This action might not be possible to undo. Are you sure you want to continue?
Advanced Series on Statistical Science & I Applied Probability ^^^A£J
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
. . . . . . . . . . . . .. . . . . . 279 X Simulation methodology 281 1 Generalities . . . . . . . 261 4 Models with dependent input . . 292 6 Sensitivity analysis . . . . . . . . . . . .. 287 4 Importance sampling for the finite horizon case . . .. . . . . . . . . . . . . . . . . . . . . . . .... . . . . . . . . . . . . . . . . . .CONTENTS vii IX Ruin probabilities in the presence of heavy tails 251 1 Subexponential distributions . .. . . . . . . . 281 2 Simulation via the PollaczeckKhinchine formula . . . . . . .. . . . . . 259 3 The renewal model . . . . . . . .. .. . . . . . . . . .. . 264 5 Finitehorizon ruin probabilities . . . . .. . . . . . . . . . .. . . . . . . . . . 251 2 The compound Poisson model . . . . . . . . . . . 344 AS Complements on phasetype distributions . . . . . . . . . . . . 306 4 The distribution of the aggregate claims . . . . . . . . . . . . . . . .. . . 290 5 Regenerative simulation . 326 Appendix 331 Al Renewal theory . . . . . 297 2 Further applications of martingales . . . . .. . . . . . . . . . . . . . . . . . . . . . . 316 5 Principles for premium calculation . . . . . . . . . . . 294 XI Miscellaneous topics 297 1 The ruin problem for Bernoulli random walk and Brownian motion.. . . . .. . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . 336 A3 Matrixexponentials . . . . . . . . . . . . . .. .. . . . . . . . . . . . .. . . . . 304 3 Large deviations . . . . . . . . . . . . . . . . . . . . 350 Bibliography Index 363 383 . . . . . . 323 6 Reinsurance . . . . . . 340 A4 Some linear algebra . . . . 271 6 Reservedependent premiums . . . . . . . . . . . . . The twobarrier ruin problem . . . .. . . . . . . . . . . . . . . . . . .. . . . . 285 3 Importance sampling via Lundberg conjugation . . . . .. . . . . . . 331 A2 WienerHopf factorization . . . . . . . . . . . . . . . . . .
This page is intentionally left blank .
Since I was to produce some handouts for the students anyway. However. University of Copenhagen. the book is basically mathematical in its flavour. which has in particular removed one of the standard criticisms of the area. and the series editor Ole BarndorffNielsen for their patience. this applies to longrange dependence which is intensely studied in the neighboring ix . Risk theory in general and ruin probablities in particular is traditionally considered as part of insurance mathematics. In particular. if the formulations occasionally give a different impression. Apart from these remarks. It has obviously not been possible to cover all subareas. I was invited to give a course on ruin probabilities at the Laboratory of Insurance Mathematics. it would not be fair not to say that the practical relevance of the area has been questioned repeatedly. As an excuse: many of these projects were related to the book.Preface The most important to say about the history of this book is: it took too long time to write it! In 1991. the idea was close to expand these to a short book on the subject. 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!). I have deliberately stayed away from discussing the practical relevance of the theory. One reason for writing this book is a feeling that the area has in the recent years achieved a considerable mathematical maturity. 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. that it can only say something about very simple models and questions. Let me take this opportunity to thank above all my publisher World Scientific Publishing Co. and other projects absorbed my interest. The course was never realized. Thus. it is not by intention. But the pace was much slower than expected. and my belief was that this could be done rather quickly. and has been an active area of research from the days of Lundberg all the way up to today. and the result is now that the book is much more related to my own research than the initial outline.
13. Here is a suggestion on how to get started with the book. The present book is in between these two possibilities.89. Resnick & Samorodnitsky [303] and references therein. Asmussen. the first part of 11. Another interesting area which is not covered is dynamic control. It is obvious that such a system involves a number of inconsistencies and omissions. VII. More recently. Chapters IIIVII introduce some of the main models and give a first derivation of some of their properties. an area which is becoming increasingly important. http:// www.x PREFACE field of queueing theory.lth. Hojgaard & Taksar [206]. see also Schmidli [325] and the references in Asmussen & Taksar [52]. In addition. For a second reading. I intend to keep a list of misprints and remarks posted on my web page. Good luck! I have tried to be fairly exhaustive in citing references close to the text. The book does not go into the broader aspects of the interface between insurance mathematics and mathematical finance.se/matstat / staff/asmus and I am therefore grateful to get relevant material sent by email to asmusfmaths . it has not been possible to incorporate more numerical examples than the few there are. another by method.g. Hojgaard & Taksar [35] and Paulsen & Gjessing [284].13 and IX. 111.1.maths . The rest is up to your specific interests. IV. IV.2. Concerning ruin probabilities.13.15. X.5. VII. the standard stochastic control setting of diffusion models has been considered.2. incorporate 11. read Chapter I. In the classical setting of CramerLundberg models.3. VI. [381]). e.14. The main motivation comes from statistical data for network traffic (e. some basic discussion can be found in the books by Biihlmann [82] and Gerber [157]. I regret that due to time constraints. VIII.se Lund February 2000 Soren Asmussen .lth.g.4a. IV. some papers not cited in the text but judged to be of interest are included in the Bibliography.2 more properly). One is by model. for the effects on tail probabilities. see in particular Michna [259]. Willinger et al. see e. For a brief orientation.g. 111.13 and XI.6 (to understand the PollaczeckKhinchine formula in 111. Finally. A book like this can be organized in many ways. IX. 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).45. for which I apologize to the reader and the authors of the many papers who ought to have been on the list.
More substantial remarks. Section VII .8 . Fig. Parts of II.4 from Asmussen.PREFACE xi The second printing differs from the first only by minor corrections. A number of other figures were supplied by Christian Geisler Asmussen . supported by Center for Mathematical Physics and Stochastics (MaPhySto). IV. 1 is almost identical to Section 2 of Asmussen [26] and reprinted with permission of Blackwell Publishers. many of which were pointed out by Hanspeter Schmidli . not least the more complicated ones. were produced by Lone Juul Hansen . 5. Fig.3 are reprinted from Asmussen & Rubinstein [46] and parts of VIII.1 and X. of which there are not many at this stage . 3 is reprinted from Asmussen & Nielsen [39] and parts of IX. Aarhus.5 from Asmussen [21] with permission from CRC Press. Section VIII.6. Schmidli & Schmidt [47] with the permission from Applied Probability Trust . . 111 . Lund September 2001 Soren Asmussen Acknowledgements Many of the figures .2 by Rafal Kulik .1 by Bjarne Hojgaard and the table in Example 111.6 is reprinted from Asmussen & Schmidt [49] and parts of IX. as well as some additional references continue to be at the web page.6 by my 1999 simulation class in Lund. Parts of X. 5 from Asmussen & Kliippelberg [36] with the permission from Elsevier Science .
This page is intentionally left blank .
2) (O<t<T Ro=ul. The probability O(u) of ultimate ruin is the probability that the reserve ever drops below zero.1) We also refer to t/) ( u) and 0(u.Rt. T) as ruin probabilities with infinite horizon and finite horizon . results and topics to be studied in the rest of the book. respectively. (1. is a model for the time evolution of the reserves of an insurance company. Letting T(u) = inf {t > 0 : Rt < 0} = inf It > 0 : St > u}. We denote throughout the initial reserve by u = Ro. and give a very brief summary of some of the models. it is frequently more convenient to work with the claim surplus process {St}t>0 defined by St = u .4) O<t<oo O<t<T 1 . They are the main topics of study of the present book. t/i(u) = P (infRt < 0) = P (infR t < 0 t>0 t>0 The probability of ruin before time T is t. we introduce some general notation and terminology. A risk reserve process { Rt}t>o. For mathematical purposes.T) = P inf Rt < 0 I . (1. (1.i(u. as defined in broad terms .Chapter I Introduction 1 The risk process In this chapter . MT = sup St.3) sup St. M = (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 following setup will cover the vast majority of the book: • There are only finitely many claims in finite time intervals. • Premiums flow in at rate p.. Figure 1. Putting things together. We denote the interarrival times of claims by T2. T3. . and T1 is the time of the first claim.b(u) = P (r(u) < oo) = P(M > u).5) i.6) Sofar we have not imposed any assumptions on the risk reserve process.. However. t] is finite. per unit time. (1.T) = F (MT > u) = P(r(u) < T).i(u.2 CHAPTER I.pt. Thus. respectively. the number Nt of arrivals in [0.1.E Uk. INTRODUCTION be the time to ruin and the maxima with infinite and finite horizon. 1. St = E Uk .1 . (1. (1. the time of arrival of the nth claim is an = T1 + • • • + Tn. we see that Nt Nt Rt = u + pt . say. 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. the ruin probabilities can then alternatively be written as .
s.1 Assume that (1. and hence .1. and hence O(u) < 1 for all sufficiently large u.s. a basic references is Gerber [127]. We shall discuss Brownian motion somewhat in Chapter IV. since any modeling involves some approximative assumptions. . allowing a countable infinity of jumps on Fig.1 the slope of {Rt} should depend also on the level). 1. however.(. one could well replace Rt by Rtnr(u) or RtA.. Thus. 1. for example. we shall.8) holds. however. The models we consider will typically have the property that there exists a constant p such that Nt a E Uk k=1 p. If 77 < 0. rl= pP P It is sometimes stated in the theoretical literature that the typical values of the safety loading 77 are relatively small. 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. immaterial.8) The interpretation of p is as the average amount of claim per unit time. If 77 > 0.1. We shall not deal with this case either. then M < oo a.) V 0. VII. say 10% . • General Levy processes (defined as continuous time processes with stationary independent increments) where the jump component has infinite Levy measure. not discuss whether this actually corresponds to practice. one may well argue that Brownian motion in itself could be a reasonable model..e.1. It would appear obvious. that the insurance company should try to ensure 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. Some main examples of models not incorporated in the above setup are: • Models with a premium depending on the reserve (i. and the basic ruin probabilities are derived in XI.b(u) = 1 for all u. For the purpose of studying ruin probabilities this distinction is. • Brownian motion or more general diffusions. However. though many results are straightforward to generalize from the compound Poisson model. then M = oo a. and in fact: Proposition 1. We study this case in Ch. t * oo. on Fig. (1. of course.20%. but as an approximation to the risk process rather than as a model of intrinsic merit.
However. in connection with risk processes in a Markovian or periodic environment (Chapter VI). The simplest concrete example (to be studied in Chapter III) is the compound Poisson model..s. are i. St In concrete models.v.2 (Cox PROCESSES) Here {Nt} is a Poisson process with random rate /3(t) (say) at time t.d.8).10) is a property which we will typically encounter. not all models considered in the literature have this feature: Example 1. Nt)}.T) = i. .. INTRODUCTION Proof It follows from (1.6EU (on the average.i. then this limit is > 0 which implies St a$ oo and hence M = oo a.. . Thus p may well be random for such processes. and independent of {(0(t).4 CHAPTER I. Proposition 1. are i.10) hold with p constant. However. tb(u) = 1 for all u holds also when rl = 0.d. this needs to be verified in each separate case.11) . we obtain typically a somewhat stronger conclusion. _ St __ k =1 Uk pt a4. zP(u . k=1 (1.8) that F N. if {(3(t)} is nonergodic.s. and that . If u oo. corresponding to the Pdlya process. (1.Tp). and independent of {Nt}. M < oo a.8) is given by ^t p = EU • lim it (3(s) ds t. (1. namely that M = oo a. t t p  p' t ^ oo. rl > 0. and here (1. If U1.i.b(u) < 1 for all u when rl > 0. where {Nt} is a Poisson process with rate . with the most notable special case being V having a Gamma distribution. 0 We shall only encounter a few instances of a Cox process. This case is referred to as the mixed Poisson process. namely.T) for {Rt} is given by V)(u) = t/i (u).Q claims arrive per unit time and the mean of a single claim is EU) and that also Nt t aoo t lira EEUk = p. If 77 < 0.oo t 0 J (provided the limit exists). then similarly limSt/t < 0. Then the connection between the ruin probabilities for the given risk process {Rt} and those ^(u).Q (say) and U1. . Here it is easy to see that p = . U2.i(u. (1.3 Assume p 54 1 and define Rt = Rt1p. it is not too difficult to show that p as defined by (1..s. 0(u. U2.. The simplest example is 3(t) = V where V is a r .10) Again.
Note that when p = 1. Schmidt & Teugels [307] and Seal [326]. Rolski. the assumption > 0 is equivalent to p < 1. Gerber [157]. The Swedish school was pioneering not only in risk theory. [101]. some main texts (typically incorporating some ruin theory but emphasizing the topic to a varying degree) are Bowers et al. in particular. [76]. we shall be able to identify p with the traffic intensity of an associated queue. Note that life insurance (e. Daykin. but in probability and applied probability as a whole. Embrechts et al. U2. Daykin et al. see e . [134]. De Vylder [110]. For mixed Poisson processes and Polya processes. lighttailed distributions (sometimes the term . often referred to as collective risk theory or just risk theory. Some main later textbooks are (in alphabetical order) Buhlmann [82]. Schmidli. Some early surveys are given in Cramer [91]. CLAIM SIZE DISTRIBUTIONS 5 The proof is trivial. which is feasible since in most cases the process { Rt } has a similar structure as {Rt} (for example. Since { Rt } has premium rate 1.. Buhlmann [82]. while the first mathematically substantial results were given in Lundberg [251] and Cramer [91].. the recent survey by Grandell [173] and references therein. Mitteilungen der Verein der Schweizerischen Versicherungsmathematiker and the Scandinavian Actuarial Journal. Segerdahl [334] and Philipson [289]. The term risk theory is often interpreted in a broader sense than as just to comprise the study of ruin probabilities. Grandell [171]. many results and methods in random walk theory originate from there and the area was ahead of related ones like queueing theory. Hipp & Michel [198].g. Notes and references The study of ruin probabilities. the research literature is often published in journals like Astin Bulletin . An idea of the additional topics and problems one may incorporate under risk theory can be obtained from the survey paper [273] by Norberg. another important early Swedish work is Tacklind [373]. Pentikainen & Pesonen [101]. Cox processes are treated extensively in Grandell [171]. Straub [353]. In the even more general area of nonlife insurance mathematics. and in fact p < 1 is the fundamental assumption of queueing theory ensuring steadystate behaviour (existence of a limiting stationary distribution). in a number of models.. Taylor [364]. Insurance: Mathematics and Economics. and we do not get near to the topic anywhere in this book. Besides in standard journals in probability and applied probability. the role of the result is to justify to take p = 1. was largely initiated in Sweden in the first half of the century.2. the claim arrivals are Poisson or renewal at the same time). Heilmann [191]. We roughly classify these into two groups . see also Chapter XI. 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.g. Gerber [159]) has a rather different flavour. Some of the main general ideas were laid down by Lundberg [250].. [330]. Sundt [354].
a fact which turns out to contain considerable information. B is heavytailed if b[s] = oo for all s > 0. and can also be interpreted as the (constant) failure rate b(x)/B(x). where B(bo. Example 2 .1 (THE EXPONENTIAL DISTRIBUTION) Here the density is b(x) = beax (2. and heavytailed distributions. the m.1) The parameter 6 is referred to as the rate or the intensity.O(u) can be found in closed form. INTRODUCTION 'Cramertype conditions' is used). then the conditional distribution of X .f. Equivalently. for the compound Poisson model with exponential claim sizes the ruin probability . The crucial feature is the lack of memory: if X is exponential with rate 6.6 CHAPTER I. For example in the compound Poisson model. 2a Lighttailed distributions Example 2. the exponential distribution is by far the simplest to deal with in risk theory as well.x given X > x is again exponential with rate b (this is essentially equivalent to the failure rate being constant).g. but different more restrictive definitions are often used: subexponential.g.f. As in a number of other applied probability areas. i. B[s] is finite for some s > 0. a simple stopping time argument shows that this implies that the conditional distribution of the overshoot ST(u) . s<8.8. On the more heuristical side. regularly varying (see below) or even regularly varying with infinite variance. (2. if 1 °O AB Jbos x B(dx) > 0. P B[s]= (8Is ) .u at the time of ruin given r(u) is again exponential u with rate 8. 6 has density r(p)xPleax b(x) P and m.B(x) satisfies B(x) = O(e8x) for some s > 0. Here lighttailed means that the tail B(x) = 1 . In particular.e.2 and /LB is the mean of B.3) .2 (THE GAMMA DISTRIBUTION) The gamma distribution with parameters p. one could mention also the folklore in actuarial practice to consider B heavytailed if '20% of the claims account for more than 80% of the total claims'.2) = 0. In contrast.
X2.c. B(x) = r(p) Asymptotically. the squared coefficient of variation (s. are i. P b(x) = r` aibiea. . The exact form of the tail B(x) is given by the incomplete Gamma function r(x.. where X1.i. > 1 for p < 1 and = 1 for p = 1 (the exponential case).. 0 < ai < 1.v. u Example 2 ...d. . then X v Xl + • • • + X. i = 1.y i=1 where >i ai = 1.) VarX1 (EX )2 p is < 1 for p > 1. p) = J tPletdt. the Gamma density (2. . An appealing feature is its simple connection to the Poisson process: B(x) = P(Xi + • • • + XP > x) is the probability of at most p . x] so that B(x) = r` e. An important property of the hyperexponential distribution is that its s. CLAIM SIZE DISTRIBUTIONS 7 The mean EX is p/b and the variance Var X is p/b2. among others. p..c.. u .1 Poisson events in [0. In particular.2. or just the Erlang(p) distribution.2) can be considered as the pth power of the exponential density (2. is > 1. In particular.1) (or the 1/pth root if p < 1). 0. if p is integer and X has the gamma distribution p. 2.). p).v.. Ruin probabilities for the general case has been studied.ate (b2 ): L• i=o In the present text. p) °° where r (x. and exponential with rate d. by Grandell & Segerdahl [175] and Thorin [369]. we develop computationally tractable results mainly for the Erlang case (p = 1. This special case is referred to as the Erlang distribution with p stages. JP 1 B(x) r(p ) XP ie ax In the sense of the theory of infinitely divisible distributions. one has r(bx.3 (THE HYPEREXPONENTIAL DISTRIBUTION) This is defined as a finite mixture of exponential distributions.
See XI. equivalently.e. . Equivalent characterizations are that the density b(x) has one of the forms q b(x) j=1 = cjxienbx.6 (DISTRIBUTIONS WITH BOUNDED SUPPORT) This example (i. We give a more comprehensive treatment in VIII.4 (PHASETYPE DISTRIBUTIONS) A phasetype distribution is the distribution of the absorption time in a Markov process with finitely many states. B(x) > 0 for x < xo) is of course a trivial instance of a lighttailed distribution. Example 2 .g.7) q1 b(x) = cjxieWWx + djxi cos(ajx)ea'x + > ejxi sin(bjx)e`ix .8) are realvalued. of which one is absorbing and the rest transient.8 CHAPTER I. it is notable from a practical point of view because of reinsurance: if excessofloss reinsurance has been arranged with retention level xo. However. The density and c. are b(x) = aeTxt. 1)' is the column vector with 1 at all entries. the Erlang and the hyperexponential distributions. but the current trend in applied probability is to restrict attention to the class of phasetype distributions.. Example 2 . T) is called the representation. The couple (a. which is slightly smaller but more amenable to probabilistic reasoning.1 and defer further details to u Chapter VIII. a.d. (or. INTRODUCTION Example 2 .f. The parameters of a phasetype distribution is the set E of transient states. 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. resp. B(x) = aeTxe where t = Te and e = (1 .. there exists a xo < oo such that B(x) = 0 for x > xo. Important special cases are the exponential.6. q2 q3 (2. We give some theory for matrixu exponential distribution in VIII. a rational Laplace transform) if B[s] _ p(s)/q(s) with p(s) and q(s) being polynomials of finite degree. This class of distributions is popular in older literature on both risk theory and queues. the restriction T of the intensity matrix of the Markov process to E and the row vector a = (ai)iEE of initial probabilities.(2.7) are possibly complexvalued but the parameters in (2.8) j=1 j=1 j=1 where the parameters in (2.f. 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 . T) or sometimes the triple (E.5 (DISTRIBUTIONS WITH RATIONAL TRANSFORMS) A distribution B has a rational m.6.xo)+ is covered by the reinsurer).
However. the mean u is eµ+a /2 and the second moment is e2µ+2o2.13) u . or equivalently as the distribution of a°U+µ where U . the tail is B (x ) 2 x. (2. one being B(x) (1 + X)b(x) (1 + x)a+1' x > 0. (2.9 (THE PARETO DISTRIBUTION) Here the essence is that the tail B(x) decreases like a power of x.N(0. (2. p is defined as the distribution of ev where V . b(x) _ A(1 + (x a The pth moment is finite if and only if p < a .7 (THE WEIBULL DISTRIBUTION) This distribution originates from reliability theory.9) which is heavytailed when 0 < r < I. There are various variants of the definition around. a)/A)a+1' x > a. Writing c = d/r. x < a.8 (THE LOGNORMAL DISTRIBUTION) The lognormal distribution with parameters a2. In particular. CLAIM SIZE DISTRIBUTIONS 9 2b Heavytailed distributions Example 2.12) Sometimes also a location parameter a > 0 and a scale parameter A > 0 is allowed.p a 1 (2. It follows that the density is 't (1ogX .11) ex log logx 2r p 1 1 2 ( a ) f 1 (lox_P)2} (2. Here failure rates b(x) = b(x)/B(x) play an important role. All moments are finite. the exponential distribution representing the simplest example since here b(x) is constant.pl = 1 W (logx . b(x) = crx''le`xr.1. u Example 2 . in practice one may observe that b(x) is either decreasing or increasing and may try to model smooth (incerasing or decreasing) deviations from constancy by 6(x) = dx''1 (0 < r < oo).u l b(x) = d dx or J ax lor 1 exp Asymptotically. we obtain the Weibull distribution B(x) = eCx'.10) The loinormal distribution has moments of all orders.N(p. a2). Example 2 . and then b(x) = 0.2.1).
L( x ). the density is { 3 (1 . oo) is slowly varying . x 4 oo (any L having a limit in (0.e.10 CHAPTER I. B(x) = O(xP). where Y is Pareto distributed with a = (p . in particular. The simplest examples correspond to p small and integervalued.15) x2 + 16x3 ) a3x/2) 3 (1 . (2. 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).11 (PARETO MIXTURES OF EXPONENTIALS) This class was introduced by Abate. satisfies L(xt)/L(x) 4 1.2). Choudhury & Whitt [1] as the class of distributions of r. { s () 1s+3s29s3log(1+2s I p=3. 6 is defined as the distribution of et' where V has the gamma density (2. i.v. u .'s of the form YX.1)/p. A = 1 and X is standard exponential. in particular. For p = 1.12 (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) .14) The pth moment is finite if p < 5 and infinite if p > 5. the loggamma distribution (with exponent 5) and a Pareto mixture of exponentials. x + 00. (2. In general. INTRODUCTION Example 2.16) 11 Example 2.12) (here L (x) * 1) and ( 2. the loggamma distribution is a Pareto distribution. Thus.(1 + 2x + 2x2)e2x) p = 2 (2. The density is 8p(log x)pi b(x) . examples of distributions with regularly varying tails are the Pareto distribution (2.x6+lr(p) (2.10 (THE LOGGAMMA DISTRIBUTION) The loggamma distribution with parameters p. u Example 2 . another standard example is (log x)').13).17) where L (x) is slowly varying.(1 + Zx + $ p = 3.
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. By far the most prominent case is the compound Poisson (CramerLundberg) model where {Nt} is Poisson and independent of the claim sizes U1. We give some discussion on standard methods to distinguish between light and heavy tails in Section 4f.. Thus. it will be seen that we obtain completely different results depending on whether the claim size distribution is exponentially bounded or heavytailed. Also. which each have a ( timehomogeneous) small rate of experiencing a . Namely.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. this phenomenon represents one of the true controversies of the area. (2.. the subexponential class of distributions provide a convenient framework for studying large classes of heavyu tailed distributions. 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. for example the lognormal distribution is subexponential (but not regularly varying). one may argue that this difficulty is not resticted to ruin probability theory alone. 3 The arrival process For the purpose of modeling a risk process . The reason is in part mathematical since this model is the easiest to analyze. the claim size distribution represents of course only one aspect (though a major one). and so is the Weibull distribution with 0 < r < 1. and based upon such information it seems questionable to extrapolate to tail behaviour.13 (THE SUBEXPONENTIAL CLASS OF DISTRIBUTIONS) We say that a distribution B is subexponential if xroo lim B `2^ = 2.1. However. but can never be sure whether this is also so for atypical levels for which far less detailed statistical information is available. From a practical point of view. but the model also admits a natural interpretation : a large portfolio of insurance holders . Similar discussion applies to the distribution of the accumulated claims (XI. When studying ruin probabilities.. though the proof of this is nontrivial.1) that any distribution with a regularly varying tail is subexponential. U2.18) B(x) It can be proved (see IX. We return to a closer study in IX.3.. THE ARRIVAL PROCESS 11 Example 2. the knowledge of the claim size distribution will typically be based upon statistical data.
so that . Cox processes are. The one we focus on (Chapter VI) is a Markovian environment : the environmental conditions are described by a finite Markov process {Jt }too. found the Poisson distribution to be inadequate and suggested various other univariate distributions as alternatives . Historically.(3. However . radioactive decay (a huge number of atoms each splitting with a tiny rate ) and many other applications. the negative binomial distribution. IV (and. Mathematically. Nevertheless . In order to prove reasonably substantial and interesting results . A more appealing way to allow for inhomogeneity is by means of an intensity . 5..e. with the extension to premiums depending on the reserve. This model . its basic feature is to allow more variation (bursty arrivals ) than inherent in the simple Poisson process. epidemics in life insurance etc. however. in Chapter VII). The point of view we take here is Markov dependent random walks in continuous time (Markov additive processes ). such that 8(t) = . An obvious example is 3(t) depending on the time of the year (the season).12 CHAPTER I.. when Jt = i.8 (t) is a periodic function of t. it is more questionable whether it provides a model with a similar intuitive content as the Poisson model. Some of them have concentrated on the marginal distribution of NT (say T = one year ). to be studied in Chapter V. has some mathematically appealing random walk features . This model can be intuitively understood in some simple cases like { Jt} describing weather conditions in car insurance . gives rise to an arrival process which is very close to a Poisson process.6. which facilitate the analysis. 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 . getting away from the simple Poisson process seems a crucial step in making the model more realistic. the periodic and the Markov modulated models also have attractive features .3(t) fluctuating over time. and also that the ruin problem may be hard to analyze . e.. To the author 's knowledge . where {/3 (t)}too is an arbitrary stochastic process . we study this case in VI . too general and one neeed to specialize to more concrete assumptions . with a common term {Nt} is a Markovmodulated Poisson process . see 11.. INTRODUCTION claim . not many detailed studies of the goodnessoffit of the Poisson model in insurance are available . Another one is Cox processes. This applies also to the case where the claim size distribution depends on the time of the year or . The compound Poisson model is studied in detail in Chapters III. but with a general not necessarily exponential distribution ). are i. in particular to allow for certain inhomogeneities.i. 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.d. T2.g. the first extension to be studied in detail was {Nt } to be renewal (the interarrival times T1 . in just the same way as the Poisson process arises in telephone traffic (a large number of subscribers each calling with a small rate). I. In others.
1).1) where V is the limit in distribution of Vt as t + oo. it is a recurrent theme of this book to stress this connection which is often neglected in the specialized literature on risk theory. interacting particle systems. point processes and so on. Mathematically. Thus. In the setting of (4.T) = P(VT > u). extreme value theory. dam/storage processes.'s like V is available. with Poisson arrivals and constant release rule p(x) = 1. others being branching processes. time series and Gaussian processes. The ones which appear most related to risk theory are queueing theory and dam/storage processes. methods or modeling ideas developed in one area often has relevance for the other one as well. The study of the steady state is by far the most dominant topic of queueing and storage theory. It should be noted. 0(u) = P(V > u). stochastic differential equations. and a lot of information on steadystate r. More generally.4. genetics models. however.1) permitting to translate freely between risk theory and the queueing/storage setting. that quite often the emphasis is on computing expected values like EV. 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. A general release rule p(x) means that {Vt} decreases according to the differential equation V = p(V) in between jumps. this amounts to Vo having the stationary distribution of {Vt}). this gives only f0 O°i (u)du which is of limited .6) .1) holds as well provided the risk process has a premium rule depending on the reserve. (4. stochastic geometry. others are quite different. The M/G/1 workload process { Vt } may also be seen as one of the simplest storage models. queueing theory. and the limit t 4 oo is the steadystate limit. A SUMMARY OF MAIN RESULTS AND METHODS 13 the environment (VI. A stochastic process {Vt } is said to be in the steady state if it is strictly stationary (in the Markov case. R = p(R) in between jumps. and which seems well motivated from a practical point of view as well.v. and here (4. Similarly. 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 .0 (u. Some of these have a certain resemblance in flavour and methodology. reliability. ruin probabilities for risk processes with an input process which is renewal. Markovmodulated or periodic can be related to queues with similar characteristics. it is desirable to have a set of formulas like (4. In fact.
14 CHAPTER I.3.1) in the setting of a general premium rule p(x): the events {VT > u} and {r (u) < T} coincide when the risk process and the storage process are coupled in a suitable way (via timereversion ).1). INTRODUCTION intrinsic interest .1.3. Thus . see Corollary III. 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.3. The fact that Theorem H. the two areas.g. see Boxma & Cohen [74] and Abate & Whitt [3]. Vi(u. as is typically the case. The qualifier 'with just a few phases ' refers to the fact that the diagonalization has to be carried out numerically in higher dimensions.2). have to some extent a different flavour. Here Vi(u) is given in terms of a matrixexponential function ( Corollary VIII. • The compound Poisson model with some rather special heavytailed claim size distributions. e . p = 0/8 and y = 8 .6. Here ?P(u) is explicit provided that . see Corollary VII. though overlapping.8.p(y) y^ Jo p(x) can be written in closed form. 3. The infinite horizon (steady state ) case is covered by letting T oo. • The compound Poisson model with premium rate p(x) depending on the reserve and exponential claim size distribution B.1 . 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. • The compound Poisson model with a claim size distribution degenerate at one point.. the ideal is to be able to come up with closed form solutions for the ruin probabilities 0(u).1 is a sample path relation should be stressed : in this way the approach also applies to models having supplementary r. .T). Similarly. B(x) = ebx. which can be expanded into a sum of exponential terms by diagonalization (see.'s like the environmental process {Jt} in a Markovmodulated setting. Example VIII. the functions w x f d 1 exdx () . • The compound Poisson model with constant premium rate p = 1 and B being phasetype with a just few phases . which gives a sample path version of (4. A prototype of the duality results in this book is Theorem 11. 3.v. Here O(u) = peryu where 3 is the arrival intensity. 4b Exact solutions Of course .3.
For the finite horizon ruin probability 0(u.1) is the explicit form of the ruin probability when {Rt} is a diffusion with infinitesimal drift and variance µ(x). esuTb( u. . Given this can be done.7. say the fast Fourier transform (FFT) as implemented in Grubel [179] for infinite horizon ruin probabilities for the renewal model. A SUMMARY OF MAIN RESULTS AND METHODS 15 • The compound Poisson model with a two step premium rule p(x) and B being phasetype with just a few phases. T) du dT 0 TO 00 in closed form than the ruin probabilities z/'(u). Here are some of the main approaches: Laplace transform inversion Often. Abate & Whitt [2]. it is easier to find the Laplace transforms = f e8 . Embrechts. the second best alternative is a numerical procedure which allows to calculate the exact values of the ruin probabilities. but are somewhat out of the mainstream of the area . T) themselves. Grubel & Pitts [132] and Grubel & Hermesmeier [180] (see also the Bibliographical Notes in [307] p. (u. T). 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 .Lef$er function. T) can then be calculated numerically by some method for transform inversion. f {eXp U LX 2.ff 2µ(y)/a2(y) dy} dx . Also Brownian models or certain skip free random walks lead to explicit solutions (see XI .2) is the natural scale. We don't discuss Laplace transform inversion much. where Furrer [150] recently computed ii(u) as an infinite series involving the Mittag. However. see VIII. O(u.u(y)/a2(y) dy} 4c Numerical methods Next to a closedform solution.1) are so complicated that they should rather be viewed as basis for numerical methods than as closedform solutions. the formulas ( IV.b(u)du . relevant references are Grubel [179]. a2 (x): Ip (u) = where S(u) = f °O exp {. Ab(u).f f 2µ(y)/a2(y) dy} dx  (4. A notable fact ( see again XI. [s. 191).S(u) 1S(oo) f °D exp {.4. • An astable Levy process with drift . 1).
it states that i/i(u) . 0(u) is then given in terms of a matrixexponential function euu (here U is some suitable matrix) which can be computed by diagonalization. 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. as the solution of linear differential equations or by some series expansion (not necessarily the straightforward Eo U'u/n! one!). However. 4d Approximations The CramdrLundberg approximation This is one of the most celebrated result of risk theory (and probability theory as a whole).3. 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. and carry out the solution by some standard numerical method. dt] most often leads to equations involving both differential and integral terms.3) in the compound Poisson model which is an integral equation of Volterra type. Differential. For the compound Poisson model with p = 1 and claim size distribution B with moment generating function (m.Ce"u.1) and y > 0 is the solution of the Lundberg equation (4.3) where C = (1 . (4. and in particular the naive idea of conditioning upon process behaviour in [0. .p)/(13B'[ry] . see VIII. In the compound Poisson model with p = 1.4) 00['Y]1)'Y = 0. which can equivalently be written as f3 [7] = 1 +13 .) B[s].7.g.or integral equation. One example where this is feasible is the renewal equation for tl'(u) (Corollary III. T) as the solution to a differential. and in quite a few cases (Chapter VIII). U is explicit in terms of the model parameters. whereas for the renewal arrival model and the Markovian environment model U has to be calculated numerically. most often it is more difficult to come up with reasonably simple equations than one may believe at a first sight.16 CHAPTER L INTRODUCTION Matrixanalytic methods This approach is relevant when the claim size distribution is of phasetype (or matrixexponential).and integral equations The idea is here to express 'O(u) or '(u. u * oo.f.
incorporating correction terms may change the picture dramatically. other approaches are thus required. 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. However. This list of approximations does by no means exhaust the topic. For example. in such cases the evaluation of C is more cumbersome. . T). in some cases the results are even more complete than for light tails. The CramerLundberg approximation is renowned not only for its mathematical beauty but also for being very precise. In particular.7 and IV. the claim size distribution should have an exponentially decreasing tail B(x). However. corrected diffusion approximations (see IV. In fact. when the claim size distribution is of phasetype. See Chapter IX. for the compound Poisson model ^(u) p pu In fact . 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. but typically not very precise in their first naive implementation. J B dx.6) 4e Bounds and inequalities The outstanding result in the area is Lundberg's inequality (u) < e"lu. In the case of heavytailed distributions. the exact solution is as easy to compute as the CramerLundberg approximation at least in the first two of these three models. some further possibilities are surveyed in 111 . It has generalizations to the models with renewal arrivals.2. T) for large u are available in most of the models we discuss.4. u > oo.6) are by far the best one can do in terms of finite horizon ruin probabilities '(u. often for all u > 0 and not just for large u. Large claims approximations In order for the CramerLundberg approximation to be valid. Approximations for O(u) as well as for 1(u. (4. Diffusion approximations are easy to calculate. a Markovian environment or periodically varying parameters. and use the fact that first passage probabilities are more readily calculated for diffusions than for the risk process itself.
