models .(7) .1 ) was given already in II. (2.(T(u)AT) r.17). and here are alternative martingale proofs of the rest . yu) where W (ay) = y. y > .9 ) and optional stopping applied to the stopping time r(u) A T.(. correponding to piecewise linear paths or . (2. is currently an extremely active area of research.304 CHAPTER XI.2) C_e7u < t(u) < C+e _7u. Lo I. 7y = ay . Lo is a martingale (cf. with the drift and the variance depending on an underlying Markov process .(a) (2.5) A martingale proof of (2.)AT .
dr) 1 = 1 I0 /o C+ C+ From this the upper inequality follows. RT(u)_) given r(u) < oo. (2.6) with = 'y that eyu . yu))• Letting T + oo yield e_ayu > eyur4ay)(0(u)  Notes and references See II.4). . u Proof of (2.yu) Y Similarly for (2. Proof of ( 2. we have tc(ay) > 0 and we can bound (2. Let H(dt. Equivalently. (B(y) .T)  V. so that i/1(uL yu) < eayu .yu))• b(u.7R. dr JO Zoo ) f e7'B(r + dy) B(r) Jo ^00 ^00 H(dt.)r(u)r. For (2. y > r.6). FURTHER APPLICATIONS OF MARTINGALES 305 (we cannot use the stopping time r(u) directly because P(r(u) = oo) > 0 and also because the conditions of the optional stopping time theorem present a problem). eyuk (ay) = e7yu e > eyu"(ay ) ij(u.d.( u ) I T(U) < 00] . Rt has distribution B(r + dy)/B(r).4): We take a = ay in (2.1 . when Rt_ = r.(u.2): As noted in Proposition II.1.3).B(r))/B(r). dr) e 7( yr)B(dy) B(r) f oo o 0 r > H(dt.6) below by 1 E Le7Rr(. dr) denote the conditional distribution of (T(u).1.yuk (ay)(u&(u.T(u)K(ay) I yu < r(u) < T] F(yu < r(u) < T) > e. and the proof of the lower inequality is similar. we have ic(ay ) < 0 and use the lower bound E [e7Rr („).. Hence E [e7Rr (u) Jr(u) < ool ^00 H( dt.3).E [e.(ay)I T(u) < yu] P(r(u) < yu) (using RT(u) < 0). A claim leading to ruin at time t has c.2.f. it follows easily from (2.
1) is an example of sharp asymptotics : .1). The advantage of the large deviations approach is. Thus. e. og For sequences fn.1) does not capture the \ in (3.1 We will go into some more detail concerning (3. Cramer considered a random walk Sn = X1 + . if x > EX1. gn with fn + 0 .nn or C2e.. The last decades have seen a boom in the area and a considerable body of applications in queueing theory.1) amounts to the weaker statement lim 1 log P I Sn > x I = 17. 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 .(B) = log EeOX 1 is defined for sufficiently many 0. then P C S. we will write fn 1. cle . 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.. logarithmic asymptotics is usually much easier to derive than sharp asymptotics but also less informative . Example 3.means (as at other places in the book) that the ratio is one in the limit (here n * oo). the parameter will be u rather than n). 1) but only the dominant exponential term . and gave sharp asymptotics for probabilities of the form P (S. . . However .?n typically only give the dominant term in an asymptotic expression .2). The classical result in the area is Cramer's theorem. + X. The limit result (3. For example.1) where we return to the values of 0.2) can be rewritten as F (Sn/n > x) 1g a'fin. which in the setting of (3.. however . such that the cumulant generating function r.g. (3.^ e nn 1 > x n 0o 2xn (3. gn 4 0./n E I) for intervals I C R. and that a considerable body of theory has been developed.3na with a < 1. Thus . not quite so much in insurance risk.gn if nioo lim 109 fn = 1 log gn (later in this section. Accordingly. (3. logarithmic asymptotics . v2 later.306 CHAPTER XI. ri.the correct sharp asymptotics might as well have +.. its generality. large deviations results been. nroo n n /// Note in particular that (3.
q = rc* (x). exponential change of measure is a key tool in large deviations methods. More precisely.425. V > 0 e.(0)) e 307 (other names are the entropy. of P(X1 E dx) = E[e9X1K.sseo f which in conjunction with (3. Most often. replacing Sn in the exponent and ignoring the indicator yields the Chernoff bound P Sn > x 1 < e°n (3.4) immediately yields (3. nx < Sn < nx + 1. P with mean nx and variance no.3) is put equal to x. and hence for large n P(Sn/n > x) > E [e. S rtn > x 1. since Sn is asymptotically normal w. which is a saddlepoint equation . if we replace Sn by nx + o / V where V is N(0.r.9S„+n' ( 9).1).(e)i XI E dx]. we have P(nx < Sn < nx + 1. 2 where o2 = o2(x) = rc"(0). i.4) n Next.the mean rc'(0) of the distribution of X1 exponentially tilted with 0.96o /] > 0.3.2).r.960/) * 0.4 enn +1. the LegendreFenchel transform or just the Legendre transform or the large deviations rate function).e. rc*(x) = sup(Ox . the sup in the definition of rc* can be evaluated by differentiation: rc*(x) = Ox . (3. we get P(Sn/n > x) E [e9nx +nK(9)9" '.rc(0) where 0 = 0(x) is the solution of x = rc'(0). LARGE DEVIATIONS Define rc* as the convex conjugate of rc.tin f o') o e9o^y 1 1 ey2/2 dy 21r = etin 1 Bo 27rn . Define . Since P nn > x) = E {e_8 ' ( 9).t. In fact.
p > 7 < zn. Mogulskii's theorem which gives path asymptotics. we shall concentrate on a result which give asymptotics under conditions similar to the GartnerEllis theorem: Theorem 3 ... We further write µ = tc'(ry). (iii) #c (8) = limn./^ >7 < zn n for n n0.'s. r(u) = inf {n : Sn > u} and o(u) = P('r(u) < oo).s. integrates to 1 by the definition of Icn).308 CHAPTER XI. 1]. we introduce a change of measure for X1. 1) and no such that Sn .1). however..3. see Jensen u [215] or [APQ] p..v.dxn) = 05nKn(7)Fn(dx1.3 For each i > 0.h. We shall need: Lemma 3 .. be a sequence of r.. X2. . MISCELLANEOUS TOPICS which is the same as (3..dxn) where Fn is the distribution of (X1i . Sanov's theorem which give rare events asymptotics for empirical distributions.2 (GLYNN & WHITT [163]) Let X1. which is of similar spirit as the dicussion in VII.. is differentiable at ry with 0 < K'(y) < 00. (ii) lim supn. and write Sn = X1 + • • • + Xn. .e < 8 < y + e..e < 8 < y + e.'(u) )Ng a"u. 260 for details.. Xn) and sn = x1 + • • • + xn (note that the r. For the proof. Pn Sn1 . (iv) tc(ry) = 0 and r. The substitution by V needs. commonly denoted as is the saddlepoint approximation. and the WentzellFreidlin theory of slow Markov walks. In the application of large deviations to ruin probabilities. .... there exists z E (0. Then i/. n Icn(0) exists and is finite for ry .. which is a version of Cramer's theorem where independence is weakened to the existence of c(O) = limn. Assume that there exists 'y. Further main results in large deviations theory are the GartnerEllis theorem.. . Xn given by Fn(dxl. e > 0 such that (i) Kn (0) = log Ee°Sn is welldefined and finite for 'y .o log Ee9Sn /n... to be made rigorous. that is. . asymptotics for probabilities of the form P ({S[nti/n}o<t<l E r) for a suitable set r of functions on [0. Ee9X n < oo for e < 0 < e.
h.2.2. We first show that lim inf„_.YS. in particular the r. it is easy to see that the r.077 n^oo n and by Taylor expansion and (iv ).s.ne(p+ 17). This proves the existence of z < 1 and no such that Pn (Sn/n > µ.. is of order .s. Let r7 > 0 be given and let m = m(77) = [u(1 + 77)/µ] + 1.77) follows by symmetry (note that the argument did not use µ > 0).Bµ .71 < e and jq9j < e.n e(µ +o)w"(7) [Eep(B +7)Sn]1 /p [EegoX.r (7) n = e.W. h.> .+r7) < zn for n > no. This establishes the first claim of the lemma . can be chosen strictly negative by taking p close enough to 1 and 0 close enough to 0.n m µ 1 + rl . mµ Sm > u] km e7Sm+n. For Sn1i we have Fn(Sn 1/n > µ+r7) < ene(µ+ 1?)EneeS„1 = ene ( µ+n)EneeSneX„ eno(µ +n) Ee(e+7)Sn ex„ wn (7) < e.. S. Clearly. P n(Sn/n > {c+77) < e no(µ 309 +n)Enees n +n)elcn(B +7). The rest of the argument is as before..]1/q = e. can be chosen strictly negative by taking 9 small enough.µ?7 .. for Sn.3. 0.91) + o(O ) as 0 J.n > u ) = [ Em [em Em 1e. ( U) P(S. S.+r. Since I EeqOX „ ] 1/q is bounded for large n by (ii). The corresponding claim for Pn(Sn/n < µ .h. u Proof of Theorem 3. log zl'(u)/u > 'y.y) .ne(µ limsup 1 log Pn (Sn/n > µ + 17) < ic(9 + ry) . > 1 +17] m(7)..m(7). 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.s. LARGE DEVIATIONS Proof Let 0 < 9 < e where a is as in Theorem 3. the r . Then V.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 +.
. I > IL exp `S.. 3.. logO(u)/u > ry.6) for some z < 1 and all n > n(E).. and since Ic.log z) /2 and Sn Fn\ n >lb+S) <Zn.n(ry)/u 4 0andm/u* (1 + r7)/µ.7) so that n(b) I1 < e'Yu E en. 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).YS +^c CHAPTER XI. this is possible by (iii). n=1 n=n(b)+1 00 Lu(1 +6) /µJ 13 F( T (u) = P(T(u) n). Pn \ > la+ 8 I < zn (3.0 log i'(u )/u < 'y.(•) goes to 1 by Lemma 3. Sn > u] < eYu+Kn(7)pn(Sn > u) (3. 0 yields liminfu __.. P(T(u) = n) < P(Sn > u) = En [e7S. I2 = F(T(u) = n). we get lum inf z/i(u) 1 +12r7 >_ ry + 77 Letting r7 J. MISCELLANEOUS TOPICS (7).I < µl1 1+77 I M 1_ 1+277 S.n Yµ 1 + m + r ('Y) } U n \ 77 m µ µ7 1 < 1+ 77 ) Here E.. n=1 .(Y). (iv) and Lemma 3.3.310 ]Em I e. Obviously.+wn(7). 14 = = E Lu(16)/aJ+1 Lu(1+6)/µJ+l = n) and n(S) is chosen such that icn('y )/n < 6 A (.. For lim supu.
C 26u `p / +1 I e6u(1+6)/µ (3. S. µ n=n(6)+1 \ 1u(16)/µ1 00 1 zn < e7u E Z n/2 < e(U xn/2 E n=n(6)+1 n=0 eYu = 1 . u . > u) Lu(1+6) /µJ +l 00 )^n 'YSn+kn (7) .' 1 + b) n e7u x 1 /2 1 n x n / 2x (3.10) 00 I4 < E F(Sn_1 < u. we get lim sup log u/00 O (U) < y + b(1 + b) U Letbl0.11) [u(1+6)/µJ+1 1  Thus an upper bound for z/'(u) is n(6) e'Yu n=1 eKn (7) + 2 + (28U + 1) e6u(1+6)/µ Fi 1 zl /2 and using (i).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. LARGE DEVIATIONS Lu(16)/µJ 311 I2 < e"u n=n(6)+1 e'n(Y)P(Sn > u) < Lu(16)/µJ ^. 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 . Sn1 C U.3. eryu en logz/2p n nt n.
z 1/z For I1. (7 + a) < 2arc'(7). ryue«iu . Corollary 3.b)/i(7). 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. u(1 + b)/i(7)) Proof Since V. it suffices to show that for j = 1. IV.2.11 ) can be sharpened to x 4 [u(1+6)/µJ /2 1 .xl/2 to give the desired conclusion.4 Under the assumptions of Theorem 3. we have rcn (a + 7) < 2n^c(7 + a) < 4narc' (7). say n n1.u(1+b)/rc'(7)).4. 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 . I2.4/3rc'(y) > 0.. 13 = P(T (u) E (u(1 b)l^ (7). MISCELLANEOUS TOPICS The following corollary shows that given that ruin occurs. Then for n large.(u) = I1+I2+I3+I4'^ ery( u). we need to redefine n(b) as L. the last steps of (3. Letting c11 = maxn<n. For 14. For I. u . e'.8) by P(S. the typical time is u/rc'(7) just as for the compound Poisson model. For 12.. 2. it holds for each b > 0 that 0(u) 1' g F(T(u) E (u(1 .7' a"ju. this is straightforward since the last inequality in (3.Q is so small that w = 1 . > u) < e"' E eIsn = ectueKn (a+'Y)Kn(7) where 0 < a < e and a is so small that r.3ui where .312 CHAPTER XI. cf. we replace the bound P(Sn > u ) < 1 used in (3.('+'Y).
V(s) with m. Let {Nt}t>0 be a possibly inhomogeneous Poisson process with arrival rate .12) k=0. r. The problem is whether this is also the correct logarithmic asymptotics for the (larger) ruin probability O(u) of the whole process. Obviously many of the most interesting examples have a continuous time scale.v. and we conclude that Theorem 3 .'(y) > 0. (iv) becomes existence of a limit tc(9) of tct(9) _ log Ee8S° It and a y > 0 with a(y) = 0.LARGE DEVIATIONS 313 Example 3 . t] is Rt = E V (Un) n: o„ <t . Thus the total reward in the interval [0. It is then wellknown and easy to prove that Sn has a normal distribution with mean np and a variance wn satisfying i lim wn = wz = Var(X1 ) + 2 E Cov(Xl. Xk+l) k=1 00 naoo n provided the sum converges absolutely.2 then immediately yields the estimate log F( sup Skh > u) a7u (3. 11 Inspection of the proof of Theorem 3..2 shows that the discrete time structure is used in an essential way.. Assuming that the further regularity conditions can be verified.. An event occuring at time s is rewarded by a r.1. Hence z z\ 2 z nrn(9) _ n Cn0p+BZn/ * . whether P ( sup St > u ltg a ^" 0<t<oo // (3. The reader not satisfied by this gap in the argument can easily construct a discrete time version of the models! The following formula (3. but nevertheless..f.1. 09(9).3(s) at time s.. for the ruin probability z/'h(u) of any discrete skeleton {Skh}k=0. To verify these in concrete examples may well present considerable difficulties. the key condition similar to (iii). If {St}t> 0 is the claims surplus process. i..e.. 2 is in force with y = 2p/wz.3. criteria are given in Duffield & O'Connell [124].(O) = 9µ+02 for all 9 E R.13) One would expect this to hold in considerable generality.5 Assume the Xn form a stationary Gaussian sequence with mean p < 0. and in fact.14) is needed in both examples .. we shall give two continuous time examples and tacitly assume that this can be done. Theorem 3.g.
i. Since the remaining conditions of Theorem 3. 7 Given the safety loading 77.. It is interesting and intuitively reasonable to note that the adjustment coefficient ry for the shot . Kt (0) t (Ee9U"it8i J0 ..t. e. (3. Example 3. At = . leading to St = At(1+77) Joo t S8 ds. if the nth claim arrives at time a.9t. it contributes to St by the amount Un(t .1) ds . but that a claim is not settled immediately. <t which is a shotnoise process. one would take p(t) = (1 + rt)At/ t.s). then the payments from the company in [on. the best estimator of /3µB based upon Ft. An apparent solution to this problem is to calculate the premium rate p = p(t) at time t based upon claims statistics . MISCELLANEOUS TOPICS are the event times.'`1 U. 0 Example 3 .0 and assume there are y.9t = /3 J t (Ee8U° i8l . We let ic (9) = 3(EeWU° . we conclude that Cu) log e7 u (cf. Un represents the total payment for the nth claim). is At . We further assume that the processes {U1(s)}8>0 are i.14) (to see this . derive . the CramerLundberg model implicitly assumes that the Poisson intensity /3 and the claim size distribution B (or at least its mean µB) are known..1) .14).1) ds rt (3. O'n +S] is a r . nondecreasing and with finite limits Un as s T oo ( thus.g. assuming a continuous premium inflow at unit rate. n: o. Then logEeOR° = J0 /3(s)(^8(9) . Most obviously. this is not realistic .314 where the an CHAPTER XI.6 We assume that claims arrive according to a homogeneous Poisson process with intensity 0 .. Thus. Of course. the Cramer. Un(s). we have rct (9)/t 4 ic (9).Q„) . a differential equation in t). (9) < oo for 9 < 'y + C.Lundberg model has the larger ruin probability.15) .v. Of course . If the nth claim arrives at time Qn = s. Thus by (3. More precisely.1) ds .noise model is the same as the one for the Cramer Lundberg model where a claim is immediately settled by the amount Un. e > 0 such that ic('y) = 0 and that r.It.2 are trivial to verify. = U„ ( t . . Thus.d. where Ft = a(A8 : 0 < s < t). 0 and since EeOUn(8) + Ee°U^ as s * oo. we have S. the above discussion of discrete skeletons).
standard exponential . 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./3. and since the remaining conditions are trivial to verify.b(u) IN a'Yu (cf.d.16) i=1 o i=1 Let ict (a) = log Eeast . one has y > y' (3. which yields eau f 1 t(1+n )audtl = E r Ee°Y = E [O(1+n)aueaul = E [eau J L Jo J L1+(l+r))aUJ .d. It then follows from (3.3.18) Thus (iii) of Theorem 3.1) or . typically the adaptive premium rule leads to a ruin probability which is asymptotically smaller than for the CramerLundberg model .(1 + r7) log t (3.21) 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 . LARGE DEVIATIONS With the Qi the arrival times.1) . (3.14) that rt _ 13 Jo _ (a [1_( i+77)log]) ds_flt = t (a) (3.i. i.i. (3. uniform (0. To see this . we conclude that t.20) (3. equivalently.(1 +i) f > i= 1 s ds = E Ui 1 . Ui Nt / t 01i 315 St = Ui .19) with equality if and only if U is degenerate.e. rewrite first rc as te(a) _ /3E 1 1 +(1+77)aUJ eau 1 . we have Nt t N. Indeed.log Oi are i.17) K(a) f o 1 O (a[I + (1 + 77) log u]) du )3. the Vi = .(1 + 17)0µB = 0.2 hold. Thus.
though we do not always spell this out. This implies n(y*) < 0. so there exists a unique zero xo = xo(r7) > 0 such that k(x) > 0.xo.7.2. the study is motivated from the formulas in IV. and k(x) < 0. 4 The distribution of the aggregate claims We study the distribution of the aggregate claims A = ^N' U. .316 CHAPTER XI. this in turn yields y > y*. MartinL6f [256]. 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. = P(N = n) = e(3an However.20) is due to Tatyana Turova. see Nyrhinen [275] and Asmussen [25].19). In addition to Glynn & Whitt [163]. For Example 3. using that Ek(U) = 0 because of (3. the function k(x) = e7*x . The main example is Nt being Poisson with rate fit.(1 + ri)y*x is convex with k(oo) = 00. we then take t = 1 so that p.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. Further applications of large deviations idea in risk theory occur in Djehiche [122]. k(0) = 0. are i. [245].d. Lehtonen & Nyrhinen [244].. assuming that the U. For notational simplicity. Further. x > x0. with common distribution B and independent of Nt. we are interested in estimating P(A > x) for large x.i. In particular. see also Nyrhinen [275] for Theorem 3. MISCELLANEOUS TOPICS Next.2 expressing the finite horizon ruin probabilities in terms of the distribution of A. at time t.1 . say one year. much of the analysis carries over to more general cases. rc*' (0 ) < 0. [257] and Nyrhinen [275]. Dembo & Zeitouni [105] and Shwartz & Weiss [339]. 0 < x < x0. Therefore e7'U _ k(U) E [1+(1+77)y*U] . y = y* can only occur if U . 11 Notes and references Some standard textbooks on large deviations are Bucklew [81]. k'(0) < 0. a* (s) are convex with tc'(0) < 0 . the proof of (3. and since tc(s). Further.