This is proved for the compound Poisson model in 111. lower bounds etc. the difficulty comes in when drawing inference about the ruin probabilities.3). In practice. However. This procedure in itself is fairly straightforward. e. 4f Statistical methods Any of the approaches and results above assume that the parameters of the model are completely known. How do we produce a confidence interval? And. .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. though not too many precise mathematical results have been obtained. given NT.k (U(`) . one expects a model with a deterministic claim size distribution B. . 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. in the compound Poisson model. it has the advantage of not involving approximations and also. . 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. as a general rule. more importantly. .. empirical evidence shows that the general principle holds in a broad variety of settings. fitting a parametric model to U1. However.U(k)) i =k+ i . The standard suggestion is to observe that the mean residual life E[U .d.i. they have however to be estimated from data. it is a general principle that adding random variation to a model increases the risk. T]. INTRODUCTION Compared to the CramerLundberg approximation (4. and to plot the empirical mean residual life 1 N .18 CHAPTER I.. obtained say by observing the risk process in [0. . of being somewhat easier to generalize beyond the compound Poisson setting.8. For example. When comparing different risk models. one may question whether it is possible to distinguish between claim size distributions which are heavytailed or have an exponentially decaying tail. which is a standard statistical problem since the claim sizes Ui. UNT are i. UNT may be viewed as an interpolation in or smoothing of the histogram). to have smaller ruin probabilities than when B is nondegenerate with the same mean m.) at various places and in various settings. say degenerate at m. We return to various extensions and sharpenings of Lundberg's inequality (finite horizon versions. . For example. can we trust the confidence intervals for the large values of u which are of interest? In the present author's opinion..g.
claims U1.v's) which can be generated by simulation.i. We look at a variety of such methods in Chapter X. and in fact methods from that area can often be used in risk theory as well . The problem is entirely analogous to estimating steadystate characteristics by simulation in queueing/storage theory. Simulation may be used just to get some vague insight in the process under study: simulate one or several sample paths.3) and Section VI. where U(1) < .3). in all other chapters than VI where we just write . the more typical situation is to perform a Monte Carlo experiment to estimate probabilities (or expectations or distributions) which are not analytically available.3 (or just VI. Truncation to a finite horizon has been used. formula VI. good methods exist in a number of models and are based upon representing the ruin probability zb(u) as expected value of a r. reference [14].5. and also discuss how to develop methods which are efficient in terms of producing a small variance for a fixed simulation budget. this is a straightforward way to estimate finite horizon ruin probabilities.2. and look at them to see whether they exhibit the expected behaviour or some surprises come up. CONVENTIONS 19 as function of U(k). (or a functional of the expectation of a set of r.. because it appears to require an infinitely long simulation. The infinite horizon case presents a difficulty. in this book referred to as [APQ]). Klnppelberg & Mikosch [134]. and of course the method is relevant in risk theory as well.3) or Section 3 of Chapter VI are referred to as Proposition VI. 5 Conventions Numbering and reference system The basic principles are just as in the author's earlier book Applied Probability and Queues (Wiley 1987. but is not very satisfying.d.e.. formula (5. Still. 4g Simulation The development of modern computers have made simulation a popular experimental tool in all branches of applied probability and statistics. For example. The chapter number is specified only when it is not the current one.. Thus Proposition 4. . < U(N) are the order statistics based upon N i.. respectively. A main problem is that ruin is typically a rare event (i.2..4. having small probability) and that therefore naive simulation is expensive or even infeasible in terms of computer time. to observe whether one or the other limiting behaviour is apparent in the tail.(5. . UN. See further Embrechts. However.v.
then for 1s < 5).s. EX2/(EX)2.The same symbol B is used for a probability measure B(dx) = P(X E dx) and its c. (moment generating function) fm e82B(dx) of the distribution B.g.o. In particular.B(x) = P(X > x) of B. B(dy).g. moment generating function.29) refer to the Appendix.Used in asymptotic relations to indicate that the ratio between two expressions is 1 in the limit. say a heuristic approxi1 + h + h2/2. random variable s.4. if B(x) . and for a defective probability distribution IIGII < 1.f. right hand side (of equation) r.g. B(x) the tail 1 . b[s] is defined always if Rs < 0 and sometimes in a larger strip (for example. see under b[s] below.d. References like Proposition A.e. oo). E.20 CHAPTER L INTRODUCTION Proposition 4. A different type of asymptotics: less precise.t.f. IIGII the total mass (variation ) of a (signed ) measure G .h. The Laplace transform is b[s]. independent identically distributed i.s. r.f. formula (5.d.v.p. n! 27r nn+1/2en.r. w. .v.3) or Section 3.i. for a probability distribution IIGII = 1.g. Abbreviations c. cumulative distribution function P(X < x) c. log E[s] where b[s] is the m.g. h + 0.2. cumulant generating function.h. B is concentrated on [0. B[s] the m. n i oo. i.c.ceax. mation.f. as for typical claim size distributions. B(x) = P(X < x) = fx. If. E expectation.f. . squared coefficient of variation. with respect to w. or a more precise one like eh . i.d. infinitely often l.f. left hand side (of equation) m. with probability Mathematical notation P probability. (A.
Xt_ the left limit limstt X8f i.e. though slightly more complicated. Matrices and vectors are denoted by bold letters. N(it. 7r. I(A) the indicator function of the event A. row vectors have lowercase Greek letters like a.e. Thus. i. oo) the space of Rvalued functions which are rightcontionuous and have left limits. Usually. 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. only have finitely many jumps in each finite interval. . . often the term 'cadlag' (continues a droite avec limites a gauche) is used for the Dproperty. Unless otherwise stated. (the dimension is usually clear from the context and left unspecified in the notation). an example or a remark. R(s) the real part of a complex number s. F o r a given set x1. 21 D [0. (xi)diag denotes the diagonal matrix with the xi on the diagonal (xi)row denotes the row vector with the xi as components (xi). i.e. the value just before t. A.. a. CONVENTIONS {6B the mean EX = f xB(dx) of B ABA' the nth moment EXn = f x"B(dx) of B. intensity interpretation. and column vectors have lowercase Roman letters like t.. . xa. the ith unit row vector is e'i. 0 marks the end of a proof. the ith entry is 1 and all other 0. a2) the normal distribution with mean p and variance oa2. the processes we consider are piecewise continuous. Usually. 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. E[X. In the Frenchinspired literature. all stochastic processes considered in this book are assumed to have sample paths in this space. Notation like f3i and 3(t) in Chapter VI has a similar .A] means E[XI(A)].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). matrices have uppercase Roman or Greek letters like T. of numbers.5.
g. I. 111.22 CHAPTER L INTRODUCTION B the claim size distribution.1.5. though slightly more complicated. ry The adjustment coefficient. cf. . cf. cf. I. Notation like BE and B(t) in Chapter VI has a similar. e. 'q the safety loading .5.1. VI. J the rate parameter of B for the exponential case B(x) = eby. interpretation. p the net amount /3pB of claims per unit time. or quantities with a similar time average interpretation. FL. EL the probability measure and its corresponding expectation corresponding to the exponential change of measure given by Lundberg conjugation.
however. in part. The likelihood ratio approach in Section 2 is basic for most of the models under study. 23 . More precisely. 5) are. When encountered for the first time in connection with the compound Poisson model in Chapter III. The topic is. The general theory is. The duality results in Section 3 (and. The reader should therefore observe that it is possible to skip most of the chapter. All results are proved elsewhere . however. not crucial for the rest of the book. strictly speaking. however. 5 on random walks and Markov additive processes can be skipped until reading Chapter VI on the Markovian environment model. fundamental ( at least in the author' s opinion) and the probability involved is rather simple and intuitive. used in Chapter VI on risk processes in a Markovian (or periodic) environment. a parallel selfcontained treatment is given of the facts needed there. in particular at a first reading of the book. Due to the generality of the theory. somewhat more advanced than in the rest of the book. the relevance for the mainstream of exposition is the following: The martingale approach in Section 1 is essentially only used here. in most cases via likelihood ratio arguments. Sections 4. the level of the exposition is.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. Sections 4.
T(u) < cc] = e7uE {e7f(u) I T(u) < cc] z/. StUit. Our first result is a representation formula for O(u) obtained by using the martingale optional stopping theorem . using r(u) A T invokes no problems because r(u) A T is bounded by T). however.u denote the overshoot. V) (u. claim size distribution B and p = .2) As T > oo. (1. Example 1 .QµB < 1. As usual. 1 Martingales We consider the claim surplus process {St} of a general risk process. the time to ruin r(u) is inf It > 0 : St > u}.24 CHAPTER II. and in the limit (1. (b) St a$ oo on {T(u) = oo}.5 can be skipped.(u).1 Assume that (a) for some ry > 0. T(u) < oo] + 0 = eryuE [e7Vu). {e'YS° }t>0 is a martingale. We get 1 = Ee7So = E e'Y S(.1 is basic for the study of the compound Poisson model in Chapter III. Proposition 1. Thus N.0.2) takes the form 1 = E [e'ys(). The more general Theorem 6..2 Consider the compound Poisson model with Poisson arrival rate . and the ruin probabilities are ip(u) = P (T(u) < oo).)AT = E [e7ST(°). Let e(u) = ST(u) . f1 . T(u) < T] + E [eryST . the second term converges to 0 by (b) and dominated convergence (e7ST < eryu on {r(u) > T}). SOME GENERAL TOOLS AND RESULTS The ladder height formula in Theorem 6. 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] . T) = P(T(u) < T).
Q and the U.2(c)). 1.5 For the compound Poisson model with B exponential.1) . Thus 00 E [e'rt (") I T(u) < oo] = I e5e  dx = f 5edx . condition (a) of Proposition 1. Eeas° = e"(°) where n(a) = a2a2/2 . Thus. From this it is readily seen (see III. the conditions of Proposition 1. Thus.1. the martingale property now follows just as in Example 1. Then {St } is Brownian motion with variance constant o2 and drift p < 0. The available information on this jump is that the distribution given r(u) = t and S.1.Q. u Corollary 1.3 Assume that {Rt} is Brownian motion with variance constant o.Q(B[a] . the ruin probability is O(u) = pe.2 and drift p > 0.Ft = a(S" : v < t).2. O(u ) < e7". of the normal distribution.4 (LUNDBERG ' S INEQUALITY ) tion 1 . and (b) follows from p < 1 and the law of large numbers (see Proposition III.(„)_ = x is that of a claim size U given U > u .ap. and thus by the memoryless property of the exponential distribution .1 are satisfied.f. it follows that E [e7st+v I J] = e"rstE [e7(st+vSt) I Ft] = e7StEe"rs° = elst where .1) . with common distribution B (and independent of {Nt}).d. Under the conditions of Proposi Proof Just note that C(u) > 0.. Proof Since c(a) = /3 (B[a] .r" where y = S .a = a . Since {St} has stationary independent increments.6a for details) that typically a solution to the Lundberg equation K(y) = 0 exists. u Corollary 1. are i. B(x) _ edx.a.1 is satisfied. A simple calculation (see Proposition III. the conditional distribution of the overshoot e(u) = U .i.x. and thus Ee7s° = 1. and thus Ee'rs° = 1.g. MARTINGALES 25 where {Nt} is a Poisson process with rate .u + x is again just exponential with rate S.3/6 < 1.1) shows that Eels. Now at the time r(u) of ruin {St} upcrosses level u by making a jump . Since {St} has stationary independent increments.1.= e"(') where K(a) = .a it is immediately seen that y = S . By standard formulas for the m. From this it is immediately seen that the solution to the Lundberg equation ic(y) = 0 is y = 2p/a2./3. and p =. Example 1 .
cf.Ft}too and the Borel afield F.3 below).1) 'though not always: it is not difficult to construct a counterexample say in terms of transient Markov processes. Embrechts. on (DE.6 If {Rt} is Brownian motion with variance constant a2 and drift p > 0. oo). The interesting concept is therefore to look for absolute continuity only on finite time intervals (possibly random.6 N S = { lim Nt I t +00 t gJ are both in F. P are then singular (concentrated on two disjoint measurable sets). A]. However. A E Ft. (2. Grandell & Schmidli [131]. P correspond to the claim surplus process of two compound Poisson risk processes with Poisson rates /3. u Notes and references The first use of martingales in risk theory is due to Gerber [156]. and in analogy with the theory of measures on finite dimensional spaces one could study conditions for the RadonNikodym derivative dP/dP to exist.e.F). Grandell [171]. the parameters of the two processes can be reconstructed from a single infinite path. which we equip with the natural filtration {. Example 2 . and by the law of large numbers for the Poisson process .v.26 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Corollary 1. Thus the sets S = I tlim +oot t =. B. F(S) = P(S) = 1. and F. hence so is Nt = limfyo N2`i. 2 Likelihood ratios and change of measure We consider stochastic processes {Xt} with a Polish state space E and paths in the Skorohod space DE = DE[0. Proof Just note that ^(u) = 0 by continuity of Brownian motion. The number Nt F) of jumps > e before time t is a (measurable) r. as shown by the following example this setup is too restrictive: typically'. Theorem 2. Delbaen & Haezendonck [103] and Schmidli [320]. F). we look for a process {Lt} (the likelihood ratio process) such that P(A) = E[Lt. More recent references are Dassios & Embrechts [98]. I. P on (DE. Two such processes may be represented by probability measures F. But if a $ ^ . [172]. then z/'(u) = e7" where 'y = 21A/a2. 0 and claim size distributions B. then S and S are disjoint . A somewhat similar u argument gives singularity when B $ B.1 Let F.. . and is further exploited in his book [157].
F the Borel o•field and P a given probability measure on (DE.t. A].t.1) holds. ({. ({Ft} . . P) such that LLt = 1. then {Lt} is a nonnegative martingale w. Then P(A) = E[Lt. using the martingale property in the fourth step. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 27 (i. . G l ] = E [_I(G)E[LTIFT] ] = E { _I(G)Lr ] = P(G). A] = E[Lt. Then Ft (A) = E[Lt.Ft}. G C {T < oo}.2. Proposition 2. A E F8.A] = EE[LtI(A)IF8] = EI(A)E[LtIFB] = EI(A)L8 = PS(A)..e.F). If r is a stopping time and G E PT. (i) If {Lt}t> o is a nonnegative martingale w. ELt = 1 follows by taking A = DE in (2. A E F. define P by Pt (A) = E[Lt. then there exists a unique probability measure Pon . then { 1 P(G) = EG .Y) such that P(A) = Pt(A). Lt < 0] can only be nonnegative if P(A) = 0.r. By the martingale property. Lets < t. 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. The truth of this for all A E Y.2(i). F) such that ELt = 1. Then Lt > 0 and ELt = 1 ensure that Pt is a probability measure on (DE. (ii) Conversely.1) and nonnegativity by letting A = {Lt < 0}. P be as in Proposition 2.F such that (2.Tt) is absolutely continuous w. A E Ft . under the assumptions of (ii) we have for A E rg and s < t that A E Ft as well and hence E[L8.r. Hence E [_ . Proof Under the assumptions of (i). the restriction of P to (DE.1) holds.Pt)) The following result gives the connection to martingales.r.3 Let {Lt}. that the restriction of P to (DE. we have E [ LTIFT]1 = LT on {T < T}. u The following likelihood ratio identity (typically with r being the time r(u) to ruin) is a fundamental tool throughout the book: Theorem 2 . This proves (i).t. Finally.2) Proof Assume first G C {T < T} for some fixed deterministic T < oo. 1 J (2. A].2 Let {Ft}t>o be the natural filtration on DE.F8] = L8 and the martingale property.Pt}adapted process {Lt}t>o (2. _. Conversely. implies that E[LtI. G ] = E [LT . Ft).
Xt = St. u From Theorem 2.4 Under condition (a) of Proposition 1. say. we need the concept of a multiplicative functional.Rt} the claim surplus process and {Jt} a process of supplementary variables possibly needed to make the process Markovian. 1 Since everything is non negative. and this problem will now be studied. t. Lr(u) 11 The advantage of (2.r.1) in Proposition 1.1: Corollary 2.1) is that it seems in general easier to deal with the (unconditional) expectation E[eryVu). (2. Now just rewrite the r. The crucial step is to obtain information on the process evolving according to F. is Markov w.O(u) = eryuE[e'YC(u).1). the natural filtration {. one would typically have Xt = Rt. applying (2. first in the Markov case and next (Sections 4.2) follows by monotone convergence.1. each F.Gn {r <T} . both sides are increasing in T. 5) for processes with some randomwalklike structure.. is nonnegative and has Ey Lt = 1 for all x.. The problem is thus to investigate which characteristics of {Xt} and {Lt} ensure a given set of properties of the changed probability measure.28 CHAPTER II.t. In the context of ruin probabilities.r.h. To this end. A change of measure is performed by finding a process {Lt} which is a martingale w.t.4) compared to (1. where {Rt} is the risk reserve process. r(u) < oo] occuring there than with the (conditional) expectation E[e'r{(u ) Jr(u) < oo] in (1.s. . (2. Xt = (Jt.3) to G of{r < T} we get 1111 F(Gn {r <T}) = E[ 1 . in continuous time (the discrete time case is parallel but slightly simpler).3 we obtain a likelihood ratio representation of the ruin probability V) (u) parallel to the martingale representation (1. {St} = {u . we assume for simplicity that {Xt} has Dpaths. we have F(G) = V )(u). First we ask when the Markov property is preserved. Consider a (timehomogeneous) Markov process {Xt} with state space E. of (2. SOME GENERAL TOOLS AND RESULTS In the general case . St).Ft} . and letting T * oo. For the definition. Rt) or Xt = (Jt.2) by noting that 1 = ersr(„) = e1'ue7Ou).4) Proof Letting G = {r(u) < oo}. T(u) < oo].
Ex[Lt+8Zt(Y8 o et)] = Ex[LtZt(Y8 o 0t)(L8 o Bt)].7) for any Ftmeasurable Zt and any . or.v.'s of the form fl' f. this in turn means Ex[Lt+8Zt(V8 oet)] = Ex[LtZtExt[L8Y8]].v. since Zt • (Y8 o Ot ) is . t and let Px be the probability measure given by t.Ft] = Ex.(Xtitl) with all t(i) < t + s. (2.v. By definition of Px.YB] for any Ftmeasurable r..[Y. oo).(A) = Ex [Lt. t. (2.2. A].Ft }.5) is equivalent to Ex[Lt+8Vt+8] = E8[Lt • (L8 o 91)Vt+8] for any .Ft+8measurable. which is the same as Ex[Zt(Y8 o Bt)] = E8[ZtEx. Y8. where Ot is the shift operator.v. t] * [0.5 Let {Xt} be Markov w.7).5) are Tt+e measurable.F8measurable r.7).5) Pxa.8) which is the same as (2.'s of the form Zt • (Y8 o 0t) comprises all r.Ftmeasurable. o 9tI. the natural filtration {Ft} on DE. The converse follows since the class of r. Similarly. (2. 0 .Y8f t < s. and then L.v. Lt has the form Lt = 'Pt ({x }0<u<t) for some mapping cot : DE[O. the Markov property can be written E. Proof Since both sides of (2. let {Lt} be a nonnegative martingale with Ex Lt = 1 for all x. for all x.r.s.Pt+8measurable r.6) for any . Zt. LIKELIHOOD RATIOS AND CHANGE OF MEASURE 29 on DE and define {Lt} to be a multiplicative functional if {Lt} is adapted to {. The precise meaning of this is the following: being . Indeed. which in turn is the same as Ex[Lt+8Zt • (V8 o Bt)] = Ex[Lt • (L8 o 91)Z1 • (Y8 o et)] (2. Vt+e. since Ext [L8Y8] = E[(Y8 o et)(L8 o 8t)I.Ft].t. nonnegative and Lt+8 = Lt•(Lso9t) (2. o 9t = V.T9measurable Y8. s. ({Xt+u} 0<u<8) Theorem 2..6) implies (2. (2. Then the family {Px}xEE defines a timehomogeneous Markov process if and only if {Lt} is a multiplicative functional.
Thus R = Ro + f p(R8) ds . SOME GENERAL TOOLS AND RESULTS to define a timehomogeneous Markov process.Ft] = LtE[L8 o 9t I. and thus for the moment no parametric assumptions (on say the structure of the arrival process) are needed. < aN < T. The result is a sample path relation. and the time to ruin is 7(u) = inf {t > 0: Rt < 0}.. Ro = u (say).30 CHAPTER H. aN where or* = T UN_k+l. In between jumps. t] = LtExt L8 = Lt... we shall establish a general connection between ruin probabilities and certain stochastic processes which occurs for example as models for queueing and storage.. UN..1) The initial condition is arbitrary. .At where At = k: vk <t U. T] in the following setup: The risk process {Rt}o<t<T has arrivals at epochs or... the arrival epochs are Qi.e. More precisely . (using the Markov property in the second step) so that the martingale property is automatic. 3 Duality with other applied probability models In this section.6 For {u . The formulation has applications to virtually all the risk models studied in this book.5 can be found in Kuchler & Sorensen [240].. . . } E[Lt+B I.. . Indeed. In between jumps. the premium rate is p(r) when the reserve is r (i. . see Dynkin [128] and Kunita [239]. The corresponding claim sizes are Ul. reflection at zero and initiar condition Vo = 0. it xEE suffices to assume that {Lt} is a multiplicative functional with Ex Lt = 1 for all x. R = p(R)). We work on a finite time interval [0. u Notes and references The results of the present section are essentially known in a very general Markov process formulation. A more elementary version along the lines of Theorem 2. 0 < vl < . then Remark 2. {Vt} . CN. (3... with a proof somewhat different from the present one. and just after time or* {Vt} makes an upwards jump of size UU = UN _k+l. The storage process {Vt }o<t<T is essentially defined by time reversion. t.
3) (u) Proof Let rt' denote the solution of R = p(R) subject to r0 = u. 3...T) = P(VT > u).1 Define r(u) = inf It > 0: Rt < 0} (r(u) = oo if Rt > 0 for all t < T) and let ii(u. That is.___ . (3.. Note that these definitions make {Rt} rightcontinuous (as standard) and {Vt} leftcontinuous.__. Then rt°) > rt°) for all t when u > v. . {Vt} remains at 0 until the next arrival).____•_.AT_t.1) we have Vt = At  f P(Vs)ds where A= U= AT . DUALITY WITH OTHER APPLIED PROBABILITY MODELS 31 decreases at rate p(r) when Vt = r (i.T) = inf Rt < 0 P (O<t<T P(r(u) < T) be the ruin probability.1._: 1} 0 011 =T01N ^N3 To 0 011 014 01N Figure 3..3.e. instead of (3.x.2) k: ok <t and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0.._.1 The events {T(u) < T} and {VT > u} coincide.. V = p(V)).. V)(u.11 4.__... :. Theorem 3. The sample path relation between these two processes is illustrated in Fig. (3. In particular.
the distinction between right. Hence if n satisfies VVN_n+1 = 0 (such a n exists. . if and only if O(u) < 1 for all u. 3.1 and its proof is from Asmussen & Schock Petersen [50].d. Then Vo. Proof Let T ^ oo in (3.and left continuity is immaterial because the probability of a Poisson arrival at any fixed time t is zero). Thus we may think of {Vt} as having compound Poisson input and being defined for all t < oo.2 from Harrison & Resnick [188]. this represents a model for storage.1 with Ro = u = ul). we have RQ„ < 0 so that indeed r(u) < T. Some further relevant more general references are Asmussen [21] and Asmussen & Sigman [51].3). Nevertheless. say of water in a dam though other interpretations like the amount of goods stored are also possible.32 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Suppose first VT > u (this situation corresponds to the solid path of {Rt} in Fig. Suppose next VT < u (this situation corresponds to the broken path of {Rt} in Fig. say V. Historically. The arrival epochs correspond to rainfalls. and so on.U1 = Rol. Resnick & Tweedie [79]. Then similarly VVN = r0.i. The results can be viewed as special cases of Siegmund duality. and a general premium rule p(r) when the reserve is r.2 Consider the compound Poisson risk model with a general premium rule p(r). Corollary 3. u A basic example is when {Rt} is the risk reserve process corresponding to claims arriving at Poisson rate . if nothing else n = N).Ul < roil  Ul = RQ„ Va1V_1 < RQ2. one may feel that the interaction between the different areas has been surprisingly limited even up to today. If VaN > 0.T l . and since ruin can only occur at the times of claims. u Notes and references Some main reference on storage processes are Harrison & Resnick [187] and Brockwell. Hence RQ„ > 0 for all n < N. Theorem 3.3 and being i. 3. the connection between risk theory and other applied probability areas appears first to have been noted by Prabhu [293] in a queueing context. and then '0 (u) = P(V > u). Then the storage process {Vt} has a proper limit in distribution. with distribution B.1 with Ro = u = u2). we can repeat the argument and get VoN_1 > Ra2 and so on. see Siegmund [344].U1 > roil . 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. Historically. = r(VT) . We get: Corollary 3. we have r(u) > T.
. . Z2.. . where the Yi are i .. there is an analogue of Theorem 3.2 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.2). .. R valued sequence Z1.2) satisfies the same recursion as in (4..... WN be the Lindley process generated by Z1 = YN..1. For a given i. (c) The Lindley process {WN} generated by Zl = Y1. Xo = 0. 1} is often referred to as simple random walk or Bernoulli random walk)... hits (oo.h. is defined by assigning Wo some arbitrary value > 0 and letting Wn+1 = (Wn + Zn+1)+• (4... I.1..1)). ..min n=0. can be viewed as the reflected version of the random walk with increments Z1.4... n=0..N From this the result immediately follows. . N min (Y1 + • • • + YNn) n=0.. of (4. = Xo + Y1 + • • • + Y.. 0 Corollary 4. evolves as a random walk with increments Z1i Z2.Yl min (YN ..1 Let r(u) = inf In: u + Y1 + • • • + Yn < 0}.1 in terms of Lindley processes ... In particular. N min (Y1 + + Yn).d.. W2.. Z2 . Here F is a general probability distribution on R (the special case of F being concentrated on {1. 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. {Wn}n=0. W1. ...Y1 according to Wo = 0. i.... with common distribution F (say)..d.. . Most often.... Proof By (4. Then the events {r(u) < N} and {WN > u} coincide.YNn+1) n=0..1. .1) Thus {Wn}n=o. Let further N be fixed and let Wo. Theorem 4. Z2 = YN_1 i .. For discrete time random walks ... the Lindley process Wo. W1.1. as long as the random walk only takes nonnegative values.1. .i... generated by Z1. .s.w..2) (for a rigorous proof.....1. and is reset to 0 once the r.... 0). just verify that the r . if Wo = 0 then (Z1+•••+Zn) WN = Zl+•••+ZN. (b) 1/i(u) = P(•r(u) < oo) > 0 as u * oo. Z2.. WN = YN .N (4.e. Z2 = Y2. ZN = .
. either M = oo a. Remark 4 .1 is equivalent to WN D MN = (Z1 + . then the restrictions of Fx.1 is actually not necessary .. the Y1. a sufficient condition for (e) is that EY is welldefined and > 0.=o. . ... (e). Thus the assertion of Theorem 4. .s. e.1 have the same distribution for n = 0. (Z1 + • • • + Zn) = m and P(W > u) = P(M > u) = i (u ). assumption on the Z1. . Next consider change of measure via likelihood ratios... By Kolmogorov's 01 law. SOME GENERAL TOOLS AND RESULTS (d) m = inf. Similarly... 0 By the law of large numbers. .s .the result is a sample path relation as is Theorem 3.l..1. In that case . The converse follows from general random walk theory since it is standard that lim sup (Y1 + • • + Yn) = oo when Y1 + • • • + Yn 74 oo. Proof Since (YN.. w.ooa. ZN or.. N.g. YN). W v m and P(W > u) = P (m > u) = 0(u).d.5 does not necessarily lead to a random walk: if.3 The i.34 CHAPTER II. .o. . (e)Yi+•••+Yn 74 . doubly u infinite (n'= 0.. .. 176) but appears to be rather intractable.. .1...s. the Lindley processes in Corollary 4.. YN in Theorem 4..g.. .) and defines Zn = Yn. ±2... In general. Y1) has the same distribution as (Y1. + Z.i. 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.1.2..N so that WN _P4 M = supra=0.. One then assumes Yn to be a stationary sequence.. equivalently.1. For a random walk. ±1.. (Yi + • • • + Yn) > oo a. (d) #. the condition 00 F(YI+•••+ Yn<0)<00 n=1 is known to be necessary and sufficient ((APQ] p. Px to Fn are equivalent (have the same null sets) so that the likelihood ratio Ln exists.) sup n=0. The following result gives the necessary and sufficient condition for {Ln} to define a new random walk: . there is a more general version of Corollary 4. a Markovian change of measure as in Theorem 2. Combining these facts gives easily the equivalence of (a)(d). Clearly..s.l.2 and Theorem 4.. or M < oo a.
e. for some function h with Eh(Y) = 1. we get L2 = L1 (L1 o91 ) = h(Y1)g(X1.5 corresponds to a new random walk if and only if Ln = h(Y1) . 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.nrc(a )} (4.5 corresponds to a new random walk with changed increment distribution F(x) = e'(a) Jr e"'F(dy) . RANDOM WALKS INDISCRETE OR CONTINUOUS TIME 35 Proposition 4. cf.3) 1Pxa.s. In particular.. Y) = h(Y ) a. Conversely.g. h(Yn) (4. Since L1 has the form g (Xo. and define Ln by (4. Y1). Y ) f (Y)] = E[g(O. (4. 100 ).. of F). and so onforn =3.s.f.. the changed increment distribution is F(x ) = E[h(Y). For n = 2. u A particular important example is exponential change of measure (h(y) = e°y'(") where r.3) holds for n = 1. In that case...4.4) ({Ln} is the familiar Wald martingale .4.4). the random walk property implies Ex f (Y1) = Eo f (Y1 ).5 Consider a random walk and an a such that c(a) = log F[a] = log Ee° ' is finite. (a) = log F(a] is the c. this means E(g(x.Y2) = h(1'i)h(I'a). Y < x]..3) holds. Proof If (4. The corresponding likelihood ratio is Ln = exp {a (Y1 + • • • + Yn) . Then the change of measure in Theorem 2. Y) f (Y)] for all f and x. implying g (x. We get: Corollary 4.4 Let {Ln} be a multiplicative functional of a random walk with E_L..g. y ). Breiman [78] p. . = 1 for all n and x. Then the change of measure in Theorem 2. where h (y) = g(0.
however.1). Xt(2)_t(l). (4. {Xt} is a random walk. {Xt} can be written as the independent sum of a pure drift {pt}. a Brownian component {Bt} (scaled by a variance constant) and a pure jump process {Mt}. However. In continuous time. The traditional formal definition is that {Xt} is Rvalued with the increments Xt(1)_t(o). Xt (n)t(n1) being independent whenever t(O) < t(1 ) < . they arise as models for the reserve or claim surplus at a discrete sequence of instants. Roughly.Ft] = Eof (X. or imbedded into continuous time processes . we are . (4. v2 = 0 and v = 3B). i. the claim surplus process for the compound Poisson risk model . this description needs some amendments. a positive measure on R with the properties e J x2v(dx) < oo. see Chapter V. Xt =Xo+pt+oBt+Mt.Xn < x). corresponds to a process with stationary independent increments and u = p. with premium rate p.e.36 CHAPTER II. An equivalent characterisation is {Xt} being Markov with state space R and E [f (Xt+e . say by recording the reserve or claim surplus just before or just after claims (see Chapter V for some fundamental examples). < t(n) and with Xt( i)_t(i_l) having distribution depending only on t(i) . The simplest case is 3 = JhvMM < oo. which corresponds to the compound Poisson case: here jumps of {Mt} occur at rate 0 and have distribution B = v/0 (in particular .Xt)I. but we omit the details ). A general jump process can be thought of as limit of compound Poisson processes with drift by considering a sequence v(n) of bounded measures with v(n) T v. the pure jump process is given by its Levy measure v(dx).7) for all e > 0. .. In risk theory. SOME GENERAL TOOLS AND RESULTS Discrete time random walks have classical applications in queueing theory via the Lindley process representation of the waiting time ... say the beginning of each month or year .6) More precisely.5) Note that the structure of such a process admits a complete description. e f x:IxJ>e} v(dx) < oo (4. given by a the increment distribution F(x) = P(Xn+l . . In discrete time. the tradition in the area is to use continuous time models.). the interpretation is that the rate of a jump of size x is v(dx) (if f of Ixlv (dx) = oo.. The appropriate generalization of random walks to continuous time is processes with stationary independent increments (Levy processes).t(i .
where VT is the virtual waiting time at time T in an initially empty M/G/1 queue with the same arrival rate /3 and the service times having the same distribution B as the claims in the risk process. First assume in the setting of Section 3 that {Rt} is the risk reserve process for the compound Poisson risk model with constant premium rate p(r) = 1. cf. WT = XT .s. Proposition 4.v. [The condition for V < oo a.min Xt (4.1. Furthermore.1)v(dx) (4. and the reflected version is then defined by means of the abstract reflection operator as in (4.3 and decreases linearly at rate 1 in between jumps. defined as a system with a single server working at a unit rate. 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.2). Now consider reflected versions of processes with stationary independent increments. T) = P(VT > u).Q and distribution B of the service times of the arriving customers. 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. A different interpretation is as the workload or virtual waiting time process in an M/G/1 queue. is easily seen to be f3pB < 1.] Processes with a more complicated path structure like Brownian motion or jump processes with unbounded Levy measure are not covered by Section 3. then Ee'(xtxo) = Eoeaxt = etx(a). Then the storage process {Vt} has constant release rate 1.6).e. Chapter III. i. VT + V for some r. has upwards jumps governed by B at the epochs of a Poisson process with rate . Corollary 4. V E [0.7 If {Xt} has stationary independent increments as in (4.4. O(u.8) O<t<T (assuming Wo = Xo = 0 for simplicity). where c(a) = ap + a2a2/2 + f 00 provided the Levy measure of the jump part {Mt} satisfies f". 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). jxJ v(dx) < oo. and b(u) = P(V > u). (ex .6 In the compound Poisson risk model with constant premium rate p(r) . oo].10) . having Poisson arrivals with rate .
Xt +B . e.f. if Lt = e9(xt . In particular.. t.6) are µ = µ + Oo2 . we use the characterization (4.xo)tk ( e). Theorem 4 .1 that E eaMt = exp fmoo In the general case . 8 Assume that {Xt} has stationary independent increments and that {Lt} is a nonnegative multiplicative functional of the form Lt = g(t. 5 has stationary independent increments as well. let e" (a ) = Eoeaxl. Eea(µt + QBt) = et{aµ +a2OZ/2}.4 o) aµ + ((a + 0 ) 2  0 2 )o 2 /2+ r w J 00 (e (a + 9)x  a 9x )v(dx) 00 a(µ + O 2) + a2a2 / 2 + J (eax .Xt)I'Ftl = E [f(Xt+B . (4. X8) = Eof (X8)L8 = Eof (X8)• For the second.5) and get E [f(Xt+B .38 CHAPTER IL SOME GENERAL TOOLS AND RESULTS Proof By standard formulas for the normal distribution.Xt)I Ftl = Eof (X8)g(s. This is of course no coincidence since the distribution of Xl . Then l e" (a) = Eo [ Li ea "] = eK (9)Eo {e ( a+9)x1 J I = er(a+o )K(B) R(a) = K(a + 0) .1 ) v(dx) . of an infinitely divisible distribution (see.1. Xt Xo) with E2Lt = 1 for all x. Q2 = v2.g.Xt)g(s. Then the Markov process given by Theorem 2.Xt)L8 o 0tIFt] = E [f (Xt+s . we show in the compound Poisson case ( IlvIl < oo) in Proposition III.Xo is necessarily infinitely divisible when {Xt} has stationary independent increments.g.11) (eax . v(dx) = e9xv (dx).10) is the LevyKhinchine representation of the c. . use the representation as limit of compound Poisson processes. u Note that (4.1)eexv(dx). Proof For the first statement . Chung [86]). By explicit calculation . then the changed parameters in the representation (4.
. b = a = 0) the changed process is the claim surplus process of another compound Poisson risk process with Poisson rate . <t whenever the RadonNikodym derivative dB/dB exists (e. v(dx) _ . is defined as a bivariate Markov process {Xt} = {(Jt. the structure of MAP's is completely understood when E is finite: 2and only there .g. 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.. U2. 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 . MAP stands for the Markovian arrival process discussed below. and let the Px refer to the claim surplus process of another compound Poisson risk process with Poisson rate.3B[B]. MARKOV ADDITIVE PROCESSES 39 Remark 4.3 and claim size distribution B. . one reason is that in parts of the applied probability literature. where . we write Pi.1) For shorthand . (5.3 =.(3B(dx). Thus (since µ = p = 1. Ei instead of P2.Ft] = Ejt.St)g(Jt+s)I. the corresponding claim sizes .3 and claim size distribution B # B. b with b(x) > 0 for all x such that b(x) > 0).l3 and claim u size distribution B.5. abbreviated as MAP in this section2.4). u 5 Markov additive processes A Markov additive processes.8. B(dx) = B[9] B(dx).9 If X0 = 0.10 Let Xt be the claim surplus process of a compound Poisson risk process with Poisson rate . Example 4 . it is then easily seen that Lt = H dB(Ui) i:o.11 For an example of a likelihood ratio not covered by Theorem 4. are the arrival times and U1.. Example 4 .0 in the following.(3 = . then the martingale {eex(t)tk(e)} is the continuous u time analogue of the Wald martingale (4.2. B have densities b.0. Ei.. Recalling that U1. . Then we can write v(dx) _ /3eOxB(dx) = / (dx). a = 0. 0. As for processes with stationary independent increments . corresponding to p = 1..o[f (S8)g(J8)]. dB/dB = b/b when B.
t+s) where Jt .o(Ji = j. SOME GENERAL TOOLS AND RESULTS In discrete time. which we omit and refer to Neveu [272] or cinlar [87]. In continuous time (assuming Dpaths). the converse requires a proof.40 CHAPTER H. {Jt} is specified by its intensity matrix A = (Aij)i. Y2. oo). with the Y„ being interpreted as interarrival times. a jump of {Jt} from i to j # i has probability qij of giving rise to a jump of {St} at the same time.Sr_1.1 For a MAP in discrete time and with E finite.. v. If all Fij are concentrated on (0. the distribution of which has some distribution Bij... let {Jt} be standard Brownian motion on the line.J1=j)= Fij (dx) Pij In simulation language. 1 J1 ='^])iJEE = (Fij[a])i . Fn[a] = F[a]n where P[a] = P . {St} evolves like a process with stationary independent increments and the parameters pi.6) depending on i. a MAP is specified by the measurevalued matrix (kernel) F(dx) whose ijth element is the defective probability distribution Fij(dx) = Pi. consider the matrix Ft [a] with ijth element least Ei .. .it = A.i. a MAP is the same as a semiMarkov or Markov renewal process.9 EE = (iii&ij[a])i j EE .jEE• On an interval [t.g. An alternative description is in terms of the transition matrix P = (piA. Jn = j..[a) = (Ei[easl. (That a process with this description is a MAP is obvious. In addition. by generating Yn according to Hij when J„_1 = i.) If E is infinite a MAP may be much more complicated. As an example. Y1 E dx) where Y„ = S„ . As a generalization of the m. Proposition 5.f. Then a Markov additive process can be defined by letting t St = lim 1 I(IJB1 < e)ds E1o 2d o be the local time at 0 up to time t. this means that the MAP can be simulated by first simulating the Markov chain {J„} and next the Y1.jEE (here pij = Pi(J1 = j)) and the probability measures Hij(dx)=P(Y1 EdxlJo=i. vi(dx) in (4.
1)) . By PerronFrobenius theory (see A.5.qkj + k?^j qkj Bkj [a] } = Ei [east. up to o( h) terms. Jt = j] is given by etK[a]. Jt = k] { xk kEE j la] .4c). MARKOV ADDITIVE PROCESSES Proof Conditioning upon (Jn. Then. we infer that in the discrete time case the . In matrix formulation . qij. j E E) and So = 0. vi(dx). assume that the Markov chain/process {Jt} is ergodic. Proof Let {Stt) } be a process with stationary independent increments and pa rameters pi . which in conjunction with Fo[a] = I implies Ft[a] = etK[a) according to the standard solution formula for systems of linear differential equations. Jt = j] Ejesh'^ + E Ak j hEi [ease .ijgij(Bij[a] . u In the following. Jt = j] (1 + htc (j) (a)) j + Ak j qk j (Bk +h E Ei [east . Jt = k] { 1 . Then the matrix Pt[a] with ijth element Ei [east. 013 .1) } (recall that qjj = 0).1 )v(dx). vi(dx) (i E E). Bij (i. this means that F't+h [a] = Ft[a] II+h(rc(i)(a)) +hA+h(Aijgij(Bij[a]1)) I. J1 = A which in matrix formulation is the same as Fn+1 [a] = Fn[a]F[a]. \ diag Ft[a] = Ft[a]K. a= . aSt h = (1 + Ajjh) Ei [east . u Proposition 5. where K[a] = A+ (r. Jn+1 = A] = 41 Ei[ e 5„.2 Let E be finite and consider a continuous time Markov additive process with parameters A. kEE Jn = k]Ek[e"Y" . 00 r(i) (a) = api + a2ot /2 + f (e° . Sn ) yields Ei[easn+ '. pi.(')(a)) diag + ().
just note that [a]h(a) = eietK (a)h(a) = etK(a)h(a). Eie"sth^a) = e'Pt[a]h( a) = e. cf. SOME GENERAL TOOLS AND RESULTS matrix F[a] has a real eigenvalue ic(a) with maximal absolute value and that in the continuous time case K[a] has a real eigenvalue K(a) with maximal real part. and write k = k(°). and we shall take V(a)h(a) = 1. a. of a random walk. we are free to impose two normalizations. 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 . Corollary 5. Proof By PerronFrobenius theory (see A.h(a)vva)etw(a). (5.3 Ei [east.4c). Corollary 5. Yrh(a ) = 1.5 EiSt = tK'(0) + ki .e=e°tk. cf. We also get an analogue of the Wald martingale for random walks: Proposition 5. . Furthermore.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). Jeast. Since v(").42 CHAPTER II.2) where 7r = v(°) is the stationary distribution.f.7. In particular.t. Proposition 5. The function ic(a) plays in many respects the same role as the cumulant g. its derivatives are 'asymptotic cumulants'.tK(a)h(a) J jj it L o is a martingale. as will be seen from the following results.Eikjt = ttc'(0) + ki . Proof For the first assertion. h(") are only given up to a constants.etx It then follows that E feast+^(t+v)K(a)h(a) I ^tl l .4. h(") may be chosen with strictly positive components.c(a) (and h(")).4 Eie"sth(a) = h=a)et?("). Corollary 5. Jt = j] . Then h(°) = e.r. The corresponding left and right eigenvectors v(").
In the same way. ] = t2tc (0)2 + 2tK'(0)vk + ttc"(0) + O(1) . there is typically a function h = h(") on E and a ^c(a) such that Ey a"st t" (") * h(x).5.") }) . E=ST = tc'(0)E7. the existence of exponential moments is assumed ) but can be made rigorous by passing to characteristic functions.5. Ee"st typically grows asymptotically exponential with a rate ic(a) independent of the initial condition (i. (5.a) + ttc (a)2hia ) Multiplying by v=. [E. subtraction yields Vary St = tic"(0) + O(1). Corollary 5. 43 Ei [Steast h(a) + east k^a)1 = et"(a) (kia) + tic (a)hia)) . u The argument is slightly heuristic (e. Remark 5 . t im v^"St = '(0) Proof The first assertion is immediate by dividing by tin Corollary 5. one obtains a generalization of Wald's identity EST = Er • ES.St]2 = t2/c'(0 ) 2 + 2ttc (0)vk .Eikjr . For the second .2ttc (0)Evkjt + 0(i). MARKOV ADDITIVE PROCESSES Proof By differentiation in Proposition 5. summing and letting a = 0 yields E„ [St + 2Stkj. More precisely.g.6 For any stopping time T with finite mean. we differentiate (5..3) Let a = 0 and recall that h(°) = e so that 0=°) = h(o) = 1.7 No matter the initial distribution v of Jo. the distribution of Jo). Squaring in Corollary 5. 8 Also for E being infinite (possibly uncountable ).5 yields + W (a)k. for a random walk: Corollary 5. Since it is easily seen by an asymptotic independence argument that E„ [Stkjt] u = trc'(0) E„kjt + O(1).4. t a oo. (5.. tam E tSt a (0).+ k.4) .3) to get Ej [St a " st h i(a ) + 2Ste"st k(a) + e"st k^a) J etI(a) (kia )' + ttc (a)ki") + t {ic"(a)h.e. .
let ha(i. V. xEE . 0) = h(i )( 1 + ttc(a)). St)} be a MAP and let 0 be such that h(Jt) OStt.6) We shall not exploit this approach systematically.44 CHAPTER IL SOME GENERAL TOOLS AND RESULTS for all x E E.. this is. 1) (i.f (x) tyo t provided the limit exists. gha(i. and the family {f LEE given by Theorem 2.s.10 Let {(Jt. In view of this discussion .4 that { h(Jt) easttK(a) L o (5. some extra conditions are imposed. (5. h(Jo) Lo is a Px martingale for each x E E.5 defines a new MAP. however.. An example beyond the finite case occurs for periodic risk processes in VI.6. St) } as follows. 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). For t small .5. Remark 5. inconvenient due to the unboundedness of ea8 so we shall not aim for complete rigour but interpret C in a broader sense. First.6. Then {Lt } is a multiplicative functional.3b and Remark VI. we take the martingale property as our basic condition below (though this is automatic in the finite case). s) = ea8h(i). 0) = n(a) h(i). however.5) is a martingale can be expressed via the infinitesimal generator G of {Xt } = { (Jt. G is defined as Gf (x) = lim Exf (Xt) .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. see. Usually.5) is a martingale . 0 Proposition 5.for the present purposes. u forsEE).9 The condition that (5. Given a function h on E.(9) {Lt}t>o = . this leads to h(i) + tcha( i.e. Jt = (s+t) mod 1 P8a. From (5. in particular that f is bounded.
u Theorem 5. Ai = µi + 0Q. That 0 < qij < 1 follows from the inequality qb <1. this gives a direct verification that A is an intensity matrix: the offdiagonal elements are nonnegative because Aij > 0. qij = r. vi (dx) = f3 Bi(dx) with . one can directly verify that (5. 1 + q(b . then also vi (dx) is compound Poisson with e Ox ^i = /3iBi[0]. Bi. In the infinite case . MARKOV ADDITIVE PROCESSES Proof That { Lt} is a multiplicative functional follows from L8 ogt = h(Jt+s) es(St+ . Then the MAP in Proposition 5.1) eft ea' f ij (dx) = Hij (dx) Hij [0] . if vi(dx) is compound Poisson. Here Oh(e) is the diagonal matrix with the h=e) on the diagonal.1) holds for the P. and by A = Oh(°) K [0]Oh(e ) vi(dx) = e"xvi (dx)..1) .11 below in the finite case.7) In particular. 0<q<1.Qi < oo and Bi a probability measure. . in the discrete time case.ic(0)e = ic(0)Oh e) h(e) .. 0 < qij < 1 and Bij [0] > 0. 0<b<oo.11 Consider the irreducible case with E finite.5.c(0)e = tc(0)e . In particular. (5. ^i = of qij Bij [0] 1 + qij ( Bij [0] . We omit the details.tc(0)e = 0 . Bi [0] Remark 5.10 is given by P = eK(e) Oh e) F[e]Oh(').12 The expression for A means h(e) Aij = hie) Aij [1 + gij(Bij[0] i 0 j.St)sl(e) h(Jt) 45 The proof that we have a MAP is contained in the proof of Theorem 5.7(dx) Bij [0] Bij(dx) in the continuous time case . Bi(dx) = Bi(dx). That the rows sum to 1 follows from Ae = Oh(e) K[O]h(B) .(0)j.
Jt = j] = hie) . F:j with a density proportional to eei . Similarly. v. First note that the ijth element of Ft[a] is etK(e)Ej [e(a+B)st E:[east Jt = j] = Ej[Lteas' .tc(0)' )Ah() = Oh(o) K[a + 0]Oh() . it follows that indeed the normalizing constant is H1 [0]. Jt = A. . this implies k[a] = A 1 ) (K[a + 0] . a = 0 in (5. this means that Ft[a] = etw ( e)Ohc) Ft[a + 9]oh (e) (5. Jl = j] :(Yi E dx. Letting a = 0 yields the stated expression for A. (dx). In matrix notation .tc (') (0) corresponds to the stated parameters µ.8).46 CHAPTER II. since Hij. H1.K [O])Oh(e) (0) l + ( A + (tc(') (a + 0) . Yi E dx. Here the stated formula for P follows immediately by letting t = 1.. Hence the same is true for H=j and H.11. This shows that F. Jl = j) = Ei[Lt.t.13) for matrixexponentials . Further Fib (dx ) = P=(YI E dx.8. (dx) of a process with stationary independent increments follows from Theorem 4.r.. .tc(0)I.tc(') (8)/ d)ag h 7 Aiiii (Bii[a + 0] . SOME GENERAL TOOLS AND RESULTS Proof of Theorem 5.e) Consider first the discrete time case .. is absolutely continuous w.8) h(. Now we can write K[a] =A+A ) ( K[a + 0] . v= .8) yields et'[a] = Ohie )et (K[a +e]K(e)I)Oh(°) By a general formula (A. are probability measures . Ji = j) h(e) eeyK(B)p h(8) h(e) eexK ( h=e) e)Fi.Bay [0]) That k(') (a + 0) . in continuous time ( 5.
Much of the pioneering was done in the sixties in papers like Keilson & Wishart [224].1) = Aij4ij(Bij[a] . hardly a single comprehensive treatment. [261]. i. Though the literature on MAP's is extensive. which. Notes and references The earliest paper on treatment of MAP's in the present spirit we know of is Nagaev [265]. oo). there is. Note that G+ is concentrated on (0. and is typically defective. has no mass on (oo. THE LADDER HEIGHT DISTRIBUTION 47 Finally note that by (5.7). the literature on the continuous time case tends more to deal with special cases. 0]. is slightly less general than the present setting. [262] in discrete time. h.)Ajjgij(Bij[a+0] . . Write r+ = T(0) and define the associated ladder height ST+ and ladder height distribution by G+(x) = 11 (S..(u) = inf {t > 0 : St > u} to ruin in the particular case u = 0 .6.Bij[0]) = hjel)ijgijBij[0](Bij[a] .6. Conditions for analogues of Corollary 5. however. an extensive bibliography on aspects of the theory can be found in Asmussen [16].. IIG+II = G+(oo) = P(T+ < oo) = 0(0) < 1 when 77 > 0 (there is positive probability that {St} will never come above level 0). 7+ < oo). [225]. 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].+ < x.3 for an infinite E are given by Ney & Nummelin [266].1). [226] and Miller [260]. 6 The ladder height distribution We consider the claim surplus process {St } of a general risk process and the time 7. < x) = 11 (S. see also Fuh & Lai [149] and Moustakides [264]. For the Wald identity in Corollary 5.e. however.
Theorem 6 . In other cases like the Markovian environment model. the sum of all the ladder steps (if rl > 0. i. On Fig.00 ). 6. For the proof of Theorem 6. it follows that for g > 0 measurable.48 CHAPTER K. they have a semiMarkov structure (but in complete generality.ST+(1) and so on.1) The interpretation of R+(A ) is as the expected time {St} spends in the set A before T+. The first ladder step is precisely ST+.. define the prer+occupation measure R+ by R+(A) = E f o "o I(St E A. the second ladder height (step) is ST+(2) . there are only finitely many). by approximation with step functions .2.d. where basically only stationarity is assumed. R+ is concentrated on (oo. Recall that B(x) = 1 . g(y)R+(dy) = E f g(St)dt. 0 f T+ (6.1 The term ladder height is motivated from the shape of the process {Mt} of relative maxima. a fact which turns out to be extremely useful. 6.1. see Fig. Thus.1.(3B(x ) = pbo(x) on (0.1. at present we concentrate on the first ladder height.i.2) . To illustrate the ideas.B(x) denotes the tail of B.e. which gives an explicit expression for G+ in a very general setting. the ladder heights are i. G+ is given by the defective density g + (x) =. has no mass on ( 0. oo). 0]. In simple cases like the compound Poisson model.. the dependence structure seems too complicated to be useful). SOME GENERAL TOOLS AND RESULTS M ST+(2) Sr. and the maximum M is the total height of the ladder. Also. we shall first consider the compound Poisson model in the notation of Example 1. 1 For the compound Poisson model with p = 01LB < 1. In any case.T+ > t)dt = E f 0T+I(St E A) dt. The main result of this section is Theorem 6. o 00 (6. the second ladder point is ST+(2) where r+(2) is the time of the next relative maximum after r+(1) = r+.5 below. i. Here bo(x) _ B(x)/µB. = ST+(1) Figure 6.e.
6. see Fig.ST<St. That is. since the distribution of the Poisson process is invariant under time reversion.ST_t.O<t<T) = P(STEA.2(b): r+ < t Thus.T+>T) = P(STEA.6. {St }o<t<T is constructed from {St}o<t<T by timereversion and hence. THE LADDER HEIGHT DISTRIBUTION Lemma 6 .O<t<T).0<t<T) = F(ST E A.2 R+ is the restriction of Lebesgue measure to (00.2(a): T+ > t Figure 6.ST<St. 0 < t < T. 49 Proof Let T be fixed and define St = ST . . 0]. 0 < t < T) P(STEA. has the same distribution as {St}o<t<T. P(STEA.2. St S* t a Figure 6.St<0. ST < ST_t.
Figure 6. T+ > t] 0 _ /3 f 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 . But since St 4 oo a.50 CHAPTER II.t dT. E A} precisely when r+ > t.. and since the jump rate is /3. for A C (0. SOME GENERAL TOOLS AND RESULTS Integrating w. That is.2) in the last). we get G+ (A) = f 00 /3 dt E[B(A .y) (here we used the fact that the probability of a jump at u t is zero in the second step. oo). this is just the Lebesgue measure of A. s. .r.T+ > t] dt 0 T+ _ /3E f g( St) dt = 0 f g(y) R+(dy) 0 00 where g(y) = B(A .St). and (6.St _)I(r+ > t). oo).. U + St_ E A.3 where the bold lines correspond to minimal values.St _).3 G+ is the restriction of /3R+*B to (0. Fig. G+(A) = Q f 0 B(A . 6.y)R+(dy) 00 Proof A jump of {St} at time t and of size U contributes to the event IS. cf.3 Lemma 6 . The probability of this given { Su}u<t is B(A .
Uk) (k = 1.e.) where ak = Ti + • • • + Tk .1. this is equivalent to the risk process {St*} being stationary in the sense of (6.s > 0).. obviously. . The marked point process . i. as a point process on [0. h]} /h (by stationarity. U k)} k=1 a is as a marked point process M *. The traditional representation of the input sequence {(TT. THE LADDER HEIGHT DISTRIBUTION 51 Proof of Theorem 6.Q f r+(x . 6 .4) are (ak. i.z)B(dz) _ f I(x < z)B(dz) _ f (x). 0 Generalizing the setup..e. oo) x (0. 6 . 4 (the points in the plane are (ak . we consider the claim surplus process {St }t>o of a risk reserve process in a very general setup.:T1 +•••+Tk <t}. we define the arrival rate as E# { k : ak E [0 .6. {St+8 .T+ < oo). Uk) for those k for which ak .s. Nt St =>Uk k=1 t where Nt = max{k = O.. The points in the plane (marked by x on Fig. We call M * stationary if M* o B8 has the same distribution as M* for all s > 0. cf. assuming basically stationarity in time and space. 2. 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. Lemma 6. Fig.* ) and the second the mark (the claim size Uk ).1 With r+(y) = I(y < 0) the density of R+.M o 08 shifted by s is defined the obvious way. the first component representing time (the arrival time o. 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. .S8 )t> o = {St }t>o for all s > 0.. oo). In the stationary case. this does not depend on h)..3 yields g+ (x) = .4).
of (6.. vi(dx) = . Note also that (again by stationarity) the Palm distribution also represents the conditional distribution of M* o Ot given an arrival at time t. k: vk E [0. . This more or less gives a proof that indeed (6. Section 5) which has pure jump structure corresponding to pi = a = 0. most often one takes h = 1)... Example 6 . letting h J.g. V(M* o eak ). i. h. Assume {Jt} irreducible so that a stationary distribution 7r = (1i)iGE exists. 0.4 Given a stationary marked point process M*.4 Consider a finite Markov additive process (cf. o.5) represents the conditional distribution of M* given vi = 0.e.52 CHAPTER II. h] and the sum approximately ^o(M*)I(ul < h).s. where T is the first arrival time > 0 of M and h > 0 an arbitrary constant (in the literature.QiBi(dx). the r. and let T = T2 denote the first proper interarrival time .5) does not depend on h. where TI = 0.. As above . Sigman [348] for these and further aspects of Palm theory. = 0 . Uk) k=1. SOME GENERAL TOOLS AND RESULTS M* U.2. The two fundamental formulas connecting M* and M are Eco(M) = aE E. See. e. We represent M by the sequence (Tk. Oh becomes the approximate probability F(ri < h) of an arrival in [0. i 1 U2 Us 1_ 0 or Q2 $ U3 *1 L 0 7 X I 11 1 Figure 6. we define its Palm version M as a marked point process having the conditional distribution of M* given an arrival at time 0 . h] Eco(M*) = 1 E f co(M o Bt)dt.
v.6. 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 . THE LADDER HEIGHT DISTRIBUTION 53 Interpreting jump times as arrival times and jump sizes as marks. Then the ladder height distribution G+ is given by the (defective) density g+(x) = ./. If Jt_ = i. Note in particular that the Palm distribution of the mark size (i. let U0 be a r.oo a. After that. we get a marked point process generated by Poisson arrivals at rate /3i and mark distribution Bi when Jt = i. Jt = j is iri(3i /.O fo "o F(x)dx = . the distribution of Ul) is the mixture B = E aii Bi + aij Bij J = j#i !i J. This follows by noting that iP*(0) = IIG+JI = J0 "o g+(x)dx = . It follows that we can describe the Palm version M as follows . the probability aij of Jt . Before giving the proof.= i.OF(x).OEU0. Jo) w. let the arrivals and their marks be generated by {Jt} starting from Jo = j. aij for (i.O for i = j and iriAijgij/. we note: Corollary 6.p.5. j) and let the initial mark Ul have distribution Bi when i = j and Bij otherwise. Assume that St * .e. and that p = 0EU0 < 1.6 Under the assumptions of Theorem 6.p. qij when {Jt} jumps from i to j and have mark distribution Bij. the ruin probability .6iBi + Aijgij Bij j#i iEE iEE 0 Theorem 6 .s.p. 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 . . an arrival for M* occurs before time t + dt w.. 5 Consider a general stationary claim surplus process {St }t>o. and by some additional arrivals which occur w.*(0) with initial reserve u = 0 is p = /3EU0. First choose (Jo_.O for i # j.
2. . Let p(t) be the conditional probability that ST+ E A. are point processes on (oo .).. Then clearly * G+ (A) = P(ST+ E A) = Consider a process { f p(t)f3dt..s. 105) shows that one can assume w.5.1] here the r . .Su_ <0.St<Su.A. moves down linearly at a unit rate in between jumps and starts from S0 = U. has a very simple interpretation as the average amount of claims received per unit time .(left limit) when 0 < it < t and is illustrated on Fig . 0).Mt). (k = St}t>o 1.o.. that M* and M have doubly infinite time (i.St <.54 By (6. the arrival times 0 < 0'1 < Q2 < .Su<0.o<u<t where a claim arrives at time t and has size Uo. in (oo. Proof of Theorem 6.0<u<t) = P(StEA. h. A standard argument for stationary processes ([78] p. 6.l.Su< 0. { Su}0<u<t is distributed as a process {Su} .0<u<tIAt) = P(St EA.$St_ u.Q_k and has size U_ k. oo)).5). CHAPTER H. SOME GENERAL TOOLS AND RESULTS V` (0) = E E Uk k: ak E [0.0<u<t) = P(St EA.. oo ) and the arrival times 0 > 0_1 > a_2 > . We then represent M by the mark (claim size ) Uo of the arrival at time 0.St*_ u. The last property is referred to as insensitivity in the applied probability literature.. the mark at time Qk is denoted by Uk. and the kth preceding claim arrives at time t .. 0<u<t) = P(St EA. which makes an upwards jump at time .e.5. 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..o. Now conditionally upon At . oo) x (0 . It follows that for A C (0. oo) p(t) = P(St EA. The sample path relation between { Su } and { Su } amounts to S„ = St . in (0.0<u<tIAt) = P(St EA. T+ = t given the event At that an arrival at t occurs .. .g.
NIt)dt .s. G' (A) = 3 f P(St E A. and we let L(dy) be the random measure L(A) = fo°° I(St E A. 6. t a oo. the left endpoint of the support is oo. Mt)dt = i3EL(A) o"o .5. and since by assumption St * oo a. Fig. THE LADDER HEIGHT DISTRIBUTION 55 { A Su}0<u<t U0 U0 \t tt u>0 N U_1 Figure 6. Since So = U0. Uo]. cf. 0 < u < t } is the event that { Su } has a relative minimum at t .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.5 where it = { St < Su.6. Thus. In Fig. 6. the support of L has right endpoint U0. time instants corresponding to such minimal values have been marked with bold lines in the path of { St}. 2 therefore immediately shows that L(dy) is Lebesgue measure on (oo.. A sample path inspection just as in the proof of Lemma 6 .
[147].6 is Bjork & Grandell [67]. A further relevant reference related to Corollary 6.2.1).56 CHAPTER II.5 is due to Schmidt & coworkers [48]. . [263] (a special case of the result appear in Proposition VI. SOME GENERAL TOOLS AND RESULTS = OE f 0 I(Uo>y)I (yEA)dy = Q f IP (Uo>y)dy A 0o a fA P(y) dy• 0 Notes and references Theorem 6.
say. Panjer's recursion ( Corollary XI. St = uRt = EUi t.6) and simulation methods ( Chapter X). U2. 3). For finite horizon ruin probabilities . exact matrixexponential solutions under the assumption that B is phasetype (see further VIII. {Rt} and the associated claims surplus process {St} are given by Nt Nt Rt = u+t EUi.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. i=1 i=1 An important omission of the discussion in this chapter is the numerical evaluation of the ruin probability. i. with common distribution B. and assume that • { Nt}t>o is a Poisson process with rate j3. . A common view of the literature is to consider such processes as perturbed compound Poisson risk processes . • the premium rate is p = 1. It is worth mentioning that much of the analysis of this chapter can be carried over in a straightforward way to more general Levy processes .. • the claim sizes U1. 4. Thus . i. being of the form Rt = Rt+Bt + Jt where {Rt } is a compound 57 . are i. and independent of {Nt}.d. Some possibilities are numerical Laplace transform inversion via Corollary 3.4 below .. see Chapter IV.e.
Proof It was noted in Chapter I that p .1 is the expected claim surplus per unit time. For (c).1).f. and this immediately yields (a).s. we get Ee8st = 00 e8t c` Ee8 (U1+. [324].Rt. A more formal proof goes as follows: Nt r Nt ESt = E > U k .g. of the claim surplus St . THE COMPOUND POISSON MODEL Poisson risk process. e .6pBa).1) .+Uk)P(Nt = k) k=O e8t k=O B[s]k .t = t(p . (b) Var St = t. m.1). where K(k) (0) is the kth derivative of is at 0.u .t = E E [ U k k=1 k=1 Nt . say stable Levy motion. and Schlegel [316]. and that B(k)[0] = Pak).)3t (fit' k t} = etk(8) exp {st '3t + B[s]f Finally. cumulants .t = E[Ntµs] . 0 . P = PAB = 1/(1 + rl) Proposition 1. {Bt} a Brownian motion and {Rt} a pure jump process. for (d) just note that the kth cumulant of St is tic(k) (0). We do not spell out in detail such generalizations. Write pB^) = EUn' YB = Pali = EU.g. 1 Introduction For later reference.58 CHAPTER III..1) = t(p . Schmidli [319].t = fltpB . (c) Ee8St = et" (8) where c(s) = f3(B[s] .1 (a) ESt = t(13µ$ .'s etc. Furrer [150]. we shall start by giving the basic formulas for moments. Dufresne & Gerber [126]. 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).. See e. (d) The kth cumulant of St is tf3p(k) for k > 2.
. (d) If 17 = 0. then St> SnhV>Snhh. For example. lim supt. We return to this approach in Chapter V. we need the following lemma: Lemma 1. The connections to random walks are in fact fundamental. 1. u + v]. so that {Sok } is a random walk with mean EUET = EU. cf. v > 0.4.. meaning that the increments are stationary and independent. we get a discrete time random walk imbedded in the claim surplus process {St}.1 = . (b) If 77 < 0. obviously 0(u) = F(maxk Sok > u). the Uk .. then St 00. The right hand inequality in (1.. In particular.. we have Sok .2 (DRIFT AND OSCILLATION) St/ta3'p1 ast >oo. Proof We first note that for u. II.Tk. then St 4 co. rather to view {St} directly as a random walk in continuous time. then lien inft. where Tk is the time between the kth and the (k . (a) No matter the value of 77. St = oo.i. Obviously.3 If nh < t < (n + 1)h. 2.1)th claim. In this way.d. For the proof.1.1 is the same as if {St} was a random walk indexed by t = 0. The point of view in the present chapter is.S„ attains its minimal value when there are no arrivals in (u. (c) If 77 > 0. if t = nh + v with 0 < v < h. which is often used in the literature for obtaining information about {St} and the ruin probabilities. INTRODUCTION 59 The linear way the index t enters in the formulas in Proposition 1. Sn+0 . Here is one immediate application: Proposition 1. then Snh . and there are at least two ways to exploit this: Recalling that ok is the time of the kth claim.Sok_l = Uk .3EU01 = 1µs where rt is the safety loading.3) is proved similarly.Tk are i. Indeed.V.h < St < S(n+1)h + h. however. St = oo. S„+V > S„ . and the value is then precisely v.
and < 1 for all u when 77 > 0.1. then {St} upcrosses level 0 a.3. Assuming that each risk generates claims at Poisson intensity /3 and pays premium 1 per unit time. u 307).. However. 1 since St 4 oo) and repeating the argument..t . if P(M > 0) = 1.. Snh u = 00 (the lemma is not needed for (d)). This contradicts u St400. Part (d) follows by a (slightly more intricate) general random walk result ([APQ]. If rl > 0. or by a general result on discrete skeletons ([APQ] p. For any fixed h.2.4 The ruin probability 0(u) is 1 for all u when 77 < 0.p. Considering the next downcrossing (which occurs w.1).60 CHAPTER III. 0 Snh = 00.. 2h.. THE COMPOUND POISSON MODEL Proof of Proposition 1._.6 Often it is of interest to consider size fluctuations. 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 . at least once. The general case now follows either by another easy application of Lemma 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).o. h. hence by induction i.1. is a discretetime random walk for any h > 0.. {Snh}n=o. Notes and references All material of the present section is standard. Snh/n a4' ESh = h(p . this case can be reduced to the compound Poisson model by an easy operational time transformation u T1(t) where T(s) = )3 fo M(t)dt. 169) stating that lim infra.. .2. where the size of the portfolio at time t is M(t). {Snh}n=o.1. Thus using Lemma 1. p. and hence by the strong law of large numbers. it suffices to prove 4'(0) = F(M > 0) < 1.2: Proposition 1.3. Proof The case of 17 < 0 is immediate since then M = oo by Proposition 1..1(b)) that the assertion holds as t 4 oo through values of the form t = 0. Remark 1 ..s. and hence it folz lows from standard central limit theory and the expression Var(St) = tf3pB (Proposition 1.. (c) are immediate consequences of (a). h A similar argument for lim sup proves (a). and (b). is a discrete time random walk.5 The limiting distribution of St . There is also a central limit version of Proposition 1. lim supn_. it is seen that upcrossing occurs at least twice. Corollary 1.
(2. but we shall be able to extract substantial information from the formula. Theorem 2. 1e.1) is not entirely satisfying because of the infinite sum of convolution powers. IV. we can rewrite the PollaczeckKhinchine formula as 00 (u) = P (M > u) = (1 . the ladder heights are i. B(x)/aB. n=0 (2. p < 1. the formula for the distribution of M follows . It is crucial to note that for the compound Poisson model. Note that the distribution B0 with density bo is familiar from renewal theory as the limiting stationary distribution of the overshoot (forwards recurrence time ). nevertheless.2. where G+ is given n=0 by the defective density g+ (x) = 3B (x) = pbo(x) on (0.6.6. equivalently. cf. 1 The distribution of M is (1. 11.just before ruin is again B0.1. Summing over n. 0 Alternatively. As a vehicle for computing tIi(u). and we shall here exploit the decomposition of the maximum M as sum of ladder heights. i. The expression for g+ was proved in Theorem 11. Combined with i/i(u) = P ( M > u). 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}.IIG +II)EG+ . we may view the ladder heights as a terminating renewal process and M becomes then the lifetime. and we further get information about the joint conditional distribution of the surplus and the deficit. d. Thus . oo ).1 provides a representation formula for 0(u). Note that this . The following results generalizes the fact that the conditional distribution of the deficit ST(o) just after ruin given that ruin occurs (i.e. This follows simply by noting that the process repeats itself after reaching a relative maximum.34 or A.IIG+II) (the parenthesis gives the probability that there are no further ladder steps after the nth ). We assume throughout rl > 0 or.P) E PnBon(u) . Fig... which we henceforth refer to as the PollaczeckKhinchine formula.1) representing the distribution of M as a geometric compound. Here bo(x) _ Proof The probability that M is attained in precisely n ladder steps and does not exceed x is G+ (x)(1 . [APQ] Ch.1. that r(0) < oo) is Bo: taking y = 0 shows that the conditional distribution of (minus) the surplus ST(o). The decomposition of M as a sum of ladder heights now yields: 00 Theorem 2 .