Vare(A) = s. e9x+K(°) P(A > x) B 2ir /3 B" [9] Proof Since EBA = x. Hence P(A > x) = E e [e9A+ ic(9). we define the saddlepoint 9 = 9(x) by EBA = x. A > x) eex+K(e ) ee AB°[ely 1 ev2/2 dy 0 2^ 00 9x+p(e) e ezez2/(2BZpB „[9)) dz 9 27r/3B" [9] fo eex+w ( e) oo z x)] ] 0 27r /3B" [9] o e 9 2 /3B" [9] J eex+w(B) dz . A > x)] = eex+K( e)E9 [e .e.[s])3/2 = 0. Then Ee"A = e'(") where x(a) _ 0(B[a] . (4.3e(bo[a] .3B"[9]. 818' where s' = sup{s : B[s] < oo}. Proposition 4.4.1).1 Assume that lim8T8.1.3B[9] and Be is the distribution given by eox B9(dx) = B [9] B(dx). B"[s] = oo.x)//3B"[9] is standard normal.1). no(a) = logE9e'A = rc(a + 9) . In particular.1) where )30 = . The analysis largely follows Example 3.9(Ax). A E dx] ."(0) = . THE DISTRIBUTION OF THE AGGREGATE CLAIMS 317 4a The saddlepoint approximation We impose the Poisson assumption (4.2) implies that the limiting Pedistribution of (A . Then as x * oo. For a given x. B"' [s] lim (B". only with 0 replaced by a9 and B by B9.ic(9) = . This shows that the Pedistribution of A has a similar compound Poisson form as the Fdistribution. K'(0) _ ic'(9) = x. The exponential family generated by A is given by Pe(A E dx) = E [eeA K(9). i.
1 goes all the way back to Esscher [141].(D X . [138].(3µB)/(0µB^))1/2 has a limiting standard normal distribution as Q ^ oo. large x. either of the following is sufficient: A. For details. more generally. In fact. 2). b(x) = q(x)eh(z). bounded with b(x) .Q{AB (4. The present proof is somewhat heuristical in the CLT steps. A covers the exponential distribution and phasetype distributions. it holds that EA = .e. just the same dominated convergence argument as in the proof of Theorem 2.x') where x' = sup {x : b(x) > 0}. the distribution of A is approximately normal . and (4. B covers distributions with finite support or with a density not too far from ax° with a > 1.l3pB.2) is often referred to as the Esscher approximation. The (first order) Edgeworth expansion states that if the characteristic function g(u) = Ee"`}' of a r. leading to P(A > x) :. b is logconcave. it is quite questionable to use (4.EN B(x). In particular. b is gammalike. 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 .2i and that (A . Thus .3) and related results u for the case of main interest .1).v.2 If B is subexponential and EzN < oo for some z > 1. or. For example. Jensen [215] and references therein. where q(x) is bounded away from 0 and oo and h (x) is convex on an interval of the form [xo. For a rigorous proof. Var(A) _ ^3p.4) . Remark 4 .318 CHAPTER XI. i.3) The result to be surveyed below improve upon this and related approximations by taking into account second order terms from the Edgeworth expansion. Notes and references Proposition 4. see Embrechts et al. Y satisfies 9(u) ti eu2/2(1 + ibu3) (4. then P(A > x) . Furthermore 00 b(x)Sdx < oo for some ( E (1. MISCELLANEOUS TOPICS It should be noted that the heavytailed asymptotics is much more straightforward. 1 . under the Poisson assumption (4. For example.ycix °ie6x B.1 yields: Proposition 4. some regularity of the density b(x) of B is required.
i 6 r 1 3 so that we should take b = ic3/6 in (4. (4. Let Y = (A . zl_e be the 1 .h.c2i.3& (y).99. one needs to show that 163. as u2 u3 u4 9(u) = Ee'uY = exp {iuci . If this holds . ylE should be close to zl_E (cf. (4.e..2 2 . and from this (4.EY)3. . K3 = E(Y .: EA + zl_E Var(A) .2 ^ \1 . .3!). u5.equantile in the distribution of Y.. THE DISTRIBUTION OF THE AGGREGATE CLAIMS where b is a small parameter. If the distribution of Y is close to N(0. f °o 9(y) = 1 e'uye u2/2(1 + iSu3) du 27r _ cc(y) .2X2 . which is often denoted VaR (the Value at Risk). s.i 3 K3 } Pt^ exp . are the cumulants . defined as the the solution of P(A < yle) = 1 . are small.l = EY.5) may be negative and is not necessarily an increasing function of y for jyj large. .5).5 (y3 . however. of (4...s. so that 1(u) 3 exp { .6) . one expects the u3 term to dominate the terms of order u4. Remark 4.2K3 + 4i 64 + .5) follows by integration. the CLT for Y = Y6 is usually derived via expanding the ch.4. the NP (normal power) approximation deals with the quantile al_E.. in particular.f. Thus if EY = 0.. A particular case is a. Heuristically. Rather than with the tail probabilities F(A > x).. In concrete examples . Var(Y) = 1 as above . and so as a first approximation we obtain a1_E = EA + yle Var(A) .. the standard normal distribution. resp. where Kl .y2)^P(y)• 319 Note as a further warning that the r. the density of Y is 1 °° _ eiuy f(u) du 2x _. K2 = Var (Y). K4 . then P(Y < y) 4(y) .5) is obtained by noting that by Fourier inversion.EA)/ Var(A) and let yl_E.6(1 .1)..
4c Panjer 's recursion Consider A = constants a. 21 .5) by noting that the 4.zlE )w(zl _E) = which combined with S = EY3/6 leads to q^ 1 Y1 . as required ..y2)cp( y) term. This leads to t( yl E) .EA)3 a1_E = EA + z1_E(Var (A))1/2 + 1 Var(A) Under the Poisson assumption (4. and assume that there exist n ) Pn_i .yi.. n = 1. . this holds with a = 0. b such that EN 1 U%. let pn Pn = (a+ = P(N = n). MISCELLANEOUS TOPICS A correction term may be computed from (4. that [101] distinguishes between the NP and Edgeworth approximations.1)! n ^eQ .zlE)W(zlE) 1 .zl E)V(zl_E) . K5 .E ..S(1 .zl E)^o(zl E) .1).1) E (A .(y) terms dominate the S(1 . We can rewrite (4.1)EY3.E = z1E + S(zi_E . [101]. k3 is small for large /3 but dominates 1c4.k = /3µB^1 / (..6pBki) d/2.S(1 .1)^ 2) µ'E Notes and references We have followed largely Sundt [354].5(1 .3ni /3 .320 CHAPTER XI.E)A1 l E) 1 E 4)(yl E) ^' .. Using Y = (A .. b = /3 for the Poisson distribution with rate /3 since Pn = Pn1 n! n (n . Another main reference is Daykin et at. Note.E )Azl E) 4(z1E) + ( ylE .E(/3PB^1 )1^2 + s(z1E . the kth cumulant of A is /3PBk' and so s.EA ) / Var(A).zi.7) as 1 (3) a1E = Qµa +z1 .. this yields the NP approximation 6(Z1 _E . In particular .E + (yl. however. For example.6 (1 .
Since the sum over i is na + b. j = 0. 2.11) Remark 4. the complexity (number of arithmetic operations required) is O(j3) for (4.10) f o = po. .4. n..14) is therefore a + b/n.9). THE DISTRIBUTION OF THE AGGREGATE CLAIMS 321 Proposition 4. and calculating the gj*n recursively by 9*1 = 9j. u Proof of Proposition 4. then j (a + b!) 1ag k_1 3 gkfj... the value of (4. ..13) but only O(j2) for Proposition 4.5 The crux of Proposition 4.12) where g*n is the nth convolution power of g.4.. fj = E (a+ b k =1 )9kfi_k .. . j = 1.. . . if go = 0.. E[a +bU=I >Ui =j l i=1 J (4.} and write gj = 2 . .k . The expression for fo is obvious. By symmetry. which would consist in noting that (in the case go = 0) fj = pn9jn n=1 (4.12) we get for j > 0 that fj n a b + n p nlgj *n 00 U I n 1 *n = E a+bUi=j pn19j n=1 j i=1 CC) n Ui EE n=1 Ia +b Ul i=1 =j pn_1 . 2. 2.4 is that the algorithm is much faster than the naive method.1. fj = P(A = j). (4.14) is independent of i = 1. n = k=n1 9k(n1 )9j k • (4.. . .13) Namely. Hence by (4.12). (4.4 Assume that B is concentrated on {0. (4.4. j = 1. 1. Then fo = >20 9onpn and fi = 1 E In particular... (4. j1 g.. .
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).
we get EX = EN • EU.i. This follows by taking Xl = U A b.6).d.4). Assuming for simplicity that the Ui are i. X2 = U .5) holds trivially.6 Assume the Ui are i.P. Applying the inequality Ecp(Y ) > EW(E (YIX )) (with W convex ) to W(y ) = eryy. Then for any local retention function u . and so on.5) reduce quite a lot. Proof As in the proof of Proposition 6. as often local as global.h( UU) = EN • E[U . u But since ry > 0. however.h(Ui)P JJJ l:='l {ry ] or.3).b)+ = Eh(U) (and the same P) satisfies 71 > ry.h(Ui)] .4).P } < 1 = Eexp E[Ui.. ry* > 0 because of (6. Local reinsurance with h(u) = (u .6.b)+ is referred to as excessofloss reinsurance and plays a particular role: Proposition 6.P I = EC [7]N.P > EexP{7[X . The arrangement used in practice is.4). N E X .6) u where C[ry] = Ee'r(u4(u)). this implies 7* > 7. REINSURANCE Proof Define N 329 h* (x) = E > h(Ui) X = x .h * (X) .4) and u g(x) = e7x in Ohlin's lemma. y = Ei [Ui .b)+ with b determined by E(U .5 Because of the independence assumptions . it suffices to show that Eexp {ry ii 'UiAb.4. Eexp 7 [E [Ui . that 01[ry] < 0[y] where 0[y] = Ee'r(U^') . expectations like those in (6.h(Ui)] . (6. i.d. we get N 1 = Eexp ry E[Ui ii .h(Ui)] . the excess ofloss rule hl (u) = (u . then (6. Remark 6. .h(U) (as in the proof of Proposition 6.h(U)].P]}. (6. ' ii (6.h(u) and any P satisfying (6. appealing to (6.
see also Sundt [354]. The present proof is from van Dawen [99]. Bowers et at. . [76].330 CHAPTER XI.g. The original reference for Ohlin's lemma is Ohlin [277]. MISCELLANEOUS TOPICS Notes and references The theory exposed is standard and can be found in. Heilman [191] and Sundt [354]. e. See further Hesselager [194] and Dickson & Waters [120].many texts on insurance mathematics.
. are independent and Y1. 2h. t +a]). when t is large. The number max k : Tk_j < t of renewals in [0..1) (here U(t) = U([0..Appendix Al Renewal theory la Renewal processes and the renewal theorem By a simple point process on the line we understand a random collection of time epochs without accumulation points and without multiple points. stating that U(t+a)U (t) ^ a. If F satisfies the stronger condition of being spreadout (F*' is nonsingular w . . t] is denoted by Nt...r. t]) so that U(t + a) ... Lebesgue measure dt normalized by the mean to of F.} for any h > 0.. = T„ . Y2. i. The point process is called a renewal process if Yo.U(t) is the expected number of renewals in (t. That is. Lebesgue measure for some n > 1)..T„_1). The associated renewal measure U is defined by U = u F*" where F*" is the nth convolution power of F. of interarrival times and the time Yo = To of the first arrival (that is. . of epochs or the set Y1. + U2 where U1 is a finite measure and U2(dt) = u(t)dt where 331 .. the distribution of Yo is called the delay distribution.t. some condition is needed: that F is nonlattice. Y1. Then Blackwell 's renewal theorem holds. denoted by F in the following and referred to as the interarrival distribution. The mathematical representation is either the ordered set 0 < To < T1 < . Technically. the renewal process is called zerodelayed.e. Y2. note in particular that U({0}) = 1. If Yo = 0.. not concentrated on {h. The renewal theorem asserts that U(dt) is close to dt/µ. Y. all have the same distribution. t 00 (A. then Stone 's decomposition holds : U = U.. U(A) is the expected number of renewals in A C R in a zerodelayed renewal process. .
4) that z is Lebesgue integrable with limZ.R. (A. that z(u) has a limit z(oo) (say) as u 4 oo. stating that U(t)/t > 1/p. and that F has a bounded density2. then it suffices for (A. IV). but suffices for the present purposes . z(u) a known function.1 if F is nonlattice and z (u) is directly Riemann integrable (d. A weaker (and much easier to prove) statement than Blackwell's renewal theorem is the elementary renewal theorem. Both result are valid for delayed renewal processes. in convolution notation Z = z + F * Z. i. see [APQ] Ch. 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 . (A. oo). (A.2 Assume that Z solves the renewal equation (A. Then Z(u) 4 z(oo).i. and F(dx) a known probability measure . µF (A.2) Z(u) = J0 u z(x)U(dx). z(x) = 0.9.2) has the unique solution Z = U * z. U Z(u . the asymptotic behavior of Z(u) is given by the key renewal theorem: Proposition A1. the statements being EN(t + a) . resp. wee shall need the following less standard parallel to the key renewal theorem: Proposition A1.out.x)F(dx). Note in particular that F is spreadout if F has a density f.a. In 111. IV).2).EN(t) .4) If F is spread.i".5) 2This condition can be weakened considerably . then Z(u) i f0 z(x)dx .. Equivalently. u u PF 4 00.332 APPENDIX u(t) has limit 1/µ as t 4 oo.3) Further.e. (A. Under weak regularity conditions (see [APQJ Ch.
. .. this covers discrete Markov chains where we can take the Tn as the instants with Xt = i for some arbitrary but fixed state i.t))u(ut) dt 0 0 J f z(oo) • 1 dt = z(OO). The kth cycle is defined as {XTk+t}o<t<Yk . the postTk process {XT. of Yo. A stochastic process {Xt}t>0 with a general state space E is called regenerative w. multiply (A. {Tn} if for any k. Y2. To this end. Note.2) by e7x to obtain Z = z +P * Z where Z(x) = e'Y'Z(x)..k+t }t>o is independent of To. Z(u) U = 1 u 1 u f z(u . The distribution F of Y1.r.. z(x) = e7xz(x). T1. However.d.x)u(x) dx = z(u( 1 . . where the Tn are the instants where a customer enters an empty system (then cycles = busy cycles). we let µ denote its mean. equivalently. asymptotic properties can easily be obtained from the key renewal equation by an exponential transformation also when F(dx) does not integrate to one.5a. this expression is to be interpreted as a random element of the space of all Evalued sequences with finite lifelengths.APPENDIX 333 Proof The condition on F implies that U(dx) has a bounded density u(x) with limit 1/µF as x * oo. however. Y1 . ...e.} be a renewal process. Here the relevant F does not have mass one (F is defective). F(dx) = e7xF(dx). refer to the zerodelayed case.i. is called the cycle length distribution and as before. We let FO. the present more general definition is needed to deal with say Harris recurrent Markov chains. Yk ). a basic reason that renewal theory is relevant is the renewal equation II. The simplest case is when {Xt} has i. Hence by dominated convergence.. . or many queueing processes. A regenerative process converges in distribution under very mild conditions: . 0 PF µF 11 In risk theory. • .3) satisfied by the ruin probability for the compound Poisson model.. 1c Regenerative processes Let {T. . and its distribution does not depend on k. Tk (or. Eo etc. that F is a probability measure. This program has been carried out in III. T1. Tk and {Xt }o<t<Tk • For example.(3. i. Assuming that y can be chosen such that f °° Ox F(dx) = 1. cycles.t. However.. results from the case fo F(dx) = 1 can then be used to study Z and thereby Z. The property of independent cycles is equivalent to the postTk process {XTk+t}t>0 being independent of To.. that the existence of y may fail for heavytailed F.