[62]. see for example [APQ]. 2 The joint distribution of (ST(o )_. Theorem A1.5. 1) and W has distribution Fw given by dFyy/ dB(x) = x/µB.2(a) is from Dufresne & Gerber [125]. ST(o) > y. in this setting there is no decomposition of M as a sum of i. (c) the marginal distribution of ST(o)_ is Bo .V)W) where V. Beekman [61]. cf.6. the PollaczeckKhinchine formula is often referred to as Beekman 's convolution formula. W are independent.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.5. ladder heights so that the results do not appear not too useful for estimating 0(u) for u>0.1 is traditionally carried out for the imbedded discrete time random walk. We assume rt > 0 throughout.62 CHAPTER III. V is uniform on (0. Feller [143] or Wolff [384]. (d) the marginal distribution of ST(o)_ is B0. 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). Again. As shown in Theorem 11. there is a general marked point process version. Theorem 2. . f +b (b) the joint distribution of (ST( o). However. Asmussen & Schmidt [49]. Theorem 2 . In the risk theory literature.just after ruin. see Schmidli [323] and references there. The proof of Theorem 11. (1 . cf. ST(o)) is given by the following four equivalent statements: B(z) dz. For the study of the joint distribution of the surplus ST(u)_ just before ruin and the deficit ST(„). (a) 11 (ST(o)_ > x. 7r(0 ) < oo) = Q 3 Special cases of the PollaczeckKhinchine formula The model and notation is the same as in the preceding sections. ST(o )) given r (0) < oo is the same as the distribution of (VW.i. 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. where it requires slightly more calculation. and the conditional distribution of ST(o)_ given ST(o)_ = z is Bo z) The proof is given in IV.6.d. cf. the form of G+ is surprisingly insensitive to the form of {St} and holds in a certain general marked point process setup.
p)pSe a ( l v)x = p( S .p. Bon is the Erlang distribution with n phases and thus the density of M at x > 0 is (1 . Thus . then. Integrating from u to oo. I.6.O)e(b0)x. Let r ( x) be the failure rate of M at x > 0.1)1 00 ( 1 .. and hence this overshoot has the same distribution as the claims themselves . The result can. also be seen probabilistically without summing infinite series . For a failure at x.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. Thus r(x) = S(1 . hence without memory. 0 . use Laplace transforms. the formula for P(O) holds in a more general setting.e.p.1 e ax = n1 (n .2 If B is exponential with rate S.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. B0 is exponential with rate S and the result can now be proved from the Pollaczeck Khinchine formula by elementary calculations . 3b Exponential claims Corollary 3. a further relevant reference is Bjork & Grandell [67]. 1 . the current ladder step must terminate which occurs at rate S and there must be no further ones which occurs w. 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. Alternatively.p) E pn S n x n. however .p) = S .3. the result follows . As shown in 11. But claims are exponential .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.
we use the PollaczeckKhinchine formula in Chapter IX to show that b(u) . E.1. then 24 1 V.3.+ >u. is ?7+ ( u).T+ <oo)+P(M> u.i(u) satisfies the defective renewal equation Z(u) = 1 . and weights 1/2 for each.+ = y yields P(M>u. THE COMPOUND POISSON MODEL In VIII. u + oo. if 3 = 3 and B is a mixture of two exponential distributions with rates 3 and 7.y)G+(dy) For the last identity in (3. 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 .3) below. Then the first term on the r. and conditioning upon S.1) For a heavytailed B.g. T+ <00) (3. Corollary 3.s.y)/3B (y) dy. (3.p + f u Z(u .64 CHAPTER III.T+ <oo). cf.+ <U.S.3).3 The ruin probability Vi(u) satisfies the defective renewal equation ik (u) = 6+ (u) + G+ * 0(u) = Q f B(y) dy + u 0 f u 0(u . the survival probability Z(u) = 1 . (3.4) zu P(M > u . u . 2 is one of the main classical early results in the area. (3.+ <u.3)). (Example VIII.p + G+ * Z(u) = 1 .2).2) Notes and references Corollary 3.y)f3 (y) dy. (3. 0 Proof Write o (u) as P(M>u) = P(S.3) Equivalently. II.h.4) can be derived by elementary algebra from (3.1 p pBo(u).3. (u) 35eu + 35e6u. (b) use stopped martingales . 3c Some classical analytical results Recall the notation G+(u) = f^°° G+(dx).4) is similar (equivalently. The case of (3.y)G+(dy ) = f U V(u . just insert the explicit form of G+.S. We mention in particular the following: (a) check that ip (u) = pe (60)u is solution of the renewal equation (3.
5). 191)./3B[s] which is the same as (3. In fact.4) can be derived by elementary but tedious manipulations. In view of (3.5). g. either of these sets of formulas are what many authors call the PollaczeckKhinchine formula. Embrechts. e.3 is standard .p)s .(3 .)3B[s]) (3.1 Bo[s] = f oc.p)2 3(1 . which yields the survival probability as 00 f u }t Z(u) = f f3eRtdt 0 from which (3. Also (3.5 The first two moments of M are 2 EM .5 can be found in virtually any queueing book. Of course.s . eau B(u) du = f PB 3PB SPB 0 o (3. The approach there is to condition upon the first claim occuring at time t and having size x .8) Proof This can be shown.P)PB 2(1 . Griibel & Pitts [132]./3B[s] .Ee8M) f ao e8' ( u)du = a8uP (M > u)du = 0 o 1 ( 1+ (1 ..s ..p) E p"Bo[s]" = 1 . it is not surprising that such arguments are more cumbersome since the ladder height representation is not used.(3B[s] 1 . 111112 or Feller [143].P)pB' (3.Ps s(. [APQ] pp. by analytical manipulations (L'Hospital's rule) from (3. SPECIAL CASES OF POLLACZECKKHINCHINE Corollary 3.3.PPB2) EM2 = PPB) + QZPBl 2(1 .f. Some relevant references are Abate & Whitt [2]. Corollary 3. for example. We omit the details (see. Griibel [179] and Thorin & Wikstad [370] (see also the Bibliographical Notes in [307] p.3 .7).3 .5) Proof We first find the m. .pBo[s] no (1 . Bo of B0 as m e8u B(du) = B[s] .g.7) and Corollary 3. see e. [APQ] pp. (3.g. 0 Notes and references Corollary 3.p)s s /3 . 206207).7) s +.6) 00 = (I .4 The Laplace transform of the ruin probability is 65 fo Hence Ee8M 00 e8uiP(u)du .p = (1 . numerical inversion of the Laplace transform is one of the classical approaches for computing ruin probabilities.
66 CHAPTER III. we may assume p = 1 so that the stated formula in terms of the survival probability Z(u) = 1 .u + 1 )]k = QZ(u) .6 If B is degenerate at p.9) shown for n .u)]k k! (1 L3) 1: e_O(ku) NIN (k (k .1)! k=1 u1 .u)]k1 ku+1) [/3( k . Z^ =eR(k.u)]k d 1 u) _ a) n ( du ( k! (1  .3+ 18+ J0 Z(uy).u/p)]k ko k! Proof By replacing {St} by {Stu/p} if necessary. THE COMPOUND POISSON MODEL 3d Deterministic claims Corollary 3./3Z(u . differentiation yields Z'(u) _ /3Z(u) which together with the boundary condition Z(0) = 1/3 yields Z(u) _ (1/3) eAu so that (3.1).y<1)dy 0<u<1 1 < u < oo uu ulhu 1a+/3 J0 uZ(y)dy U Z(y) dy 113+0 For 0 < u < 1.9) follows for 0 < u < 1.)3(1 .u) [N(k ./32(u .u) [p(k . eO('u) [)3(k .h.Q) 3e.1). differentiation yields Z(u) _ /3Z(u) .Q (k 1 k= n  [O(k . . of (3. Assume (3. then p) 1: ep(k u/. For n < u < n + 1.s.u)]k k! k0 The renewal equation (3.1 < u < n and let Z(u ) denote the r.3I( 0<y<1)dy Z(y)/3I(0<u.Q) k=0 k! E e0( = /32(u) .9).u) a)Qea" + (1 .4) for Z( u) means f lhu Z(u) = 1.z/'(u) takes the form Z(u) L^J L.
but will now be repeated for the sake of selfcontainedness.(9) is welldefined.rc(9) = . (4.Qe(Bo[a] . we set up . F and c.3B = . Formalizing this for the purpose of studying the whole process {St}.2) shows that the solution is Ox [O]0]. 0 Notes and references Corollary 3. and define rce by (4.4) . corresponding to a compound Poisson risk process in the sense that for a suitable arrival intensity 00 and a suitable claim size distribution BB we have no(a) = rc(a + 9) .2) (Here 9 is any such number such that r. (4. 00 the standard definition of the exponential family {F9} generated by F is FB(dx) = e°xK(e)F(dx). See also Iversen & Staalhagen [208] for a discussion of computational aspects and further references.d.2).r.4.4) works as well. B9(dx) = B[9] B(dx).4. The answer is yes: inserting in (4.(9).6 is identical to the formula for the M/D/1 waiting time distribution derived by Erlang [139]. The question then naturally arises whether ie is the c. We could first tentatively consider the claim surplus X = St for a single t. (4.1) .f.) The adaptation of this construction to stochastic processes with stationary independent increment as {St} has been carried out in 11.1) .f.g. it follows that Z(u) = 2(u) for n<u<n+1. in terms of the c. we just have to multiply (4.3) by t.1) or equivalently.1 that c(a) = /3(B[a] .f.g.a.g. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 67 Since Z(n) = 2(n) by the induction hypothesis. and thus (4. or equivalently BB[a] = B[^+ Repeating for t 54 1.a.3B[9]. co(a) = rc(a + 9) . say t = 1: recall from Proposition 1.f. 4 Change of measure via exponential families If X is a random variable with c. K(a) = logEe'X = 109f 00 eaxF(dx) = logF[a]. of F9.
t < T.1) and multiply from 1 to n). n) for a given n.f. G]. with T taking the role of n) is the analogue of the expression exp{8(x1 + • • • + xn) . in particular the expression (4. and thus (4.6) F(G) = Po (G) = EB [exp {BST + Ttc(0)} .2. (4.d. for G E FT. = exp {BST . G C {T < oo}. .t. the PBT) are mutually equivalent on.1. .. with common c.8) follows by discrete exponential family theory. THE COMPOUND POISSON MODEL Definition 4.g.t.3 and claim size distribution B. and define 09.8) By standard measure theory. Then the Xk are i. The identity (4.. BB by (4.d.FTn) = Q(SkTIn : k = 0. Z is measurable w.r.1 Let P be the probability measure on D[0. Then P(G) = Fo(G) = EB [exp {BST + TK(O)} .68 CHAPTER III. Xn). EeeBSt + tk(B) = 1.0e and claim size distribution Be. (4. .2 For any fixed T. Proposition 4.4).. (4.9) Proof We first note that for any fixed t.r. and PBT) the restriction of PB to FT. (4.Tic (0)} .7) Proof We must prove that if Z is FTmeasurable. But let Xk = SkT/n . oo) governing a given compound Poisson risk process with arrival intensity. (4. The following result (Proposition 4. . .i.10) . Then FB denotes the probability measure governing the compound Poisson risk process with arrival intensity.7) now follows by taking Z = eBST+TK(e)I(G) u Theorem 4 .5) for the density.(9)} (4. Ti(a)/n.S(k_1)Tln. Let FT = o(St : t < T) denote the o•algebra spanned by the St. replications from Fe (replace x by xi in (4.FT. it suffices to consider the case where Z is measurable w. and dP(T) dP^T) That is. the corresponding expectation operator is E9.i. G].5) for the density of n i. v(Xi.nr. then EBZ = E [Ze9ST _T"(9)I . .3 Let T be any stopping time and let G E FT.
.10). Now consider a general G. Ee [exp { BST +Trc(9)} I(G) FT)] = 1.ST) + (T . Thus. Then G E FT. so that PG = EeE0 [exp { 9ST+Trc(9)}I(G)I FT)] = Ee [exp { BST + rrrc(O)} I(G)EB [ exp {9 (ST . Letting T t oo and using monotone convergence then shows u that (4. t = T .1) .FT]] = EB [exp { BST + Trc(9)} I(G)] .r is deterministic.f. The behaviour at zero is given by the first order Taylor expansion c(a) r.(Y) = 13(B['Y] .9) holds with G replaced by GT. Then GT = G n Jr < T} satisfies GT E FT.9) holds for G as well.1(a). (0) + rc'(0)a = 0 + ES1 a = a (p .1 It is seen that typically) a ry > 0 satisfying 0 = r.r)rc(9)}I .5. 77 Thus. GT C_ Jr < T}. . 5 Lundberg conjugation Being a c. the typical shape of rc is as in Fig. Given FT.7 1 Some discussion further supporting this statement is given in the next section. (4. 5.g. c(a) is a convex function of a. according to what has just been proved. Thus by (4. subject to the basic assumption ij > 0 of a positive safety loading. and hence (4.7) holds. (a) rc (a) (b) KL(a) 'Y 'Y Figure 5. LUNDBERG CONJUGATION 69 Now assume first that G C Jr < T} for some deterministic T.1) _ 1 + a.
(5. the Lundberg exponent. the claim surplus process has positive drift > 0. b[s] = 5/(b . Note that KL (a) = /L (BL [a] .1) is precisely what is needed for one of the terms in the exponent . (5.4) ELS1 = #L(0) cf. 5. 5. Equation (5.1) .1(b).2 is B(7) = 1 + ^. THE COMPOUND POISSON MODEL exists . 5.a = i(a + 7). Taking T = r(u). (5.4) yields /3L = b and that BL is again exponential with rate bL =. Fig. It is then readily seen that the nonzero solution of (5. e.1) is known as the Lundberg equation and plays a fundamental role in risk theory . u It is a crucial fact that when governed by FL.2 s As support for memory.3.1(b).3.g.QL instead of /37 and so on in the following . Example 5 . An established terminology is to call y the adjustment coefficient but there are various alternatives around.2)) is 7 = 5/3. and (4. Thus.1 Consider the case of exponential claims. Fig.1) (or (5. G = {T(u) < oo} in Theorem 4. Thus B[7] = 6/.s).3) cf. we write FL instead of F7.70 CHAPTER III. an equivalent version illustrated in Fig.3. . we further note that ( 5. Lundberg conjugation corresponds to interchanging the rates of the interarrival times and the claim sizes.2) 7 Figure 5.
3 takes a particular simple form. 0 Theorem 5 .G+L)(x)) dx ry^+L) J 00 f 0 (1 .3 (THE CRAMERLUNDBERG APPROXIMATION) i'(u) . we can rewrite this as 0(u) = e"ELe7^(u). we therefore have ELe7t(u) + C where C ELe7 (00) = µ+) f e7(1 .4). G = {S. ST+ E A] . (5.Ce7u as u 4 oo. To this end. Letting e(u) = ST(u) . which shows that G(L) (dx) = e7xG +(dx) = e7x /3 (x) dx.2 (LUNDBERG'S INEQUALITY) For all u > 0. (5. (5.1.7) 0 and all that is needed to check is that ( 5.+ E A} in Theorem 4.t. PL ) with density 1 .8) .u be the overshoot and noting that PL(T(u) < oo) = 1 by (5. LUNDBERG CONJUGATION 71 to vanish so that Theorem 4.(oo) (in the sense of weak convergence w.e7x)G+(dx). Proof Just note that e(u) > 0 in (5. e(u) has a limit i.1 (5. where C .7) is the same as (5.(u)} . Since a7' is continuous and bounded.5) Theorem 5 . Then P(ST+ E A) = EL [exp { 7S?+} .r.ascending ladder height distribution and µ+ its mean. T(u) < oo] . V)(u) < e7u. T = T+.6) Proof By renewal theory.G+ L) (x) G+L) (x) IL(+) µ+L) L) where G+L) is the FL. V) (u) = P(T(u) < oo) = EL [exp {ryS.3.5.5).P Y j o' xeryxOB (x) dx /3k [Y] .6 ).1p . see A .1e. take first 0 = ry.
e. that 7 = S .12) Example 5 .1 = ^7 The accuracy of Lundberg's inequality in the exponential case thus depends on how close p is to one.7). Using (5.1)) and 7µ+L) = 'y/3 [7] 7 1/0 = /3B ['y] .8) yields +L) J0 xel'B ( x) dx (5.3 (this was found already in Example 5.1) (5. . THE COMPOUND POISSON MODEL In principle.72 CHAPTER III.1 . we get L where 00 (1 . Then 0(u) = pe(a_Q)u where p = /3/S. Noting that SIG(L)II = 1 because of (5. u . From this it follows.4).1 above) and that C = p. of course.11) so that I 7B ['Y](B[7]1) BI [7]Q VP (7) 72 7 (using (5.")G + (dx ) = 1  J0 00 3B(x) dx = 1p. or equivalently of how close the safety loading 77 is to zero. 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. but some tedious (though elementary) calculations remain to bring the expressions on a final form. this solves the problem of evaluating (5. (5.4 Consider first the exponential case b(x) = Seax.10) VW = JI c* e° (x) dx = a (B[a] .
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)
.
7) that Ee$(Amex /j)M _ 1p _ 1p Eo [s (0max 1 . the premiums exceed only slightly the expected claims.8µBo  Ss' u where 6 = µB/µBo = 2µa/µB 2) . [174]) shows that it may produce surprisingly good results.p = (/3max 0)µB.3 )1 } _ 1p 1 . p* _ . Though of course it is based upon purely empirical grounds. or equivalently that /3 is only slightly smaller than /3max = 1/µ8. we shall represent this situation with a limit where /3 T fl but B is fixed. (/3max .80 CHAPTER III. That is. Letting Bo be the stationary excess life distribution. Mathematically.Ps(/3max . Grandell [171] pp.PBo [s (/3max . numerical evidence (e. Proposition 7.1 As /3 f Nmax. 1924.1. 7c The heavy traffic approximation The term heavy traffic comes from queueing theory. THE COMPOUND POISSON MODEL These three equations have solutions 9/3µB2)3 30µa2)2 3µa2) (3) P+ (3) ' 0 . but has an obvious interpretation also in risk theory: on the average.(bA*)". and hence the ruin probability approximation is b(u) e(bAln)u.)3 )PBo µB  . the approximating risk process has ruin probability z. Notes and references The approximation (7.g. heavy traffic conditions mean that the safety loading q is positive but small./3)M converges in distribution to the 2a exponential distribution with rate S = B' Proof Note first that 1 .2) was suggested by De Vylder [109].1./3)] 1 . Proposition 1. cf.3 and Corollary 3.P . we have according to the PollaczeckKhinchine formula in the form (3.(3)2 P PB 2µB 2µB Letting /3* = /3/P.b(u) = p*e.3* /S./3)PBo PB .s(/3max .p + p { 1 1p ti 1 .p .
Q T /3max. the first results of heavy traffic type seem to be due to Hadwiger [184]. However .1 1 . and hence 2µ2B 1 . as well as it is needed for the interpolation approximation to be studied in the next subsection. then P(u) 4 e6„ Proof Write z'(u) as P((/3max . the term light traffic comes from queueing theory. [APQ] Ch. We return to heavy traffic from a different point of view (diffusion approximations) in Chapter IV and give further references there . in risk theory heavy traffic is most often argued to be the typical case rather than light traffic . the premiums are much larger than the expected claims .p _ 2rl11B PB p. In the setting of risk theory./3)u * v. but has an obvious interpretation also in risk theory: on the average .3) is reasonable for g being say 1020% and u being small or moderate.7. Numerical evidence shows that the fit of (7. VIII).0)u. .2 If .ryu .l3)M > (/3max . or equivalently that 0 is small compared to µB ./3)u). 7d The light traffic approximation As for heavy traffic . It is worth noting that this is essentially the same as the approximation (2) z/i(u) Ce. That is . VARIOUS APPROXIMATIONS FOR THE RUIN PROBABILITY 81 Corollary 7. This follows since rl = 1/p . we shall represent this situation with a limit where 3 10 but B is fixed.B AB ) 6()3max _'3) = However . The present situation of Poisson arrivals is somewhat more elementary to deal with than the renewal case (see e .g. These results suggest the approximation Vi(u) e6(0_. Mathematically.4) suggested by the Cramer Lundberg approximation and Proposition 6. obviously Corollary 7. u * oo in such a way that (3max . light traffic is of some interest as a complement to heavy traffic .ze a2unµB laB (7. while the approximation may be far off for large u. light traffic conditions mean that the safety loading rl is positive and large .2. Of course.p. Notes and references Heavy traffic limit theory for queues goes back to Kingman [230]. 2 provides the better mathematical foundation.
and hence 00 (U) /3pBBo (u) = 0 / B(x)dx. Asmussen [19] and references there.(3 10. 7e Interpolating between light and heavy traffic We shall now outline an idea of how the heavy and light traffic approximations can be combined. ao n=1 00 n=1 (u) P) anllBBon(U) onPaBon(u) • Asymptotically.5) u Proof According to the PollaczeckKhinchine formula. [97]. For a more comprehensive treatment. cf.T > u) = J o" B(x + u)/3eax dx . Omax max m.3 As . see Daley & Rolski [96]. i. 0 u Notes and references Light traffic limit theory for queues was initiated by Bloomfield & Cox [69]. Light traffic does not appear to have been studied in risk theory.u. .e. the Poisson case is much easier than the renewal case. z/' (u) convergence P(U . The crude idea of interpolating between light and heavy traffic leads to 0 (u) C1 . (7. u Note that heuristically the light traffic approximation in Proposition 7. THE COMPOUND POISSON MODEL Proposition 7. 10 ( u The alternative expressions in (7.3 is the same which comes out by saying that basically ruin can only occur at the F(U . U > u] = /3iE(U .T > u). 0(u) /3 J B(x)dx = /3E[U .5) follow by integration by parts.Q limIP ( u) + Q lim z/'(u) Amax &0 amax ATAm.82 CHAPTER III. Indeed. 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 . by monotone time T of the first claim . ( 3 J O B dx. En'=2 • • • = O(/32) so that only the first terms matters.= 1 aJ 1 a 0+ 1 = = p.u)+. Sigman [347]. Again.
COMPARISONS OF CLAIM SIZE DISTRIBUTIONS 83 which is clearly useless . Let OLT) (u) denote the light traffic approximation given by Proposition 7.6) is . [84]. Another main queueing paper is Whitt [380]. f / Qmax B(x)dx 00 eQmaxxdx 4/ Qmax 00 QmaxQ amaze" and the approximation we suggest is J B(x) dx = cLT(v) (say).3). to get nondegenerate limits . Instead. no empirical study of the fit of (7.VHT) ( ax QmQ ) h (B) ( . available. Thus ./3)) . ^IE) exist: 1 (B) HT QmsxQ hm J e e6" 2µE/µE2)'" = e(1 6)" =  Q1Qm.3n. with rate 1/µB = /3max. we see that the following limits HT) (u').O(E)(u) 1 (1 . ^ LT Q maxQ m"^ Qlo V LT) ( CHT(v) (say). where further references can be found .x .Wmax f(x ) dx + pee6mQ. ) M. Al . B(2). 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. . (7. Notes and references In the queueing setting . The adaptation to risk theory is new. 8 Comparing the risks of different claim size distributions Given two claim size distributions B(1). 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. . "/Qmex Cu) CLT(u ( /3max 0) + O16 CHT( U(Qmaz . that is. however.O0 M. Substituting v = u(. (U).3 and use similar notation for %(B) (u) = (u). the idea of interpolating between light and heavy traffic is due to Burman & Smith [83 ]. _(E) (u).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. one may hope that some correction of the heavy traffic approximation has been obtained.8. even if the safety loading is not very small. z/i(E) (u) = pe(QmaxQ)u.
Rather than measuring difference in size. B(') <i. B(') <d B(2)) if B(1)(x) < B(2)(x) for all x. In terms of the time to ruin. this ordering measures difference in variability. this implies St T(l)(u) > r(2)(u) for all u so that 17(I) (U) < oo} C_ {T(2)(u) < oo}. whereas (consider x2) B(2) has the larger variance. U(2) such that U(l) has distribution B('). B(2)) in the increasing convex order if f BM (y) dy < f 00 Bi2i (y) dy x x for all x. then Bill = B(2).ill(u) < V)(2) (U) for all u. . XI. the proof is complete. cf. Proposition 8. for more detail and background on which we refer to Stoyan [352] or Shaked & Shantikumar [337]. then . Proof According to the above characterization of stochastical ordering. In particular (consider the convex functions x and x) the definition implies that B(1) and B(2) must have the same mean. Proposition 8. an equivalent characterization is f f dB(') < f f dB (2) for any nondecreasing convex function f.1 is quite weak. most often the term stoploss ordering is used instead of increasing convex ordering because for a given distribution B. 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). Recall that B(') is said to be stochastically smaller than B(2) (in symbols. Bill is said to be convexly smaller than B(2) (in symbols.' 1)(u) < V)(2) (U) for all u. Finally. THE COMPOUND POISSON MODEL To this end. one can interpret f x°° B(y) dy as the net stoploss premium in a stoploss or excessofloss reinsurance arrangement with retention limit x. we have the convex ordering.6. Here convex ordering is useful: Proposition 8. Taking probabilities. equivalent characterizations are f f dB(') < f f dB (2) for any nondecreasing function f.84 CHAPTER III. U(2) distribution B(2) and U(1) < U(2) a. we shall need various ordering properties of distributions. then i. u Of course. or the existence of random variables U(l). B(2) and PB(1) = µB(2). B(' <.1 If B(') <d B(2). we can assume that 1) < St 2l for all t. In the literature on risk theory.2 If B(') <j. A weaker concept is increasing convex ordering: B(1) is said to be smaller than B(2) (in symbols. B(2)) if f fdB(1) < f fdB(2) for any convex function f.s.
2 is the following: Proposition 8. This u implies that D <. Corollary 8. from which the result immediately follows. A general picture that emerges from these results and numerical studies like in Example 8. Proof If f is convex. we have by Jensen 's inequality that E f (U) > f ( EU).3 If B(1) <.(1) (.u) = (1 _ P) E /3npnBo( 1):n(u) n=1 00 < (1. B(2).3 provides another instance of this.8.6 below is that (in a rough formulation) increased variation in B increases the risk (assuming that we fix the mean). The problem is to specify what 'variation' means.5 If '0(1)(u) < p(2) (U) for all u and a. and here is one more result of the same flavor: Corollary 8.1. then /'(')(u) < 0(2)(u) for all u. Bo1) <_d Bo2) which implies the same order relation for all convolution powers.. then B(1) <. The heavy traffic approximation (7. say to p. Proof Consider the light traffic approximation in Proposition 7.1 and µB at 1 so that the safety loading 11 is 10%.4) certainly supports this view: noting that. Then V. u We finally give a numerical example illustrating how differences in the claim size distribution B may lead to very different ruin probabilities even if we fix the mean p = PB. B. . A first attempt would of course be to identify 'variation' with variance. Hence by the PollaczeckKhinchine formula . B(2).4 Let D refer to the distribution degenerate at 'LB . Example 8. (D) (u) < O(B) (U ) for all u. and consider the following claim size distributions: B1: the standard exponential distribution with density ay. larger variance is paramount to larger second moment. we have Bol) (x) f ' B(1) (y) dy < ' f' B(2) (y) dy = Bo2) (x)• µ 85 I.p ) E /3"µ"Bo2)* n(u) _ V(2) (u) n=1 = Corollary 8. A partial converse to Proposition 8.6 Fix /3 at 1/1. with fixed mean. COMPARISONS OF CLAIM SIZE DISTRIBUTIONS Proof Since the means are equal.. it is seen that asymptotically in heavy traffic larger claim size variance leads to larger ruin probabilities.e.
0. with the hyperexponential distribution being more variable than the exponential distribution and the Erlang distribution less. 1/)(u. 9 Sensitivity estimates In a broad setting. B.1%.4142.86 CHAPTER III.01%.001 u0.9A2e'2r where A. 0.1358. and consider a = 5%.0' U0. A standard example from queueing theory is . B3 the comparison is as expected from the intutition concerning the variability of these distributions.4. Kluppelberg [234]. In terms of variances o2. A2 = 3. Let ua denote the a fractile of the ruin function. We return to ordering of ruin probabilities in a special problem in VI. = 0. THE COMPOUND POISSON MODEL B2: the hyperexponential distribution with density 0.lA. B4: the Pareto distribution with density 3/(1 + 2x)5/2. all distributions have mean 1.01%. which appears to be smaller than the range of interest in insurance risk (certainly not in queueing applications!). 11 Notes and references Further relevant references are Goovaerts et al. 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). 1%. For B1i B2..) = a. i. 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. One then obtains the following table: U005 U0. B3: the Erlang distribution with density 4xe2x.e'\1x + 0. and this is presumably a consequence of a heavier tail rather than larger variance. we have 0r3 = 2 < or2 = 1 < 02 = 10 < 04 = 00 so that in this sense B4 is the most variable.000. in comparison to B2 the effect on the ua does not show before a = 0. [166]. van Heerwarden [189]. the behaviour of which is governed by a parameter 9. Pellerey [287] and (for the convex ordering) Makowski [ 252]. Note to make the figures comparable.e. However.
t]. with 0 the vector of service rates at different nodes and routing probabilities.01/2u. and hence a _ e(60)u + u e(60)u = ( i + which is of the order of magnitude uV. Assume for example that 8 is known. SENSITIVITY ESTIMATES 87 a queueing network. obtained say in the natural way as the empirical arrival rate Nt/t in [0. and s(9) the expected sojourn time of a customer in the network.3. where the partial derivatives are evaluated at p = 1. and hence the effect of changing p from 1 to 1 + Ap corresponds to changing /3 to /3/(1 + Op) /3(1 .2 Consider a risk process { Rt} with a general premium rate p. a2/t).9. For example. Thus. the premium rate p and the claim size distribution B. the standard deviation on the normalized estimate ^/1' (the relative error ) is approximatively . while /3 = j3 is an estimate. Then ib = Pe(613)u. Then a p ao = 00 Qa/. a0 as ao 80 19P . Then if t is large . where Q2 = fl ( l2 1113 / _ Ou2v)2. we may be interested in a'/ap for assesing the effects of a small change in the premium. it follows that ' is approximatively normal N(0.e.19P a/ . 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. Thus at p = 1. Proof This is an easy time transformation argument in a similar way as in Proposition 1. Example 9. Then the arrival rate /3(P) for { R(P) } is )31p.Ap).(u) for large u. In particular . say estimated from data. if = a e(6A)u. a/3 0 .. or we may be interested in aV)/0/3 as a measure of the uncertainty on '0 if 0 is only approximatively known.1 Consider the case of claims which are exponential with rate 8 (the premium rate is one). Let R(P) = Rtli. Similar conclusions will be found below.1. In the present setting. increasing in u. i. u Proposition 9.
Similar notation for partial derivatived are used below.2) (see Remark 9.0 = t/'(u) and the CramerLundberg constant C.((dx ) = exp {Ox + (t(x) . x > 0 (9. (9.()YC = 1 +y/ /3 \ Q2 From this (9.6 below for some discussion of this assumption). mathematically a proof is needed basically to show that two limits (u * oo and the differentiation as limit of finite differences) are interchangeable.3.1 or Proposition 9. The most intuitive approach is to rely on the accuracy of the CramerLundberg approximation . we can rewrite the Lundberg equation as w(9+ y. Viei '0(. ^) . (9.w(6. (3+'y)PC (0+7. THE COMPOUND POISSON MODEL As a consequence. () Proof According to (9.10) below.5) are similar. but must look for approximations for the sensitivities 0. 4) (9 . so that heuristically we obtain '00 50ryu = Coe"u . e. Differentiating w. ()} p(dx) . namely that of a twoparameter exponential family of the form Bo. Consider first the case of 8/8/3: . 9. we cannot expect in general to find explicit expressions like in Example 9. /3 yields w e(e + Y. for the ruin probabilities .t.4).uypCe7u urypO.3) follows by straightforward algebra.6) As will be seen below. In the case of the claim size distribution B. this intuition is indeed correct. 5) (Q+'Y)[we(0+7.w(O. Consider first the adjustment coefficient y as function of 3.3 70 = 'Ye = = 7 /3(1 we(e +'y. but we shall concentrate on a special structure covering a number of important cases. 3) ( 9 .g.()(0 +'0) ' (9 . u Now consider the ruin probability 0 = 0 (u) itself. Of course. various parametric families of claim size distributions could be considered.r.3. it suffices to fix the premium at p = 1 and consider only the effects of changing . and the proofs of (9.88 CHAPTER III. Proposition 9.Owe (9.()wC(e. (. However .(/3 + y)we(9 + 7. () = log(1 + y//3).()^ 1 .^)] 1(/3+y)we(9+'y.3 or/and B. and write yp = 8y/8/3 and so on .
x)B(x)dx.(e"U = = wS(O. u 0 Then Z = z + F * Z and F is a proper probability distribution .3(x) dx.4t (U)e°`U = which are wellknown and easy to show (see e.x). () . z2(U) = e7" J u b(u .8) Letting cp = e0/e/3 and differentiating (9.8). the proof is complete. z2(u) _ 1 ^ e'ri`i7i( u . () . By dominated convergence. Combining these estimates . Hence by a variant of the key renewal theorem (Proposition A1. F(dx) = e'yy/3B(x)dx.2 of the Appendix ). it holds that 89 a ue ryu a/3 Q(1 P) 7C2 Proof We shall use the renewal equation (3. Further write de = [we (9 +'y. ()] exp {w (O + y. BarndorffNielsen [58]). Be. () . O} (9.we(9 .9.12)). we multiply by e7" and let Z(u) = elt" cp(u). u 0 Proceeding in a similar way as in the proof of the CramerLundberg approximation based upon (9. Z= zl + z2 where zl (u) = e7u J m B(x)dx.w(O.x) F(dx ) f u J C F(dx) = C as u 4 oo.8) (Section 5).4 As u oo.10) (9. But from the proof of Theorem 5. and alsoo zl(u) + 0 because of B['y ] < oo.([a] = exp {w(9 + a.w(9. () exp {w(9 + a. PF = (1 .St (U) Ee. (9. w((9 + a.11) Ee.x). SENSITIVITY ESTIMATES Proposition 9.w(9. 0(u) = /3 Ju"O B(x) dx + f 0 0(u .QB(x) dx.9) (9.g. ()} . ()} . we get p(u) = J "O B(x) dx + J U O(u . () . Z(u)/u a C//3PF where PF is the mean of F.C). we note the formulas Ee.3) for z/'(u).x)B(x) dx + J U W(u .3 (see in particular (5. 11 For the following.p)/C'y.
wc(O.x)f3 f ^[t(y) .2) holds.e7x/3 f 00 [t(y) .x).w (9. oo z2 (u) f C .w( (0.we (0. 01 (i+) do = +'Y. ()](e7v .6C do 89 1p 8( 1p Proof By straightforward differentiation.1) B(dy) 'f '[t(y) . ^)} [wc (0 + 7. u Z2(U) = e7° f u ^/i(u . 0 x Multiplying by e7" and letting Z(u) = e"uV(u). this implies Z = z + F * Z. ()]B(dy) dx x 0 0C T ON O .lB(x) dx = e7uzl(u) + e7°zz(u) + V(u T where zl (u) = .8) that cp(u) . ()} 1z(dy) = f [t(y) .12) f exp {O y + (t(y) .9)(9.11). z = zl + z2. 8^ ue7u. C)] (1 + 7 ) Proposition 9. ^) .6e7u f "o f[t(y) .5 Assume that (9. C) . )}B(dy)• Letting cp it thus follows from (9.w( (0. By dominated convergence and (9. THE COMPOUND POISSON MODEL [we(e+7. 2 z 07P N ue7u (3C de .wc(9.wc (O. 8 8() 8( (9.w(e.QB(x)dx. ()]e7vB(dy) 'fCd 7 c . ()]B(dy) dx. F(dx) = e7x.90 CHAPTER III. ())B(dy) dx.w(0. Then as u > oo.
() = log r(a) . and the proof of the first one u is similar.) log(9) = %F(a) logs where %1 = F'/]F is the Digamma function.2) holds with p(dx) = xldx.w((9. SENSITIVITY ESTIMATES as u 3 oo.. .12) takes the form y) alp a . we (9.ry) 5a1 cry (5 .Y)a+1 ' (9. Example 9.6 Consider the gamma density b (x ) = Sa xa. and also zj (u) 4 0 because of f Hence.13) (9.rye) S 5rya. Z(u) /3C 91 o c'o e11(t (y) . () = C/9 = a/S. It follows after some elementary calculus that p = a)3/5 and. 9 = S.1 ..17) (9.12) follows.1edz = 1 exp {Sx + a log x . t(x) = logx.a log S = log r(c) . ())B(dy) < oo. Here (9.C log(9)./35' a/i'y + aryl 625ry.15) (9. () ='I'(t.16) (9.yu/3C2do u86 89 1p' az/) = 8z/. < = a.(log r(a) a log S)} • r(a) 1.9. that C = a. w(e.3ary tog('Finally.14) de = d( 7!3 76 = 7e = log ( \ ( \5a_ / \SSry ) 72 .. by inserting in the above formulas.pa+1 . U 7µF from which the second assertion of (9. a /(S . We get w( (0. ( 9.QS 1 .18) (05 + 57 _'3_y .Sry a/32 + a/37 + /37 . ue_Yu 'C2d( 8a 8( 1 p .a/35a&y' ' (9.