6) id Cumulative processes Let {Tn} be a renewal process with i. and q(t) = sup It ... is given by Eg(Xoo) = 1 E0 f Ylg (Xt)dt..334 APPENDIX Proposition A1.r. assume that p < 00 and define Un = ZT}1 . oo).t. µ 0 If F is spreadout.3. {Tn}. Then it (ii. If p = oo. cycles (we allow a different distribution of the first cycle)...0 is called cumulative w. C).oo (i. under the condition of Blackwell's renewal theorem. [0. An example is Zt = fo f (X8) ds where {Xt} is regenerative w.r.e.v.ZT }0<t<Y„+.t : t < Tk}.4 Let {Zt}t^. Then Xt Di X. Otherwise . Then {e(t)}.ZT Then: (a) If E sup I ZTo+t .t. for n = 1. This is the case considered in [APQ] V. then e (t) . just the same proof as there carries over to show: Proposition A1.'s by e.d. r. {Tn}. resp . i. in total variation.i.tEU1/µ)/f has a limiting normal distribution with mean 0 and variance Var(Ui) + (!)2Var (Yi)_ 2EU1 Cov(U1. (A.0 be cumulative w. then Xt . then (Zt .d. {i7(t)} are Markov with state spaces (0. C(t) and ij (t) both have a limiting stationary distribution F0 given by the density F (x)/p.. {Tn} if the processes {ZT +t . We denote the limiting r.. Then {Zt}t^.. 0<t<Yi then Zt /t a$• EU1/µ.t. are i..3 Consider a regenerative process such that the cycle length distribution is nonlattice with p < oo.e. 2. (b) If in addition Var(Ul ) < oo.. oo). fi (t) = inf {Tk .r. e(t )) .Tk : t < Tk} as the age.+ X. P(C ( t) < a) 4 0 for any a < oo) and ij (t) * oo.. but in fact. and we have: holds more generally that (rl(t)..i. where the distribution of X. Y1) le Residual and past lifetime Consider a renewal process and define e ( t) as the residual lifetime of the renewal interval straddling t.ZTOI < 00.
(t). W are independent. if in addition EYo < oo. Y1i Y2.y)P(Yo E dy) . Since the maximum Mn of n i. Y1 > t] 4 0. 0 If Markov renewal theory By a Markov renewal process we understand a point process where the interarrival times Yo .U(x) (c < oo because it is easily seen that U(x + 1) .6 Consider a renewal process with µ < oo. we can bound e(t) by M(t) = max {Yk : k < 2t/p}. and the conditional distribution of given 17 = y is the overshoot distribution R0(Y) given by FO(Y) (z) = Fo (y+z)/Fo(y).5 Under the condition of Blackwell's renewal theorem.i. Yo > 0] + f Eo^ (t . = z is Foz) The proof of (a) is straightforward by viewing {(r. the sum is o(t) so that Eo£(t)/t + 0 . and the conditional distribution of ri given l.APPENDIX 335 Theorem A1. U(x + 1) . Since z ( k) < E[Yi . l:) is the same as the distribution of (VW. Yl > t].i. (1 V)W) where V. V is uniform on (0.d. (b) the joint distribution of (ri.d.t. ^ > y) = 1 f +Y (z)dz. Then fi(t)/t a4' 0 and.v. Proof The number Nt of renewal before t satisfies Nt/t a4' p. . Hence for t large enough.t. For the second. ^) is given by the following four equivalent statements: (a) P (77 > x.U(x) < U( 1))..4.. we used: Proposition A1.^(t))} as a regenerative process.. but governed by a Markov chain {Jn} (we .y) = f U(t . the first statement follows.'s with finite mean satisfies Mn/n a$• 0 (BorelCantelli). (c) the marginal distribution of q is FO. are not i. Hence t t lt ) = f U(dy)z(t . r. the joint distribution of (rl. and the equivalence of (a) with (b)(d) is an easy exercise. EC(t)/t + 0. 1) and W has distribution Fw given by dFw/dF(x) = x/pF. (d) the marginal distribution of ^ is FO. Then Eo^(t) satisfies a renewal equation with z(t) _ E[Y1 . assume first the renewal process is zerodelayed. In IV.dy )z(y) < c ^ l z(k) Eoe(t 0 0 k=o where c = sup. In the general case. use t E^(t)/t = E[Yo .
IT. .t. oo).. the conditional distribution of {XT„+t}t>o given Yo.jEE is a family of distributions on (0. Jn = i is the same as the P. Yn. T_=inf{n>0: Sn<0}.336 APPENDIX assume here that /the state space E is// finite) in the sense that P(Y. in [APQ].7 Consider a nonlattice semiregenerative process. X2..} is nonlattice (it is easily seen that this definition does not depend on i).. .. Jo. .. Further: Proposition A1. Let X1. G+(x) = P(S.r. 0] or (0 . distribution ofjXt}t>o itself where Pi refers to the case Jo = i.+ < x. The semiregenerative process is then regenerative w. r+ < oo). namely {Twk } where {Wk } is the sequence of instants w where Jo. the Markov renewal process if for any n. . where the distribution of X. Jn_1. . Sn = X1 + • • • + Xn the associated random walk.i .. Notes and references Renewal theory and regenerative processes are treated..and regenerative processes.g. < yIJ) = Fij( y) on {Jn= i. = io for some arbitrary but fixed reference state io E E.T_ < oo). Alsmeyer [5] and Thorisson [372]. J1 i .. .}. the semiregenerative process is called nonlattice if {T. . be i. Jn +1=j} where J = a(JO. Then Xt 4 Xo..r. A2 WienerHopf factorization Let F be a distribution which is not concentrated on (oo. A stochastic process {Xt}t>o is called semiregenerative w. We call r+ (T_) the strict ascending (weak descending) ladder epoch and G+ (G_) the corresponding ladder height distributions..t. oo). These facts allow many definitions and results to be reduced to ordinary renewal. . Y1. e. ...) and (Fij )i.. For example. with common distribution F. . Assume that uj = EjYo < oo for all j and that {J„} is irreducible with stationary distribution (v3)jEE.. and define r+=inf{n>0: Sn>0}. G_(x) = P(ST_ < x. A Markov renewal process {Tn} contains an imbedded renewal process.d. is given by Eg(X00) = 1 YO vjEj f g(Xt) dt µ jEE o where p = ujEEViAj..
G+ * G_: (b) G_ (A) = f °° F(A . 0]. F(A) is the contribution from the event {T_ = 1} = {X1 < 0}. oo). (d) R+ = U_.. On {T_ > 2}. define w as the time where the preT_ path S1.T_=n} = {S. oo) (A. >0.x)R+(dx). 0<j<m.1 . A C (0..x)R_ (dx). Proof Considering the restrictions of measures to (oc. A C (oo. m<j<n}.7). n=0 The basic identities are the following: Theorem A2. . G_.7) (A. oo). S. 0]).. F(A . (A.7) follows since G+(A) = 0 when A C (oo. n 0 R_(A) = E I(Sn E A). 0). (e) R_ = U+. we may rewrite (a) as G_ (A) = G+(A) = F(A) + (G+ * G_)(A). the renewal measures U+=>G+. 0] and (0.8) (e. F(A) + (G+ * G_)(A).1 (a) F = G+ + G_ .and r_ preoccupation measures T+1 r_1 R+(A) = E E I(Sn E A). n=0 n=0 00 00 and the T+.r. .S. Sr_ _1 is at its minimum . (c) G+(A) = f °.>0.g. we consider the last such time (to make w unique) so that {w=m.S. In (A. A C (oo. More rigorously.APPENDIX 337 Probabilistic WienerHopf theory deals with the relation between F. U. G+. u .=n w=m i Figure A.=EGn. A C (0.
Sn_1Edx.>0. Aso. m < j <n. clearly (Sj Sm>0.+ E du) E P(S. ST_ E A) P(T+ = m.u) f0m m=1 n=m+1 00 J0 OO P(S. It follows that for n > 2 F (7. . S.. m=1 f S mming over n = 2. Sn1 E dx) n=1  F(A . ST_ E A . .8) is similar.XnEAx) 00 f 0 f 0 00 00 1: F(A .3 8 APPENDIX Reversing the time points 0..du) (G+ * G)(A)• C llecting terms._ E A ..0<k<ri .m.7) follows. (A.. m it follows (see Fig.3. 0<j<m.1. A._ = n . S.. E du) = P(T_=nm. ST+Edu).x)P(Sk < 0. SnEAIS. A. and reversing the order of summation yields P(T_ > 2.+ E du)P(S.= n. and the proof of (A.F(r_n_mSrEA_u). r+ = n) n=1 n=1 0  C0 E fF(Sk< 0.1).1) that P(Sj Sn._ E A) n1 f P(r_=nw=m Sm EduSrEA) m=1 n1 F(r+=mSr+Edu). Sr_ E Adu) (s ee again Fig . (b) follows from 00 G+ (A) _ E F(Sn E A. 0 < k < n. ...x)R+(dx).. SmEdu) = P(T+=m.
T+> n) = P(Sk < O. u Notes and references In its above discrete time version. we can rewrite (a) as 1 . Nevertheless.2 In terms of m. and the proof of (e) is similar. and G+. u Remark A2. if {St} is Brownian motion.4). it serves as model and motivation for a number of results and arguments in continuous time. the analogue of a random walk is a process with stationary independent increments (a Levy process.g. 11. and similarly H_ (s) = 1 . there are direct analogues of Theorem A2. In this generality of.F[s] = (1 . G_ are trivial. being concentrated at 0.SnEA) is the probability that n is a weak descending ladder point with Sn E A. In continuous time.1. Another main extension of the theory deals with Markov dependence.SnEA) = P(Sn<Sk. and sometimes in a larger strip.g.6. The classical analytical form of the WienerHopf problem is to write 1 .1(a) is from Kennedy [228].9) whenever F[s]. Sk = X1 + • • • + Xk = Sn .O<k<n. Then for A C (oo. 6+ [s].APPENDIX 339 and the proof of (c) is similar.s. However. Again. In discrete time. cf.SnEA) = P(Sn<Sk. Summing over n yields R+ (A) = U_ (A).f. then T+ = inf It > 0 : St = 0} is 0 a. there is no direct analogue of Theorem A2. see e.1.g.'s. this holds always on the line its = 0. consider a fixed n and let Xk = Xn_k+l..SnEA) = P(SnSn_ k. Since G+ is concentrated on (0. is based upon representing G+ as in (b). P(SnEA . G_ [s] are defined at the same time.0+[s])(1 . WienerHopf theory is only used at a few places in this book. The present proof of Theorem A2. H+ (s) = 1G+[s] is defined and bounded in the halfplane Is : ERs < 0} and nonzero in Is: Rs < 01 (because IIG+lI _< 1). which is basic for the PollaczeckKhinchine formula. the derivation of the form of G+ for the compound Poisson model (Theorem 11. such developments motivate the approach in Chapter VI on the Markovian environment model. E. .G_ [s] is defined and bounded in the halfplane is : ERs > 01 and nonzero in Is : ERs > 0}.0<k<n..Sn_k. the survey [15] by the author and the extensive list of references there.G_[s]) (A. For example. and using timereversion as in (d) to obtain the explicit form of R+ (Lebesgue measure). see for example Bingham [65]. For (d). 0].0<k<n.P as a product H+H_ of functions with such properties. oo).O<k<n.1). a number of related identities can be derived.
1. however . if m is s fficiently large. 0 . 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). three of the c rrently most widely used ones: xample A3. _I 0 (A.12) eA'AO = Ale AA (A. Here are. whereas there is no similar single established a proach in the case of matrix exponentials. Some fundamental properties are the following: sp(eA) = {e' : A E sp(A)} (A. ere A is the eigenvalue of largest absolute value.5 that when handling phase type distributi ons.13) henever A is a diagonal matrix with all diagonal elements nonzero. and eQ can then be computed as the mth power (by squaring if = 2). 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. Thus. It is seen from Theorem VIII.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. Here it is standard to compute matrixinverses by GaussJordan el imination with full pivoting .10) d dteAt = AeAt = eAtA (A. To circumvent this. one needs to compute matrix inverses Q1 and matrix exponentials eQt ( r just eQ ). write eQ = (eK)m where = Q/m for some suitable integer m (this is the scaling step).11) A f eAtdt = eA. Eo Kn/n! converges rapidly and can be evaluated without p oblems.
In practice. i. what is needed is quite often only Zt = TreQt (or eQth) with it (h) a given row (column) vector. some jumps are dummy in the sense that no state transition occurs ). 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. p different eigenvalues Aj i .2 (UNIFORMIZATION) Formally.3 i (A.3 (DIFFERENTIAL EQUATIONS) Letting Kt = eQt..APPENDIX 341 Example A3.14) 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}..e. To this end. The probabilistic reason that (A. .e. vp be the corresponding left . letting P = I + Q/i and truncating the series in the identity = e17t 00 Pn(. 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 . Zo = a (Z = QZ.7t) n=0 n! u °O n Pn (to see this. assume that Q is the intensity matrix for {Xt} and choose q with rt > max J%J = max qii• 1.15) Then it is easily checked that P is a transition matrix .14) holds is therefore that the tstep transition matrix for {fft} is eQt = E ent (. . construction of {Xt} by realizing the jump times as a thinning of a Poisson process {Nt } with constant intensity 77.. condition upon the number n of Poisson events in [Olt])  Example A3.. The approach is in particular convenient if one wants eQt for many different u values of t. the procedure consists in choosing some suitable i > 0. Zo = h).4 (DIAGONALIZATION) Assume that Q has diagonal form..]t)n (A. 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. However . One then can reduce to p linear differential equations by noting that k = ZQ. Let vi. Ap. i. Here is a further method which appears quite appealing at a first sight: Example A3 .
we have an explicit formula for eQt once the A j. (A. Then vihj = 0. Example A3. vi. i= 1 i=1 P P (A. however. some cases remain where diagonalization may still be appealing. Nevertheless. and we need to have access to software permitting calculations with complex numbers or to perform the cumbersome translation into real and imaginary parts. and hence A2 is so because of A2 = tr(Q). hp. and writing eQt as eQt = He°tH1 = H (e\it)di.18) Namely. D = ) 2 2 . say Al. Complex calculus : Typically.. we can take H as the matrix with columns hl. Qhi = vihi. (A. and vihi ¢ 0. In view of this phenomenon alone care should be taken when using diagonalization as a general tool for computing matrixexponentials.5 If Q= ( 411 ( q21 q12 q22 is 2 x 2. not all ai are real. and we may adapt some normalization convention ensuring vihi = 1.. the eigenvalue. this last step is equivalent to finding a matrix H such that H1QH is a diagonal matrix.. Everything is nice and explicit here: 411+q2+D' )12_g11+q2^^ where (411422z + 4412421.. v5Q = Aivi. hi have been computed. Then P P Q = > Aihivi = E Aihi (9 vi. i # j.17) eQt = E e\`thivi = E ea:thi ® vi. i=1 i=1 Thus. hp the corresponding right (column) eigenvectors.. two serious drawbacks of this approach: u Numerical instability : If the A5 are too close.342 APPENDIX (row) eigenvectors and hl.18) contains terms which almost cancel and the loss of digits may be disasterous. The phenomenon occurs not least when the dimension p is large. There are.16) (A.. say A = (Ai)diag. under the conditions of the PerronFrobenius theorem).g H1... of largest real part is often real (say.
e.21) Here the first term is the stationary limit and the second term thus describes the rate of convergence to stationarity.APPENDIX 343 Write 7r (= v1) for the left eigenvector corresponding to a1 and k (= hl) for the right eigenvector. The other eigenvalue is A = A2 = q1 . 1) .q. Then Al = 0 and the corresponding left and right eigenvectors are the stationary probability distribution 7r and e. h2 = Thus. b are any constants ensuring//Irk = 1.6 A particular important case arises when Q = q1 qi ) q2 q2 J is an intensity matrix.7 Let 3 9 2 14 7 11 2 2 . Then 7r = (ir1 7r2 ) = a (q21 Al . where (A.19) Example A3 .k1). l ab (g12g21 + (A1  411) 2) = 1. eqt = eNlt ( ir1ki i2k1 \ ir1 k2 72 k2 + e azt 7r2k2 i2k1 7ri k2 7r1 k1 (A. k  C k2 ) =b ( A1 q 1 Q11 / where a . i. v2 and h2 can be computed in just the same way.Q2i and after some trivial calculus one gets eQt = 7r 1 112 + eat 7r1 7r2 / (7fl 7r2) = ( 7r2 1r2 7r1 IF. it is easier to note that 7rh2 = 0 and v2k = 1 implies v2 = (k2 .20) ir = q2 ql qi +q 2 9l +q2 (A. replacing ai by A2. Of course. However. u Example A3.
2 2 1=ab(142+(1+2)2 ) = tab.344 Then D= 2+ 11)' 7 T4 2 =52. (A.11/2 .5 . e_6u A4 Some linear algebra 4a Generalized inverses A generalized inverse of a matrix A is defined as any matrix A. APPENDIX x1 3/2 .22) Note that in this generality it is not assumed that A is necessarily square.satisfying AAA = A. They are most often constructed by imposing some additional properties .. but only that dimensions match . and a generalized inverse may not unique.6. Generalized inverses play an important role in statistics. (A+A)' = A+A. for example AA+A = A. (A. ir =a(2 9 9 14 2 1 3 2 2)' k=b 14 =b 1+ 2 ir1 k1 ir2 k1 _ 9 2 10 5 7 9 70 1 ' 7r1 k2 7r2 k2 10 9 9 10 10 + 7 1 10 10 10 1 10 7 10 9 70 9 10 0 e4" = e_. A+AA+ = A+.. (AA+)' = AA+.11/2 + 5 1.23) . A2 = 3/2 .
Then for some b > 0.1 goes under the name fundamental matrix of the Markov chain). and can define /ail 0 0 0 0 0 0 A+ = C A' 0 0 0 C' .D + O(ebt). and exists and is unique (see for example Rao [300]).1Q = Q(Q . one is also faced with singular matrices .eir )1. = 0 where m < p is the rank of A.ew. then there exists an orthogonal matrix C such that A = CDC' where 0 0 D = AP Here we can assume that the A . Assume that a unique stationary distribution w exists .eir ).. . Rather than with generalized inverses .. one then works with Q = (Q .eir)1 = I . if A is a possibly singular covariance matrix (nonnegative definite).= (I .23) is called the MoorePenrose inverse of A. _ A.. E.P).e.25) . (A.P + e7r ). These matrices are not generalized inverses but act roughly as inverses except that 7r and e play a particular role . lt o eAx dx = te7r + D(eAt . (I . and define D = (A . .e ® 7r)1.APPENDIX 345 A matrix A+ satisfying (A. Am > 0. ( Q . Am+1 = ..g. are ordered such that Al > 0.I) (A.P + e7r)1 (here ( I .1 Let A be an irreducible intensity matrix with stationary row vector it. Here is a typical result on the role of such matrices in applied probability: Proposition A4.. 0 01 In applied probability.24) = te7r . most often either an intensity matrix Q or a matrix of the form IP where P is a transition matrix.g..
resp.I) (A. Then A(O) _ B(O) = 0. (A.h..27) Proof Let A(t).DZ(ent .e. I. . the rows are proportional to it. respectively.26) follows by integration by parts: t f t /' xeAx dx = [x {xe7r + D(eAx .26) 2 = 2 e7r + tD . For example. .h. h as 1 x m and k x 1 matrices. Interpreting 7r.s. Note that h ® it has rank 1.D + D2 + O(ebt). o Finally. in block notation i2h A®B= ( a11B a21 B a12B a22 B Example A4.I) . ()®(6 f 6/ 7f 8^ 7 8 )=! ^)( 6 7 8 )=(6^ 7^ 8^) \ u Example A4. the r. it follows that h ® it is the k x m matrix with ijth element hi7rj . the formulas involving O(e6t) follow by PerronFrobenius theory. 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. h ® it reduces to hit in standard matrix notation. (A.2e7r .I)} dx. of (A. Equivalently.91a(2) . B(t) denote the l.eir)eAt = eAt = A'(t).24).s. and the columns to h.J {xe^r + D(e .3 Let 2 A= 4 3 Vf' N7 5 )' B= ( 8 ). 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. and in fact any rank 1 matrix can be written on this form. B'(t) = e7r + DAeAt = eir + (I .346 t APPENDIX 2 xe Ax dx = eir + t(D + e7r) + D(eAt .I)}. see below.2 Let it be a row vector with m components and h a column vector with k components.
(A. then v1B1h1 and v2B2h2 are real numbers.3v'6.29).50 6 7 6 4f 4. and the number of such factors is precisely given by the relevant binomial coefficient. Using (A. (A B)' = eA®B e! L 1=0 0 .4vf.APPENDIX 347 Then A®B = 2 f 20. and v1B1h1 • v2B2h2 = v1B1h1 ® v2B2h2 = ( v1(&v2 )( B1(&B2 )( h1(&h2 ) . each of which is A ® I or I ® B.A9. it follows that e® ® e B An _ 0o oo oo Bn 7 I F n! = ` k! (I .31).30) eA+B = eAeB function generalizes to Kronecker notation (note that in contrast typically only holds when A and B commute): Proposition A4.4 eA® B = eA ®eB.31) Indeed.3vV/72f 20.3V8.28) In particular. (AED B)1 = (A®I+I(9 B)l is the sum of all products of t factors.5v/. if A ® I occurs k times.(A. C2 = h2 are column vectors. such a factor is Ak (&B 1k according to (A.k)! ( n0 n=0 t=0 k=0 J _ ® Ak ®Blk r ^. A2 = v2 are row vectors and C1 = h1. 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.29) If A and B are both square (k1 = ml and k2 = m2). then the Kronecker sum is defined by A(1) ®A(2) = A(1) ®Ik2 + k ®A(2). if Al = vi. (A.5v'8 5vf9 11 A fundamental formula is (A1B1C1) ®(A2B2C2) = (A1 (9 A2)(B1 (9 B2)(C1®C2).3f 4v/.
(A. Let P8f P(Sl).32).s. Then 2 0 ire At h • ve Bt kdt = (^®v)(A®B)1(e A®Ba . independent Markov chains. 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 +. Ps 1) = exp {sQ ( 1) } > p(2 ) = exp {sQ(2) } can therefore be rewritten as Taking s = 1 for simplicity . Q(2). the same time. From what has been said about matrices of {Yt( 1). { 1't(1) }. Yt(2) where independent Markov processes with intensity matri{y(2) } are {Y(1) }. in the definition (A. where transition matrix of the bivariate Markov chain {X n1). k any column vectors. P8 = exp {sQ} = exp {s (Q(1) ®Q(2)) } . p = P(1) ® {X }. { On the other hand. represents ces Q( 1). Thus .4 can easily be obtained by probabilistic be the sstep transition reasoning along the same lines . and Q = Q(1) ® Q (2) = Q(1) ® I + I ® Q(2) (A. and the form of the bivariate intensity matrix reflects the fact that Yt(2) } cannot change state in both components at due to independence .348 APPENDIX Remark A4. P(t) Yt(2) }. A special case of Proposition A4. h.6 Suppose that A and of B. {Yt(1). Yt(2 ) }. X ) }.5 Many of the concepts and results in Kronecker calculus have p(2) is the intuitive illustrations in probabilistic terms.I)(h ® k). resp . P(2). n2 n1 ) {X(2) } are independent Markov chains with transition matrices P(1). v whenever a is an eigenvalue of A and 0 is an eigenvalue be any row vectors and h. the {Yt(2) } transitions in the {Yt(1) } component and the second transitions in the component .3 < 0 Lemma A4 . first term on the r . Let further it.32) is the intensity matrix of the bivariate continuous Markov process {Yt(1).33) . we have P8 = Pal) ® p(2).
. j = 1. h can be chosen with 3By this..3 whenever a is an eigenvalue of A and 3 is an eigenvalue of B. and the corresponding left and right eigenvectors v. then IN < Ao for all A E sp(A). That is.. [APQ] X.7 Let A be a p x pmatrix with nonnegative elements. f o r each i. .29). which can be found in a great number of books. we have AO = 1. Here is the PerronFrobenius theorem. A is called aperiodic if the pattern of zero and nonzero elements is the same as for an aperiodic transition matrix. we mean that the pattern of nonzero offdiagonal elements is the same as for an irreducible intensity matrix. > 0 for k = 1. n..The PerronFrobenius theorem has an analogue for matrices B with properties similar to intensity matrices: Corollary A4.12). in such that io = i. and if we normalize v. h such that vh = 1. h can be chosen with strictly positive elements. .. (b) if in addition A is aperiodic. E (0. (A. We call A irreducible if the pattern of zero and nonzero elements is the same as for an irreducible transition matrix. the integrand can be written as ( 7r (9 v)( eAt ® eBt )(h ®k ) = ( 7r ®v)(eA (DBt)(h (& k). and appeal to (A. and the corresponding left and right eigenvectors v. 4c The PerronFrobenius theorem Let A be a p x pmatrix with nonnegative elements. see e.8 Let B be an irreducible3 p x pmatrix with nonnegative offdiagonal elements. Similarly. = j and atk_li. il. . so that by asssumption A ® B is u invertible. Then the eigenvalue Ao with largest real part is simple and real. . p there should exist io. ao). i. then An = Aohv+O(µ") = Aoh®v+O(µ") for some u. Then: (a) The spectral radius Ao = max{JAI : A E sp(A)} is itself a strictly positive and simple eigenvalue of A. Now note that the eigenvalues of A ® B are of the form a +. h = e and v = 7r (the stationary row vector)... .g. . .. .34) Note that for a transition matrix.1 and references there (to which we add Berman & Plemmons [63]): Theorem A4.APPENDIX 349 Proof According to (A.
(ti)diag where Q = T + (ti)diag is a proper intensity matrix (Qe = 0). For example. h such that vh = 1. let t = (ti)iEE # 0 have nonnegative entries and define T(°) = aQ . relate the eigenvalues of B to those of B via (A.1. 10) and use the formula me at e Bt = e 00 Antn = e . h = e and v = 7r (the stationary row vector). note that we can write the phase generator T as Q . if we normalize v. To this end. Ao).(3.350 APPENDIX strictly positive elements.35) for some p E (oo.e. I. it was shown that under mild conditions the tail of a phasetype distribution B is asymptotical exponential.2).(ti)ding. Example A3. Note that for an intensity matrix.8.. Corollary A4. the analogy of this procedure with unformization. then eBt = ea0thv + O(eµt) = eA0th ® v + O(et t) (A. we have A0 = 0. The content is that B is approximately exponential if the exit rates ti are small compared to the feedback intensities tij (i # j).n t AL n=0 n! (cf. T(°)) is asymptotically exponential with parameter t* _ r EiEE aiti as a 4 oo. Then for any (3. A5 Complements on phasetype distributions 5a Asymptotic exponentiality In Proposition VIII. 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). Furthermore. The next result gives a condition for asymptotical exponentiality. Bi° (x) + at*x Proof Let { 4 } be the phase process associated with B(a) and (°) its lifelength. the phasetype distribution B(a) with representation (. but is an easy consequence of the PerronFrobenius theorem.8 is most often not stated explicitly in textbooks. let {Yti°i } be a Markov process with initial distribution a and intensity .1 Let Q be a proper irreducible intensity matrix with stationary distribution a. not only in the tail but in the whole distribution. Proposition A5. the condition is that t is small compared to Q.
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}.APPENDIX 351 ((1) etc. We can assume that Jta) = Yt(°). a' = a . Hence O ((a) aa. from which it is easily checked that the limiting stationary distribution is (aiti/t*)iEE• Now let a' 4 oo with a in such a way that a' < a. Then {Ix} is a Markov process with to = Yo. We can think of ( ( a) as the first event in an inhomogeneous Poisson process ( Cox process ) with intensity process matrix aQ . a'/a + 1. Conditioning upon whether { Yt} changes state in [0. fo tY dv/t a$' t*. dx/ti] or not.Yj(av) = j f . and this easily yields a(x)/x a' 1/t*. has a limit distribution: Proposition A5. = YQ(x). J(()) _ = i) + at•x t tt' . Since JJ(. and write Yt = Yt(1).aE where 0 < e < 1). We shall . Then a(a'V)/a (aV) a' 1. {t Y( a) } v>0 . it states that the state. a .x (1 . in fact . t < (a).a' + oo (e.g. By the law of large numbers for Markov processes . In addition to the asymptotic exponentiality.2 Pi (c(a) > x.bij) Hence the intensity matrix of { Ix} is (qij/ti)i.YQ(av) = j) Pi ( ci(a'V) > x. and that Yt(a) = Yat for all t. from which the phase process is terminated . prove a somewhat more general result which was used in the proof of Proposition VI.9.jEE.(a) > x . we get dx F (Idx = j) = (1 + qij t )Sij + qij dt. v/ t. Proof Assume first ti > 0 for all i and let I.)_ = Y(a) = 1'aS(a) = Ya(av)^ it follows that Pi ((. Let further V be exponential with intensity V and independent of everything else.1. J^O)_ = j) Pi (v(aaV) > x.
.zP)'p. (c) the nth moment k 1 k"bkis 1)"n!aP"p. . = 0 for one or more i. P. Example A5.. the simplest discrete phasetype distribution: here E has only one element. so is the geometric distribution.+ at*x • a't' L ` at t* t* J Reducing the state space of {Ix } to {i E E : t.2 do not appear to be in the literature. k = 1. Then P is substochastic and the vector of exit probabilities is p = e . A distribution B on {1. u Notes and references Propositions A5.. 2.. a = b = (bk)k=1.} is said to be discrete phasetype with representation (E.4 Any discrete distribution B with finite support. (b) the generating function b[z] _ E' . Keilson [223]. > 0}. .. However.Pe..1 and A5.3 As the exponential distribution is the simplest continuous phasetype distribution. ' pk 0 k>1 11 Theorem A5.. a) if B is the lifelength of a terminating Markov chain (in discrete time) on E which has transition matrix P = (p. K}. is discrete phasetype. .5 Let B be discrete phasetype with representation (P. Indeed... See also Korolyuk. Then: (a) The point probabilities are bk = aPklp. with point probabilities bk = (1 .j) and initial distribution a.g.p)k1 p. say bk = 0. so we shall be brief. k>1.x k > K. zkbk is za(I . Gnedenko & Kovalenko [164] and Glasserman & Kou [162]).352 rr Ia(a'V) Ei I ( > x) P APPENDIX L at (Yo (aV) .. let E and Pkj j=k1. and thus the parameter p of the geometric distribution u can be identified with the exit probability vector p. Example A5. 1 k=1 1 0 otherwise.. a). Penev & Turbin [238]. Et II I a(a^V) > x) at' . these results are in the spirit of rare events theory for regenerative processes (e. 2. 5b Discrete phasetype distributions The theory of discrete phasetype distributions is a close parallel of the continuous case. an easy modification of the argument yields finally the result for the case where t.
. 11 Example A5. a' .T(2)).APPENDIX 353 5c Closure properties Example A5. Jt t > U1 + U2. { Jt 2) } with lifetimes U1 . Then {Jt} has lifetime U1 + U2 . initial distribution a and phase generator T.7 (THE NEGATIVE BINOMIAL DISTRIBUTION) The most trivial special case of Example A5. (E(2). The discrete counterpart is the negative binomial distribution with point probabilities bk k1) (1 k = r.. A. a.6 (CONVOLUTIONS) Let B1. T= ( 0 T(2) ) (A. _ i E E(1) T(1) t(1)a(2) i E E(2) .T(1)). B2 be phasetype with representations (E(1). and a=1). resp. as is seen by minor modifications of Example A5. r . Then the convolution B = B1 * B2 is phasetype with representation (E. A reduced phase diagram (omitting transitions within the two blocks) is am E(1) t(1) a(2) (2) t(2) Figure A.r + 1. and piece the processes together by it = 41) 0<t<U1 U1 < t < U1 + U2 2U. and hence the negative binomial distribution is discrete phaseu type.{ 0. .6 is the Erlang distribution Er which is the convolution of r exponential distributions.a(1).a(2).36) in blockpartitioned notation (where we could also write a as (a (1) 0)). T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2)..1 This corresponds to a convolution of r geometric distributions with the same parameter p.2 The form of these results is easily recognized if one considers two independent phase processes { Jt 1) }. U2.. resp.6.
A reduced phase diagram is 0a(1) E(1) A . this means that a = (Oa(1) (1 . Then the mixture B = 9B1 + (1 . Equivalently.E) where a(°) = fAa(a)v(da). To obtain a phase process for C. and consider B(") = fA B(a) v(da) where v is a probability measure on A. resp. B2 be phasetype with representations (E(1). are i.a(1). a reduced phase diagram is f a E t Figure A..3 In exactly the same way. P(N = n) = (1 .37) (1) (1 .O)B2 (0 < 0 < 1) is phasetype with representation (E.d. U2. if U1. we need to restart the phase process for B w. Thus. i E E(1) T 0 I (A.. one obvious interpretation of the claim u size distribution B to be a mixture is several types of claims.T(2)).p. Let B(") be the corresponding phasetype distribution. T) and C = EO°_1(1 .. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2).0)a(2))).').T. i E E(2) 0 T(2) =IT (in blockpartitioned notation. (E(2). p at each termination. with common distribution and N is independent of the Uk and geometrically distributed with parameter p.9 (INFINITE MIXTURES WITH T FIXED) Assume that a = a(°) depends on a parameter a E A whereas E and T are the same for all a. a mixture of more than two phasetype distributions is seen to be phasetype.354 APPENDIX Example A5.p)pn1B*n.0)a(2) E(2) Figure A. then C is the distribution of Ul + • • • + UN. a. Example A5.8 (FINITE MIXTURES) Let B1.10 (GEOMETRIC COMPOUNDS) Let B be phasetype with representation (E.0)ai2). Then it is trivial to see that B(") is u phasetype with representation (a(").p)pn1. In risk theory. a. Example A5.a(2). and o'i Oa.4 .T(1)).i.
say with distribution F. Example A5 .aF[T]. a(2).. +UN 2..X)+ is zeromodified phasetype with representation (E.x)+.T + pta). if {Jt} is a phase process for U. T + ta. Corollary VIII. j E F}. B2 of phasetype with representations (E('). let the initial vector be a ® v and u let the phase generator be I ® T + P ® (ta). a(1) ® a(2 ). Example A5 . Example A5.. cf. then C is the distribution of U1 + • • • + UN. If U1 has a different initial vector.. X independent of U. are i. resp.aeTx.TWWW). f2. a.2.°.v.. { 4 } as exit of {Jt}. E).T) where F[T] = J0 "o eTx F(dx) u is the matrix m.T) if U is phasetype with representation (E.9) that (U . a.g.. v. Equivalently.7.11 (OVERSHOOTS) The overshoot of U over x is defined as the distribution of (U . but the same T. Minor modifications of the argument show that 1. be the point probabilities of a discrete phasetype distribution with representation (E. it follows by mixing (Example A5. resp. then U1 + • • + UN is zeromodified phasetype with representation (a.d. if U1. 13 (MINIMA AND MAXIMA ) Let U1. To obtain a phase representation for C .2.1. To see this. if B is defective and N + 1 is the first n with U„ = oo.f.a(1). Proposition VIII. i. of F. . If we replace x by a r. with common distribution B and N is independent of the Uk with P(N = n) = f.. Indeed. (E(2). T(2) ). 12 (PHASETYPE COMPOUNDS ) Let fl. we then let the governing phase process be {Jt} _ {(411 Jt2))} 2) interpreting exit of either of {4 M }. Note that this was exactly the structure of the lifetime of a terminating renewal u process. then U1 +• is phasetype with representation (E. For U1 A U2. U2. Then the minimum U1 A U2 and the maximum U1 V U2 are again phasetype. then Jy has distribution aeTx. Thus the representation is (E(1) x E(2). T(1) ® T(2)). T + pta). T) and C = F. let the phase space be E x F = {i j : i E E.°_1 f„ B*?l. U2. { Jt2) } be independent with lifetimes U1.. let {Jtl)}. cf. a.APPENDIX 355 and C is phasetype with representation (E. It is zeromodified phasetype with representation (E. say v. v. . T). let B be a continuous phasetype distribution with representation (F. U2 be random variables with distributions B1. P).