C = . w(e.2a) } Thus the condition B[a*] > 1 + a* /. THE COMPOUND POISSON MODEL Example 9.CZ try)} 1 C C2 try .log c = 2 In particular. t(x) _ .22.1 = eXP {c(C .1 16 +ry c C22ry 2( = + 70 We (e.([Y] .9) (() .S[a] = exp {w (9 + a. ()} = exp {c (C . Be. C) . 9 = .2) with µ(dx) = 2x3zrdx.3Ee.l3 of Section 6a needed for the existence of ry becomes e^Q > 1+62 / 2. further yield . C) = B = Yc = de = do = .2 log (0.21og 2. Straightforward but tedious calculations . which we omit in part .7 Consider the inverse Gaussian density ( b(x) Zx37 exp This has the form (9.2 .w(9."62 . for a < a* = z (.92 CHAPTER III. () = Cc .3.
if 1 PT = /3TNT(U1+. That it is no restriction to assume one of the ti(x) to be linear follows since the whole setup requires exponential moments to be finite (thus we can always extend the family if necessary by adding a term Ox). we can just fix k . 10 Estimation of the adjustment coefficient We consider a nonparametric setup where /3. thus. Van Wouve et al. Also. sj=1 and let YT be defined by IKT('ryT) = 0. the main tool is simulation.10.1) . or Ct(x). (9.. In general.8 The specific form of (9. That it is no restriction to assume k < 2 follows since if k > 2. queueing networks) are typically much more complicated than the one considered here. BT [a]= NT ^` e"U.. Note that if NT = 0. Notes and references The general area of sensitivity analysis (gradient estimation) is currently receiving considerable interest in queueing theory.7 and references there. the exponent is either Ox. However .cue_7u)3C P Remark 9.a.. in which case we can just let t(x) = 0. and hence explicit or asymptotic estimates are in general not possible. then BT and hence ryT is undefined. ESTIMATION OF THE ADJUSTMENT COEFFICIENT Finally. the models there (e. we have assumed k = 2 and ti (x) = x. Thus. ae t 1lEY u S _ . in u which case the extension just described applies. To this end. Finally if k = 1.12) takes the form a = a 93 ar. B are assumed to be completely unknown. [379] consider a special problem related to reinsurance. then ryT < 0. and we estimate y by means of the empirical solution ryT to the Lundberg equation. the results presented here are new. by the LLN both F (NT = 0) and F (PT > 1) converge to 0 as T . to our knowledge. let NT 16T = ^T . the exponent of the density in an exponential family has the form 01 tl (x) + • • • + 9ktk (x).oo. Thus. kT (a) = /T (BT [a] . for which we refer to X.2) is motivated as follows. . However.g.+UNT) > 1.2 of the parameters. Comparatively less work seems to have been done in risk theory.3C2de 1p' z a = c .
THE COMPOUND POISSON MODEL Theorem 10 .B[7]2 n Hence ( 10.B[7]2 V2 .2) rT(7) N N (0.: N ()3.: N 0.B[7]) 0+ Iv/o(b[y].'s. 1) r.94 CHAPTER III.1) 'YT .)vl+ N CO.b[Yp'V21 T { (E[7] . we need a lemma.2) follows from NT/T a4' .'Y . it is easy to see that we can write \ V1 1 l _ .1)2 + E[27] .3) Proof Since Var(eryU) = we have B[7].2 As T * oo. For the proof.Q and Anscombe 's theorem.v. Lemma 10 . Hence KT(7) = (F' + (OT a(B[7l 0))((BT [7] . a2 where a2 = /3r. 7T a4' 7./^ B[27] . (10.1) . N ( n[7].a BT[7] I B[7] I + .7 + (. 16T where V1. More generally.B[7]2 }) ( T 0 .(27)/K'(7)2. B[2'Y]  /3T ) .If .3T . . vfoVFB[2y].B[7]) + B [7] . then (10.T y . If furthermore B[27] < oo. since NT /T . B [7]2 (10.1 As T 4 oo.3/ T).i3)(B[7] 1) + (3(BT[7]  .1) .. B[27] . V2 are independent N (0.
Combining ( 10. 7T E (y .4) + E). OT a 95 u 4 /3.E) < 4T(7T) < 4T(7 + E).3).2. If ryT E (7  we have KT(7 .1 can be used to obtain error bounds on the ruin probabilities when the parameters . Now write KT(7T)  kT(7) = 4T(7T)( 7T 7).'T(a) = 1 E Uie°U' a$' EUe "u = B'[a]. and the truth of this for all e > 0 implies ryT at 'y. By the law of large numbers. lcT(a ) 4 /c(a).e. To this end . NT BT [a] Hence r.(ry + e) and hence KT(7 .KT(7) kT(7) K'(7) . °7IT) . I.10.e) < 0 < r. 6"Y (10.. Let 0 < E < ry.E) < 0 < kT(7 + E) for all sufficiently large T .E ) < 4T(7T) < (7 +0' which implies 'T(ry4) a$' r. Then r.c'(7) N (0' T (2(7) / N (0.Q. y + E) eventually. ESTIMATION OF THE ADJUSTMENT COEFFICIENT which is the same as (10. Theorem 10.(ry . it follows that 7T7 KT(7T) .4) and Lemma 10. Proof of Theorem 10. where ryT is some point between ryT and ry.1 By the law of large numbers. 0 are estimated from data . BT[a] 3 B[a].'(y). first note that e7TU N (e7U u2e27Uo'2/T) 7 .e. NT i =1 n'(a) for all a so that for all sufficiently large T K7 .
.. 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.i.3 or equivalently p > 1/2 or 11 < 100%. Embrechts & Mikosch [133]. 6 < 2.1 is from Grandell [170]. = 1..5%). ft. For this reason . the nth busy cycle is then [Wn1. Notes and references Theorem 10.) = a (e.f.g. and the known fact that the Y„ = max Vt tE[W„1.g.96 CHAPTER III.T = 3TKT ( 21T)IKT (^T)2 is the empirical estimate of vy and fc. satisfies b(.info< „< t S. Frees [146]. For example .. if B is exponential with rate 8 so that ry = 8 . Griibel & Embrechts [292].96 if a = 2. V..e.e. Hipp [196]. A major restriction of the approach is the condition B[2ry] < oo which may be quite restrictive. various alternatives have been developed. Vt = St . i ...0) < 5. i.d. Herkenrath [192]. > 0 for some t E [Wn_ 1..T VIT where r7ry. it means 2 (8 .ueryuU ". This approach in fact applies also for many models more general than the compound Poisson one. Deheuvels & Steinebach [102].1 : Vt = 0. t]}. U2. 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. wn = inf{t > W. Further work on estimation of y with different methods can be found in Csorgo & Steinebach [94].Q. Csorgo & Teugels [95]. [197].Wn) are i . Wn). with a tail of the form P(Y > y) . Letting Wo = 0.C1e"a ( see e. Mammitzsch [253] and Pitts. Asmussen [23]) can then be used to produce an estimate of ry..
Further let 'Yo be the unique point in (0. the Poisson intensity is 0 and the claim size distribution is B with m. it is assumed that i > 0 and that the adjustment coefficient (Lundberg exponent) y. See Fig. Unless otherwise stated. the premium rate is 1. 97 .1 where p = 13µB. exists. The safety loading is q = 1/p . 'y) where c(a) attains it minimum value.f. 0.s. In particular.g. T) = P( /r(u) <T) \ = PI inf Rt <OIRo=u1 /\0<t<T PI sup St>ul 0<t<T Only the compound Poisson case is treated. The notation is essentially as in Chapter III.Chapter IV The probability of ruin within finite time This chapter is concerned with the finite time ruin probabilities 0(u. B[•] and mean AB.1 (the role of ryy will be explained in Section 4b). generalizations to other models are either discussed in the Notes and References or in relevant chapters.1) . defined as solution of c(ry) = 0 where ic(s) _ /3(B[s] .
PL = 6/0 = 1/p > 1). (u) is exponential with rate 0 w. using that the overshoot l. Var[T(u) I T(u) < 00] = VarL T( U) . FL and independent of T(u).1 In the compound Poisson model with exponential claims with rate S and safety loading 77 > 0.u is the overshoot. 1 FL.98 CHAPTER IV.5 .(4. By the likelihood identity III.) = e7u ELe'Y^(u) ELT(U)k = e'Yu b ELT(u)k = O(u)ELT(u)k.(U) < 00] = ELT(u)ke'YS. In particular.. 7. E[T(u) I T(u) < 00 ] = ELT (U). EL refer to the exponentially tilted process 3 with arrival intensity S and exponential claims with rate / (thus . PROBABILITY OF RUIN IN FINITE TIME Figure 0. 1 Exponential claims Proposition 1. the conditional mean and variance of the time to ruin are given by E[r(u) I T (u) < oo] Var [T ( u) I T( u) < oo] /3u+1 J )3 _ 2/3Su+/3+S (S)3)3 (1. the time of ruin is T(u) and ^(u) = ST(t&) .r. 2 that E [T(u)k.1) (1. . we have for k = 1.2) Proof Let as in Example 111.(.9).1 The claims surplus is {St}.t.
12 Thus the l.(yo) = 2V .6a = 0 with solution 0 (the .1)T(u) are independent with QL the same mean . This means that /3(6/(6 .2) is aLELT( u) .1 (6)3)2 which is the same as the r. u + ELe(u) _ PL .B = a. T(u) < oo] fora > r.V/ is as asserted.3) B = 0(a) = + (6/3a)2+4a6 2 and hence that the value of ic(yo) Proof It is readily checked that yo = 6 .h. where = eBu I 1 .h. Wald's second moment identity yields 2 EL (Sr(u) ./3 .6. the 1. Let 0 > yo be determined by ^c(0 ) = a.1) . the Laplace transform of the time to ruin is given by Eea7( u) = E [eaT (u).(PL .h.2 In the compound Poisson model with exponential claims with rate 6 and safety loading rl > 0.1)ELT(u).1 /3u + 1 u + 1 //3 = 6/3 6/01 For (1. is V1rLSr( u) +VarL ((PL .s.I (1.1//32 (6/)3 1)2 26(/3u + 1)/(6 .1)2VarLT(u) + 2 Ca 1I VarLT(u).1)T(u))2 = UL where = s. we have by Wald's identity that (note that ELSt = t(pL .2). 1). 0 Proposition 1.s.1)) ELST(u) ELT(u) (PL .6 + a)0 ./3) .1)T(u)) = VarLe(u) + (PL .1. Since Sr (u) and (PL . of (1. which leads to the quadratic 02 + (/3 .0) .s. EXPONENTIAL CLAIMS For (1 . ."(ry) = 26//32.
9ST(u) +T(u)!c(0)} .3.. More precisely.4.v.100 CHAPTER IV. Using 5 = 6 .v.'s with rate 5..4) The interpretation of this that T(u) can be written as the independent sum of T(0) plus a r.. . T(u) < oo] = EB [exp {aT(u ) .1 . Note that it follows from Proposition 1..0. Y(u) belonging to a convolution semigroup .Y1 Y2 Figure 1. M(u) T(u) = T + E Tk k=1 where T = T(0) is the length of the first ladder segment . are the ladder heights which form a terminating sequence of exponential r. the result follows.T+ Ti a t U T I 1 a i F. Ti.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. (1...3 that we can write EeaT( u) = eeuEe 017(o). But by the fundamental likelihood ratio identity ( Theorem 111. Cf. are the lengths of of the ladder segments 2. T(u) < oo] = e. Y2.3) we have E [e«T(u ).. and M(u)+1 is the index of the ladder segment corresponding to T(u).1 where Y1. . St Ti F.. PROBABILITY OF RUIN IN FINITE TIME sign of the square root is + because 0 > 0). 1. Fig. T2 .
Since U1 .d.1 )!. .. 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 . U2. For j = 0. UN. the following formula is convenient by allowing t.I ex cos B cos j O dB fo " . . T) to be evaluated by numerical integration: Proposition 1.. Proof We use the formula . then VT = U1.T) 1 I fl(O)h(0) fdO where (1. T. 1).3 Assume that claims are exponential with rate b = 1.v. UN.T.ST). Corollary 11.cos (u/. density xN lex/(N .T are conditionally i. Let {Qt} be the queue length process of the queue (number in system.i (u.1.T the service times of the customers awaiting service .T is the residual service time of the customer being currently served and U2 . EXPONENTIAL CLAIMS 101 For numerical purposes . EN has an Erlang distribution with parameters (N.4. cf.. including the customer being currently served).6) fl(9) f2(0) = = fexp {2iTcos9(1+/3)T+u(/cos91) cos (uisin9) . 2. If QT = N > 0. where U1.6.1.6(u) = Vfl/j l(Su..T + • • • + UN. the conditional distribution of VT given QT = N is that of EN where the r.T) = P(VT > u) where {Vt } is the workload process in an initially empty M/M/1 queue with arrival rate 0 and service rate S = 1.T..1(u.. [4]) 00 (x/2)2n+3 Ij (x) OnI(n+j)! .. and exponential with rate S = 1.e. . Then V(u.T. .k + 1).3 sin0 + 29) f3(0) = 1+/32/cos9. .i. i.0. let (cf. Note that the case 6 # 1 is easily reduced to the case S = 1 via the formula V.
ie . PROBABILITY OF RUIN IN FINITE TIME denote the modified Bessel function of order j.1 00 ok+lR 00 j=k1 +1)/2e .1 R [.44). 8789) 00 E aj j= 00 = 1. k k2 + $k+1 E bj 00 t j . f3(0) . (1. 912.i(k +1)e R [/3( klal/2e:0 (01 /2 e .3(k +1)/ 2ei(k + l)6 (.(31/2 cos (( k + 2)9) .8 ) yields F(QT > k + 1) .cos((k + 2)9)] d9.13(k +l)/2ei(k +1)9 R E . 00 E '3j/2 cos(je) j=k+1 00 _ j=k+1 ^j/ zeij = .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 .k + 1) = 1 k +1 + bj j=00 j=00 00 j=kk+1 j=k1 By Euler 's formulas.1)] L _112 /(k+1)/2 [. in particular equations (1.38). Then (see Prabhu [294] pp.)3k +1 tj g'(QT >. and define tj = e(1+R)Taj/2Ij(2vT T).(31/2eie . similar formulas are in [APQ] pp. let I _ j (x) = Ij (x).cos (( k + 1)0)] f3(9) Hence the integral expression in (1.112 l 1( k +1)/2 [ 31/ 2 cos(kO) .102 CHAPTER IV.31 /2eie L 1)] 1 I/31/2eie .)3k+1 = e(1+0)T e201/2Tcos 7r 0 e )3(k +l)/2 [31/ 2 cos ( kO) .
THE RUIN PROBABILITY WITH NO INITIAL RESERVE Since P(QOO > k + 1) = flk+1. or.7) that _ [^ au ak+l (30 k L. there are several misprints in the formula there. We allow a general claim size distribution B and recall that we have the explicit formula z/i(0) _ P(7(0) Goo) = p. We first prove two classical formulas which are remarkable by showing that the ruin probabilities can be reconstructed from the distributions of the St. Ui < x I / (note that P(St < x ) = F(x + t. from the accumulated claim distribution N. T) in terms of F(. 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. equivalently. the numerical examples in [12] are correct). however. F(x. E Fk. u Notes and references Proposition 1.2. Seal [327] gives a different numerical integration fomula for 1 . The rest of the proof is easy algebra. k! k=O k0 i/z Co Uk ate" o'/z e . oo (u)31/2e^e)k = )3k z cos(k9) = R k. is numerically unstable for large T.T) which. we are concerned with describing the distribution of the ruin time T(0) in the case where the initial reserve is u = 0. T). expresses V)(0. Related formulas are in Takacs [359]. going back to Cramer. k=0 103 Cu) A further application of Euler's formulas yields cc k =0 k 'ese)k __ U #kJ2 cos((k + 2)9) = R eNO ^` (u^1 L k= = eup i/z L OI = =ateU161/2 e '0+2iO COS a cos(u(31/2 sin 9 + 20). t )). The first formula.0(u. . 2 The ruin probability with no initial reserve In this section . t) = P . however.3 was given in Asmussen [12] (as pointed out by BarndorffNielsen & Schmidli [59]. it follows as in (1..e = e' COS a cos(uf31/2 sin 0).
Stv^ _ Define M(v.3) with A = (0.S„ 0 <t<Tv STS„+St_T+v Tv<t<T as the event that IS. T T o where the second equality follows from II. we define a new claim surplus process St StM NJ Figure 2. See Fig. f T lStv)} 0<t<T by a 'cyclic translation'.T)dx.i. v].1 In formulas.T))dv E^T I(M(v. resp.(.(0.1. PROBABILITY OF RUIN IN FINITE TIME Theorem 2 .T)) does not {Stv)} depend on v.T) T F(x.104 CHAPTER IV. [v. co ).b (0. ") } is at a minimum at time t.(6. Then 1 . 1 1 . meaning that we interchange the two segments of the arrival process of {St}o<t<_T corresponding to the intervals [0. 2.T))dv.0<w<t} St+v .T]. T].T)) 1 fT P(M(v. Proof For any v E [0. . and the third from the obvious fact (exchangeability properties of the Poisson process) that has the same distribution as St = { Si0)} so that P(M(v.t)= {Stv) < SM. T) = P(Tr(0) > T) = P(M(0.
v)) dv = ST T T o (note that the Lebesgue measure of the v for which {St} is at a minimum at v is exactly .T) and Sv < 0 on M(0. v < t < T} n M(0. w) for some small E. THE RUIN PROBABILITY WITH NO INITIAL RESERVE 105 Now consider the evaluation of fo I(M(v.t)dt. Hence T TE f I( M(v. then i fT I(M(v.Sv. t) denote the density of F(•.T)f(I z /)(0. in which case there is a last time o where St downcrosses level u. We claim that if M(0.v<t<T} = {ST<StSv.Tt))f(u+t.2. If ST < 0. v). T)) dv = TEST = T fP(ST < x) dx T T NT 1 f P(ST < x) dx = 1 f P Ui T .2.T)) dv f T I(M(0. T) as {ST<St+ vS. For example. T) occurs. T].T) occurs. . Indeed.2 10(u. T T o i =1 Let f (•. Obviously.T) occurs or not as long as ST < 0. It follows that if M(0 . we can take v E (w E. v). then M(v. we can write M(v. v<t<T}n{ST<STSv+St. T) = M(0. cf. ST > 0. v). 0<t <Tv}n{ST<ST Sv+St T+v. Fig 2.. v) = M(0. T Theorem 2 . T)). this integral is 0 if STv) .ST on M (0. It is then clear from the cyclical nature of the problem that this holds irrespective of whether M(0. T) occurs. 0<t<v} = {ST < St . Proof The event {ST < u} = { Ei T Ui < u + T j can occur in two ways: either ruin does not occur in [0. T)) dv. where the last equality follows from ST < St on M(0.T) = F(u+T. T. letting w = inf It > 0 : St_ = mino<w<T Sw}.xdx.. t). or it occurs. there exist v such that M(v.
ST_ _ z} .T) = . ST_ _ z}. {S t > z. 0 < t < T. 0 < t <T. Proposition 2.t). Then P(T(0) E • I T(0) < oo) = P(T_ (Z) E •).T)+ J0 T (1V.2.b(u. u which is the same as the assertion of the theorem. which occurs w. u + dt] and there is no upcrossing of level u after time t. PROBABILITY OF RUIN IN FINITE TIME u Q II T Figure 2. {St > .T) = C(z. ST_ _ z}.p.(0.2 Here o. C*(z.T) = {St < 0.u+dt]). Let Z be a r. t + dt] occurs if and only if St E [u. E [t.ST_ t_ and let A(z.2 . The following representation of T(0) will be used in the next section. For a fixed T > 0.3 Define r_ (z) = inf It > 0 : St = z}. define St = ST . Proof of Theorem 111.2. which is independent of St and has the stationary excess distribution B0. O(T .z. Hence P(ST<u) = 1 .Tt))P(StE[u. 2. The proof is combined with the proof of Theorem 111.v.106 CHAPTER IV. z > 0. 0 < t <T .
1) yields P(ST(o)_ E [z.3 But by sample path inspection (cf. THE RUIN PROBABILITY WITH NO INITIAL RESERVE Then 107 P(r(0) E [T. Hence integrating (2.2.T + dT].ST(o) >y. Figure 2. 7( 0) < oo) = P (C(z)) dT.T)).T))f3B(z) dz dT.T) = C*(z.1) z T .T(0)<oc) = f x F(U > y + z U > z) P(Sr(o)_ E [z. z + dz]. T(0) < oo) = OR(z) dz in (2.1) that P(T(0) E [T. r(0) < oo) = 3R(z) dz JP(C(z.T))dT = Off(z) dz P(T_ (z ) < oo) = 3B(z) dz. {St }o<t<T have the same distribution . (2. Thus P(Sr(o)_>x. . T(0) < oo) B(y B(z) + z) f3B(z) dz = 3 f °^ B(y + z) dz = f3 + x v f B(z) dz. u which is the assertion of Theorem 111.2. z + dz]) = P(A(z.3.T). z + dz]. It follows by division by P(ST(o)_ E [z. T + dT] I S7(o)_ E [z. z + dz]. 2.3).T)) = P(Cx. z + dz]. Fig.2. Proof of Proposition 2. we therefore have P(A(z. ST(o)_ E [z. and since {St}o<t<T. A(z.
(yo).c(r(a)) l = l er( a)se+at } u yields 1 = eyr(a)Eear(y). 3 Laplace transforms Throughout in this section. r(a) denotes the solution < 'Yo of the equation a = ic(r (a)) = .6.1.3 was noted by Asmussen & Kl(ippelberg [36].s. Theorem 2. cf. In the setting of general Levy processes.108 Hence CHAPTER IV.1 Eear( y) = eyr(a). Notes and references For Theorems 2. [329].r(a). PROBABILITY OF RUIN IN FINITE TIME ]P(7(0) E [T. a martingale proof is in Delbaen & Haezendonck [103]. Proof Optional stopping of the martingale I er (a) 9 t.T + dT] T(0) < oo) dT f ' P(C(z))P(Sr( o)_ E [z. some relevant references are Shtatland [338] and Gusak & Korolyuk [181]. Lemma 3 .2. T(0) < oo) = dTP(T_(Z) E [T.1) where a > r.2 ga(x) = Qexr(a) f "o eyr(a)B(dy) x . z + dz]. (3. ^(0) E dx] (recall that ^(0) = Sr(o)) and write ga[b] = f OD ebxga(x) dx. one based upon a result of Asmussen & Schmidt [49] generalizing Theorem 11.(3(B[r( a)] .1 and the present proof is in the spirit of Ballot theorems.3. see in addition to Prabhu [293] also Seal [326]. because of77>0. Proposition 2. 2. Tak'ecs [359]. z + dz]. r(0) < oo. who instead of the present direct proof gave two arguments. Note that T_ (y) < oo a.5 and one upon excursion theory for Markov processes (see IX.T+dT]). I L Let ga(x) be the density of the measure E[ear(°).5a).1) . T(0) < oo) 0 = dT f 0 P(C(z))P(Z E [z. Lemma 3. Let T_ (y) be defined as Proposition 2.
time T(u): u u Here is a classical result : the double m. u . LAPLACE TRANSFORMS 109 Proof Let Z be the surplus .f.r(a) The result now follows by inserting /3B[s] = ic( s) +/3+ s and ic(r(a)) =a.r(a) b .3. E[ear (o) I T(0) < oo .3.3 ga[b] = c(b) Proof + b + a .2. Further by Theorem 111.T(0) < oo] = 20[b] = za[b] (9a[b] 9a[0])/b 1 . y + dy].g.ga [b] 0 TO Using Lemma 3. It is then easily seen that Za(u) is the solution of the renewal equation Za (u) = za (u) + fo Z.r(a) = a [B[b] B[r(a)]] .1] evr(a)B(dy)[ b .ST(o)_ just before ruin . b .°° ga(x)dx.(v) = ev''(a). (Laplace transform) of the ruin Corollary 3. T(u) < oo] du = Proof Define Za(u) = E [eaT(" ).5 f 00 o a/r(a) . Hence eb"du E[eaT("). rr(0) < oo) = 1_ r(a) Proof Let b = 0. the result follows after simple algebra. (u .4 E[eaT (o). Corollary 3.2 P(Z E [y.3.x)(a) B(dy)• Lemma 3 . Then by Proposition 2. r(u) < oo).x)ga (x) dx where za(u) = f. £(0) E dx) = /3B(x + dy) dx and hence ga(x) = f e r)/3B(x + dy) _ /3 f x e(v.ic(b)/b x(b) + a eb"E[eaT(" ).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)) . Z = y] = EeaT.ga [b] 1 .
St/t 1 1/m. Theorem 4 .s.2. PROBABILITY OF RUIN IN FINITE TIME 4 When does ruin occur? For the general compound Poisson model. Then as u * oo.UProof The assumption 11 < 0 ensures that P(T(u) < oo) = 1 and r(u) a4' oo. (4. Later results then deal with more precise and refined versions of this statement. where _ 1 _ 1 1 C ML w(ry) 6B'[7J 1 . for any c > 0 P( Further. T(u) a.00 St = lim . The first main result of the present section is that the value umL.1 Assume 77 > 0.3). This proves the first assertion of (4. mu ) ( 0 m < ML '(u) 1 m > rL..2 Assume ri < 0. u 1 ET(u) 1 p1 u where Pw2 = 311B)m3• 7(u) . we need the following auxiliary result: Proposition 4. = (p .110 CHAPTER IV. P = /3µB > 1. For the proof. t T(u) T(u) T(u) t m = lim = lim = lim Utioo u + Sr(u) u+oo S.1) i.(u ) = o(u) a. the known results are even less explicit than for the exponential claims case.1)Er(u) ..e. note that by Wald's identity u + EC(u) = ES.1. Proposition A1.3LELU 1 1p' is in some appropriate sense critical as the most 'likely' time of ruin (here C is the CramerLundberg constant). uoo u using e. and take basically the form of approximations and inequalities.h(u. i. (u) t. T(u)/u mL as u + oo. cf.r(u) = Er(u) • ES. for any m T(u) u . and hence a.w ) v/. Then given r(u) < 00.s. By Proposition 111. For the second .mu D 2 4 N(0.mL > E T(u) < 00 ) 40. That is.6.
again Proposition A1.7 6  11 Proof of Theorem 4.1).6µB2) Z v m (3µB2) Z.^ N (o. If Z . this can be rewritten as u + 1(u) .h. Theorem 7.s. apB ) .2. proving (4.1). T(u) < oo f / 00) e7uE L [e_7 (t1). Tu) T( u) . For (4 . .2 of [86]) and (4.1 The l.g. cf. the result comes out not only by the present direct proof but also from any of the results in the following subsections. PL (•)+ 0.mL >E By Proposition 4.1). implying T(u) .t/m D (2) 111 .6. note first that ( Proposition 111. 4a Segerdahl's normal approximation We shall now prove a classical result due to Segerdahl.1 (by considering 0(u.1. T (u) < 00 J 0(u) e7'PL U \ I T u) . T) for T which are close to the critical value umL). 1'r(U) .1 is standard. and (4. which may be viewed both as a refinement of Theorem 4. WHEN DOES RUIN OCCUR? and that Ee(u)/u a 0. Thus.1) is T (u)  U mL P( T (u) < I > E.mu (2) '• m3/2 µB 7 .mL U > E.N(0.mu m . though it is not easy to attribute priority to any particular author. of (4.r(u)/m T(u) ti µB2) Z.4.3. the same conclusion holds with t replaced by r(u). According to Anscombe' s theorem (e. Notes and references Theorem 4. 4). and as a timedependent version of the CramerLundberg approximation.2) follows immediately from u (4.3).5) St .
a C4'(y )• ( 4. we get E[ T (u) . we can replace T(u) by r(u'). Then h(u) 4 h(oo) = E f (6(oo)). Hence Ef (Vu )) 9 (T(u. Using ( 4. g are continuous and bounded on [0.5) For the proof. oo). one has 9 (r(u)_rnu) Ef (^(u)) * E.4 (SIAM'S LEMMA) If 71 < 0. and thus in (4.3 (SEGERDAHL [333]) Let C be the CramerLundberg constant and define wL = f3LELU2mL = f3B"[ry]mL where ML = 1/(pL1) = 1/($B'[ry]1).ST( u') = u1/4 .u1/4)I(S(u') > u1 /4) h(oo) + 0. using that ul/4 .mul h(oo)Eg(Z). P because of ^(u') . O . and similarly as above we get E[f(^(u)) I Fr(u. e'°'/b (u. Then the distribution of T(u) .))I h(ul /4  ^(u)) I(6 (u') C ) f < ul /4 + f(e(u') . oo ). we need the following auxiliary result: Proposition 4.t.6) whenever f.r.w2) r.6).T ( u')] = E[ T ( ul /4 . PROBABILITY OF RUIN IN FINITE TIME Corollary 4. Proof Define u' = u .L+YWLV'U) .)mu \ h(oo)Eg (r(ul) .f ( (oo)) . resp .) is readily seen to be degenerate at zero if ST(u•) > u and otherwise that of T(v) with v = u . S( u ) < ul/4] < ET(ul / 4) = O(ul/4).v.e(u') oo w .112 CHAPTER IV.ul/4.(u.l:(oo) (recall that rt < 0).4). E9(Z) (4. letting Z be a N(0.^(T(u')). Then for any y. Let h(u) = E f (^(u)). with w2 as in (4.3).um.VU T. (oo. then e(u) and r(u) are asymptotically independent in the sense that.T(u') given F.
Notes and references Corollary 4 . y u) < e 7v" . where we used Stain's lemma in the third step and (4.oo. in practice one would trust (4.(ay) = 17 7y = ay . just substitute T = umL + ywL in (4. The precise condition for (4. For practical purposes .7) to be good. PL(T(u ) < umL + ywL) 113 4 C4(y). see also von Bahr [55 ] and Gut [182].z/)(u . 4b Gerber's time. For refinements of Corollary 4.3 ery"z/i(u . Cf.7) whenever u is large and ly(T)l moderate or small (numerical evidence presented in [12 ] indicates .4) in the last. y u) < . 3 is due to Segerdahl [333].T) Ce7"4 (T .5 '(u . Thus .1. Segerdahl 's result suggests the approximation b(u. Theorem 4. umL + ywL f) = e"P(T (u) < umL + ywL) = EL [e7V "). WHEN DOES RUIN OCCUR? Proof of Corollary 4. 10) '5(u) .5) and solve for y = y(T). y > k'(7) .yK(ay)• (4.9) ( 4 . also Hoglund [204]. CL Fig. 0. u needs to be very large). T(u) < umL + ywL f. The present proof is basically that of Siegmund [342].4. oo ) as u * oo. yy by 1 K.3 in terms of Edgeworth expansions . define ay. however .dependent version of Lundberg's inequality For y > 0.7) to be valid is that T varies with u in such a way that y(T) has a limit in (. see Asmussen [12] and Malinovskii [254]. that for the fit of (4.umL wI V"U u (4. e7v" y < ^'(7) (4 . ELe7E (") .7) To arrive at this . .8) Note that ay > 7o and that 7y > •y (unless for the critical value y = 1/ML).
a.1). In view of Theorem 4. PROBABILITY OF RUIN IN FINITE TIME Proof Consider first the case y < 1/K'(y). Then ic(ay) > 0 (see Fig .Y' (u. yu 11 < T(u) < oo j < eayu +Y UK(ay) Remark 4. the point is that we want to select an a which produces the largest possible exponent in the inequalities.h(u. yu) is e 'Yyu/ . which shows that the correct rate of decay of tp(u. 5 is due to Gerber [156 ].114 CHAPTER IV. and generalizations to more general models are given in Chapter VI. if y > 1/ic'(y).yu ) = eayuEav [e . which may be understood from Theorem 4.8 below . An easy combination with the proof of Theorem 111. see MartinLM [257] . For a different proof.8). dy) Notes and references Theorem 4 .b (u.t. we arrive at the expression in (4. and hence t. yu < T (u) < oo 1 l e ayuEav [eT ( u)K(ay).6.3 yields easily the following sharpening of (4. Numerical comparisons are in Grandell [172 ].7 i. 0. However. yu ) = < eayuEay [eay^ ( u)+T(U)K ( ay).r. u Differentiating w.5. the bound a7y° turns out to be rather crude . yy is sometimes called the timedependent Lundberg exponent. T(u) < yu] < eayu + yUr(ay) Y < eayuEav [ eT(u)K(av )L T(u) < yu} Similarly.2. yu) < C+(ay)e7a„ where l C+(ay) = sup f 00 eayR(xy)B( .v"U. f Some urther discussion is given in XI. . we have rc(ay) < 0 and get (u) . Hoglund [203] treats the renewal case.9): Proposition 4.6 It may appear that the proof uses considerably less information on ay than is inherent in the definition (4.ay4(u)+ T(u)K(ay ). From the proof it is seen that this amounts to that a should maximize ayic(a). who used a martingale argument.8).
yyu y l ay I 21ry/3B" [ay] V fU_ u + 00.i(u. This idea is precisely what characterizes the saddlepoint method. then the relevant choice is precisely a = ay where y = T/u. (4.ayuEay f eay^ ( u)+T(u)K(ay). (4. i. u 4 oo.2 yields EaT(u) u u r. yu) = e.13) . we have ryas = ay . WHEN DOES RUIN OCCUR? 115 4c Arfwedson's saddlepoint approximation Our next objective is to strengthen the timedependent Lundberg inequalities to approximations. not inequalities.ay and get Ea e ayf (00) y _ 'Ya( ayKal lay C 1 . T(u) < yu] .4.: T. We thereby obtain that T is 'in the center' of the Padistribution of T(u). the choice of ay. if we want EaT(u) .c(&) = ic(ay) is < 0. then ay > 0. The traditional application of the saddlepoint method is to derive approximations. and ii(u) .ayC() .. Proposition 4. and in case of ruin probabilities the approach leads to the following result: Theorem 4 .e.yu) c ay . [eT(u )K( ay).e.12) < yu] Here the first expectation can be estimated similarly as in the proof of the CramerLundberg ' s approximation in Chapter III.. For any a > yo. then the solution &y < ay of .ay y 'Yay  ay . Using Lemma 111.5. (0) r1 (a) ' I. Ea . and b(u.8 If y < 1/ic'(ry). (4.6 with P replaced by Pay and FL by Pay.(u.ay a. the formula 0(u.z..^3 ]1/ Bay [lay . yu ) eaauEaye . As a motivation.11) ' If y > 1/ r . yu ) ayay e ryyu ayay 27ry/3B"[ay] u Proof In view of Stam 's lemma.'(y ). T(u) suggests heuristically that l t/. it is instructive to reinspect the choice of the change of measure in the proof.
(4.4).13). and the equation ic'(a) = 1/y is easily seen to have . 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=.(ay) _ y(ay .a)2 .116 CHAPTER IV. The difficulties in making the proof precise is in part to show (4.ay + ayl /BLay] .1.a.l'B)y /(Pay . Writing r(u) and W2 = I3ay{.1)3 = y3/3B"[ay].ay)K(ay) ay ayI&YI For the second term in (4. we get heuristically that Eay Ler (u)r(ay).13).a) . where V is normal(0. V < 01 Ir 00 er(ay)"1'2"'x eyur. i B[7ay .7ruw2 Inserting these estimates in (4.1)3 = (jB"[ay]l (Pay . The proof of (4.ay) ay +.1B[ay]1 ) y(ay .I ay &y a ^c'(ay) a (1 +.c(ay)ul/2W p 2ir = eyu(ay) dz 1 rc(ay ) 2. it seems tempting to apply the normal approxiyu + ul/2wV. PROBABILITY OF RUIN IN FINITE TIME ry I i . Example 4.13) rigorously.11) follows.ay ) r.(j (1 . (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 . T(U) < yu] = eyuk (ay)E''ay (ek(ay )"1/2WV.3(5/(S .12) is 0 entirely similar.1) under Pay mation (4.9 Assume that B(x) = eay.c'(a) _ /3a/(8 .B[ay] /ay &y y(ay . .1) . Then ic(a) = .