. any distribution B on (0..B(bk) I < 1/n for n > k. the initial vector is (a(1) (& a (2) 0 0).(bk) + B(bk) for all k.(bk)'.8. That is.} of phasetype distributions such that Bn 3 B as n + oo. Proof Assume first that B is a onepoint distribution. we can assume that ID. we need to allow { Jt. Then from above. 5d Phasetype approximation A fundamental property of phasetype distributions is denseness . By the diagonal argument (subsequent thinnings). Then we must find phasetype distributions Bn with B. 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 . Thus the state space is E(1 ) x E(2) U E(1) U E( 2). elementary) Let {bk} be any dense sequence of continuity points for B(x). and vice versa.. The mean of B„ is n/Sn = b and the variance is n/Sn = b2/n. oo) can be approximated 'arbitrarily close' by a phasetype distribution B: Theorem A5. with weight pi(n) for xi(n). oo).2) } to go on (on E(2)) when { i 1) } exits. Here are the details at two somewhat different levels of abstraction: (diagonal argument . cf. Hence it is immediate that Bn 4 B.. The general case now follows easily from this. Now we can find first a sequence {Dm} of distributions with finite support such that D.356 APPENDIX For U1 V U2.. say degenerate at b. relies more on matrix algebra than the probabilistic interpretation exploited here). r # oo.n = I:pi(n)Er v ( __ ) n) ) a= 1 ..(bk) + B(bk) for all k as n * oo. and let Bn be the Erlang distribution E. there is a sequence {B. however.. Let the support of Dn be {xl(n).(n) = D.. Example A5. and the closedness of the class of phasetype distributions under the formation of finite mixtures.xq(n)(n)}. q(n) q(n) pi(n)a . i= 1 C.(Sn) with Sn = n/b. the fact that any distribution B can be approximated arbitrarily close by a distribution with finite support. see Neuts [269] (where the proof.14 To a given distribution B on (0.
for some a < oo.r. replications).. Then ICr( n ). PIT contains all finite mixtures of onepoint distributions. one would use the B given by some statistical fitting procedure (see below). But To is the class G of all distributions on [0. Since PET is closed under the continuous operation of formation of finite mixtures. Let E be the class of functions f : [0. if information on Bo is given in terms of observations (i.t.( dx) * f r f{(x)B(dx). the class CO of all discrete distributions.(x)Bf.n. oo) such that f (x) = O(e«x). say on the claim size distribution B in risk theory. the topology for weak convergence) PET of the class PET of phasetype distributions contains all onepoint distributions. oo) and any fl.. compute W(B) and use this quantity as an approximation to cp(B0). and we can take Bn = Cr(n)..APPENDIX 357 Hence we can choose r(n) in such a way that ICr( n). f2.n (bk) .D(bk)I < n. oo). k < n. It should be noted. then it is immediate that WI(B) = p2(B) for all distributions B on [0. . that this procedure should be used with care if ^p(B) is the ruin probability O(u) and u is large.. 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. x 4 oo. u 2 (abstract topological ) The essence of the argument above is that the closure (w. i. there is a sequence {Bn} of phase type distributions such that Bn Di B as n 4 oo and f ' f. Hence G C PET and L = PIT. however. For a general Bo.B(bk )I < .15 To a given distribution B on (0 . Corollary A5. E E. If Cpl (B) and ^02(B) are weakly continuous.i. In particular. oo) * [0. u Theorem A5.e. 2.n( b k ) .14 is fundamental and can motivate phasetype assumptions. and that cp is known to be continuous.. k < n. we can then approximate Bo by a phasetype B..d. oo) approximation Assume that we can compute a functional W(B) when B is phasetype. i = 1...
39) Indeed. and hence it is sufficient to show that we can obtain limsup n4oo fi(x)Bn(dx) < Jo 0 f fi( x)B(dx ).oo J fi(x)B. . i=1. n. . .358 Proof By Fatou' s lemma.f ' f (x)B(dx). By (A. i = 1...16 To a given distribution B on (0 . TO (A. if f (x ) = e°x.. \\ 0 Corollary A5. .. liminf B... . and hence we may choose r(n) such that L 9l) f (x)Cr(n).. and the case of a general f then follows from the definition of the class E and a uniform integrability argument. Bn=En z f f (x)Bn(dx) fof (x)B(dx) = ° (A.n(dx) + f 0 fi(x)Dn(dx). f00 fi(x)Cr.. for each i.39).38) We first show that for each f E E.. i = 1. i = 1.14 Dn has been chosen such that 00 1 °° f fi(x)D n(dx ) < 1++ '  o \ n o f fi(x)B(dx).n(dx) < 1+. we may assume that in the proof of Theorem A5. n B=az.f (z) = f = 1 1 1 1n/ o .f (x)B(dx). Now returning to the proof of (A. 2.38 ).. f° xtBn(dx ) * f °° x`B( dx)...(dx) > J fi(x)B(dx).. there is a sequence {Bn} of phase type distributions such that Bn Di B as n + oo and all moments converge. then cc f (x)Bn ( dx) = (?!c ) e'= .2 . n.  APPENDIX B implies that 00 o o 00 n. oo).
However.APPENDIX 359 In compound Poisson risk processes with arrival intensity /3 and claim size distribution B satisfying . the problem thus arises of how to fit a phasetype distribution B to a given set of data (1. the remaining results may be slightly stronger than those given in the literature. .. Proof Let fi(x) = el'r+E. We shall formulate the problem in the slightly broader setting of fitting a phasetype distribution B to a given set of data (1i . For practical purposes. I. 0 as i * oo. and in part from the fact that many of the algorithms that we describe below have been formulated within the setup of fitting distributions. . /3) = ry for all n./3). . there is a sequence {B.17 To a given /3 > 0 and a given distribution B on (0. lim sup ryn < 7. then Bn['Y + ei] * B[y + ei] > 1 + 7 Q implies that 'yn < ry + ei for all sufficiently large n .e. The adjustment coefficient is a fundamental quantity. from a more conceptual . . . but are certainly not unexpected. e ) and ei J.. (N. This is motivated in part from the fact that a number of nonphasetype distributions like the lognormal.18 In the setting of Corollary A5. . O We state without proof the following result: Corollary A5. . the loggamma or the Weibull have been argued to provide adequate descriptions of claim size distributions. oo) with B[y +e] < oo for some e > y = 7(B.14 is classical.> y for some sequence {ei} with ei E (0. the adjustment coefficient 'y = 7(B.16. The present section is a survey of some of the available approaches and software for inplementing this. one can obtain 7(Bn./3) is defined as the unique solution > 0 of B[y] = l+y/j3.l3µb < 1. there is substantial advantage in assuming the claim sizes to be phasetype when one wants to compute ruin probabilities.} of phasetype distributions such that Bfz + B as n * oo and Yn 4 ry where ryn = y(Bn.. (N or a given distribution Bo. lim inf > is proved similarly. and therefore the following result is highly relevant as support for phasetype assumptions in risk theory: Corollary A5. Notes and references Theorem A5.3). 5e Phasetype fitting As has been mentioned a number of times already. If ei > 0.
Schmickler (the MEDA package. [202]. we do not not want to perform matrix calculus in hundreds or thousands dimensions).g. three for a mixture of two Erlangs ). It seems therefore a key issue to develop methods allowing for a more general phase diagram.g. and in practice this sets a limitation to the usefulness (the curse of dimensionality .f. and used a nonlinear programming approach .g. . Johnson & Taaffe considered a mixture of two Erlangs (with different rates ) and matched (when possible ) the first three moments .d.. risk theory. reliability or queueing theory. 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. d. and we next describe two such approaches which also have the feature of being based upon the traditional statistical tool of like maximum likelihood. Of course.} of phasetype distribution such that Bo. and this is what matters when using phasetype distributions as computational vehicle in say renewal theory. cf. e . Asmussen & Nerman [38] implemented maximum likelihood in the full class of phasetype distributions via the EM algorithm . The characteristics of all of these methods is that even the number of parameters may be low (e. The observation is that the statistical problem would be straightforward if the whole ( EAvalued) phase process { Jtk)} o<t<( k associated with each observa . defined by the absence of loops in the phase diagram . and as fitted distribution we may take B. g.f. . In a series of papers (e.360 APPENDIX point of view the two sets of problems are hardly different : an equivalent representation of a set of data (1 .. giving mass 1 /N to each S=. 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.g . [317] ) has considered an extension of this setup. [70]) restrict attention to acyclic phase type distributions . the number of phases required for a good fit will typically be much larger.'s). The earliest such reference is Bux & Herzog [85] who assumed that the Erlang distributions have the same rate parameter. a program package written in C for the SUN workstation or the PC is available as shareware. A method developed by Bobbio and coworkers (see e. The likelihood function is maximized by a local linearization method allowing to use linear programming techniques. A number of approaches restrict the phase type distribution to a suitable class of mixtures of Erlang distributions . we have constructed a sequence { B. the L1 distance between the c . [216] ). for some suitable large n.. at a a number of selected points .. (N is the empirical distribution Be. B„ The problem is that the constructions of {B„} are not economical : the number of phases grows rapidly. one could argue that the results of the preceding section concerning phasetype approximation contains a solution to our problem : given Bo (or Be).
since this is parameterdependent. it is easy to see that N (k Ea(n).T(n) (Ti ^^ 1. . N Ti = I(J= i) dt. (N) tJk Ea ( n). In practice...g. = j) f k=1 k =1 tE[0. Nii = = .T (n)(TiI(1. jEEA.APPENDIX 361 tion Sk was available. (N ) (^ 54 k )+ and similarly for the cn+1) The crux is the computation of the conditional expectations. eieT(n)((k.g... it seems open whether the restriction to the acyclic case is a severe loss of generality... E.(k] (Ti is the total time spent in state i and Nii is the total number of jumps from i to j).. the methods of [70] and [38] appear to produce almost identical results.(N) = E Ea(n). EN where ai = N 1 I ((k) = i) tii=i iEE.T(n) (Nik IC1. .x)t(n) 1 and this and similar expressions are then computed by numerical solution of a set of differential equations.. . e. . The general idea of the EM algorithm ([106]) is to replace such unobserved quantities by the conditional expectation given the observations.. (n+1) _ Ea (n). .T(n) k=1 I (Jti) dt o \f a(n)eT(n )(kt(n) N f:i a(n)eT(n)xei . one is lead to an iterative scheme. In fact. Thus. then the estimators would be of simple occurenceexposure type.
This page is intentionally left blank .
Stoch. The Mathematical Scientist 14. Appl. Anantharam (1988 ) How large delays build up in a GI/GI11 queue. J. Asmussen (1984) Approximations for the probability of ruin within finite time . 53100.A. Aktuar Tidskr. B. [5] G. Skand. Transactions XVth International Congress of Actuaries. Appl. 363 . [10] K. New York. 311338. 20. Probab. 1985. Queueing Systems 5 . 38. New York.G. with applications to risk reserve processes and the GI /G/1 queue. Skand. II. ed. [9] G. Asmussen (1985) Conjugate processes and the simulation of ruin problems. Queueing Systems 10. Act. 587.Bibliography [1] J. 213229. 1984 . 3157 . 253267. Abramowitz & I. Whitt (1994) Waitingtime tail probabilities in queues with longtail service time distributions . Abate & W. [3] J. [2] J. [12] S. Stuttgart. Teubner. Abate & W. [4] M. Alsmeyer (1991) Erneuerungstheorie. [11) S. [7] E. pp. New York. Abate. Asmussen ( 1982 ) Conditioned limit theorems relating a random walk to its associate. Whitt (1998) Explicit M/G/1 waitingtime distributions for a class of longtail service time distributions . G. Optimizations Software. Arfwedson ( 1955) Research in collective risk theory.). Asmussen (1987) Applied Probability and Queues. John Wiley & Sons. [14] S. [6] V. 14. Queueing Systems 16. Arfwedson ( 1954) Research in collective risk theory. [15] S. 219229. Dover. Scand. Adv.L. 191223. 101116. ibid. [8] G. The case of equal risk sums. 143170. Arndt (1984) On the distribution of the supremum of a random walk on a Markov chain . Asmussen (1989a) Aspects of matrix WienerHopf factorisation in applied probability. Whitt ( 1992) The Fourierseries method for inverting transforms of probability distributions . Proc. 345368. Sparre Andersen (1957) On the collective theory of risk in the case of contagion between the claims . AT&T. 37. 57. Aktuar Tidskr. In: Limit Theorems and Related Problems (A. Borokov. Preprint.). [13] S. Stegun ( 1972 ) Handbook of Mathematical Functions (10th ed. Chichester New York. Choudhury & W.
CRC Press. Appl. Ann. Dshalalow ed. Random Processes 2. Proc. 79102. Asmussen. Probab. Asmussen & K. Floe Henriksen & C. 1989 . J. eds. Asmussen & M. 20. Bernoulli 6. Bladt (1996) Phasetype distributions and risk processes with premiums dependent on the current reserve. 37. [19] S. 313341. [32] S. Act. Asmussen (1995) Stationary distributions via first passage times. [17] S. Asmussen & B. Kliippelberg (1994) Large claims approximations for risk processes in a Markovian environment. Asmussen (1992c) Stationary distributions for fluid flow models and Markovmodulated reflected Brownian motion. Asmussen (2000) Matrixanalytic models and their analysis. Rolski & V. Asmussen. Ann. Asmussen (1992a) Phasetype representations in random walk and queueing problems. [27] S. [30] S. Asmussen (1992b) Light traffic equivalence in single server queues. 106119. Asmussen (1998a) Subexponential asymptotics for stochastic processes: extremal behaviour. Riga Aviation University. Scand. 27. 772789. Bladt (1996) Renewal theory and queueing algorithms for matrixexponential distributions.). 297318. 25. 189201. . stationary distributions and first passage probabilities. Probab. [22] S. Binswanger (1997) Simulation of ruin probabilities for subexponential claims. Asmussen. Appl.364 BIBLIOGRAPHY [16] S. Asmussen & M. eds. Asmussen (1989b) Risk theory in a Markovian environment.M. Scand. 2943. 555574. [28] S. [31] S.). Asmussen (1998b) Extreme value theory for queues via cycle maxima. K. 137168. 2. Extremes 1 . 193226. Stoch. Andronov. [34] S. Kalashnikov & A. 313326. MatrixAnalytic Methods in Stochastic Models (A. Alfa & S. 1936. SIAM Review 40. Appl. 8. 96107. Scand. Proc. [21] S. [24] S. Asmussen (1998c) A probabilistic look at the WienerHopf equation. Probabilistic Analysis of Rare Events (V. [33] S. Asmussen (1999) On the ruin problem for some adapted premium rules. 315. Appl. Advances in Queueing: Models. J. Hojgaard (1996) Finite horizon ruin probabilities for Markovmodulated risk processes with heavy tails. 69100. 2149. Asmussen & B. Th. 354374. 54. J. [29] S.S. [26] S. Probab.K. Astin Bulletin 27. [23] S. L. 1996. A. Binswanger & B. Frey. Stochastic Models 11. Methods V Problems (J. [20] S. 303322. Hojgaard (2000) Rare events simulation for heavytailed distributions. Marcel Dekker. Ann.). 4966. T. New York. Chakravarty. Schmidt (1995) Does Markovmodulation increase the risk? ASTIN Bull. Boca Raton. Asmussen (1991) Ladder heights and the Markovmodulated M/G/1 queue. 1999 . Scand. J. [25] S. Statist. Stoch. Act. Florida. [18] S. Hojgaard (1999) Approximations for finite horizon ruin prob abilities in the renewal model. Act.
Schmidli & V. Proc. Asmussen. Finance and Stochastics 4. Sigman (1996) Monotone stochastic recursions and their duals. Asmussen & R. Kliippelberg (1996) Large deviations results for subexponential tails. 419441. Insurance: Mathematics and Economics 10. Asmussen & M. Asmussen & G. 913916. 31. Asmussen & C. 19. Asmussen & C. 30. Methods 8 Problems (J. Rubinstein (1995) Steadystate rare events simulation in queueing models and its complexity properties.A. 429466. Marked point processes as limits of Markovian arrival streams. Opns.). Asmussen & S. Boca Raton. Res. Koole (1993). CRC Press. J. 259274. [40] S.Y. Adv. Asmussen & R. Prob 32 . Rubinstein (1999) Sensitivity analysis of insurance risk models. Advances in Queueing: Models. [39] S. 422447. Stochastic Models 8. 299324. H. Appl. Read eds. Probab. Statistica Neerlandica 47. Asmussen & C. Taksar (2000) Optimal risk control and dividend distribution policies . O'Cinneide (2000/01) On the tail of the waiting time in a Markovmodulated M/G/1 queue. O'Cinneide (2000/2001) Matrixexponential distributions [Distributions with a rational Laplace transform] Encyclopedia of Statistical Sciences. Nielsen (1995) Ruin probabilities via local adjustment coefficients . Appl. Fitting phasetype distributions via the EM algorithm. Rolski (1994) Risk theory in a periodic environment: Lundberg's inequality and the CramerLundberg approximation. Appl.M.BIBLIOGRAPHY 365 [35] S. .). Hojgaard & M. Taksar (1997) Controlled diffusion models for optimal dividend payout. Asmussen & T. Johnson. 421458. Perry (1992) On cycle maxima. 23. Insurance: Mathematics and Economics 20. 64.Y. Stoch. 20. Appl. Example of excessoffloss reinsurance for an insurance corporation. 736755. Probab. Wiley. [48] S. Supplementary Volume (Kotz. Proc. Asmussen & D. Math. 115. 58. [45] S. Nerman & M. J. Statist. 105119. (to appear). [36] S. [47] S. Appl. Scand. O. [44] S. Th. Eng. with applications to insurance risk Stoch. Asmussen & V. Appl. 365372. Olsson (1996). Oper. 10. [52] S. Adv. Sc. Schmidt (1995) Ladder height distributions with marks. [42] S. Asmussen . 103125. Schock Petersen (1989) Ruin probabilities expressed in terms of storage processes. Rolski (1991) Computational methods in risk theory: a matrixalgorithmic approach. Res.A. [41] S. Probab. Asmussen & H. 120. Schmidt (1999) Tail approximations for nonstandard risk and queueing processes with subexponential tails. [43] S. Dshalalow ed. Schmidt (1993) The ascending ladder height distribution for a class of dependent random walks. B. 410433. Asmussen. [51] S. [46] S. [49] S. Florida. Asmussen & T. Inf. Asmussen & V. [38] S. J. Asmussen & K. Probab. first passage problems and extreme value theory for queues. [37] S. 11251141. Management Science 45. [50] S.
1995. Wiley. [65] N.366 BIBLIOGRAPHY [53] S. [61] J. Springer . [67] T. J. de Waegenaere (1990) Simulation of ruin probabilities. J. Asmussen & J. 169186. [56] B. J. 1975.M. Beekman (1969) A ruin function approximation. Bingham. Soc. von Bahr (1974) Ruin probabilities expressed in terms of ladder height distributions. J. 832840. J. Insurance: Mathematics and Economics 6. [73] A. 9599. [59] O. Scand. Bobbio & M. J. 11811190. Scand. Teugels ( 1996) Convergence rates for M /G/1 queues and ruin problems with heavy tails . Probab. 33. New York. 1985 . C. 7. 1974. Boogaert & V. Act.H. [62] J. Appl. Act. Schmidli (1995) Saddlepoint approximations for the probability of ruin in finite time. Probab. Insurance: Mathematics and Economics 4. Baker (1977) The Numerical Solution of Integral Equations. Boogaert & A. Cox (1972) A low traffic approximation for queues. Beekman (1985) A series for infinite time ruin probabilities. Ney ( 1972) Branching Processes . Act. Actuaries 21. [70] A. Adv.B . Appl. Insurance: Mathematics and Economics 9. 129134. Bjork & J.L. New York. Telek (1994) A bencmark for PH estimation algorithms: results for acyclic PH. [71] P.J. Cambridge University Press. Berlin. . 190204. Scand. Goldie & J. Borovkov (1976) Asymptotic Methods in Queueing Theory.H. BarndorffNielsen (1978) Information and Exponential Families in Statistical Theory. 9. Bloomfield & D. Clarendon Press. Bingham (1975) Fluctuation theory in continuous time. 77111. Appl. 221232.A. Act. [57] C. 661667. Stochastic Models 10. Wiley. [63] A. 1988 . 4148. BarndorffNielsen & H. Crijns (1987) Upper bounds on ruin probabilities in case of negative loadings and positive interest rates. [60] J. Grandell (1988) Exponential inequalities for ruin probabilities in the Cox case. 148156.R. Act.A. SpringerVerlag. Scand. Plemmons (1994) Nonnegative Matrices in the Mathematical Sciences. Bjork & J. Probab. [54] K. Athreya & P. Trans. Cambridge. von Bahr (1975) Asymptotic ruin probabilities when exponential moments do not exist. New York. Scand. [72] P. [55] B. Grandell (1985) An insensitivity property of the ruin probability. J. Billingsley (1968) Convergence of Probability Measures. [68] T. 705766.L. Teugels (1987) Regular Variation. Halsted Press. Berman & R. SIAM. [69] P. 275279. Beekman (1974) Two Stochastic Processes. [58] O.H. [64] P. 610. [66] N.T. Oxford.
L. Appl. Cramer (1930) On the Mathematical Theory of Risk.J. Qinlar (1972) Markov additive processes. [75] O.J. verve. Springer. Gerber.R. [92] H. Chung (1974) A Course in Probability Theory (2nd ed. Veraverbeke (1990). Insurance : Mathematics and Economics 9. 907920.R. Bell. Cohen (1998) The M/G/1 queue with heavytailed service time distribution. 34. Amsterdam. Chandy & M. Sel. De Vylder (1986) Ordering of risks and ruin probabilities. Itasca. SpringerVerlag. [91] H. 16. Cohen (1982) The Single Server Queue (2nd ed.U. 177204. [84] D. . Geb. IEEE Trans. [80] F. Brockwell .R. Bratley. Control AC28. New York. Berlin. J.BIBLIOGRAPHY 367 [74] O. Queueing Systems 33.A. Boxma & J. 24. New York San Francisco London. Reading.L. Boxma & J. Cambr. M. Burman & D . Stockholm. [82] H.A. AddisonWesley. Tweedie (1982) Storage processes with general release rule and additive inputs . 62. [79] P. Herzog (1977) The phase concept : approximations of measured data and performance analysis.L. Broeck . Hickman. [83] D. Cottrell. Oper. Burman & D. NorthHolland.W. Z. Syst. H. Philos.) NorthHolland. S. Reiser eds. Wahrscheinlichkeitsth.W. 105119. New York.L. Adv. The Jubilee volume of Forsakringsbolaget Skandia. Cramer (1955) Collective risk theory. [93] K. Resnick & R. Cohen (1999) Heavytraffic analysis for the GI/G/1 queue with heavytailed service time distributions . 93121. Nonparametric estimators for the probability of ruin. Jr. Smith ( 1983) Asymptotic analysis of a queueing system with bursty traffic. [81] J. Fox & L.. 313319. Cox (1955) Use of complex probabilities in the theory of stochastic processes.). [77] P. Croux & N. D. 14331453. 392433. [90] D. Goovaerts & F. Tech.Y. Amsterdam. 51. 3539. II. [88] J. Computer Performance (K. Area Commun. Academic Press. Breiman ( 1968) Probability. Res. Jones & C.M. B. Illinois.Y . Insurance : Mathematics and Economics 5. Soc. [89] M. [86] K. Smith ( 1986) An asymptotic analysis of a queueing system with Markovmodulated arrivals .W. 749763.C. [76] N. Malgouyres (1983) Large deviations and rare events in the study of stochastic algorithms . Bucklew ( 1990) Large Deviation Techniques in Decision. Proc. Probab. J. [85] W. Biihlmann ( 1970) Mathematical Methods in Risk Theory. 2338. Bowers . The Society of Actuaries . Wiley. J. 14.I. 127130. Bux & U. [87] E. Aut. Skandia Jubilee Volume. Stockholm.). Schrage (1987) A Guide to Simulation. Simulation and Estimation. IEEE J. Fort & G. [78] L.C. Nesbitt (1986) Actuarial Mathematics .
Martingales and insurance risk. Res. Aktuar. [110] F. Scand.M. Math. Appl. N. 135159. De Vylder (1996) Advanced Risk Theory. Goovaerts (1988) Recursive calculation of finitetime ruin probabilities. Insurance: Mathematics and Economics 6. Res. Suppl. Stockholm. 9 .P. De Vylder (1977) A new proof of a known result in risk theory.L. Embrechts (1989). 17. [104] F. van Dawen (1986) Ein einfacher Beweis fur Ohlins Lemma. Math. [95] M. Soc. J. [106] A.. Blotter der deutschen Gesellschaft fur Versicherungsmathematik XVII. 435436. J. Daykin. 3750. [105] A. Probab. Goovaerts (1984) Bounds for classical ruin probabilities. Rubin (1977) Maximum likelihood from incomplete data via the EM algorithm. [98] A. Rolski (1991) Light traffic approximations in queues. Dassios & P. Deheuvels & J. Rolski (1984) A light traffic approximation for a single server queue. Delbaen & J. Actuarial J. Oper.D. Dempster. Editions de 1'Universite de Bruxelles. [109] F. Zeitouni (1992) Large Deviations Techniques and Applications. 7083 (1969). Comp.368 BIBLIOGRAPHY [94] M. 121131. Roy. [99] R.B. 624628. Davidson (1946) On the ruin problem of collective risk theory under the assumption of a variable safety loading (in Swedish). New York. Insurance: Mathematics and Economics 3. Steinebach ( 1991 ) On the estimation of the adjustment coefficient in risk theory via intermediate order statistics . De Vylder & M. Scand. [96] D. 227229. 201206. J. Daley & T. [111] F. Csorgo & J. Boston. Oper. Appl. . Laird & D. 22. 114119. Haezendonck (1987) Classical risk theory in an economic environment. 1990 . 27. Insurance: Mathematics and Economics 4. Pesonen (1994) Practical Risk Theory for Actuaries. Insurance: Mathematics and Economics 7. Steinebach (1990) On some alternative estimators of the adjustment coefficient in risk theory. Insurance: Mathematics and Economics 10. Chapman & Hall. Jones and Bartlett. [107] L. [103] F. Springer. Tidskr. 88101. 16. Haezendonck (1985) Inversed martingales in risk theory. Daley & T. [108] F. Csorgo & J. Teugels ( 1990) Empirical Laplace transform and approximation of compound distributions .. Delbaen & J. Statist. Stochastic Models 5. English version published in Skand. [112] F. Dembo & O. [101] C. [97] D. 181217. 3. 85116. [100] A. Forsiikringsmatematiska Studier Tillagnade Filip Lundberg. Act. Devroye (1986) Nonuniform Random Variate Generation. De Vylder & M. Pentikainen & E. Math. 583602. [102] P. J. De Vylder (1978) A practical solution to the problem of ultimate ruin probability. 5771. T.
S. 5159.C. 1975 . [120] D. O'Connell (1995) Large deviations and overflow probabilities for the general singleserver queue. Taylor ( 1975) A diffusion approximation for the ruin probability with compounding assets. 1993 .W. [119] D. Astin Bulletin 21. 1993 .B.J. with applications. 118. Gerber (1991) Risk theory for the compound Poisson process that is perturbed by a diffusion.M. Act. [123] E. Waters (1996 ) Reinsurance and ruin . [127] F. Dufresne & H. Dickson & C. [116] D. 269274. Hipp (1998) Ruin probabilities for Erlang (2) risk processes. 1984 . 1984 . Act. Act.U.M.G.C. 1741. Berlin Gottingen Heidelberg.C. J. Insurance : Mathematics and Economics 22. Insurance: Mathematics and Economics 25. [115] D. J. Math. .C. Grandell & H. Scand.J. Scand. Dickson & C.R. Camb. Scand. Proc. [122] B. 107124. Schmidli (1993) Finitetime Lundberg inequalities in the Cox case . 1. Dickson (1992) On the distribution of the surplus prior to ruin . Act. Dufresne & H.C. British Actuarial J. Emanuel . Insurance : Mathematics and Economics 11. 6180.U. 4962. Act. Scand. Dufresne. J. Act.C. Waters (1999) Ruin probabilities wih compounding. [121] D. [124] N. 174186. Letters 7. Scand. 193199. 3745.M. Gray (1984) Approximations to the ruin probability in the presence of an upper absorbing barrier. 191207. Hipp (1999) Ruin problems for phasetype(2) risk processes. Regterschot (1988) Conditional PASTA. Dickson & J. Duffield & N. Philos. J.C. Embrechts . [114] D. 147167.M. 177192. 131138. Ruin estimates for large claims . [129] D .BIBLIOGRAPHY 369 [113] D. [118] D.M.M. J. Dickson & H. Djehiche (1993) A large deviation estimate for ruin probabilities. J. 2000 . [125] F. Scand.C.C. and the amount of claim causing ruin. van Doorn & G.M. 105115. [126] F. J. Embrechts ( 1988 ). Insurance: Mathematics and Economics 7. J. Soc. Dynkin (1965) Markov Processes I. Dickson & J. Dickson & H. [130] P. 4259.U. Dickson (1994) An upper bound for the probability of ultimate ruin. Insurance: Mathematics and Economics 19. Gerber (1988) The surpluses immediately before and at ruin. Dickson (1995) A review of Panjer's recursion formula and its applications. Insurance : Mathematics and Economics 7. Res.R. Gerber & E. J. Insurance : Mathematics and Economics 10. 251262. H. [128] E. 1994 .M. [131] P. Act. Gray (1984) Exact solutions for ruin probability in the presence of an upper absorbing barrier. Scand.M.K.R. [117] D. Springer . 363374.C. 229232.R.M. Shiu (1991) Risk theory with the Gamma process. Oper. Harrison & A.
Jensen. Esscher (1932) On the probability function in the collective theory of risk.L. Approximations for compound Poisson and Polya processes. Arndt & V. Probab. 566580. Springer. Frees (1986) Nonparametric estimation of the probability of ruin. New York. 404422. [141] F. [135] P. D. Appl. Probab. Fuh (1997) Corrected ruin probabilities for ruin probabilities in Markov random walks. Wiley. Reprinted 1948 as 'The theory of probabilities and telephone conversations' in The Life and work of A. A bootstrap procedure for estimating the adjustment coefficient. 38. Insurance: Mathematics and Economics 10. [138] P. Embrechts. 35. Esary. Pitts (1993) Some applications of the fast Fourier transform in insurance mathematics . Embrechts. Math. 181190. [137] P. Embrechts & T. F. Kliippelberg & T. Griibel & S. Astin Bulletin 16. C. [149] C. Danish Academy Tech. 269274. [148] C.M. [144] P. Adv. 29. Skand. Schmidli (1994) Ruin estimation for a general insurance risk model. 14661474. Insurance: Mathematics and Economics 7. 8190. [134] P. Sci. Fitzsimmons (1987). New York. On the excursions of Markov processes in classical duality. 26.D. J. Wiley. M. 695712. Embrechts & J.D. 1998 . Konig. Embrechts & H. 17. [145] P. Mikosch (1997) Modelling Extremal Events for Finance and Insurance. Probab. first passage times and moments of ladder variables in Markov random walks.). Walkup (1967) Association of random variables. Frenz & V.370 BIBLIOGRAPHY [132] P. Proschan & D.J. Appl. . Maejima & J. Adv. Appl. U. Wiley. Appl. Act. Tidsskr. Insurance : Mathematics and Economics 1.A. Appl. [133] P. [150] H. J. Fuh & T. 616624. 5572. Embrechts & N. Villasenor (1988) Ruin estimates for large claims. Rel. Feller (1966) An Introduction to Probability Theory and its Applications I (3nd ed. R. [142] W. 3339. Probab. [140] J. Furrer (1998) Risk processes perturbed by astable Levy motion. 5974. Statistica Neerlandica 47. Feller (1971) An Introduction to Probability Theory and its Applications II (2nd ed. Nyt Tidsskrift for Matematik B20. Embrechts. Erlang (1909) Sandsynlighedsregning og telefonsamtaler .). Veraverbeke (1982) Estimates for the probability of ruin with special emphasis on the possibility of large claims .K.D.W. Heidelberg. J. [143] W. Mikosch (1991). J. Teugels (1985). Erlang 2. Adv. 175195. 131137. Scand.W.L.K. Lai (1998) Wald's equations. Ann. 159178. Th. Akt. 623637. Statist. [136] P. Schmidt (1992) An insensitivity property of ladder height distributions. Probab. 29. Schmidt (1982) Queues and Point Processes. with applications. [147] M. Probab. 5975. Franken. Fields 75. [139] A. [146] E. Trans.
Glasserman & S. A. Whitt (1994) Logarithmic asymptotics for steadystate tail probabilities in a singleserver queue. NorthHolland. Dordrecht. Probab. Gerber (1973) Martingales in risk theory. Kluwer. [159] H. Kou (1995) Limits of first passage times to rare sets in regenerative processes. 3752. [156] H.).W. de Vylder & J. 105115. NorthHolland. Insurance : Mathematics and Economics 7. Skand. 424445. Appl. Scand. eds. Appl. J. [169] J. Gerber (1988) Mathematical fun with ruin theory. [157] H. Glasserman (1991) Gradient Estimation via Perturbation Analysis. Adv.U. 1977. Basel.U. Griibel (1996) Perpetuities with thin tails. Gerber (1971) Der Einfluss von Zins auf die Ruinwahrscheinlichkeit. Kaas. In Studies in Applied Probability (J. Math. Vers. [152] H. contained in the authors PhD thesis. Ver. 463480. 205210.V. 1981 . 73. Unpublished. Kovalenko (1989) Introduction to Queueing Theory (2nd ed. Haezendonck (1984) Insurance Premiums. Springer. Birkhauser . Michna & A. Gani. [162] P. Goovaerts.S.. Gerber (1981) On the probability of ruin in the presence of a linear dividend barrier.G. London. Schweiz. [164] B. Amsterdam. Aktuarietidskrift 1970 . Schweiz. Insurance: Mathematics and Economics 20. 6370. Mitt. [168] J. Appl. [161] P. [166] M. Insurance: Mathematics and Economics 15. Gerber (1970) An extension of the renewal equation and its application in the collective theory of risk. J. Weron (1997) Stable Levy motion approximation in collective risk theory. Act.E.U.M. Boston. Gnedenko & I. J. [154] H. [153] H. Galambos & J.N. Vers.71.. [158] H. Act. Schmidli (1994) Exponential inequalities for ruin probabilities of risk processes perturbed by diffusion. Gerber (1986). 5. 7778.BIBLIOGRAPHY 371 [151] H. Life Insurance Mathematics. ETH Zurich. Ann.U. [155] H. Gerber (1979) An Introduction to Mathematical Risk Theory. Huebner Foundation Monographs. [163] P. Amsterdam.U.J.U. van Heerwarden & T. Probab. .U. 1523. 131156. Suppl. Scand. 31A. Grandell (1978) A remark on 'A class of approximations of ruin probabilities'. Furrer & H. [165] M. [167] C.J. Probab. Bauwelinckx (1990) Effective Actuarial Methods. R. Goldie & R. 28. J. University of Pennsylvania. [160] H. Furrer (1996) A note on the convergence of the infinitetime ruin probabilities when the weak approximation is astable Levy motion. Mitt. Furrer. 205216. Z. Act. Ver. S. F. Glynn & W. 2336.). 97114. Scand. Goovaerts. Math. Grandell (1977) A class of approximations of ruin probabilities. 1978 .