0 Notes and references Theorem 4.1.tcp) Lo {Wo ( t)}t>0 . ./4 ^y for 1/i (u.8 is from Arfwedson [9]... The mathematical result behind is Donsker's theorem for a simple random walk {Sn}n=o.g.1) . in discrete time: if p = ES.p.because the c.11) gives the expression '31/4 ( .3+52 1+/351/y' sy 7 B ii[ay] 25 _ 251/2(1 + y)3/2 (5 . (5.i )( v s vc ('3 + s _2 / . 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 =. c a 00. A related result appears in BarndorffNielsen & Schmidli [59]. is the drift and o. y) a''y" L '3 _ fl ) 51 /4(1 +1IY)3/4 \. and next to note that such an approximation in particular implies that the first passage probabilities are close. then { __ . is undefined for a > 5). DIFFUSION APPROXIMATIONS solution ay=5 117 V 1 (the sign of the square root is .f.5. 5 Diffusion approximations The idea behind the diffusion approximation is to first approximate the claim surplus process by a Brownian motion with drift by matching the two first moments. yu) when y < 1/ic'('y) = p/1 ..ay)3 0 3/2 and (4. 2 = Var(Si ) the variance.= (s.
and this can be obtained under the assumption that the safety loading rt is small and positive. cf. 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.p.a = Snp) and the inequalities Sn )C . for the purpose of approximating ruin probabilities the centering around the mean (the tcp term in (5. of which a particular case is the claim surplus process (see the proof of Theorem 5. We want an approximation of the claim surplus process itself.. we have o {i!t s: .3. (5. and consider the limit p j p. 0 . oo)).118 CHAPTER IV.t} _ {W_1(t)} .2) t>o where p = pp = p .1)) is inconvenient. St = EN` U= . this is an easy consequence of (5.tcpp) y = { WC (Sct) pct) } {Wo( t)}t>o (5. It is fairly straightforward to translate Donsker's theorem into a parallel statement for continuous time random walks (Levy processes). where p is the critical premium rate APBTheorem 5 .1.3) takes the form LI S(P) { a2 to2/µ2 + t LI S (P) { a2 ta2/µ2 {W0(t)}.tp).e.1 As p J. p. This is the regime of the diffusion approximation (note that this is just the same as for the heavy traffic approximation for infinite horizon ruin probabilities studied in III.7c). Lemma 111. a2 =/3µB2) Proof The first step is to note that { WC (St P) .z } {W_1(t )}t>o (5. Mathematically. n/c < t < (n + 1)/c.1 below). However..3) whenever c = cp f oo as p 1 p. + {Wo(t ) .1) with S.p/c < St(p) < S((n+l)/ c + Pp/c. we shall represent this assumption on 77 by a family {StP) L of claim surplus processes indexed by the premium rate p. such that the claim size distribution B and the Poisson rate a are the same for all p (i. Indeed . Letting c = a2/pp.
Because of the direct argument in Chapter III. ulpl /a2) = e2"1µl / or2.6) This is the same as the heavy traffic approximation derived in III. (5.r. the continuity argument above does not generalize immediately.4) Note that IG(. u). Corollary 5 .s. (ua2 To2 op \ IPI > IG ( T . is 1/ip (ua2 /IpI.. TS(u)=inf{t>0: WW(t)>u}. 199. the continuous mapping theorem yields sup W Sz2 to lP 4 sup Wi(t)• O<t<T O<t<T a2 Since the r. C. (.s. w.(u) ti IG(oo.s.e..1. 263) that the distribution IG(•.u). 1. DIFFUSION APPROXIMATIONS Now let Tp(u) = inf{t>0: S?)>u}. (5.1 I 7= .2 suggests the approximation u 0(u. since ti(u) has infinite horizon . ulpI/a2).Ta2 /p2). we omit the details .f I \\\ J \ (5. However. 119 It is wellknown (Corollary XI. 196. u) is defective when < 0.5.5). For practical purposes .1 .( ^ I + e2( \ I .T) IG(Tp2/ a2). ('.h.5) Note that letting T * oo in ( 5. has a continuous distribution. we obtain formally the approximation V. 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..h. . [169] or [APQ] pp.h. ^ p2 Proof Since f 4 SUP0<t<T f (t) is continuous on D a. any probability measure concentrated on the continuous functions. see Grandell [ 168]. and the r. is IG(T.1. and in fact some additional arguments are needed to justify (5. u) of r( (u) (often referred to as the inverse Gaussian distribution) is given by IG(x. u) =PIT( (u) < x) = 1 .8 or [APQ] p.2 As p j p.t. Corollary 5.7c.6) from Theorem 5.
Po = 09µB6 . Assume further that 039µB6 < pe.. the B9. in the next subsection we shall derive a refinement of (5. we have ^A. The first application in risk theory is Iglehart [207].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. The picture which emerges is that the approximations are not terribly precise.120 CHAPTER IV. However. that 00 4090.1. and two further standard references in the area are Grandell [168].Pe. . e. the simplicity of (5. Furrer. Then as 0 _+ 90. B0 * Boo. a2 = ae = 00µa6 Notes and references Diffusion approximations of random walks via Donsker's theorem is a classical topic of probability theory. In view of the excellent fit of the CramerLundberg approximation. for more general models it may be easier to generalize the diffusion approximation than the CramerLundberg approximation.3 Consider a family {Ste) } oc claim surplus processes indexed by a parameter 9. the claim size distribution B9 and the premium rate p9 depends on 0. in particular for large u.6) are presented. However. All material of this section can be found in these references. Further relevant references in this direction are Furrer [151] and Boxma & Cohen [75]. [169]. and which is much more precise. Theorem 5. In contrast. The proof is a straightforward combination of the proof of Theorem 5. We conclude this section by giving a more general triangular array version of Theorem 5. See for example Billingsley [64]. pe . (5. PROBABILITY OF RUIN IN FINITE TIME Checks of the numerical fits of (5.5) for the compound Poisson model which does not require much more computation. as an example of such a generalization we mention the paper [129] by Emanuel et al. 0) { 2 StQ2 /µ2 D { W_ i(t)}t>o t>o D 2 where p = pe = pe .6 of [APQ].00µB6 + 0. in Asmussen [12]. on the premium rule involving interest. pt? 4 peo. as 0 * 00 and that the U2 are uniformly integrable w.r.6) therefore does not appear to of much practical relevance for the compound Poisson model. Michna & Weron [152] suggested an approximation by a stable Levy process rather than a Brownian motion. such that the Poisson rate Oe.g.5) and (5. For claims with infinite variance.1 and Section VIII.t.
6. P9(r (u) < oo) = 1 for 9 > 0. The objective of the corrected diffusion approximation is to take this and other deficits into consideration. Bo(dx) = B[eo]B(dx). it is more convenient here to use some value 9o < 0 and let 9 = 0 correspond to n = 0 (zero drift). Then EOU' = Boki[0] = Biki[eo]/E[9o] and "(s) = k(sBo)k(9o). PB('r(u ) < oo) < 1 for 9 < 0.90] B(dx). In this setup. this means the following: 1. and we are studying b(u. .Q (B[s] .1) .6. this is because in the regime of the diffusion approximation . CORRECTED DIFFUSION APPROXIMATIONS 121 6 Corrected diffusion approximations The idea behind the simple diffusion approximation is to replace the risk process by a Brownian motion (by fitting the two first moments ) and use the Brownian first passage probabilities as approximation for the ruin probabilities. In terms of the given risk process with Poisson intensity . However . Determine yo > 0 by r. 2. 77 is close to zero. this idea ignores (among other things) the presence of the overshoot e(u). and we want to consider the limit 77 10 corresponding to Oo f 0. claim size distribution B ./c(9 . 9o T 0. .c(s) = . Since Brownian motion is skip free.90) .1 > 0. whereas there we let the given 3B.9(s) = Ico ( s + 9) .s and p = /3µB < 1.90) and the given risk process corresponds to Poo where 90 = 'yo. B9(dx) =Bale] Bo(dx) e9z keo)z = B[9 .T) = Peo(r(u) < T) for 90 < 0. 77 = 1/p .'(yo) = 0 and let 90 = 'Yo.4. which we have seen to play an important role for example for the CramerLundberg approximation . For each 9.ao (0) _ /c(s + 9 . Then r. let P9 refer to the risk process with parameters Q9 = QoB0[9] = QB[9 9o]. The setup is the exponential family of compound risk processes with parameters ( B9 constructed in III. risk process with safety loading 77 > 0 correspond to 9 = 0 . Let PO refer to the risk process with parameters e9oz Qo = QB[90]. 0(0) = 0. 3.
.122 CHAPTER IV.S. the solution of r.1) .. C) = 2A + (2 . tu2 ) i IG (t.2) .Varo S1 = f30Eo U2 = S1.e. u) = euh(a . C .. (01. Theorem 5. _ ^(u) = ST .C. () where h (A.(y) = 0. i. for brevity. 1) • Since L eatIG (dt. 9otc0" (0) = 0061 = ul.T) 1+u2 (6.3) this implies (take u = 1) Ego exp { . bl IG(t81. (. Vargo S.. One has (6. (..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. write r = T(u). PROBABILITY OF RUIN IN FINITE TIME Recall that IG(x. 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.() The idea of the proof is to improve upon this by an O (u1) term (in the following. The corrected diffusion approximation to be derived is (u. . means up to o(u1) terms): . The first step in the derivation is to note that µ = k (0) = r0 (00) . u) = IG(x/u2.1) IG(x.7(u)/u2} eh(A. (6.2' where as ususal ry > 0 is the adjustment coefficient for the given risk process. IGu+u2. and Si = QoEoU2 = Q B"'['Yo Eo U3 ]. C. u) denotes the distribution function of the passage time of Brownian motion {W((t)} with unit variance and drift C from level 0 to level u > 0. 0o to. (U.u.
(6.2). p = 0.5) according to (6.'yu /2)(1 + b2/u)} + Aug 1I J . A numerical illustration is given in Fig. the formal Laplace transform inversion is heuristic: an additional argument would be required to infer that the remainder term in (6. To arrive at (6. calculated using numerical integration and Proposition 1.2 ). it holds for any fixed A > 0 that Ego exp { Ab1rr(u)/u2} .2) is indeed o(u1). and the dotted line the corrected diffusion approximation (6. we get by formal Laplace transform inversion that C 2 u. of a (defective) r. however .f. the r. distributed as Z . In ( 1) and (2). 9o T 0 in such as way that C = Sou is fixed. 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. that the saddlepoint approximation of BarndorffNielsen & Schmidli [59] is a serious competitor and is in fact preferable if 77 is large] .1 + 629.3 = 0.5) Once this is established . of (6.7. that whereas the proof of Proposition 6.1 As u + oo.3. we have p =. in (3) and (4). But the Laplace transform of such a r. CORRECTED DIFFUSION APPROXIMATIONS 123 Proposition 6.h. is the c.s. however. u is Eeazead2/++ Eeaz[1 + ab2/u] where the last expression coincides with the r.3). The initial reserve u has been selected such that the infinite horizon ruin probability b(u) is 10% in (1) and (3).ry2 .1 below is exact. .z . .6. just replace t by Tb1/u2. . .s.52/u where Z has distribution IG (•.1 + u2 I Indeed. bl I IG I t +2 .4. 1% in (2) and (4).exp { h(A.d. Note. 1.v. The solid line represents the exact value .v.h. 6 . which is based upon exponential claims with mean µB = 1.
.^) . BarndorffNielsen & Schmidli [59] and Asmussen & Hojgaard [34].011 L1 60 T IM 11. the fit at p = 0.1 It is seen that the numerical fit is extraordinary for p = 0. PROBABILITY OF RUIN IN FINITE TIME 0. see Asmussen [12].1 proceeds in several steps.0 0. (Inc 0s 0. OM 0.19)2 11 20 20 i0 T 1n0 Figure 6..08 a.2 e.T) 111 0.199 0.T1 00.124 0. The proof of Proposition 6.4 may not be outstanding but nevertheless.07 0. Note that the ordinary diffusion approximation requires p to be close to 1 and '0 (u) to be not too small.05{ 0.(061 0. A51 7(SAT 3 3 h(X.08 0.T) 0.114 0.EB 0 p ex p ( 7 S h ^)u . For further numerical illustrations.W21 0.00 0.OOIi O. and all of the numerical studies the author knows of indicate that its fit at p = 0.1 W IU.u2 2u3 (e .TI CHAPTER IV.01 0.7 or at values of Vi(u) like 1% is unsatisfying.02 I 90 120 160 2W A0 Z WT 40 80 120 160 100 240 280 T 111 WI.7.() Lemma 6. it gives the right order of magnitude and the ordinary diffusion approximation hopelessly fails for this value of p.111 W(U.aa1 . Similarly. .
+ h (A. + a1b2 + .4 Ea.co ((/u)) } Let 8 = (2a + (2)1/2 = h(). the formulas Po(C(0) > x) Po(C(co) > x) imply 1 °° Po(ST(o) > x) = EIU fIP0 (U>y)dy . () + C and note that 2 KO (0) = 102.C2 = 2).r0 (00)) } Replacing B by 8/u and Bo by C/u yields e(B() = E eo exp { (e . 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 lim Eof (u) = EoC(oo) = a2 Ep = 3EoU2 uroo Proof By partial integration . (6.(3) J t _ aa1T l + eh(A.7) 2 2 .co (e) .6) u U3 Lemma 6 . (6... () 62 Eeo exp u u2 J .. in Lemma 6.2u (B3 . CORRECTED DIFFUSION APPROXIMATIONS Proof For a>0.(3)Eea LauT exp i 3J .1) h(A.3 EoU2 + 103OoEoU3 + " 2 6 Using d2 .C)C/u .61a2T (B3 .2 behaves like C l Eeo eXp r _ ^81T 1 Sl u2 1 u 2u3 [1+h(AC) S . 1 = PB(T < oo) = Eo0 exp 125 {(B . exp ue } al 1J 3 exP I [2)..00)(u +C)  'r (.6.4) that the r.T (co (8/u) . 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.h.s. the result follows.
and the correction terms which need to be added cancels conveniently with some of the more complicated expressions in Lemma 6. 5 exp { _h(A) (1 + / y u J)) exp 1.2u [2A+ (2 3 . yields +90 62 0 + O(u 3) 2u2 +O(u 3). () . we get the correct asymptotic exponential decay parameter ^/ in the approximation ( 6.e h(aS)h (^^ 262 exp {_h(.7) and using eh(a.\ + () 1 2 / .4.S) d e 62 .h (A. 2 + 00 = . PROBABILITY OF RUIN IN FINITE TIME The last term is approximately (e 3 (3) 27.h.2. yu/2). yu/2) h(A.1 (y/2 + Oo)u .126 CHAPTER IV.2) for O(u) (indeed .6  d h(A..6) and 7co (Oo) = ico('y + Bo) to get 0 = 21 (^/2 + 2y90) + 1112 (_Y3 + 3_Y200 + 3y9o) + O(u4). we get h(A.\+ (2 (3 e 2u [ (2. [2+ (2 . 2 and (6.() . () by h(\.() I 1 + u2 ) y . There are two reasons for this : in this way. C) ( 1+ u2 The result follows by combining Lemma 6 . and inserting this and 9o 2 = S/u on the r. 0 The last step is to replace h(A. letting formally T * oo yields 7/)(u) C'e7u where C' = e7a2).x. Thus by Taylor expansion around ( = 90u. Thus a2 y = 290 + O (u2). yu/2) 11+ 62 I} S 1 \\\ u/11 l 62 (3 2u 2A Proof Use first (6. l Lemma 6 .s.2 (^/2 + 3y9o + 390) + O(u3).(2A + ()1/21 exp S h(A.
4. 'yu/2) 127 ( i+ M pz^ exP { h (A. In Siegmund's book [346]. () I 1 + u2 )I 2u L 2A+C2_(2 exp { _h. that is. We shall now generalize this question by asking what a sample path of the risk process looks like given it leads to ruin. The answer is similar: the process behaved as if it changed its whole distribution to FL. the approach to the finite horizon case is in part different and uses local central limit theorems. ()} 3 h (A. His ideas were adapted by Asmussen & Binswanger [27] to derive approximations for the infinite horizon ruin probability 'i(u) when claims are heavytailed. the analogous analysis of finite horizon ruin probabilities O(u. 7 How does ruin occur? We saw in Section 4 that given that ruin occurs. 0 1 Proof of Proposition 6.T) has not been carried out and seems nontrivial.1 (y/2 + Oo)u )} 1 (i + U ) [2+ C2 2u 62 S Pt^ exP { J 62(2 exp { h(A.5 in Lemma 6. .(i+ 62 exP{ 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.7. HOW DOES RUIN OCCUR? exp { h (x. i. Hogan [200] considered a variant of the corrected diffusion approximation which does not require exponential moments. The corrected diffusion approximation was extended to the renewal model in Asmussen & Hojgaard [34]. Fuh [148] considers the closely related case of discrete time Markov additive processes. u Notes and references Corrected diffusion approximations were introduced by Siegmund [345] in a discrete random walk setting. () (i+a ) 2A + (2 .e. and to the Markovmodulated model of Chapter VI in Asmussen [16]. the 'typical' value (say in sense of the conditional mean) was umL. () I 1 + u 2 ) } S 1 . with the translation to risk processes being carried out by the author [12].1: Just insert Lemma 6. ()} . The adaptation to risk theory has not been carried out. the same as for the unconditional Lundberg process.
FT(u) = o' (T(u ). r(u) < oo) .(u) is not .1 Let {F(u)}u>0 be any family of events with F(u) E F. We are concerned with describing the F(u) distribution of {St}o<t<T(u) (note that the behaviour after rr(u) is trivial: by the strong Markov property. the numerator becomes e'ruELe7^ (u)PL(F( u)t) = e7uCFL (F(u)°) when F(u) E . Recall that . F(u)c] ti e' ru]PL (F(u)`) > 0. u * oo. Recall that 13L = (3B[ry] and BL(dx) = e'rxB(dx)/B[7]. Theorem 7 .ST(u)}t> o is just an independent copy of {St}t>o). then P(u) and FL coincide on .2 If B is exponential.vi(u) Ce'Yu Corollary 7. we give a typical application of Theorem 7. + TMOO ).. .3L and the claim size distribution from B to BL.F. Note that basically the difference between FT(u) and .t.(u)_ and similarly the denominator is exactly equal to Ce7u.. so in the in the proof.EL[e7S.(u)_ and ^(u) are independent .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.r.(u)_ is that i.TT(u) _measurable. {ST(u)+t . F(u)c] P(r(u) < oo) ?P(U) < EL[e7u. the Poisson rate changes from . Then also P(u)(F(u)) + 1. ^(u) is exponential with rate 8 w.3 to .(u).t. Proof P(u) (F(u)c) = F(flu)c.128 CHAPTER IV.1. FL As example. .r.F. In the exponential case. P(u) and rate = aL w. and let M(u) be the index of the claim leading to ruin (thus T(u) = Ti + T2 + . {St}0< t<T(u)) Proof Write e'rsr(u ) = e'rue'r£(u). In fact. PROBABILITY OF RUIN IN FINITE TIME changed its arrival rate from 0 to /3L and its claim size distribution from B to BL.T. (u) and satifying PL(F(u)) * 1. stating roughly that under F(u). .
the queueing results are of a somewhat different type because of the presence of reflection at 0. HOW DOES RUIN OCCUR? Corollary 7. A somewhat similar study was carried out in the queueing setting by Anantharam [6]. the subject treated in this section leads into the area of large deviations theory.eALx) M(u) k=1 u The proof of the second is similar. From a mathematical point of view.eaLx. 129 M(u) >2 I(Tk < x) M(tu) p(u) M(u) >2 I(Uk < x) BL(x).3. Proof For the first assertion. take I(Tk < x) . .3 M(u) pcu) 1 . This is currently a very active area of research.(1 . see further XI. who also treated the heavytailed case. Notes and references The results of the present section are part of a more general study carried out by the author [11].7. however.
This page is intentionally left blank .
. In the socalled zerodelayed case. D'2. the claim sizes U1. and M is the maximum of {St}.. r(u) the time to ruin..1). Thus the premium rate is 1. U2. the one corresponding to the stationary case by 00)(u).1 (T1 = a1). Proposition 1. with common distribution B. Var(St) = 11Ba2A + I4AaB lim t goo t PA 131 .. A different important possibility is Al to be the stationary delay distribution A° with density A(x)/µA.(1.1 Define p = !µ. AA t*oo lim St = lim ESt t tioo t = p .. and the one corresponding to T1 = s by 1/i8 (u).. of the risk process form a renewal process: letting Tn = Qn .. The ruin probability corresponding to the zerodelayed case is denoted by 1/'(u).d. the Tn are independent. see A. . 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 distribution Al of T1 is A as well.7). with Nt = # {n: Un <t} the number of arrivals before t.i.1. T3.. . with the same distribution A (say) for T2. We use much of the same notation as in Chapter I. {St} is the claim surplus process given by I.Then no matter the distribution Al of T1i B.Chapter V Renewal arrivals 1 Introduction The basic assumption of this chapter states that the arrival epochs O'1.Q. are i. .
after E. For (1 . However . Nt = Var(PBNt) + E(4Nt) Q2 2 0`2 A tpB B + o(t). Example 1 . RENEWAL ARRIVALS lim E [St+a . stating that E[Nt+a . CHAPTER V.1) ENt/t + 1/µA. The simplest case is of course the Poisson case where A and Al are both exponential with rate 0.t. Here are two special cases of the renewal model with a similar direct interpretation: Example 1.3 (SWITCHED POISSON ARRIVALS) Assume that the process has a random environment with two states ON. The renewal model is often referedd to as the Sparre Andersen process. This has a direct physical interpretation (a large portfolio with claims arising with small rates and independently).1).Nt] * a/PA. Thus. one could imagine that the claims are recorded only at discrete epochs (say each week or month) and thus each U. the . 3) follows similarly by Blackwell 's renewal theorem. From this ( 1. A. such that no arrivals occur in the off state.2 (DETERMINISTIC ARRIVALS) If A is degenerate. Nt ESt = E E UI Nt t = ENt•pB . Nt + EVar U. OFF. but the arrival rate in the ON state is . the definition 77 = 1/p . t 4oo Proof Obviously. Sparre Andersen whose 1959 paper [7] was the first to treat renewal assumptions in risk theory in more depth. If the environment is Markovian with transition rate A from on to off and u from OFF to ON.a is really the accumulated claims over a period u of length a.132 Furthermore for any a > 0. s + t µA PA 0 Of course. say at a. Proposition 1. 2).1 of the safety loading appears reasonable here as well.1 gives the desired interpretation of the constant p as the expected claims per unit time. and (1 .0 > 0. we get similarly by using known facts about ENt and Var Nt that Nt Var(St) = Var E Nt U.St] = a(p . by the elementary renewal theorem (cf.1) follows .
s.d..T between a claim U and an interarrival time T. u For later use. INTRODUCTION 133 interarrival times become i. S o<t<oo n=0. A is phasetype (Example 1. Therefore.y)B(dy).} with {S(d)} a discrete time random walk with increments distributed as the independent difference U .. Ao.oFF}.i.4) with phase space {oN. 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 . (an arrival occurs necessarily in the ON state. U1 .r. 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^. arrival times. the relevance of the model has been questioned repeatedly. initial vector (1 0) and phase generator 11 However.s > u) of ruin at the time s of the first claim whereas the second is P(r(u) < oo. . we have From this the result immediately follows.s < u). (1. if for nothing else then for the mathematical elegance of the subject. integrate (1. The following representation of the ruin probability will be a basic vehicle for studying the ruin probabilities: Proposition 1.1.1.4) fo Indeed. However.4) w.t.1.. More precisely. 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 .. and for historical reasons. For the stationary case. we feel it reasonable to present at least some basic features of the model. the first term represents the probability F(U1 .. and the present author agrees to a large extent to this criticism. in general the mechanism generating a renewal arrival process appears much harder to understand.. the fundamental connections to the theory of queues and random walks.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. and then the whole process repeats itself).2. Proof The essence of the argument is that ruin can only occur at claim times.
Ut. the remaining part of the prepayment (if any ) is made available to the company. RENEWAL ARRIVALS 2 Exponential claims. That is .d. b=1 !=1 where {Nt } is a Poisson process with rate . the claims and the premium rate are negative so that the risk reserve process . At the time of death .134 CHAPTER V. < 1. (2.1) +ry. are independent of {Nt} and i. 00). then 0*(u) = e 'r" where ry > 0 is the unique solution of 0 = k*(ry) = *(B*[ry] . each of which receive a payment at constant rate during the lifetime .3* (say ) and the U. The compound Poisson model with negative claims We first consider a variant of the compound Poisson model obtained essentially by signreversion . U Figure 2. t.1 If. with common distribution B* (say) concentrated on (0. i. St = t . we shall be able to compute the ruin probabilities i(i* (u) for this model very quickly (.1. If . A typical sample path of {Rt } is illustrated in Fig.0* (u) = P (rr* (u) < oo) where rr* (u) = inf It > 0: Rt < 0} ) . 2. Using Lundberg conjugation .3* pB. The initial reserve is obtained by prepayments from the policy holders. Theorem 2 .1 r* (u) One situation where this model is argued to be relevant is life annuities.1) . resp . then 0 * (u) = 1 for all u > 0. A simple sample path comparison will then provide us with the ruin probabilities for the renewal model with exponential claim size distribution.a*PB• > 1. the claim surplus process are given by Nt Nt Rt = u+^U.
T_ ( u) < 001 e7"P(T_ (u) < oo) = e"V)* (u). 0 Now return to the renewal model.2(a).s.2 Assume now . 2. Fig. B.f. Then the c. and thus 1 = P(T. EXPONENTIAL OR NEGATIVE CLAIMS [Note that r. Hence T_(u) < oo a. Proof Define 135 St =u . Let B(dx) = ^e7x B*(dx). of {St} is c(a) = is*(a7). B* [7] and let {St} be a compound Poisson risk process with parameters . B*.0. > 1 .Rt. the safety loading of { St} is > 0.. Then { St } is the claim surplus process of a standard compound Poisson risk process with parameters 0 *.* (a) = log Ee'st I. T_ (u) = inf { t > 0 : St = u 'r* (u). Define T_ (u) = inf It > 0 : St = u} . cf. . If I3*pB* < 1. Then the function k* is defined on the whole of (oo.1. 0) and has typically the shape on Fig.3*. 2.(u ) < oo) = E {e7sr_ (u). Hence y exists and is unique. St=Rtu=St. and the Lundberg conjugate of {St} is { St } and vice versa.g.2 sup St = inf St = 00 t>o t>o and hence 0* (u) = 1 follows.UB.2(b). (a) is*(a) (b) . Since ic'(0) < 0. then by Proposition 111.2.(a) 7 Figure 2.
Ui . To + M(d) in the notation of Proposition 1.Ti = U1. However.1.1) + ry = 0 which is easily seen to be the same as (2.7r+ 7r EeTo b/(Sa) + +.•.g. Now the value of {St*} just before the nth claim is To +T1* +. Then B* = A. RENEWAL ARRIVALS Theorem 2 ..Y] (2. 1) means that 8(A[ry] .1. T2 = U2. the distribution of M(d) is a mixture of an atom at zero and an exponential distribution with rate parameter ry with weights 1 ... and hence the failure rate ..+Tn U1 Un.Un } = max St = t>0 n=0.4 goes as follows: define 7r+ = P(M(d) > 0) and consider {St*} only when the process is at a maximum value. the failure rate of this process is y. alternatively termination occurs at a jump time (having rate 8).f..Tr+.• • • . which has the advantage of avoiding transforms and leading up to the basic ideas of the study of the phasetype case in VIII.'s and noting that V)*(u) = P(M* > u) so that Theorem 2.. with rate S (say). 3* = 6.u+ and lr+.e. Taking m. 2..Y a I.2 If B is exponential.1 it is seen that ruin is equivalent to one of these values being > u.2) 7 and7r+=1Proof We can couple the renewal model { St} and the compound Poisson model {St*} with negative claims in such a way the interarrival times of { St*} are To .4.. and from Fig .Tn} n=0. then . To + max {Ul+•••+UnTI. with the probability that a particular jump time is not followed by any later maximum values being 1 .2).. we get Ee'M(d) = Ee°M* _ Y/(.136 CHAPTER V. respectively..)(u) _ 1r+e7" where ry > 0 is the unique solution of 1 = Ee'Y(uT ) = S 8 A[. Hence M* max {To + Ti + • • • + Tn .. and (2 .1.a) = 1 . A variant of the last part of the proof.1 means that M* is exponentially distributed with rate ry.. According to Theorem 2.. and 5PA > 1. u Hence P(M(d) > u) _ 1r+e'r"..
the imbedded discrete time random walk and Markov additive processes.5. 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^. Thus a ladder step terminates at rate b and is followed by one more with probability 7r+.7r+) = ry and hence P(M(d) > u) = P(M(d) > 0)e7u = 7r+e'r". Hence the failure rate of M(') is 6(1 . This follow since. letting P(d) refer to the renewal risk model with these changed parameters .2.6. a ladder step is the overshoot of a claim size. Furthermore. consider instead the failure rate of M(d) and decompose M(d) into ladder steps as in II. : S(d) > u} .2.4. we see that ry = 6(1. 3a The imbedded random walk The key steps have already been carried out in Corollary 11.. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 137 is b(1..3 A[a] B[a] F( d) [a +)3] F(d) [a] = Fad) [^] Letting M(u) = inf in = 1. the relevant exponential change of measure corresponds to changing the distribution F(d) of Y = U . 0 3 Change of measure via exponential families We shall discuss two points of view. However. which states that for a given a. B^d) where Aad> (dt) = ^[ a] A(dt). Bads (dx) = ..3. resp.7r+) and hence r+ = 1.B(dx). we have ] A[a )3] E«d'efl' = Bad> [a] A ad> [Q] = B[a +.y/b. Putting this equal to y. hence exponential with rate b. 111.7r+).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] .. The probability that the first ladder step is finite is 7r+.
00)(u) ..3 For the delayed case Tl = s. let 7 > 0 be the solution of r. In fact.2 p. This is known to be sufficient for ^(O) ]p (d) ([APQ] Proposition 3. the evaluation of C is at the same level of difficulty as the evaluation of i/i(u) in matrixexponential form. It should be noted that the computation of the CramerLundberg constant C is much more complicated for the renewal case than for the compound Poisson case where C = (1 .. 187) and thereby for ^(u) to be nonlattice w. VIII. cf. Corollary 3. Consider now the Lundberg case. i .. 7µA . RENEWAL ARRIVALS be the number of claims leading to ruin and ^(u) u = SM(u) . provided the distribution F of U .r.(u) . (a) '(u) < eryu.e. (b) V)(u) .C(°)eryu where C(O) = C0[7] .4. to converge in distribution since p(yd) (r(0) < oo) = 1 because of r (d)' (y) > 0.T is nonlattice. just note that F7(d) is nonlattice when F is so .Ce"u where C = limu. Proof Proposition 3.C8e7u where Cs = Ce78B[7]. (d) (7) _ 0. in the easiest nonexponential case where B is phasetype.p)/($B'[7] .1 For any a such that k(d)' (a) > 0. ik. For claim (b). we get: Proposition 3.2 In the zerodelayed case. For the stationary case.138 CHAPTER V.1). O(u) = eauE (d)ea{ (u)+M(u)K (d)(a) . E(d)e 1' (u).t.1 implies Cu) = e«uE ( 7d)e«^(u) . and claim (a) follows immediately from this and e (u) > 0. We have the following versions of Lundberg' s inequality and the CramerLundberg approximation: Theorem 3 .u the overshoot .1) is explicit given 7.
Equating this to tch(s) and dividing by h(s) yields h'(s)/h(s) _ .Sdt] = Ee'uh(T) means 1 = f ' e°^B(dy) f ' h( s)A(ds). St)} and h.a . 0) = ah (s) . IPA 0 Of course./c. (s. Sdt) = h(s . 0 0 . Here K. where G is the infinitesimal generator of {Xt} = {(Jt. The underlying Markov process {Jt} for the Markov additive process {Xt} = {(Jt. Let P8f E8 refer to the case Jo = s. y) = e°yh(s). To determine boundary 0. According to Remark 11.3.1) (normalizing by h(0) = 1). CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES Proof Using (1.St)} can be defined by taking Jt as the residual time until the next arrival. we get r u +8 e"8(u) 139 e7uB(u + s) + 4 0 + L 00 J e7(v8)e7(u+8v). 0) = tc(a)h(s)..dt ) eadt = h ( s) .1) = C(O). E8h0 (Jdt. h(s) = e(a +x( a))8 (3.y) B(dy) 0 For the stationary case. 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] . delayed version of Lundberg's inequality can be obtained in a e7u.(°) ( Ce8B[7] Ao(ds) similar manner.5.h'(s).(s. 3b Markov additive processes We take the Markov additive point of view of II.0 ) = Eo[ha ( Jdt.dt(ah ( s) + h'(s)) so that Gha ( s.. For s > 0. (u + s .9. B(x) = o(e7x) and dominated convergence. we look for a function h(s) and a k (both depending on a) such that Gh.4). we invoke the behavior at the 1 = h«(0. The expressions are slightly more complicated and we omit the details.5.
[e1U1 + 6T2ea ( U1s)stc ( a)e(a+K(a ))( T2s)I B[a +.c(a)] B[a] Proof Pa. we can now for each a define a new probability measure Pa. An important exception is.s is the probability measure governing a renewal risk process with Jo = s and the interarrival distribution A and the service time distribution B changed to Aa.tK( a)e. .c(a)] which shows that U1.S„):0<v< )be Lt = eaSt tK(a) h(Jt) = east .rc(a)] = B = Ba[13]Aa[5].. i. 5 For the compound Poisson case where A is exponential with rate . rc(a) = 0 (B[a] . the determination of the adjustment coefficient ry where the defining equations rc(d) (ry) = 0 and rc(ry) = 0 are the same. .. J n+1 u are independent with distribution Aa for the Tk and Ba for the Uk.a . U. Remark 3 . A[a . T2. ] = E. Ba(dx) = B(dx).rc(a)] B[a] A[a . (3. . resp.8.e. since JT. .2) means 1 = B[a]/3/(/3+a+rc (a)).13]A[b . Note that the changed distributions of A and B are in general not the same for Pa..2) As in 11.4 The probability measure Pa. resp. [a + /3] A[b .s governing {(Jt. Further.. Ease AU1+6T2 [ AU1+6T2 = Ea a LT. An easy extension of the argument shows that U1.s(Jo = s) = 1 follows trivially from Lo = 1. RENEWAL ARRIVALS B[a]A[a . Aa as asserted . . = J8 = T2..e(«+k(a))t esy A(dt).1)a in agreement u with Chapter III.140 CHAPTER V..a . (3. Proposition 3.( a+r' (a))(Jt s) h(s) where c(a) is the solution of (3.5. . St)}too by letting the likelihood ratio Lt restricted to Yt = a((J. T2 are independent with distributions Ba.2). Ba where Aa (dt) .s and P(d).rc(a)] = 1. however.