355365. Probab. Grigelionis (1996) Lundbergtype stochastic processes. 5766. Grandell (1997) Mixed Poisson Processes. 143158. [177] B. [175] J. Korolyuk (1969) On the joint distribution of a process with stationary increments and its maximum .243255. [184] H . Opns .V. M. Winter School on Stochastic Analysis and Appl .and Sozialforschung 6.I. [189] A. Appl. [173] J. 54. Liet. Informatica 2. Appl. 3 . Opns .. Matem. Silvestrov (2000) Cram€rLundberg approximations for nonlinearly perturbed risk processes . Lemoine (1977) Limit theorems for periodic queues. [172] J. Griibel & R. [178] B. 5 . 566576. Proc. Preprint. Grandell & C. Resnick (1977) The recurrence classification of risk and storage processes . M. [174] J. Hadwiger (1940) Uber die Wahrscheinlichkeit des Ruins bei einer grossen Zahl von Geschiiften . [186] J. mathematische Wirtschaft . 347358. Grandell (1979) Empirical bounds for ruin probabilities.). Resnick (1976 ) The stationary distribution and first exit probabilities of a storage process with general release rule . . Astin Bulletin 29. Stoch. Grandell (1990) Aspects of Risk Theory. 6779. [171] J. SpringerVerlag. M.J. 33. 14. Gyllenberg & D. [179] R. Grigelionis ( 1993) Twosided Lundberg inequalities in a Markovian environment. [185] J. 400409. Grandell ( 1992) Finite time ruin probabilities and martingales . E. [183] M. 3041. M. Archiv v. Statistical Algorithm 265.A. Res. Gut (1988 ) Stopped Random Walks . Appl. [188] J. Hermesmeier (1999) Computation of compound distributions I: aliasing errors and exponential tilting. 14. Aktuar Tidskr.O. ed. 197214. Amsterdam. Grandell (1999) Simple approximations of ruin functions. Chapman & Hall. [181] D. Theory and Applications. Tinbergen Institute Research Series 20. Grnbel (1991) G/G/1 via FFT. 8.S. [182] A. KTH.372 BIBLIOGRAPHY [170] J. Appl. Th. 1. Berlin. Ibragimov. Harrison & S. New York. [187] J. Math. Applied Statistics 40. [176] B. Insurance: Mathematics and Economics 26. SpringerVerlag. Probability Theory and Mathematical Statistics (I. J. van Heerwarden (1991) Ordering of Risks: Theory and Actuarial Applications. 167176. Segerdahl (1971) A comparison of some approximations of ruin probabilities. Rink. Probab. Proc . AkademieVerlag. Proc. [180] R. Math. 332. Grigelionis ( 1992) On Lundberg inequalities in a Markovian environment. Gusak & V.I.131135. Harrison (1977) Ruin problems with compounding assets . 7590. Res. Harrison & A. Stoch. Harrison & S. Skand.
Iversen & L. Proc. 29. 6. 8996. Asmussen & 0. Probab. Probab. [207] D. [204] T. 285293. Hoglund (1974) Central limit theorems and statistical inference for Markov chains. Hipp (1989b) Estimators and bootstrap confidence intervals for ruin probabilities. 8095. Scand. [194] 0. Iglehart (1969) Diffusion approximations in collective risk theory. The ruin problem for finite Markov chains.. Appl. Nachrichten 30. Ann. Mathematical Statistics. [205] T. Z. Karlsruhe. Hogan (1986) Comment on corrected diffusion approximations in certain random walk problems. eds. Statist. Wiley. J. 12981310. Staalhagen (1999) Waiting time distributions in M/D/1 queueing systems. Herkenrath (1986) On the estimation of the adjustment coefficient in risk theory by means of stochastic approximation procedures. Astin Bulletin 19. Karlsruhe.a program for fitting phasetype distributions.M. J. [208] V.L. Electronic Letters 35.V. Hogg & S. Versicherungswirtschaft. Foruth Prague Symp. 377381. 1990 .B. Huskova. Scand. [200] M. No. Hesselager (1990) Some results on optimal reinsurance in terms of the adjustment coefficient. Insurance: Mathematics and Economics 5.BIBLIOGRAPHY 373 [190] P. 11631174. J. Ann. 4385.V. Heidelberger (1995) Fast simulation of rare events in queueing and reliability models.L. Klugman (1984) Loss Distributions. Appl. 3. Studies in Statistical Quality Control and Reliability 1992 :4. 5770. [203] T. [197] C.A. [193] U. Math. 25. ACM TOMACS 6. 18. S. Haggstrom. [191] W. Taksar (2000) Optimal proportional reinsurance policies for diffusion models. Hipp (1989a) Efficient estimators for ruin probabilities. [201] L. [198] C. Hoglund (1991). 285292. 19. Hojgaard & M. Probab. Hermann (1965) Bin Approximationssatz fiir Verteilungen stationarer zufalliger Punktfolgen. [192] U. Mandl & M. Michel (1990) Risikotheorie: Stochastische Modelle and Statistische Methoden. Nerman (1992) EMPHT . 259268. Heilmann (1987) Grundbegriffe der Risikotheorie. 378389. J. Horvath & E. Geb. 23. New York. . Ann. [206] B. Hill (1975) A simple general approach to inference about the tail of a distribution. verve. [202] 0. 1998 . Willekens (1986) Estimates for the probability of ruin starting with a large initial reserve. Chalmers University of Technology and the University of Goteborg. Insurance: Mathematics and Economics 5. Prob. on Asymptotic Statistics (P.). [195] B. Wahrscheinlichkeitsth. [199] R. Hipp & R. [196] C. Hoglund (1990) An asymptotic expression for the probability of ruin within finite time.R. 123151. Act. Verlag Versicherungswirtschaft e. 305313. Act. 166180.
Oxford.M. [222] S. Taylor (1981) A Second Course in Stochastic Processes. 259281. Reinhard (1985) Probabilities de ruine pour une classe de modeles de risque semiMarkoviens. 37. [216] M.M. 15. 325247. [217] R. Chicago. Jelenkovic & A. Jensen (1953) A Distribution Model Applicable to Economics.J. 61.374 BIBLIOGRAPHY [209] D. Kluwer. J. [219] V. Jagerman (1991) Analytical and numerical solution of Volterra integral equations with applications to queues. . [211] J. Johnson & M.G. [218] V. 8793. Keilson (1966) A limit theorem for passage times in ergodic regenerative processes. Cambridge Philos.A. Lazar (1998) Subexponential asymptotics of a Markovmodulated random walk with a queueing application. Soc. J. Ann. [224] J. Cambridge Philos.L. Keilson & D. 6. 3.J. 6. [225] J. J. 547567. Stochastic Models 1. [220] V. Taaffe (1989/90) Matching moments to phase distributions. 283306. Manuscript. Statist. Karlin & H.B. 35. [227] J. Wishart (1964) Boundary problems for additive processes defined on a finite Markov chain. 60. 31. Academic Press. Wishart (1964). Astin Bull. 63. Soc.C. J.G. Keilson & D. New York. Jensen (1995) Saddle Point Approximations. [215] J. 173190. [214] A. Jagerman (1985) Certain Volterra integral equations arising in queueing. N. Kennedy (1994) Understanding the WienerHopf factorization for the simple random walk. Kalashnikov (1996) Twosided bounds of ruin probabilities. 11. 187193. Soc. 711743. ibid. [226] J. 116129. Kalashnikov (1999) Bounds for ruin probabilities in the presence of large claims and their comparison. Act. Math. Wishart (1964) A central limit theorem for processes defined on a finite Markov chain. Janssen & J. Kao (1988) Computing the phasetype renewal and related functions. Addenda to for processes defined on a finite Markov chain. [213] P. [223] J. N. [212] J. Astin Bull. Cambridge Philos. Goovaerts (1986) General bound on ruin probabilities. Appl. [210] D.P. Kaas & M.M. Kemperman (1961) The Passage Problem for a Markov Chain.M. [228] J. 118. 165167. Kalashnikov (1997) Geometric Sums: Bounds for Rare Event with Applications.H. Scand. 1996. Stochastic Models 5. Princeton. [221] E. Proc. Proc.R. Keilson & D. 239256. Probab.M. Clarendon Press. Insurance: Mathematics and Economics 5. 4151. ibid. University of Chicago Press.G. Appl. NEC Research Institute. Amer. 123133. 866870. Janssen (1980) Some transient results on the M/SM/1 special semiMarkov model in risk and queueing theories. Act. Proc. Copenhagen. 561563. Technometrics 30. Probab. Munksgaard.
Appl. Kliippelberg & U. [231] J. Kuchler & M. J. 60 . 259264. . Soc. Turbin (1973) An asymptotic expansion for the absorption time of a Markov chain distribution. Mikosch (1995) Explosive Poisson shot noise with applications to risk retention. Klnppelberg & T. Insurance: Mathematics and Economics 8. 154168. [247] L. Stadtmiiller (1998) Ruin probabilities in the presence of heavytails and interest rates . J. Nyrhinen (1992a) Simulating levelcrossing probabilities by importance sampling. Probab. Latouche & V. B24. Proc. Scand. [233] C. 6075.F. Penev & A. Klnppelberg & T. Quart.J. 24. [246] D. J. 390397. 132141. Soc. 146. I. SIAM. J. Lipsky (1992) Queueing Theory . Lemoine (1981) On queues with periodic Poisson input. [230] J.. Adv. 383392. J. Proc. Stochastic Models 7. [236] C.F. Bernoulli 1 . Seminaire de Probabilties X . Lecture Notes in Mathematics 511. [238] V. 359361. Philos. J.S. (235] C. [242] A. 277289. 4958. [241] G. Kingman (1962) On queues in heavy traffic. Scand.F. 26. Klnppelberg (1989) Estimation of ruin probabilities by means of hazard rates. Probab. Korolyuk. 133135 (in Russian). Philos. [248] D. Quart. [243] A.BIBLIOGRAPHY 375 [229] J. Appl. Cambr. Act.J. Appl. Oxford 12. New York. Cybernetica 4. 858874. Probab. Macmillan. Math. Roy. Lehtonen & H. Mikosch (1995) Modelling delay in claim settlement. SpringerVerlag. 25. 48. Sorensen (1997) Exponential Families of Stochastic Processes. SpringerVerlag. [239] H. 1998 . [237] C. Ramaswami (1999) Introduction to MatrixAnalytic Methods in Stochastic Modelling. 18. 283284. 4477. [240] U. Lemoine (1989) Waiting time and workload in queues with periodic Poisson input. Scand.P. Lucantoni (1991) New results on the single server queue with a batch Markovian arrival process. J. Lehtonen & H. Actuarial J. Insurance : Mathematics and Economics 12. 279285. Soc. [245] T. Camb. [234] C.F. 1995 . [244] T. 125147.C. Act. Kunita (1976) Absolute continuity of Markov processes. Appl. Statist.C. Lindley (1952) The theory of queues with a single server. 889900. Kingman (1961) A convexity property of positive matrices . [232] C. Klnppelberg (1993) Asymptotic ordering of risks and ruin probabilities. Prob.C. Kliippelberg (1988) Subexponential distributions and integrated tails.a Linear Algebraic Approach. Nyrhinen (1992b) On asymptotically efficient simulation of ruin probabilities in a Markovian environment. Kingman (1964) A martingale inequality in the theory of queues.
J. MartinL6f (1983) Entropy estimates for ruin probabilities . 676705. 36. K. Englunds Boktryckeri AB. Probability and Mathematical Statistics (A. Act. Lucantoni . [258] K .S. Miller (1962 ) A matrix factorization problem in the theory of random variables defined on a finite Markov chain . [260] H. [261] H. J. J. Probab. Act. Appl.D. 22. Anal. [253] V.S. 1261 (0?)1270. Mammitzsch ( 1986 ) A note on the adjustment coefficient in ruin theory.376 BIBLIOGRAPHY [249] D. A fitting algorithm for Markovmodulated Poisson processes having two arrival rates . Ann. II Aterforsdkring av Kollektivrisker . Lundberg (1903) I Approximerad Framstdllning av Sannolikhetsfunktionen. Moustakides ( 1999) Extension of Wald 's first lemma to Markov processes. Stoch.F. 763776. [263] M . [255] V. Eigenvalue properties and limit theorems . 429448. 11 .M. 58 . Lundberg (1926) Forsdkringsteknisk Riskutjdmning. Act. Michna ( 1998) Selfsimilar processes in collective risk theory. 31. Ann. Scand. Hoist eds. Appl. 370377. 3. [251] F. Gut & J. Uppsala. Appl. 124147. F. [262] H . MeierHellstern & M . Ann. Miyazawa & V. Opns. Malinovskii ( 1994) Corrected normal approximation for the probability of ruin within finite time. 834841. Res. Probab . Scand. Ney & E. Soc. MartinL3f (1986) Entropy. 286298. Appl. Appl.). . with an application . Probab. Neuts ( 1990) A single server queue with server vacations and a class of nonrenewal arrival processes . 1986 . Stockholm. Malinovskii ( 1996) Approximation and upper bounds on probabilities of large deviations of ruin within finite time. Statist. 223235. [264] G. Miller ( 1961 ) A convexity property in the theory of random variables defined on a finite Markov chain. 1994. Appl. [259] Z . 161174. 147149. 2939. 15. [250] F.D. Schmidt ( 1993) On ladder height distributions of general risk processes . Nagaev (1957) Some limit theorems for stationary Markov chains. 58. Insurance : Mathematics and Economics 5. Almqvist & Wiksell. a useful concept in risk theory. 561592. Meier ( 1984). Miller ( 1962 ) Absorption probabilities for sums of random variables defined on a finite Markov chain . Europ. [254] V . Scand. [252] A. 29 . 4859. J. Cambridge Philos. J. Proc. 378406. Math.K. Math. Nummelin (1987) Markov additive processes I. Soc. Probab. [257] A. 32. Adv. 2. 1996 .K. J.V. Makowski ( 1994) On an elementary characterization of the increasing convex order . [265] S. [266] P.V. Th. 268285. . J. Cambridge Philos. Probab.D. Probab. Proc. [256] A.
36. Probab. [277] J. 75. 12551265. Paulsen (1993) Risk theory in a stochastic economic environment. 555564. Nyrhinen ( 1998 ) Rough descriptions of ruin for a general class of surplus processes. Omey & E. 316.F. 12. Probab. New York. Gjessing (1997b) Present value distributions with applications to ruin theory and stochastic equations. Probab. Naval Research Logistics Quarterly 25. Appl. [276] C. [270] M . Adv. J. [280] A. Proc. Proc. 311342. Appl. Appl. . Math. [268] M . Proc. O'Cinneide (1990) Characterization of phasetype distributions. Gjessing (1997a) Optimal choice of dividend barriers for a risk process with stochastic return on investments. Nyrhinen (1999) Large deviations for the time of ruin. 733746. 21. Insurance: Mathematics and Economics 20. Stochastic Models 3 . Neuts (1977) A versatile Markovian point process. Johns Hopkins University Press. [273] R. Proc. [283] J. 339353. [275] H.G. Appl. Neuts ( 1989) Structured Stochastic Matrices of the M/G/1 Type and their Applications. Stoch. Stoch. [272] J . London. J. Willekens (1987) Second order behaviour of distributions subordinate to a distribution with finite mean . Models 6. E75B . Omey & E. [278] E. Norberg ( 1990) Risk theory and its statistics environment . Willekens (1986) Second order behaviour of the tail of a subordinated probability distributions. Probab. Appl. 71. Statistics 21. 764779. [274] H . [284] J. [282] J . Paulsen & H. Stoch. Neuts (1992) Models based on the Markovian arrival process. Ohlin (1969) On a class of measures for dispersion with application to optimal insurance. Marcel Dekker. 135148.F. [285] J. Abh. Insurance: Mathematics and Economics 22. 10081026. Paulsen & H. 215223. Commun. [269] M . [279] E.K. J. Paulsen (1998) Ruin theory with compounding assets .BIBLIOGRAPHY 377 [267] M . Astin Bulletin 5 .F. 157. 46. 249266.a survey. Paulsen ( 1998) Sharp conditions for certain ruin in a risk process with stochastic return on investments. Neuts ( 1978) Renewal processes of phase type. Stoch. Baltimore . Appl. Hamburg.A. 16. 30. IEICE Trans. [271] M.K. 327361. 123144. Stoch. Appl. Neveu ( 1961 ) Une generalisation des processus a accroisances independantes. Pakes (1975) On the tail of waiting time distributions. [281] J. 445454.F. Sem. Appl. Neuts ( 1981 ) MatrixGeometric Solutions in Stochastic Models. 273299.F.
Statist. Ann. Berlin. Probab.U. [297] R. Res.M. 30. [293] N. 1989 .G.U. 568576. Inst. Prabhu & Zhu (1989) Markovmodulated queueing systems. 215246. New York. [291] S. On the ruin problem of collective risk theory. Gjessing (1997c) Ruin theory with stochastic return on investments. Appl. 46. Rao (1965 ) Linear Statistical Inference and Its Applications. Pellerey (1995) On the preservation of some orderings of risks under convolution. Prabhu (1961).U. [292] S. . Queueing Systems 5. 29A. Adv. [304] B.K. J. Opns.J. Probab. Stat. Heidelberg. Math. Queues. [296] N. 820827. Math. Griibel & P. J. Wiley. Springer. 139143. Math.R. 117133. Appl. Insurance: Mathematics and Economics 16. Appl. Austr. 2330. 147159. [298] V. Pyke (1959) The supremum and infimum of the Poisson process. 691707. Astin Bulletin XIV. [288] S. Oper. Philipson (1968) A review of the collective theory of risk. Probab . [295] N. 537555. Reinhard (1984) On a class of semiMarkov risk models obtained as classical risk models in a markovian environment. Appl. de Smit (1986) The queue M/G/1 with Markovmodulated arrivals and services. Scand. Rogers (1994) Fluid models in queueing theory and WienerHopf factorisation of Markov chains. Samorodnitsky (1997) Performance degradation in a single server exponential queueing model with long range dependence.H. [306] T. 32.. 965985. [305] C. Aktuar. 28. 19. [300] C. 337347. 4. Ripley (1987) Stochastic Simulation . [299] C. Insurance: Mathematics and Economics 3.J. Paulsen & H. Res. J. [301] G.L. Math. Regterschot & J. Tidskr. 757764. Resnick & G. 2343.A. [303] S. Adv.M. Ramaswami (1980) The N/G/1 queue and its detailed analysis. [302] J. Skand. Probab. Rolski (1987) Approximation of periodic queues.U.K. Pitts (1994) Nonparametric estimation of compound distributions with applications in insurance. Act. Prabhu (1980) Stochastic Storage Processes. Prabhu (1965) Queues and Inventories. Wiley. 12. R. Ann. 29. Wiley. Insurance Risk. 4568. Schock Petersen (1989) Calculation of ruin probabilities when the premium depends on the current reserve. Adv. Pitts. [287] F.G. Appl. [289] C. Ramsay (1984) The asymptotic ruin problem when the healthy and sick periods form an alternating renewal process. 61.378 BIBLIOGRAPHY [286] J. New York. Probab. 390413. [290] E. 11. Statist.M. Ann. and Dams. 465483. Embrechts (1996) Confidence bounds for the adjustment coefficient. Pitman (1980) Subexponential distribution functions. 235243. Ann. 222261. Soc. [294] N. Math.
Proc. Schmidli ( 1997a) Estimation of the Lundberg coefficient for a Markov modulated risk model . [315] T. [312] R. 135149. Math. [313] M. J. [319] H. Schmickler ( 1992 ) MEDA : mixed Erlang distributions as phasetype representations of empirical distribution functions . Appl. [321] H. Wiley. Stoch. Act. Insurance: Mathematics and Economics 16. Singapore. Rubinstein & A. [318] H. 5. 4068. 4857. 7. [311] R. 2001 . Probab. 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. 131156. Schmidt & J. Comp. Schmidli ( 1994) Diffusion approximations for a risk process with the possibility of borrowing and interest. Science Culture Technology Publishing . J. [323] H. 1997. H. 365388. [325] H . [309] H. Rubinstein & B. Ann. 155188. Data Anal.M. Act. Ross (1974) Bounds on the delay distribution in GI/G/1 queues. Wiley. Statist. Insurance: Mathematics and Economics 22. Shapiro (1993) Discrete Event Systems: Sensitivity Analysis and Stochastic Optimization via the Score Function Method. Schmidli ( 1996) Martingales and insurance risk. Scand. Statist. 417421. 11. Ryden (1994) Parameter estimation for Markov modulated Poisson processes. Appl. Rudemo (1973) Point processes generated by transitions of a Markov chain. Stochastic Models 10 . Schmidli.J. 5. Seal ( 1969) The Stochastic Theory of a Risk Business. Ryden (1996) An EM algorithm for estimation in Markovmodulated Poisson processes. [320] H. [326] H. [317] L. Probab. Rolski. 262286. Appi. Lecture Notes of the 8th Summer School on Probability and Mathematical Statistics ( Varna). Wiley. Stochastic Models 6. Wiley.Y. Schmidli ( 1999a) On the distribution of the surplus prior and at ruin. [310] R. 431447 [316] S. [308] S.BIBLIOGRAPHY 379 [307] T. 795829. Schmidli ( 1997b) An extension to the renewal theorem and an application in risk theory. Stochastic Models 10. Th. [322] H.Y. 227244. Schmidli ( 2001 ) Optimal proportional reinsurance policies in a dynamic setting . 21. V. 145165. Schlegel ( 1998) Ruin probabilities in perturbed risk models. . 121133. 93104. Adv. Teugels (1999) Stochastic Processes for Insurance and Finance. Astin Bulletin 29.L. Rubinstein (1981) Simulation and the Monte Carlo Method. Scand. Schmidli ( 1995 ) CramerLundberg approximations for ruin probabilities of risk processes perturbed by a diffusion . Melamed (1998) Modern Simulation and Modelling. Operationsforsch. [324] H . J.Y. Wiley. [314] T. 687699. Probab . 5. Schmidli ( 1999b) Perturbed risk processes: a review .
Versich. 61. J. Wiley. L. 2236. Verein Schweiz. [346) D . Math. [348] K . Math. Probab . 4383. Sengupta ( 1989) Markov processes whose steadystate distribution is matrixgeometric with an application to the GI/PH/1 queue. Siegmund (1979) Corrected diffusion approximations in certain random walk problems. 429442.W. Appl. Verein Schweiz Versich. 243249. 1987. [343] D. 77100.L. Shin (1987) Convolution of uniform distributions and ruin probability. Mitt. Stockholm. . [342] D. J. Probab. Shaked & J. t). Probab. Risk teory and the single server queue . Aktuar Tidsskr. Appl. 701719.L. Chapman and Hall. 483487. [345] D.S. Seal ( 1972). [329] H. Academic Press. 11. Shwartz & A. 121139. Versich. 914924. [347] K. 75. [340] E. Sengupta (1990) The semiMarkov queue: theory and applications. [331] G. 4. Skand. Sigman (1992) Light traffic for workload in queues. Appl. Th. Seal (1978) Survival Probabilities. Siegmund (1975) The time until ruin in collective risk theory. Act. Act. Shtatland ( 1966) On the distribution of the maximum of a process with independent increments. Scand. Siegmund ( 1985) Sequential Analysis . 191197. [335] B . 1955. [332] C : O. Weiss (1995) Large Deviations for Performance Analysis. [336] B. Seber (1984) Multivariate Observations. [341] E. 72.S. 279299. In Probability and statistics .380 BIBLIOGRAPHY [327] H. [344] D. t). Adv. Siegmund (1976) Importance sampling in the Monte Carlo study of sequential tests. Probab.A. [333] C: O. Wiley. 11. 171178. Ann. Mitt. Grenander). Siegmund (1976) The equivalence of absorbing and reflecting barrier problems for stochastically monotone Markov processes . Chapman & Hall. . 383413. Ann. [330] H. Seal (1972) Numerical calculcation of the probability of ruin in the Poisson/Exponential case. Segerdahl (1955) When does ruin occur in the collective theory of risk? Skand. New York. Shin (1989) Ruin probability by operational calculus.G. Seal (1974) The numerical calculation of U(w. Stochastic Models 6. 4. SpringerVerlag. [337] M. [338] E. Scand.the Harald Cramer volume (ed. 21. pp. 1974. [334] C: O. Shantikumar (1993) Stochastic Orders and Their Applications. Segerdahl ( 1942) Uber einige Risikotheoretische Fagestellungen . Verein Schweiz.F. Adv. [339] A. Insurance: Mathematics and Economics 8. [328] H. Sigman ( 1994 ) Stationary Marked Point Processes: An Intuitive Approach. Math. Almqvist & Wiksell. Segerdahl (1959) A survey of results in the collective theory of risk.L. Statist. 72. Queueing Systems 11. Aktuar Tidsskr. 673684. Mitteil. 157166. 159180. the probability of nonruin in an interval (0.W.S.
Straub (1988) NonLife Insurance Mathematics. 5776. Smith ( 1953) Distribution of queueing times. [368] O. Insurance: Mathematics and Economics 16. Proc.C. Taylor (1979) Probability of ruin with variable premium rate. 65102. 449461. [351] D .C. Thorin (1974) On the asymptotic behaviour of the ruin probability when the epochs of claims form a renewal process. Taylor (1979) Probability of ruin under inflationary conditions or under experience ratings . Ruin probabilities prepared for numerical calculations. Thorin (1986) Ruin probabilities when the claim amounts are gamma distributed.J. [353] E. Taylor (1978) Representation and explicit calculation of finitetime ruin probabilities. Act. Soc 49. [357] B. 8594.S.C. [354] B. Actuarial J. 118. NorthHolland. J. Teugels ( 1995 ) Ruin estimates under interest force . Sundt & W. Cambridge Philos.. [361] G. 21. [365] J. Daley ed. J. Scand. Adv. [356] B . . 235254. J. Teugels (1982) Estimation of ruin probabilities .and largedeviations probabilities in actuarial risk theory.J. [367] O.L. Insurance: Mathematics and Economics 1.L. Probab. 65102. Act. Teugels (1997) The adjustment coefficient in ruin estimates under interest force . 2739. Taylor (1976 ) Use of differential and integral inequalities to bound ruin and queueing probabilities . Scand. 8199.A. Sundt & J. Unpublished manuscript. [359] L. 1974. Verlag Versicherungswirtschaft e. Scand. Slud & C. 1982 . 1980 . Act. 1976.BIBLIOGRAPHY 381 [349] E. Takhcs (1967) Combinatorial Methods in the Theory of Stochastic Processes. Stoyan (1983) Comparison Methods for Queues and Other Stochastic Models (D.L. Taylor (1986) Claims reserving in NonLife Insurance. 725741. [360] G.C. Suri ( 1989) Perturbation analysis : the state of the art and research issues explained via the GI/G/1 queue. John Wiley & Sons.V. Karlsruhe. New York. [358] R.. J. Scand.). Hoesman ( 1989) Moderate. J. 1978. 114137.L.C. New York. [369] O. Astin Bulletin 16 . Thorin (1982) Probabilities of ruin. [352] D. Act. [355] B . Wiley. [350] W. Stanford & K. Thorin (1977). Stroinski ( 1994) Recursive method for computing finitetime ruin probabilities for phasedistributed claim sizes. Scand. Scand. [366] O. Proceeedings of the IEEE 77. 722. Springer.). 149162. [363] G. [362] G. [364] G. 197208. Sundt & J. Insurance: Mathematics and Economics 19. Astin Bulletin 12. Astin Bulletin 24. 163175. Appl. Sundt ( 1993) An Introduction to NonLife Insurance Mathematics (3rd ed. Act. Jewell ( 1981 ) Further results on recursive evaluation of compound distributions .
Thorisson (2000) Coupling. [371] 0. PrenticeHall. Res.R. [376] A. VazquezAbad (1999) RPA pathwise derivatives of ruin probabilities. 231246. 142. 936952. 148164. 31. Insurance: Mathematics and Economics 15. D. J. Submitted. Reidel. In: Premium Calculation in Insurance (F. Willmot & X. [372] H. Whitt (1989) An interpolation approximation for the mean workload in a GI/G/1 queue. Aktuar. Wallace (1969) The solution of quasi birth and death processes arising from multiple access computer systems. Dordrecht Boston Lancaster. Thorin & N. Astin Bulletin IX. 1983 . Wikstad (1977) Numerical evaluation of ruin probabilities for a finite period. 37. 137153. [375] N.W. Tacklind (1942) Sur le risque dans les jeux inequitables. Wolff (1990) Stochastic Modeling and the Theory of Queues. Insurance : Mathematics and Economics 13. Thorin & N. Wald (1947) Sequential Analysis. SpringerVerlag. Taqqu.382 BIBLIOGRAPHY [370] 0. [381] W. Waters (1983) Probability of ruin for a risk process with claims cost inflation. [383] G. [379] M. 4963. 743756. Skand. Probab. J. M. F. Wiley. Act. Oper. Scand. 6785. Tidskr.). [384] R. [373] S. Van Wouve. . Haezendonck eds. [374] F. [382] G. Willinger.E. Wikstad (1977) Calculation of ruin probabilities when the claim distribution is lognormal. R. Appl. Statistical Science 10. [380] W. De Vylder. 1942. 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. [377] V. Stationarity and Regeneration. thesis. Unpublished Ph. [378] H. University of Michigan. Veraverbeke (1993) Asymptotic estimates for the probability of ruin in a Poisson model with diffusion. Astin Bulletin VII. J. De Vylder & M. Lin (1994) Lundberg bounds on the tails of compound distributions. Goovaerts. 5762. M. Sherman & D. Willmot (1994) Refinements and distributional generalizations of Lundberg's inequality.E.
228229. 91.6779. 170173. 308. 141144. 3839. 1819. 40. 117127 corrected 121127 duality 1314.182. 17. 4851. 301 Kronecker product. 138139. 97. 14.272. 318319 Erlang distribution 7.203.200201.293294. 7879. 227229. 1415. 39. 239.100.281.137141. 3436. 1112. 278 gamma distribution 67.285292. 86.318320 change of measure 2630. 207 heavytailed distribution 6. 341. 5796. 2526. 7179.308 CramerLundberg model: see compound Poisson model cumulative process 334 dams: see storage process differential equation 16. 122.242. 97129.121129. 217. 316323 Bessel function 102. 5. 360 excursion 155156.314316.269. 332333 Volterra 192194. 9293.307312 compound Poisson model 4.249.217. 80 81. 89. 79.346349 383 . 135. 7079. 110113. 361 diffusion 3. 302303 diffusion approximation 17.178184.259261. 111117. 15.150.160167.203. 9396.299.4447. 189.251280 heavy traffic 76. 3334. 245248.249250 integral equation 16 Lindley 143 renewal 64. 162164. 201 Brownian motion 3 .226.and sum 221. 9496. 218 Cox process 4. 283. 7475.9899.292293 Edgeworth expansion 113. 271274. 201214.359 aggregate claims 103106. 226. 248 WienerHopf 144 interest rate 190. 196201 inverse Gaussian distribution 76. 3032.185187. 205. 323 Coxian distribution 147.86. 37.328330. 17. 117128.Index adjustment coefficient 17. 180182. 119. 8283 hyperexponential distribution 7.287292. 2425. 12 CramerLundberg approximation 1617.135.301 central limit theorem 60 .
42.215250. 267269 Panjer's recursion 320323 Pareto distribution 910. 3947. 176185. 59.227230.139141. 39.160161.384 ladder heights 4756. 7179. 175 light traffic 8183 Lindley integral equation 143 process 3334. nonlinear 155. 80.336339 . 14. 138. 157.240244. 133.161. 213214. 295. 260 Lundberg conjugation 6979 . 35. 134.234240. 25. 271274. 142 likelihood ratio : see change of measure lognormal distribution 9.336339 Laplace transform 15. 154. 144.302. 162. 261264.287291 INDEX matrix equation . 251. 99. 100. 134135.285287 queue 14 . 145187. 203 Markov additive process 12.218221. 35.304 process 2830. see also sensitivity analysis phasetype distribution 8. 7576. 86 periodicity 12. 3639. 16. 6970. 203204.161. 32. 137139.161164.339 large deviations 129. 229 M/M/1 101 Markovmodulated 185187 periodic 187 martingale 2426. 65.180. 9899.350361 Poisson process Markovmodulated 12 periodic 12.146148.269271.152160.275278. 257.108. 106108.134135.234.349 350 perturbation 172173. 37. 149.348 terminating 215216. 44. 44.201.128129.238. 15. 304305 random walk 3336. 5758. 113114.315 inequality 1718.261264.259261. 3947. 96. 16. 6162. 185187 GI/G/1 141144 M/D/1 6667 equation 16.178182.148. 245 M/G/1 13. 52 53. 230. 38. 133. 171. 269 PerronFrobenius theory 4142. 108109.288290. 4446. 112113. 108 life insurance 5. 2730.287.340350 multiplicative functional 2830. 132133. 179 NP approximation 318320 Palm distribution 5253. 306316 Levy process 3. 38.298299. 234 matrixexponential distribution 240244 matrixexponentials 14. 227228. 141144. 176185 nonhomogeneous 60 PollaczeckKhinchine formula 6167. 25. 41.123. 178 modulation 12. 71.174.297299.
359361 stochastic control x stochastic ordering 18. 279280 subexponential distribution 11. 331336 equation 64. 3032. 9693. 172173. 240. 7475. 222.273274. 168172 storage process 13. 107. 37. 244.244250. 191192. 177 timereversion 14. 238 saddlepoint method 115117. 335336 sensitivity analysis 8693.336339 workload 13. see also matrixexponential distribution regenerative process 264 268.186. 141144. 162. 256258. 229234. 251. 281296 stable process 15. 280. 317318 semiMarkov 147. 8386. 260 WienerHopf theory 144. 160. 89.INDEX 385 waiting time 141. 131144. 307308. 5455. 174. 261264 reservedependent premiums 14. 152. 11. 233234. 189214. 123. 120 statistics x. 253. 146. 213.262263. 327 .314. 223226. 1819. 186187 virtual: see workload rational Laplace transform 8. 60. 147. 333334 regular variation 10. 186187 renewal process 131. 31. 251.279280 Rouche roots 158. 260 reinsurance 8. 292294. 87. 4950. 332333 model 12. 233.154157. 338 utility 324. 12. 294296 shotnoise process 314 simulation 19. 326330 Weibull distribution 9. 257. 251280 time change 4.
extensions of the classical compound Poisson model to allow f o r reservedependent premiums. y finite horizon ruin probabilities. phasetype distributions as a computational vehicle and the connection to other applied probability areas like queueing theory.Vol. P'i yfliother approximations (e. Markovmodulation or periodicity. for heavytailed claim size distributions). It is a comprehensive treatment of the known results on ruin probabilities. worldscientific.. Some i (l I JL I J r of the topics are Lundberg's inequality.T [Ail i The book is a comprehensive treatment of  I i I \ classical and modern ruin probability theory.. I 1! Ruin Probabilities . 2 A I 11 JjVb l' i  i Yj . "This book is a must for anybody working in applied probability.g. the ^W A l \ i l ' ''' CramerLundberg approximation.Advanced Series on Statistical Science & Applied Probability .com 2779 he 9 "789810ll22293211 . Special features of the book are the emphasis on change of measure techniques. exact solutions." Short Book Reviews ISBN 9810222939 mi u inn i nun I I I I I I i in u www.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue listening from where you left off, or restart the preview.