4. . A(ds). . that is.g. for the zerodelayed case zp8(u. and U„ the service time of customer n. the amount of .r.. which is the same as the asserted inequality for 0. Proof As in the proof of Theorem IV.6 Let y < let ay > 0 be the solution of ic'(ay) = 1/y.t.y yuAa y [ay + K(ay) . 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. The virtual waiting time Vt at time t is the residual amount of work at time t.ay+ray))TM(.5.(u. T(u) < yu h(JT(u)) < eayu+yuk(ay ) ( Eia y Le(a(+k(ay))s v. defined as the single server queue with first in first out (FIFO. yu) = F'ay.4. see in particular Dassios & Embrechts [98] and Asmussen & Rubinstein [45].. the time from he arrives to the queue till he starts service. and assume that T„ is the time between the arrivals of customers n . Then J(rr(u)) = TM(u)+1 and hence Ws(u.5 Proposition 3. not after time T. The actual waiting time Wn of customer n is defined as his time spent in queue (excluding the service time). 2. u The approach via the embedded random walk is standard. [APQ] Ch. and define yy = ay .yu ) e7vu A[ay . The claim for the zerodelayed case follows by integration w. . or FCFS = first come first served) queueing discipline and renewal interarrival times. yu). see e. XII. Then "^ e(ay+w(aY))8 Ys(u.yx(ay).4. Notes and references 4 The duality with queueing theory We first review some basic facts about the GI/G/1 queue. yu ) < e7yu. that is. Let M(u) be the number of claims leading to ruin .s eaysr(")+r(u ) K(ay) h (s) . For the approach via Markov additive processes.1 and n. it is easily seen that ic(ay ) > 0.)+1 e J j e(ay+w(ay))8 e . Label the customers 1. say finite horizon ruin probabilities where the approach via the imbedded random walk yields results on the probability of ruin after N claims.rc( ay)] = e(aa+(r ))sb[a ]e7yu L y1 In particular. Using the Markov additive approach yields for example the following analogue of Theorem IV.
we have Wn = Van(left limit). 0 Applying Theorem 11..1. since customer n arrives at time on.n1 (U1 +• • •+Uk Tl . Then: (a) as n + oo. Vt converges in distribution to a random variable V. 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. Let the T there be the random time UN.3) Proof Part (a) is contained in Theorem 11.2) (b) as t * oo. The traffic intensity of the queue is p = EU/ET.4. . (4. the waiting time a customer would have if he arrived at time t.1.."^ Vi(N) (u). If W1 = 0. (4. The next result summarizes the fundamental duality relations between the steadystate behaviour of the queue and the ruin probabilities (part (a) was essentially derived already in 11. (u). equivalently. but interchanging the set .1).1 Wn+1 = (Wn + U. and obviously z/'(u) = limN. the proposition follows. on + Tn) the residual work decreases linearly until possibly zero is hit. whereas in [On . Wn converges i n distribution to a random variable W.Tn)+.4.• • • Tk ). and we have P(W > u) = V.4: Proposition 4. (4.4): Proposition 4. but we shall present a slightly different proof via the duality result given in Theorem II.1 and Corollary 11.3 Assume rl > 0 or. Thus Vos}1 _ = (Wn + Un .4.+ Un..2.3.2 Let Mnd) = maxk=o. It then jumps to VQ„ . an+1) = [on. we get: Corollary 4..Tn)+ Proof The amount of residual work just before customer n arrives is VQ„ . Also {Zt}o<t<T evolves like the leftcontinuous version of the virtual waiting time process up to just before the Nth arrival. equivalently. p < 1.1) The following result shows that {Wn} is a Lindley process in the sense of II. then Wn v M.142 CHAPTER V. Thus. and combining with (4. Then P(r(u) < T) is the probability z/iiNi (u) of ruin after at most N claims. and we have P(V > u) = ?/iiol(u). in which case {V} remains at zero until time on+1.
A. cf.5) Proof Letting n . resp ... K(x) = P(W < x).(N)(u) has the limit tp(u) for all u. by an obvious reversibility argument this does not affect the distribution . which implies the convergence in distribution and (4.oo in Proposition 4.2). { Zt}o<t < T has the same distribution as the leftcontinuous version of the virtual waiting time process so that P(s)(VT > u) = P(s)(r(u) < T).T*)+ < x) = P(W + U* . as WN.2. Then K(x) = J x00K(x . T] form a stationary renewal process with interarrival distribution A. For part (b). but this requires some additional arguments (involving regeneration at 0 but not difficult) that we omit. (4. Then the corresponding queue is M/G/1.. x > 0. Corollary 4.. convergence in distribution hold for arbitrary initial conditions . i. and we get: .. Hence for x > 0.4) 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. T1 .y)F(dy).4 The steadystate actual waiting time W has the same distribution as M(d). hence (since the residual lifetime at 0 and the age at T have the same distribution ... Letting n oo in Corollary 4. conditioning upon U* ..4.Ao in (b).. Ti) and similarly for the U. we obtain: Corollary 4. It follows that P(WN > u) =. we let T be deterministic . u Now return to the Poisson case .T* = y yields K(x) = P ((W + U* . and hence in particular ZT is distributed as the virtual waiting time just before the Nth arrival. Then the arrivals of {Rt} in [0.T* < x) fK(x_y)F(dy) (x > 0 is crucial for the second equality!).le) the same is true for the timereversed point process which is the interarrival process for { Zt}o<t < T• Thus as before . (4. where U*. TN) with (TN.5 (LINDLEY'S INTEGRAL EQUATION) Let F(x) = P(U1T1 < x).T*)+. we get W = (W + U* . T1..T* are independent and distributed as U1.e. namely W1 = 0 in (a) and Vo = 0. THE DUALITY WITH QUEUEING THEORY 143 (T1.1. However. In fact .
Proof For the Poisson case. The equation (4.5) looks like the convolution equation K = F * K but is not the same (one would need (4. the zerodelayed and the stationary renewal processes are identical. implying P(W > u) = P(V > u) for all u. Hence '(u) = Ali(°)(u). Cohen [88] or [APQ] Ch.144 CHAPTER V. That is.5) is in fact a homogeneous WienerHopf equation. Asmussen [24] and references there. VIII). W v V. .g. 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. Some early classical papers are Smith [350] and Lindley [246].g. RENEWAL ARRIVALS Corollary 4. despite the fact that the extension from M/G/1 is of equally doubtful relevance as we argued in Section 1 to be the case in risk theory. Note that (4. see e.6 For the M/G/1 queue with p < 1.5) to hold for all x E R and not just x > 0).
{Jt} describes the environmental conditions for the risk process. here it exists whenever A is irreducible which is assumed throughout. t St = E Ui .)iJEE and its stationary limiting distribution by lr. Thus. The intensity matrix governing {Jt} is denoted by A = (A.T) = Pi (T(u) < T).. dv. . Ire = 1.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 . N. i=1 0 and r(u) = inf It > 0: St > u}. M = supt>o St. 145 Oj( u. • The premium rate when Jt = i is pi. • Claims arriving when Jt = i have distribution Bi. The ruin probabilities with initial environment i are '+ki(u) = pi(T(u ) < oo) = Pi (M > u).(3i when Jt = i. and can be computed as the positive solution of WA = 0. As in Chapter I. {St} denotes the claim surplus process.f pi.
meaning that accidents occuring during icy road conditions lead to claim amounts which are different from the normal ones. and assume that weather conditions play a major role for the occurence of accidents. Example 1. cf. f3i and claim size distributions Bn. we shall assume that pi = 1. assume that the sojourn time in the icy state has a more general distribution A(i). Unless otherwise stated. Thus. say. i and corresponding arrival intensities Qn. For example. we could distinguish between normal and icy road conditions. this is no restriction when studying infinite horizon ruin probabilities. An example of how such a mechanism could be relevant in risk theory follows.4) with representation (E(i). r^ = P (1. We let p Pi = /ji/AB. /3 = Nn when j E E(n). and p is the overall average amount of claims per unit time. one expects that 3i > on and presumably also that Bn # Bi.T(n)). Cl The versatility of the model in terms of incorporating (or at least approximating) many phenomena which look very different or more complicated at a first sight goes in fact much further: Example 1.1 Consider car insurance.2 (ALTERNATING RENEWAL ENVIRONMENT) The model of Example 1. say. we can approximate A(i) with a phasetype distribution (cf. Then the state space for the environment is the disjoint union of E(n) and E(i). P = E 7riPi. cf. and we have f3. T(=)). u .11 below. According to Theorem A5. Proposition 1.1 implicitly assumes that the sojourn times of the environment in the normal and the icy states are exponential.5 below.. with rates Ani and Ain. respectively. which is clearly unrealistic.2. a(i). Example 1 . = iii when j E E(i).a(').146 CHAPTER VI. MARKOVIAN ENVIRONMENT where as usual Pi refers to the case Jo = i. Assume similarly that the sojourn time in the normal state has distribution A(n) which we approximate with a phasetype distribution with representation (E(').14. Bi. leading to E having two states n. t(i) = T(')e are the exit rates. in blockpartitioned form.1) iEE Then pi is the average amount of claims received per unit time when the environment is in state i. the operational time argument given in Example 1. the intensity matrix is A OW) T(i) T(n) t(n)a(i) where t(n) = T(n)e.
. one could for example have H = {i1.3i. 1 . let T 9(T) = f pi. say it is larger initially. and . n8}.2. Qi = . i8f n1. A('^).tEH is a transition matrix.3i/pi. Example VIII. resp. the parameters are ^ij = aid/pi. (9) where q = CHI.p. depending only on 77.3 Consider again the alternating renewal model for car insurance in Example 1.1... but assume now that the arrival intensity changes during the icy period. One way to model this would be to take A(') to be Coxian (cf. dt. St = SB=(t).Q.J017. iq (visited in that order) and letfOil >. 0 Then (by standard operational time arguments) {St } is a risk process in a Markovian environment with unit premium rate. i E E(n) }. and similarly for the normal period. t(n) = T("i)e. u From now on.. Then for example wi.n.j = . say. .3.. . where W = (w.. we assume again pi = 1 so that the claim surplus is Nt St = ?Ui_t.1...4) with states i1. The simplest model for the arrival intensity amounts to . is the probability that a long icy period is followed by a short normal one. such that the icy period is of two types (long and short) each with their sojourn time distribution A('L). and 1/ii(u) = t/ii(u). w.5 (MARKOVMODULATED PREMIUMS) Returning for a short while to the case of general premium rates pi depending on the environment i. In the car insurance example. Indeed. u Example 1 .. Approximating each A('?) by a phasetype distribution with representation (E('l). 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) .. i ) : n E H. it = Jel(t).T(n)). This amounts to a family (A(")) ?CH Of sojourn time distributions.a(n). such that a sojourn time of type rt is followed by one of type c w. T(9) +wggt(9)0.. MODEL AND EXAMPLES 147 Example 1 . u Example 1. 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.>. the state space E for the environment is { ('q..
More precisely. Nt Nt a .(3iBi(dx).e(A(Oi)d'sg)xe.5. one can associate in a natural way a standard Poisson one by averaging over the environment.A . )3*. . dx. iEE iEE )3 These parameters are the ones which the statistician would estimate if he ignored the presence of Markovmodulation: Proposition 1. B* = 1 /^* Bi. qij = 0 in the notation of Chapter 11. MARKOVIAN ENVIRONMENT We now turn to some more mathematically oriented basic discussion. N > 1(Ul < x) a4 B*(x).7 The Pidistribution of T1 is phasetype with representation (ei. the empirical distribution of the claims is B*. . t l=1 Note that the last statement of the proposition just means that in the limit. In particular. o = 0. we put )3* = E 7fi/3i. Note also that (as the proof shows) 7ri/3i//3* gives the proportion of the claims which are of type i (arrive in state i). A remark which is fundamental for much of the intuition on the model consists in noting that to each risk process in a Markovian environment. JT1 = j) = Qj • e. vi(dx) = . The key property for much of the analysis presented below is the following immediate observation: Proposition 1. the Markov additive structure will be used for exponential change of measure and thereby versions of Lundberg's inequality and the CramerLundberg approximation.8 As t oo.(Qi)diag)• More precisely.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. Pi (Ti E dx. Next we note a semiMarkov structure of the arrival process: Proposition 1.148 CHAPTER VI. 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.
has distribution (7ri)3i //3*)iEE and is independent of Ti. oo) as a 4 oo. aA. In particular.9 Consider a Markovmodulated risk process {St} with param eters Ni. Example 11. iEE 13 A different interpretation of B* is as the Palm distribution of the claim size. 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 . and let {St °i} refer to the one with parameters Pi. However .7. the Fidistribution of T1 in {St(a ) } is phase type with representation (E. N + oo Hence 1 Nt 1 N`+) Nits Nt E I ( Ut <. Hence Nt'> a . Bi. Bi(x). the limiting distribution of the first claim size U1 is B*.4.1. Proof According to Proposition 1. Conditioning . cf.. Proposition 1. {St} to the compound Poisson model with parameters 0 *.aA . and furthermore in the limit JT. 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.. 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$.(/3i)aiag). we may view Nt`i as the number of events in a Poisson process where the accumulated intensity at time t is Niti. Bi. Nt a' t t iEE Also.. ^j 7riNi.. A.2. zli( (u) . B*. In particular. this converges to the exponential distribution with rate 0* as a * oo. Then it is standard that ti lt '4' iri as t > oo. By Proposition A5. Then {St)} + {St*} in D[0.* (u) for all u.x) = Nt E > I (Uk) X) Nt Bi(x) 1=1 iEE k=1 iEE 1: t5 Bi( x) = B*(x).6. y Ni) i Nti) t a.. i. e.
0 Example 1.2. 5 5 where E5 denotes the exponential distribution with intensity parameter 5 and a > 0 is arbitrary.150 CHAPTER VI. 132=2. Claims of type E3 arrive with intensity 2 . the company even suffers an average loss.. state 1 appears as more dangerous than state 2. T) + ?P* (u. U2) are independent of .)..31µB 2 = 2 5 3 7 70 Thus in state 1 where p. which also yield O(a) (u. e. with T2 being exponential with rate . is as in {St }. thick.2}..FT.1 with periods with positive drift alternating with periods with negative drift. from Theorem 3.l3* and U2 having distribution B*. U.. Continuing in this manner shows that the limiting distribution of (T. A= (  a a ) \ a a 5 5 J 9 3 2 a1=2. the overall drift is negative since it = (2 2) so that p = 71P1 + 112P2 = 7. The fact that indeed 0(a) (U) 3 0* (u) follows.2 +2 2 = 3.10 Let E_{1. Thus. 1.. s 5 in state 2. oo).1.=1. shows similarly that in the limit (T2.s = 1o in state 2. On Fig.2. B2=1E3+4E7. we first get that 3 (3* = 2. 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 = . MARKOVIAN ENVIRONMENT upon FT. marked by thin.. lines in the path of {St}. T) for all u and T..s = o in state 1 and with intensity 1 . 9 . those of type E7 with intensity z s = 5 in state 1 and with intensity z .. 1. resp.. and (at least when a is small such that state changes of the environment are infrequent).g. there are p = 2 background states of {Jt}. B1=3E3+2E7. > 1. That is. since E3 is a more dangerous claim size distribution than E7 (the mean is larger and the tail is heavier). we may imagine that we have two types of claims such that the claim size distributions are E3 and E7. From this the convergence in distribution follows by general facts on weak convergence in D[0.1 of [145]. Computing the parameters of the averaged compound Poisson model. the paths of the surplus process will exhibit the type of behaviour in Fig.
11 (a ) ESt/t * p . iEE . That is. Hence B* = 415E3+5E7/ +4 ( 51E3 +5 E7) = 1E 3 +2E7. t + oo. 0 The definition (1. note first that EN Uk')/N a4' µgi.1 Thus. 01 /. Proof In the notation of Proposition 1.3* = 3/4 of the claims occur in state 1 and the remaining fraction 1/4 in state 2. t * oo. 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).1.(3.1 a.s.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 ..1). = P.1. the averaged compound Poisson model is the same as in III.8. Hence (i) Nti) 1 U(i) k' N(i)k=1 E t 4 St + t = iEE Nt t 1: 7ri Qi µs. a fraction r. MODEL AND EXAMPLES 151 Figure 1. For (b).
PB. and a more comprehensive treatment in Asmussen [16]. MARKOVIAN ENVIRONMENT Corollary 1.ld. EiX = 0. let some state i be fixed and define w=wl=inf{t >0:Jt_#i. Proposition 1.1)Eiw = 0. Theorem II. X 1 =Sty. 136 or A. 2 The ladder height distribution Our mathematical treatment of the ruin problem follows the model of Chapter III for the simple compound Poisson model. Eiw.Jt=i}. Since the X„ are independent . some early studies are in Janssen & Reinhard [211]. s. and hence oscillates between 0o and oo so that also here M = oo. The case 77 > 0 is similarly easy. having the Pidistribution of X..g. and hence 1/ii(u ) = 1 for all i and u..\ i and EiX1 Ei f 13 J. There seems still to be more to be done in this area. with X2. and ... limit p . In risk theory. The mainstream of the present chapter follows [16]. [APQ]. and hence M = 00.1 jEE = (p .2(a) p.Jt=i}. 38) Eiw1 = 1/ir.1(b) is essentially the same as the proof of the strong law of large numbers for cumulative processes. . also + . Now obviously the w.152 CHAPTER VI.0i(u) < 1 for all i and u. n n Thus {SWn l is a discrete time random walk with mean zero. [302]. with some important improvements being obtained in Asmussen [17] in the queueing setting and being implemented numerically in Asmussen & Rolski [43]. If 77 > 0. dt . X3. The proof of Proposition 1. [315]. then M < oo a.Eiw o'o Eiw • E ^ifjµs. then M = 00 a.12 If 77 < 0.. and involves a version of the .1 of St / t is > 0.. see the Notes to Section 7. See Meier [258] and Ryden [314].1 and the Corollary are standard.s. see [APQ] p. and so on.s. . and hence wn /n a4. Statistical aspects are not treated here. Now let r) = 0. X2 =SW2 So. 0 Notes and references The Markovmodulated Poisson process has become very popular in queueing theory during the last decade. [212]. Then by standard Markov process formulas (e... + Xn SWn ](1 a . w2=inf {t>w1:Jt_#i. Proof The case 77 < 0 is trivial since then the a.4..a form a renewal process ..
2 (a) The distribution of M is given by 00 1 . see also Example II. which represents a nice simplified form of the ladder height distribution G+ when taking certain averages : starting {Jt} stationary. 00 (2.j E E. Thus. That is.2) R(dx)S((y . only with the product of real numbers replaced by convolution of measures. Proposition 2. oo)) = f R(i.j. we define the convolution operation by the same rule as for multiplication of realvalued matrices. IIG+ II denotes the matrix with ijth element IIG+(i. oo) = J ao 0 G+(i.j.1 irG+(dy)e =. dx)/jBj(y . j. G+ is the matrix whose ijth element is E G +(i.3*B *(y)dy.2(a) below ) where the ladder height distribution is evaluated by a time reversion argument.1) 0 (b) G+ (y. T+ < oo) and let G+ be the measurevalued matrix with ijth element G+(i.a/i.EA. j. cf.x. j. we get the same ladder height distribution as for the averaged compound Poisson model.dx). However .Jr+ =j.g. •)• kEE Also.(u) = Pi(M < u) = e' E G+ (u)(I . •). •) * G +(k. but is substantially more involved than for the compound Poisson case . the definition of .A) = Pt(ST+ E A. oo)). 6.Jt=j)dt.4) we obtain the following result . e. j.. Define the ladder epoch T+ by T+ = inf It : St > 0} = r(0).2.6*.A) =ZI(St E. k. The form of G+ turns out to be explicit (or at least computable). Proposition 2. Let further R denote the preT+ occupation kernel.i. •) II = JG+(i. . j.IIG +II)e. by specializing results for general stationary risk processes (Theorem II . for i.5. (y. n=0 (2. For measurevalued matrices. and S (dx) the measure valued diagonal matrix with /3 Bj(dx) as ith diagonal element. THE LADDER HEIGHT DISTRIBUTION 153 PollaczeckKhinchine formula (see Proposition 2.x).6. B* in Section 1. let G+(i. T R(i.
2 useful . we need to invoke the timereversed version {Jt } of {Jt} . the intensity matrix A* has ijth element * 7r ^i3 7ri and we have Pi(JT = j) = 7rj P2(JT = i)7ri (2. and let further {my} be the Evalued process obtained by observing {Jt } only when {St*} is at a minimum value.IIG+II)e.3 When q > 0. we need as in Chapters II. resp. {mx} is a non terminating Markov process on E. JJ = j. St < S* for u < t. lines in the path of {St}. That is. From this (2. MARKOVIAN ENVIRONMENT Proof The probability that there are n proper ladder steps not exceeding x and (x)ej. hence uniquely specified by its intensity matrix Q (say). To make Proposition 2. G+ the probability that there are no further ladder steps starting from environment j is e^ ( I . and that the environment is j at the nth when we start from i is e . only with {Jt} replaced by {Jt } (the /3i and Bi are the same ). III to bring R and G+ on a more explicit form .2) is just the same as the proof of Lemma 11.3.6. see Figure 2. thick. The u proof of (2.1 The following observation is immediate: Proposition 2. 0  x Figure 2. marked by thin. . mx = j when for some (necessarily unique) t we have St = x.3) We let {St*} be defined as {St}.1 for an illustration in the case of p = 2 environmental states of {Jt}.154 CHAPTER VI. To this end .1) follows by summing over n and j.
(/3i)diag + T S(dx) eQx. 2. 0 mms1   ^O \ T.(/3i) diag. The definitions are illustrated on Fig.0. ( Q( n)) converges monotonically to Q. THE LADDER HEIGHT DISTRIBUTION 155 Proposition 2. s]. In general.and a jump (claim arrival) occurs at time t.2 . 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). {S. Otherwise each jump at a minimum level during the excursion starts a subexcursion. An excursion of {St*} above level x starts at time t if St = x. If there are no jumps in (t. and the excursion ends at time s = inf {v > t : S. the sequence {Q(n)} A* defined by Q(O) = .. and the excursion is said to have depth 1 if each of these subexcursions have depth 0. Figure 2. we say that the excursion has depth 0. we recursively define the depth of an excursion as 1 plus the maximal depth of a subexcursion. Furthermore.*. For example the excursion of depth 2 has one subexcursion which is of depth 1.4 Q satisfies the nonlinear matrix equation Q = W(Q) where 0 co(Q) = n* . corresponding to two subexcursions of depth 0. Q( n+l) _ ^. = x}. Proof The argument relies on an interpretation in terms of excursions. and S(dx) is the diagonal matrix with the f3iBi(dx) on the diagonal.2. } is a minimum value at v = t.2 where there are three excursions of depth 1.2.
5 R(i. = j. we first compute qij for i $ j. A) = L' U(j. Similarly. of the definition to make U be concentrated on (co. it becomes clear that pij = r [eQh] 0 ij Bi (dy) • (2. Writing out in matrix notation . Fi(mh =i ) = 1 + =hflh+Qihpii+o(h) implies qii = 'iii /i +)3ipii. Now let {m ( n) } be {mx } killed at the first time i7n (say) a subexcursion of depth at least n occurs .5) A (note that we use A = {x : x E Al on the r. A) = f Pi(mx = j) dx eie4xej dx A u (2. either due to a jump of {Jt } which occurs with intensity A= j. the subintensity matrix of {min+i ) } is cp (Q(n)) = Q(n +l) which implies that qgj +1) = \!. 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.St <S*. p1^) Define a further kernel U by f U(i.u< t). i. By considering minimum values within the excursion. mx+dx = j) occurs in two ways .4) To show Q = cp(Q). Q = W(Q) follows.. (2. or through an arrival starting an excursion terminating with J. e.156 CHAPTER VI. 7rE Proof We shall show that Fi(Jt=j.j. A).s. Similarly by induction ..4).j +/3ipij. It follows that qij = A.6) . The proof of Q = W(Q) then immediately carries over to show that the subintensity matrix of {mil) } is cp (Q(o)) = Q(l). StEA .T+>t) _ ^iF 7ri (JJ =i. Theorem 2 . h. 0)).St EA. It is clear that { mini } is a terminating Markov process and that { mio) } has subintensity matrix A* .Qi + )%pij) Now just note that t pij and insert (2.(01)diag = Q. Suppose mx = i. j. Then a jump to j (i.
e. 0 < u < t) = 7rjPj(Jt =i. consider stationary versions of {Jt}. u It is convenient at this stage to rewrite the above results in terms of the matrix K = 0'Q'A.z+>t) = P.(Jt=j.2. `` {K(n)} [the W(•) here is of course not the same as in Proposition 2.StEA. oo)) = f o' eIXS((x + z. THE LADDER HEIGHT DISTRIBUTION 157 from which the result immediately follows by integrating from 0 to oo w.St <Su. G+((z. {Jt }. St EA. St E A.g.6). Remark 2.4].1 can be rederived using the more detailed form of G+ in Corollary 2. it is readily checked that 7r is a left eigenvector of K corresponding to the eigenvalue 0 (when p < 1).r.0<u<t. We may then assume Ju=Jtu. From Qe = 0.Jo=i. S. and this immediately yields (2.St EA.6 is hardly all that explicit in general.6 (a) R(dx) = eKxdx.7 It is instructive to see how Proposition 2.(. K( n (d) the sequence converges monotonically to K. (b) for z > 0.. (c) the matrix K satisfies the nonlinear matrix equation K = W(K) where W( K) = A ( i) diag + fi J "O eKx S(dx).=StSt. 0<u<t).t. (Jo = j... dt. we shall see that nevertheless we have enough information to derive.0<u<t) = P.6(b): from 7rK = 0 we get 7rG+(dy)e = J W 7reKx(fiiBi(dy + x))diag dx • e 0 w(fiiB1(dy + x))col dx f 0 EirifiiBi(y)dy = fi*B*(y)dy• iEE 0 Though maybe Corollary 2. and get irPi(Jt =j. where A is the diagonal matrix with 7r on the diagonal: Corollary 2.. 0 +1) = cp (K( n)) defined by K(o) = A . and we let k be the corresponding right eigenvector normalized by Irk = 1. oo))dx. x < 0.S„<0. the CramerLundberg approximation (Section 3). and to obtain a simple solution in the . Jt = i.Qi)diag. St < St U.. To this end.
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
3. (3.j) * coj)(u) _ f u ^Pj(u .j. dy)• o iEE jEE . yu < r(u) < 00 h 4(u) < h(av)C+o)(y)eavuEav . We further write G(u) for the vector with ith component Gi(u) = EiEE G+(i. r(u) < yu] hiay)C+ h=av)C+ o) (y)eayu+yuw(av). exp {e() + r(u))} . However. we have ic(ay) < 0 and get 'i(u) . Chie ryu < Vi(u ) < C+hie 7u. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES 165 hiav)e _avuE. av 'i [h.V)i(u. as in the classical case (3.i I (a) exp {aye(u) + r(u)r. 1 Similarly. if y > 1lk'(ry). we shall need the matrices G+ and R of Section 2.5) will produce the maximal ryy for which the argument works. for a vector <p(u) = (cpi (u))iEE of functions .i [eT(u)K(av ).y)G+(z. r(u) yu o)(y)eavuEav. 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 3.00 su e7( ( 3.8 ) Then for all i E E and all u > 0. (u.7. we let G+ * W(u) be the vector with ith component E(G+(i. 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. oo)) and.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)..i [e*(u)K(av). yu) f h(av)e v avuE«v. hj P .9) For the proof.5).(ay)}.
= Eo G+ G. °O . dx) 100 C .& (u). j.x ) = Gi(u). dx). dy) = aj f Bj(dy .x ) R(i. 00 f C_ hj f e(Y)G+(i.(0) ] (u) < sup Jt t. j.166 CHAPTER VI. j. we have G *(N +1) * ^. _ To see that the ith component of U * G(u) equals ?Pi (u).dy). dy) : 1(u) < C+ > hj u e(1tL)G+(i.3jhj // f 00 R(i. dx) f e7( vu)Bj (dy . U = U".u IMP:°) (u) I < oo. j.12 Assume sup1. 00 Thus C+ > hj f"o e7(Yu)G +(i. Hence lim cp(n) exists and equals U * G.ery(&u+x)Bj (dy) Bj(u Bj (u . 0 G+(i.j. jEE u 0 j. Proof Write UN = EN G+ .13 For all i and u. Then cpin)(u) sit (u) as n + oo. if r+ (n) is the nth ladder epoch. Lemma 3 .u Iv 2°)(u)I Pi(rr+(N + 1) < oo) + 0. and define W(n+1) (u) = G(u) + (G+ * tp(n))(u). dy) 00 C+ ijhj f R(i.j. dx ) Bj (u . Then iterating the defining equation ip(n+1) = G + G+ * V(n) we get W(N+1) = UN * G + G+N+1) * ^(o) However. 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. MARKOVIAN ENVIRONMENT Lemma 3 .x) jEE 00 u 0 //^^ C+E.7.x) x) jEE 0 E Qj f jEE R(i. n > oo.
dy) jEE o (3.3.T) = Pi(M > u) . 14 Let yo > 0 be the solution of 'c'(yo ) = 0.tpi(u.11. from which the lower inequality follows by letting n * oo.12) Proof We first note that just as in the proof of Theorem 3. T) = Pi (7. 167 u Proof of Theorem 3.Pi(MT > u) = Pi(MT < u. j.ST). let C+(yo) be as in (3.10) C_ 1 f hje7(y.u)G+(i.n) ( u . dy) jEE u U +C_ hje7( yu)G jEE"" +(i. it follows that Vi(u) < C_(yo) h=70)e7ou. we have Vii (u) . and assuming it shown for n.13) Hence. 9 for the last equality in (3. Then 0< Vi (u )  0i(u. j.10 ) by Lemma 3 .y)G+(i. (3.13. letting MT = maxo<t<T St. j.11). Indeed.11) C_e7u 57 O+[i. T) < C+(')' o)hi7u)e7ou8T . +i . u The proof of the upper inequality is similar .10: Theorem 3 . Here is an estimate of the rate of convergence of the finite horizon ruin probabilities 'i (u. CHANGE OF MEASURE VIA EXPONENTIAL FAMILIES proving the upper inequality. We claim by induction that then cpin) (u) > C_ hie7u for all n. 13 and the second by the induction hypothesis . and using Lemma 3 . j.MT<u. this is obvious if n = 0. 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.M > u) = Pi(ST<u.13 Let first cp=°)(u) = C_ hie"u in Lemma 3. jEE estimating the first term in (3. we get Wo n +1) (u) = ? 7 i ( U ) + E J u gyp. and the proof of the lower one is similar. and let 8 = e'(70). (3. y]hj = C_ e7uhi. MT < u. taking cps°) (u) = 0. dy) (3.8) with y replaced by yo and hi by h=7o ).M>u) = Ei [VGJT (u .
2) alone just amounts to an ordering of the states. in part from the folklore principle that any added stochastic variation increases the risk.31:5)32 . The motivation that such a result should be true came in part from numerical studies. B2 <_s. It was long conjectured that 0* Vi.3) Bl <_s.. (4..1) Obviously.168 CHAPTER VI.. we define the stochastic ordering by 0' < s.. Occasionally we strengthen (4. we also assume that there exist i # j such that either /3i <. [177].o.2) (4. we refer to . this correponds to the usual stochastic ordering of the maxima M'.33 or Bi 0 Bj. where it has been observed repeatedly that Markovmodulation increases waiting times and in fact some partial results had been obtained. The Markov process {Jt} is stochastically monotone (4.. (4. but that in general the picture is more diverse. <s. and finally in part from queueing theory. Bp.5) Note that whereas (4.o. u > 0. ".o. The conditions which play a role in the following are: . V)" if z/i'(u) <'c"" (u).3). this is not the case for (4..0. MARKOVIAN ENVIRONMENT Notes and references The results and proofs are from Asmussen and Rolski [44]. < . The results to be presented show that quite often this is so.. Further related discussion is given in Grigelionis [176]. 4 Comparisons with the compound Poisson model 4a Ordering of the ruin functions For two risk functions 0'. M" of the corresponding two claim surplus proceses (note that 0'(u) _ P(M' > u).4) To avoid trivialities.3) to B = Bi does not depend on i. (4. where o*(u) is the ruin probability for the averaged compound Poisson model defined in Section 1 and . is the one for the Markovmodulated one in the stationary case (the distribution of J0 is 7r).o..3p. 0"(u) = P(M" > u)) Now consider the risk process in a Markovian environment and define i' (u) _ >iEE irioi(u). For the notion of monotone Markov processes.
. Proof of Theorem 4.3iBi(x)YPi(u . (4.x)dx _ /3*B*(u) + f u / ^ t=1 > 3 * B* ( ) + f (4..5 (cf. then j is the more risky state ..3* f uB(x) z/^. T(0) < oo) = Bi(x) dx/tcai ..2)(4. < a.4) is automatic in some simple examples like birthdeath processes or p = 2 .9) (4..10) Q*B*(u)+.6) 7r= fl*B*(u) + p> s=1 +) fu 0 b (u  x)Bt (x) /pB.10) and (4. the second follows from an extension of Theorem I1..7) and Lemma 4.3 for the second) *(u) _ /3 *B* (u) +. also Proposition 2.r (Sr(o) E dx Jr(o) = i. Comparing (4.6). 7(0) < oo) = pirf+). Conditions (4...1 for the first term in (4.1) which with basically the same proof can be found in Asmussen & Schmidt [49].x)B*(x) dx.. < bp and 7ri > 0 (i = 1. = b. Section 4. b1 < . we obtain (cf.x) of i and using Lemma 4. it follows by a standard . The first is a standard result going back to Chebycheff and appearing in a more general form in Esary. 1:7riaibi > E 7riai i=1 i=1 j=1 The equality holds if and only if a1 = . COMPARISONS WITH THE COMPOUND POISSON MODEL 169 Stoyan [352].2)(4. Proposition 2.r (JT(o) = i. then P P P 7rjbj. E 7r i Wi(u . dx (4. and it is in fact easy to show that Vii(u ) < t/j(u) (this is used in the derivation of (4.8) ^j Tri/iBd(x) ..6.4) hold.2. 3 (a) P. Conditioning upon the first ladder epoch.7) 7ri. 2 If al < .9) follows by considering the increasing functions 3iBi (x) and Oi (u . (b) P. we need two lemmas. = aP or b1 = .13* J0 u 0*(u . Lemma 4 . .1.. Lemma 4 .1 Assume that conditions (4.9 ) below)... where 7r2+) = QiµBilri/p. Proschan & Walkup [140]. note that (4. Then V.2.. Theorem 4 . .* For the proof.4) say basically that if i < j .x) dx u o i =1 i=1 (4. ^i 7ri = 1.r(u x)dx. 0 Here (4. p).4.
8) we get P P '*' (0) = 3* + /3*1* (0) _ > lri'3qqi • E 7i/ipBi . 0. As is seen. of order 101.* (0). µB. (4. For u = 0. Using (4.(0) = V. Bi as e J. (u) is not in general true: Proposition 4. they are at present not quite complete. 01 = 103.0*• i=1 But it is intuitively clear (see Theorem 3. MARKOVIAN ENVIRONMENT argument from renewal theory that tk.4 is the understanding of whether the stochastic monotonicity condition (4. that P P /^ 1r1NiµBi /^2 /^ ^i/ji pBi < 1il3i i=1 i=1 (4.s.. Then i/i*(u) < .h. Proof Since 0. except possibly for a very special situation . (u) may fail for some u. Then the l. u Here is a counterexample showing that the inequality tp* (u) < V).s.4 is not vacuous. let = ( 1/2 1/2 ) . Q2 = 1.3i.. µB2 = 104.3µi < 1 for all i. What is missing in relation to Theorem 4.170 CHAPTER VI. of (4.1 of [145] for a formal proof) that z/ii(u) converges to the ruin probability for the compound Poisson model with parameters .6). Frey.0. Rolski & Schmidt [32].1 and Proposition 4./3*. Recall that .h. this ruin probability is /3iPBi.4) is essential (the present author conjectures it is). dominates the solution 0* to the renewal equation (4.11) is of order 104 and the r.(0) < b *'(0) for e small enough. i=1 i=1 7'r(0) _ EFioiwi(0) . it will hold for all sufficiently large u. 4b Ordering of adjustment coefficients Despite the fact that V)* (u) < *. Notes and references The results are from Asmussen.r (u ) fails for all sufficiently small e > 0.4 Assume that . and from this the claim follows.6). = 102..11) i=1` and that A has the form eAo for some fixed intensity matrix A0. it is sufficient to show that 0'.2... u To see that Proposition 4.
13) (4. It is clear that the distribution of X. in particular . This implies that A is strictly convex.a.5 y < ry*.ld) with generic cycle w = inf{t>0: Jt_54 k. Now we can view {Xt} as a cumulative process (see A.4. Jt = i])' EE = vA+n(6.a = E irirci(a). (4.1) . with strict inequality unless rci (y*) does not depend on iEE. 0 . which in view of EiEE 1ibi = 0 is only possible if Si = 0 for all i E E. is nondegenerate unless bi does not depend on i E E.Jt=kI A (the return time of k) where k E E is some arbitrary but fixed state. with strict inequality unless a = 0 or bi = 0 for all i E E..4(b) that the limit in (4. Xt)} is a Markov additive process (a socalled Markovian fluid model.5.1) .5.)a.2 we have (Ei[e"X'.(a) > 0 for all a 0 0. Further (see Corollary 11.7) )i is convex with A'(0) = lim EXt tioo t = iEE 70i = 0. 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. e. Asmussen [20]) as discussed in 11. COMPARISONS WITH THE COMPOUND POISSON MODEL 171 the adjustment coefficient for the Markovmodulated model is defined as the solution y > 0 of ic(y) = 0 where c(a) is the eigenvalue with maximal real part of the matrix A + (rci(a))diag where rci(a) = ai(Bi[a] .g. Lemma 4.14) is nonzero so that A"(0) > 0.13) implies A(a) > 0 for all a. cf. Proof Define X= f &ids. and by Proposition II. (4.6 Let (di)iEE be a given set of constants satisfying EiEE iribi = 0 and define A(a) as the eigenvalue with maximal real part of the matrix A + a(bi)diag• Then )t(a) > 0. (4.g.12) iEE Theorem 4. it follows by Proposition A1. Then {(Jt. Hence if 5i 54 0 for some i E E.14) A„(O) iioo varXt t t By convexity..
15) once more and letting e = 0 we get . Thus y(e) * y* as e 10. note that y(a) + mins=1. improving upon more incomplete results from Asmussen.172 CHAPTER VI. If rci(y* ) is not a constant function of i E E. (4. Notes and references Theorem 4. h(0) = e. (4.. MARKOVIAN ENVIRONMENT Proof of Theorem 4.15) yields 0 = (Ii(y*)) diage + Aoh'(0) = (rci('Y*)) diage + (Ao . and our aim is to compute the sensitivity ay e a E=O A dual result deals with the limit a 4 oo. h'(0) = (Ao . In the case of e. this implies that the solution y > 0 of K(y) = 0 must satisfy y < y*. Hence rc (y*) > 0. Frey.. 0 = ((ri(Y))diag + ery (4{('Y))diag)h + (A0 + e(?i'Y))diag)h'. a = 1 in Lemma 4. Here we put a = 1/e.. Further a(1) = rc(y*) by definition of A(.6. whereas the . 4c Sensitivity estimates for the adjustment coefficient Now assume that the intensity matrix for the environment is Ae = Ao/ e.12) and rc*(y*) = 0. Hence letting e = 0 in (4.e7r)1 (Ici(Y*))diage.5 is from Asmussen & O'Cinneide [40]. The corresponding adjustment coefficient is denoted by ry(e).) and rc (•). the basic equation is (A + (rci(y))diag)h = 0.p yi and compute 8y 8a a=0 In both cases. Since ic is convex with rc'(0) < 0 . y.15) Normalizing h by 7rh = 0.Qi and Bi are fixed . we get rc (y*) > 0 which in a similar manner implies that u y < y*. Let bi = rci(y*). h depend on the parameter (e or a). Then > risi = 0 because of (4.eir)h'(0).5. we have 7rh' = 0. multiply the basic equation by a to obtain 0 = (A0 + e(r£i(y))diag)h.16) Differentiating (4.. where A. Rolski & Schmidt [32].
5. and may have some relevance in risk theory as well (though this still remains to be implemented).. Frey.7 8ry aE = 1 7r(ci ('Y*))diag ( Ao e7r)1(Xi(Y*))diage *=0 P Now turn to the case of a. Rolski & Schmidt [32]. 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.8 If (4. We get 0 = (aAo + ( lc&Y))diag)h. the intensity for such a transition (referred to as marked in the following) is denoted by Aii l and the remaining intensity . 0 = (Ao + ry'(ii(Y)) diag )h + (aAo + (Ki(7'))diag)h'. . Inserting (4. multiplying (4. (4. (4.18). i = 2.20) 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). and we have proved: Proposition 4. We assume that 0 < y < 7i. (4.19) holds.8 when ryi < 0 for some i is open.16) yields Proposition 4.. THE MARKOVIAN ARRIVAL PROCESS 173 0 = 27'(0)(ri(`Y *)) diage + 2(ci('Y* )) diag h' (0) + Aoh" (0) .18) 0 = 27'(0)p+27r(rs.17) by 7r to the left to get (4.19) Then 'y ^ ryl as a ^ 0 and we may take h(0) = el (the first unit vector).20) Letting a = 0 in (4. which has recently received much attention in the queueing literature. The analogue of Proposition 4. .17) (4. p. 5 The Markovian arrival process We shall here briefly survey an extension of the model.i(7' *))diagh'(0). then 8a a=o All rci (0) Notes and references The results are from Asmussen.
1 (PHASETYPE RENEWAL ARRIVALS) Consider a risk process where the claim sizes are i. where qij is the probability that a transition i * j is accompanied by a claim with distribution.i. the definition of Bi is redundant because of f3i = 0.2 for details). Bij = B. A(1) = A . that Bii = Bi . The extension of the model can also be motivated via Markov additive processes: if {Nt} is the counting process of a point process. then {Nt} is a Markov additive process if and only if it corresponds to an arrival mechanism of the type just considered. the claim surplus is a Markov additive process (cf. we may let {Jt} represent the phase processes of the individual interarrival times glued together (see further VIII. and that are determined by A = A(l ) +A(2) where A is the intensity matrix the governing {Jt}. Jt2)) (2. refer to notation) { Jt k) }. and the marked transitions are then the ones corresponding to arrivals.6i ) diag. Note that the case that 0 < qij < 1. Thus . For i = j. j(2) } be two independent environmental processes and let E(k). This is the only way in which arrivals can occur. is neither 0 or 1 is covered by letting Bij have an atom of size qij at 0.(13i )diag. We then let (see the Appendix for the Kronecker E = E(1) x E(2). T). Indeed. with common distribution B. the definition of Bij is redundant for i i4 j.2) A(1) = A(' 1) ® A(1.2). the Markovmodulated compound Poisson model considered sofar corresponds to A(l) = (. Here are some main examples: Example 5 . let { Jt 1) }. MARKOVIAN ENVIRONMENT f o r a transition i + j by A . u Example 5 . . Again . In the above setting. Jt = (Jtl).4).2 (SUPERPOSITIONS) A nice feature of the setup is that it is closed under superposition of independent arrival streams . A(l) = T. A(1'k) A(2 k1). A ( 2) = A (2`1 ) ® A.d.174 CHAPTER VI. II. A(l) = tv. B. and thus 1i = 0. we use the convention that a1i = f3i where 3i is the Poisson rate in state i.^) etc. Bii = Bi . but the point process of arrivals is not Poisson but renewal with interclaim times having common distribution A of phasetype with representation (v.
say.kl is redundant).iN. Easy modifications apply to allow for • the time until expiration of the kth policy is general phasetype rather than exponential. iN = all BOi2.. the idea of arrivals at transition epochs can be found in Hermann [193] and Rudemo [313]. E 10. Similarly. i2i . • upon a claim. iN. 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 . after which it starts afresh. The individual pays at rate pi when in state i and receives an amount having distribution Bij when his/her state changes from i to j.1i2 . Example 5 .iN = a2. INVALIDIZED... MARRIED. where ik = 0 means that the kth policy has not yet expired and ik = 1 that it has expired. assume that there is a finite number N of policies. 11.kj = Bik) B13 4k = Bak) 175  (the definition of the remaining Bij. the kth policy enters a recovering state.. In fact . However .iN are zero and all Bi are redundant. Hermann [193 ] and Asmussen & Koole [37] showed that in some appropriate . The versatility of the setup is even greater than for the Markovmodulated model. E = { WORKING....iN.iil. THE MARKOVIAN ARRIVAL PROCESS Bij.. with rate ai. claims occur only at state transitions for the environment so that AN2. superpositions of renewal processes..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). iN.. Thus. In this way we can model. This means that the environmental states are of the form i1i2 • • • iN with il..iN = C27 All other offdiagonal elements of A are zero so that all other Bii are redundant....iN C17 AilO.. all Al i2. DIVORCED.}. as the Markovian arrival process ( MAP).5.1i2. and that the policy then expires.. DEAD etc. e. Assume further that the ith policy leads to a claim having distribution Ci after a time which is exponential. possibly having a general phasetype sojourn time. WIDOWED. Bilo. more recently. RETIRED..iil.... or.4 (A SINGLE LIFE INSURANCE POLICY ) Consider the life insurance of a single policy holder which can be in one of several states..g. u Example 5 .
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 . Sengupta [336]. from an application point of view. We denote throughout the initial season by s and by P(8) the corresponding governing probability measure for the risk process. Neuts [271] and Asmussen & Perry [42]. p(t) and B(t) are defined also for t t [0. 1). we talk of s as the 'time of the year'. we may assume that the functions /3(t). . Let 1 1 /3* _ f /3(t) dt. where i f00 xB(°) (dx) _ . The basic assumptions are as follows: • The arrival intensity at time t of the year is 3(t) for a certain function /3(t).176 CHAPTER VI.1) Then the average arrival rate is /3* and the safety loading rt is 77 = (p* . 1). )3 t 1 J (6. Without loss of generality. a claim arrives with rate /3(s + t) and is distributed according to B(8+0 . one limitation for approximation purposes is the inequality Var Nt > ENt which needs not hold for all arrival streams. • Claims arriving at time t of the year have distribution B(t). For the Markovmodulated model. Lucantoni [248]. Obviously. B* = J f B(t) ((*) dt. continuity would hold in presumably all reasonable examples. [248].2) Note that p is the average net claim amount per unit time and µ* = p//3* the average mean claim size. 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. for s E E = [0. but now exhibiting (deterministic) periodic fluctuations rather than (random ) Markovian ones. 0 < t < 1. one needs to assume also (as a minimum) that they are measurable in t.3*µs • p = f /3(v) dv 0 0 (6. By periodic extension. • The premium rate at time t of the year is p(t). let the period be 1. Thus at time t the premium rate is p(s + t). Lucantoni et at. Some main queueing references using the MAP are Ramaswami [298]. p * = 0 p(t) dt.p)/p.
let . p* as an averaged version of the periodic model. RISK THEORY IN A PERIODIC ENVIRONMENT 177 In a similar manner as in Proposition 1. 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. 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 . The behaviour of the periodic model needs not to be seen as a violation of this principle. and we recall from there that the ruin probability is 24 1 *(u) _ 3 5eu + 35e6u. Thus. the conditional distribution . 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 . Section 4b). in agreement with the general principle of added variation increasing the risk (cf.1) and Example 1.1.8. u Remark 6 . we shall see that for the periodic model increasing A increases the effect of the periodic fluctuations.3(t) = 3A(1 + sin 27rt). It is easily seen that . since the added variation is deterministic. Thus .2 Define T 6(T) = p(t ) dt.t. the discussion in 111. In contrast. it turns out that they have the same adjustment coefficient. We u assume in the rest of this section that p(t) . In contrast.w(t). not random. for Markovmodulated model typically the adjustment coefficient is larger than for the averaged model (cf. one may think of the standard compound Poisson model with parameters 3*.1 As an example to be used for numerical illustration throughout this section. Many of the results given below indicate that the averaged and the periodic model share a number of main features. equivalently.3* = 3A.w(t)) dt). B*.3) Note that A enters just as a scaling factor of the time axis. the average compound Poisson model is the same as in III.6. (6. In particular. The arrival process {Nt}t>0 is a timeinhomogeneous Poisson process with intensity function {/3(s + t)}t>0 . of the periodic model as arising from the compound Poisson model by adding some extra variability. or.(3. The claim surplus process {St } two is defined in the obvious way as St = ^N° Ui .9). and thus the averaged standard compound Poisson models have the same risk for all A.10. respectively. St = SeI(t).
east B(8+t) [a] east . of the averaged compound Poisson model (the last expression is independent of s by periodicity). 3 E(8)eaSt = h(s. and the ruin probabilities are 0(8) (U) = P(s )(r(u) < 00).8). a) is periodic on R. J Theorem 6 . Jt = (s + t) mod 1 P(8) . .1]) . Notes and references The model has been studied in risk theory by. let f 8+1 tc *(a) _ (B* [a] .g.a be the c.f. The claim surplus process {St} may be seen as a Markov additive process.a.1) a = J8 . (6. with the underlying Markov process {Jt} being deterministic period motion on E = [0. of the claim surplus process.a) Proof Conditioning upon whether a claim occurs in [t. 0 (5)(u.Q(v) (B(„) [a] . and define h(s. 1). 6b Lundberg conjugation Motivated by the discussion in Chapter II. 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.(1 . Dassios & Embrechts [98] and Asmussen & Rolski [43].g. To this end.al. t + dt] or not.a ..3(v)(B(vl [a] .tc* (a)] dv then h (.5..a) = exp { . [44] (the literature in the mathematical equivalent setting of queueing theory is somewhat more extensive. i.^8 [.3(s + t)dt[B(8+t)[a] . given that the ith claim occurs at time t is B(8+t). see the Notes to Section 7).T) = P(8)(r(u) <T).5 (see in particular Remark 11. with some variants in the proofs.adt +. Daykin et.1) dv . [101] . e.4) At a first sight this point of view may appear quite artificial. we obtain E.178 CHAPTER VL MARKOVIAN ENVIRONMENT of U.e.(3(s + t)dt)e«St adt + /3(s + t)dt . r(u) _ inf It > 0 : St > u} is the time to ruin . As usual.. a) etw*(a) h(s+t. we start by deriving formulas giving the m.f.1) . The exposition of the present chapter is basically an extract from [44].s .(8) [eaSt+dt I7t] = = (1 .g.
log h(s.6. a) Thus E(8)east = h(s + t. a) h(s + t. at + f log h(s + t.3.adt +.1)dv  o h(t.4 For each 0 such that the integrals in the definition of h(t . According to Remark 11. a) .2. + v)(B([a] . a) Corollary 6. 9) is a P ( 8)martingale with mean one. u Remark 6.t}t>o = h(s. St)} and . a) = exp I f t3(v)(kv)[a) . a). RISK THEORY IN A PERIODIC ENVIRONMENT E(8)east+ dt d Et.9) eastt. h(s + t. it then suffices to note that E(8)Le.t = 1 by Theorem 6. dt log E(8)east a + f3(s + t) [B(8+t) [a] .(8)east 179 = = = = = E(8)east (1 ..t} is a multiplicative functional for the Markov process { (Jt. Proof In the Markov additive sense of (6.4).1]) .0(s + t)dt[B(8+t)[a] . 0) P(8)a.9 as follows.1)dv l og E(8) et where atetk•(a) h(t.c* (e) {Le.1]) . so that obviously {Lo.5. 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 (8)east (a +. a) et.* (a) h(s.6 .5 The formula for h(s) = h(s. a) = h(s. we can write Lo Jt. 0) exist and are finite. With g the infinitesimal generator of {Xt} = {(Jt.s. B) eoSt t.(e) Let = h( h(Jo.3(s + t)[D(8 +t)[a] .t.1]. St)} .
0) = h(s) + dt {ah(s) . it follows by Theorem II. St)} with governing probability measures Fes). of St is as for the asserted periodic risk model. That is.g. When a = y. Bet)(dx) = ^ B(t ) (dx). Proof (i) Check that m.6 ( s ) exp { 0( s )&s) [a] + tc . we put for short h(s) = h(s. (ii) use Markovmodulated approximations (Section 6c).3(s)ks)[a]h(s)} ah(s) 13(s)h(s) + h'(s) +.a .y) = eayh(s).1. That rc = is*(a) then follows by noting that h(1) _ u h(0) by periodicity.f. y solves n* (y) = 0. However. Equating this to rch (s) and dividing by h(s) yields h(s ) = h(s) = a + . Lemma 6 . For each 0 satisfying the conditions of Corollary 6.3(v)( Bi"i [a] .3(s)B(s) [a]h(s). (iv) finally. such that for any s and T < oo.T. Proposition 6. Now define 'y as the positive solution of the Lundberg equation for the averaged model.0) = Kh(s). ( iii) use approximations with piecewiese constant /3(s).tc] dv} (normalizing by h(0) = 1).2.1) . the restrictions of Plsi and Pest to Ft are equivalent with likelihood ratio Le.'y).3(s)dt • B(s)[a]h(s) = gha(s. correspond to a new periodic risk model with parameters ex .7 When a > yo.4.3(s)h(s) + h'(s) +. 0 < s < 1. MARKOVIAN ENVIRONMENT ha(s. as above E (s) ha(Jdt. Proposition 6.60(t) = a(t)B(t)[0]. That is. . see [44] for 11 a formal proof.(3(s)dt) +.3.180 CHAPTER VI. B(s). [70] .5 that we can define a new Markov process {(Jt. A further important constant is the value yo (located in (0. J s [. yo is determined by 0 = k* (70) = QB*. P(s) (T(u) < oo) = 1 for all u > 0. Sdt) = h(s + dt) eadt (1 .6 The P(s). the requirement is cha(i. ry)) at which n* (a) attains its minimum. say. cf.
B(oo)). have components with densities b(8)(x) satisfying inf sEI. a)e«uE (a iP(s) (u) = h( s)e7uE(` ) h(O(u)) To obtain the CramerLundberg approximation from Corollary 3. Corollary 6. the Markov process {(^(u). 9(u)) for any bounded continuous function (e. xEJ 0 (s)b(8)(x) > 0. The proof involves machinery from the ergodic theory of Markov chains on a general state space. q) = eryx/h(q)). a) TI h(9(u).1. (6. 1). J C R+ such that the B(8).9) 0(')(u) = h(s. and no matter what is the initial season s. which is not used elsewhere in the book. Wu). has a unique stationary distribution. Lemma 6 . u which is > 1 by convexity. e(cc)) Letting u > oo in (6. a) e«uE(8 ) e «^ . RISK THEORY IN A PERIODIC ENVIRONMENT Proof According to (6.2). we need the following auxiliary result . Here and in the following.2. considered with governing probability measures { E(8) }E[ . T(u) < (6.9 Assume that there exist open intervals I C [0.1) the distribution of (l: (oo).9) and noting that weak convergence entails convergence of E f (^(u). a) a > ry0 (6.7) h(B(u).g. 0(u)) * (b(oo).10) Then for each a. T) = h(s. s E I. and we refer to [44].6(v) dv Jo ' xe«xB (°) (dx) r^ xe«xB'(dx) = Q'B' [ a] = ^' J 0 = ^c"'(a) + 1. The relevant likelihood ratio representation of the ruin probabilities now follows immediately from Corollary 11. the mean number of claims per unit time is p« 181 = Jo 1. ^(u) = ST(u) .6.4. say s0.8 The ruin probabilities can be computed as (u)+T(u)k'(a) ^/i(8) (u.u is the overshoot and 9(u) = (T(u) + s) mod 1 the season at the time of ruin. f (x. we get: .8) (6.9(u))} u>0.
it does not seem within the range of our methods to compute C explicitly.Ch(s)ery".10) of Lemma 3. where C(o) = 1 + info < t<i h(t) . A=1/4 A=1 A=4 0 Figure 6. (6. which may provide one among many motivations for the Markovmodulated approximation procedure to be considered in Section 6c.W.16. At this stage . 6.182 CHAPTER VI.10 Under the condition (6. 10 shows that certainly ry is the correct Lundberg exponent. Vi(8) (u) .ir) } Plots of h for different values of A are given in Fig. For our basic Example 6 .11) Note that ( 6. 11 7/'O (u) < C+°)h(s) ery".) C = E1 h(B(oo)) u + oo. MARKOVIAN ENVIRONMENT Theorem 6. where e. we obtain immediately the following version of Lundberg ' s inequality which is a direct parallel of the result given in Corollary 3.1.6 for the Markovmodulated model: Theorem 6 . Among other things. Noting that ^(u) > 0 in ( 6. Theorem 6 . this provides an algorithm for computing C as a limit. illustrating that the effect of seasonality increases with A. elementary calculus yields h(s) = exp { A C 2^ cos 2irs  4^ sin 21rs + 11 cos 41rs .1 In contrast to h.1. 1.11) gives an interpretation of h(s ) as a measure of how the risks of different initial seasons s vary.9).
1 ) and all u > 0.w)e4u .47r sin 27rs + 167r cos 47rs .7e .17) (6.15) The next result improves upon the constant C+) in front of eryu in Theorem 6. C_h(s)e7u < V.12) As for the Markovian environment model. Lundberg's inequality can be con siderably sharpened and extended. r.g.4. where ay is the unique solution of W(ay) =y• (6. We state the results below. e7 ( yx)B(t)(dy) > Then for all $ E [0.13 to our basic example. we first note that the function fu° exu {w • 3e .3x + (1 . 1 (6.yr. we substitute T = yu in 0(u. whereas ay < y.0(8) (u+ yu) (6.13 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. the proofs are basically the same as in Section 3 and we refer to [44] for details.w ) • 7e u{w • 3e3x + ( 1 .w)e4u dx 9w + 7(1 .13) Elementary convexity arguments show that we always have ryy > Y and ay > ry. e.42 • exp {J_ cos 27rs . Theorem 6 . (6. Consider first the timedependent version of Lundberg's inequality. (ay).(ay) > 0 when y < 1/ic' (7).16 In order to apply Theorem 6.w) .(s)(u) < C+h(s)e7". . ay) • (6.12 Let 00)(y) 1 Then info < t<i h(t..6. RISK THEORY IN A PERIODIC ENVIRONMENT Thus.42 so that 183 tp(8) (u) < 1.14) < C+)(y)h(s) e7yu.(8) (u.11 as well as it supplements with a lower bound. in our basic example with A = 1. Theorem 6.167r I Cu. yu) 000 (u) . Just as in IV. #c( ay) < 0 when y > 1/tc'('y). T) and replace the Lundberg exponent ry by ryy = ay .7x j dx _7x } _ 6w + 6(1 . we obtain Co) = 1.
14 Let C+('yo) be as in (6.16) with 'y replaced by yo and h(t) by h(t. such a deterministic periodic environment may be seen as a special case of a Markovian one (allowing a continuous state space E = [0.0.\ = 0 .. Thus.16. MARKOVIAN ENVIRONMENT attains its minimum 2 /3 for u = oo and its maximum 6 /(7 + 2w) for u = 0.'Yo)e (6. Finally.\ 3 C+ = sup 6 exp { A (. but thereby also slightly longer.66.013. much of the analysis of the preceding section is modelled after the techniques developed in the preceding sections for the case of a finite E. .18) Notes and references The material is from Asmussen & Rolski [44]. The idea is basically to approximate the (deterministic) continuous clock by a discrete (random) Markovian one with n 'months'. .cos 27rs . with s the initial season. Thus C_ = 2 inf ex cos 2irs . Then 7oudT . completing a cycle . the nth Markovian environmental process {Jt} moves cyclically on {1.20). This observation motivates to look for a more formal connection between the periodic model and the one evolving in a finite Markovian environment.013. and let 8 = er' (Y0). exp 2^ cos 21rs .1 sin 27rs + 1 cos 47rs .19 } 0 <8<1 8 + cos 21rs Thus e. and in fact. C+ = 1.9 3 0<8<1 p 27r 47r 167r 161r 2 _ _e.184 CHAPTER VI. n}. Some of the present proofs are more elementary by avoiding the general point process machinery of [44]. 1)..181 s(u) < 1. 6c Markovmodulated approximations A periodic risk model may be seen as a varying environment model.20 •exp { 2n cos 27rs . yo). Of course. where the environment at time t is (s + t) mod 1 E [0.(8)(u.66.g.4^ sin 2irs + 16^ cos 41rs .19 I eu. 0 <'p(8)(u ) . 1) for the environment). 1/i18 1 s (u) > 0.L sin 27rs + 1 I cos 47rs .1 sin 2irs + 16_ cos 47rs .I eu. we have the following result: Theorem 6 . .T) < C+('Yo)h( s. for A = 1 (where 3 e0.
21) which serves as an approximation to 0(1)(u) whenever n is large and i/n s. since the settings are equivalent from a mathematical point of view. Let 0j. (6.20) be the claim surplus process of t>o the nth approximating Markovmodulated model. . Thus. but others are also possible. and the ruin probability corresponding to the initial state i of the environment is then Y'yn)(t) = F (M(n) > t). We want to choose the /3ni and Bni in order to achieve good convergence to the periodic model. We let {Stn)} (6. one simple choice is Oni = 0( i . This queue is commonly denoted as the Markovmodulated M/G/1 queue and has received considerable attention in the last decade.7. AE= Aii'r?/7ri• The arrival intensity is /3i when Jt = i. 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.jEE of the risk process. so that the intensity matrix is A(n) given by n n 0 ••• 0 0 n n ••• 0 A(n) _ (6. Bi. 7 Dual queueing models The essence of the results of the present section is that the ruin probabilities i/ (u).19) n 0 0 ••• n Arrivals occur at rate /3ni and their claim sizes are distributed according to Bni if the governing Markov process is in state i. M(n) = Supt>o Stn). it is desirable to have formulas permitting freely to translate from one setting into the other. 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.1 ((i 1)/n) ) and Bni = B . Notes and references See Rolski [306]. DUAL QUEUEING MODELS 185 within one unit of time on the average . To this end.
0 < t < T and that the risk process {Rt}o<t<T is coupled to the virtual waiting process {Vt}o<t<T as in the basic dualitylemma (Theorem 11. The first conclusion of that result then states that the events {T(u) < T. (7. JT = i} coincide.1).3) 7ri where (V. JT = j) = 7rjPj(VT > u. J* = i) = P.oo in u (7.3. 2 .4) where 0* = >jEE 7rj/3j. Taking probabilities and using the stationarity yields 7riPi(T(u) < T. Jo = i.2). {Jt }o<t<T• Then we may assume that Jt = JTt. Jt ). . For (7.. (VT > u I JT = 2). JT = Z).n(VT > u. JT = i) = 'P. J*) is the steadystate limit of (Vt.0i (u . and the virtual waiting time (workload) process {Vt}too are defined exactly as for the renewal model in Chapter V. ii (u) = it /3 P(W > u.1 Assume V0 = 0.1) follows. JT = j) = LjPj (VT > u. Proposition 7. • The queueing discipline is FIFO.3). Proposition 7. Now let In denote the environment when customer n arrives and I* the steadystate limit. I* )3i P(V > u.2) Oi(u) = 1. J* = i) for all j. The actual waiting time process 1W1. . just sum (7.1) 7ri In particular. JT = j} and {VT > u. T) = 7ri 1 P. In particular.. and (7.186 CHAPTER VI.1) over j. (7. JT = i) = P(V > u.T(V > u I J* = i). Proof Consider stationary versions of {Jt}o<t<T. MARKOVIAN ENVIRONMENT • Customers arriving when Jt = i have service time distribution Bi. (7. J* = i). (7. JJ = i). let T .P(V > u. Jo = j. and for (7.2 The relation between the steadystate distributions of the actual and the virtual waiting time distribution is given by F(W > u.2) and use that limF (VT > u. Then Pi(T(u) < T. I* = i)..=1 .
7) of that paper. and of these. see in particular Harrison & Lemoine [186]. and one has PI'>(rr(u) < T) = P('_T)(VT > u).1 is from Asmussen [16]. P(W >u.7.3). u Notes and references One of the earliest papers drawing attention to the Markovmodulated M/G/1 queue is Burman & Smith [84]. N * oo. p < 1 then ensures that V(*) = limNloo VN+9 exists in distribution.6) (7. The first comprehensive solution of the waiting time problem is Regterschot & de Smit [301].g. I*) with the timeaverage . Taking the ratio yields (7. on average /32TP(V > u. on average 0*T customers arrive in [0. T]. I* = i. (7. J* = i) see W > u. The relation (7.4) can be found in Regterschot & de Smit [301]. Lemoine [242]. see Regterschot & van Doorn [123]. and further references (to which we add Prabhu & Zhu [296]) can be found there.4).3) improving somewhat upon (2. With {Vt} denoting the workload process of the periodic queue.l. DUAL QUEUEING MODELS 187 Proof Identifying the distribution of (W. [243].I.5) follows from (7. In the setting of the periodic model of Section 6. that /3(t).8) For treatments of periodic M/G/1 queue. a general formalism allowing this type of conclusion is 'conditional PASTA'. Proposition 7. if T is large.7) (7. the dual queueing model is a periodic M/G/1 queue with arrival rate 0(t) and service time distribution B(') at time t of the year (assuming w. a paper relying heavily on classical complex plane methods. B(t) have been periodically extended to negative t). P(.T)(T(u) <T) = P(8)(VT > u). and Rolski [306]. with (7.o. >u. we have 1: I(W. and (7.=i) a4.. P(1')(r(u) < oo) = P(')(00) > u).I *=i). n=1 N However. .4) and (7. A more probabilistic treatment was given by Asmussen [17].
This page is intentionally left blank .
T) = FloinfTRt< OIRo=u1 denote the ruin probabilities with/initial reserve u and infinite. z/i(u) = F IinffRt< 0IRo=u 1 (1.2) tk(u.. the aggregate claims in [0.d. i&(u. are i.T) = F(T(u) < T). the premium charged is assumed to depend upon the current reserve Rt so that the premium rate is p(r) when Rt = r. and T(u) = inf {t > 0 : Rt < u} is the time to ruin starting from Ro = u so that '(u) = F(T(u) < oo). resp . 189 . {Rt} moves according to the differential equation R = p(R).6. Thus in between jumps. However .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 . finite horizon. t] are Nt At = Ui (1. U2. Zt As earlier.At + p(R8) ds. and that the claim sizes U1..1) (other terms are accumulated claims or total claims).i. Thus. with common distribution B and independent of {Nt}. . and the evolution of the reserve may be described by the equation Rt = u .
In this situation. Thus at deficit x > 0 (meaning Rt = x). there is positive probability. oo) is given by i (u + p/S). Assume 0(u) < 1 for some u.'(u)) > 0 so that V'(v) < 1. we get p(r) = p + er.2. but when the reserve comes above v.4 Either i. that {Rt} will reach level u before the first claim arrives.1 Assume that the company reduces the premium rate from pi to p2 when the reserve comes above some critical value v. RESERVEDEPENDENT PREMIUMS The following examples provide some main motivation for studying the model: Example 1 . i. Example 1. say e.2 (INTEREST) If the company charges a constant premium rate p u but invests its money at interest rate e.p2. pi > p2 and p(r) = One reason could be competition.3 (ABSOLUTE RUIN) Consider the same situation as in Example 1. when x > p/S. 1 . Hence in terms of survival probabilities. or o(u) < 1 for all u. P(r) _ p + e(r . However. Now return to the general model. we can put Rt = Rt + p/S. say at interest rate b. Proposition 1. No tractable necessary and sufficient condition is known in complete generality of the model. If Ro = v < u.p/S) r > p/S p5(p/5r) 0<r<p/5 Then the ruin problem for {Rt } is of the type defined above. That is. the payout rate of interest is Sx and absolute ruin occurs when this exceeds the premium inflow p. A basic question is thus which premium rules p(r) ensure that 'O(u) < 1.e. Proof Obviously '(u) < ilb(v) when u > v.Vi(v) u > e(1 . where one would try to attract new customers as soon as the business has become reasonably safe. rather than when the reserve itself becomes negative.i(u) = 1 for all u. but assume now that the company borrows the deficit in the bank when the reserve goes negative. Another could be the payout of dividends: here the premium paid by the policy holders is the same for all r. and the probability of absolute ruin with initial reserve u E [p/S. Example 1.190 CHAPTER VII. dividends are paid out at rate pi . it seems reasonable to assume monotonicity (p(r) is u .
1.+ p(r) exists.3.e.1. let uo be chosen such that p(r) > p = 0ILB + e for r > uo. obviously infu<uo z/'(u) > 0. Then if u > no.uo) and. and hence by a geometric trials argument. Theorem 1. { Vt} remains at 0 until the next arrival).5) and the process {Vt} has a proper limit in distribution . . appealing to Proposition 111.6) .3µB for all sufficiently large r. {Vt} decreases at rate p(v) when Vt = v (i.2(d)).2 once more.2(d). we have z/i(u) <p(u .4.. if and only if V)(u) < 1 for all u.b(u. In case (b).o(uo) = 1 so that t/'(u) = 1 for all u by Proposition 1. and P(Rt + oo) > 0. T] i n such a way that the events {r(u) <T} and {VT > u} coincide.1. instead of (1. hence Rt < uo also for a whole sequence of is converging to oo.I3IB requires a more detailed analysis and that µB < oo is not always necessary for O(u) < 1 when p(r) 4 oo. [APQ] pp. Then 0(u) = P(V > u).6 For any T < oo. Proof This follows by a simple comparison with the compound Poisson model. say V. Let Op(u) refer to the compound Poisson model with the same 0.uo) < 1. We next recall the following results. However. (1.1 and increasing in Example 1. 296297): Theorem 1. one can couple the risk process and the storage process on [0. B and (constant) premium rate p. Here {Vt}two is a storage process which has reflection at zero and initial condition Vo = 0. then l/i(u) < 1 for all u. that u zPp(u . which was proved in 11.T) = P(VT > u). (b) If p(r) > /3µB + e for all sufficiently large r and some e > 0.1. (1. then ?(u) = 1 for all u.2) we have t Vt = At . In between jumps. In particular. Proposition I1I. Starting from Ro = uo. This is basically covered by the following result (but note that the case p(r) .4) 0 and we use the convention p(O) = 0 to make zero a reflecting barrier (when hitting 0. the probability that Rt < uo for some t is at least tp(0) = 1 (cf. (1.f p(Vs) ds.5 (a) If p(r) < /. In case (a). INTRODUCTION 191 decreasing in Example 1. That is. V = p(V)).2) for r sufficiently large so that p(oo) = limr. let uo be chosen such that p(r) < p = /3µB for r > uo. cf. Hence ik(u) < 1 for all u by Proposition III.
Sx} dx. Then the ruin probability is tp (u) = f' g(y)dy.h.s.y. x + p(x)dt]). Oeax f x e'Yg (y) dy } = p) eaxa(x) . one having an atom at 0 of size 'yo. t + dt] can be neglected so that a path of {Vt} corresponds to a downcrossing in [t.8 Assume that B is exponential with rate b.h.s. (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.Sx}.Qw(x) . u Define ^x 1 w(x) Jo p(t) dt. Jo AX) (1. of (1.7 p(x)g(x) = tofB (x) + a f (x . say. and is succesful if the jump size is larger than x .192 CHAPTER VII.9) Proof We may rewrite (1. x] to (x.6w(x) . we thus need to look more into the stationary distribution G.8) as the rate of upcrossings. oo).6 applicable. where g(x) = p( ^ exp {. t + dt] if and only if Vt E [x. Considering the cases y = 0 and 0 < y < x separately. Corollary 1. B(x) = e.y)g(y) dy. RESERVEDEPENDENT PREMIUMS In order to make Theorem 1. this means that the rate of upcrossings of level x must be the same as the rate of downcrossings. oo) must be the same as the flow the other way. the flow of mass from [0. In view of the path structure of {V t }.8) Proof In stationarity.7) Proposition 1. Note that it may happen that w (x) = oo for all x > 0. and the other being given by a density g(x) on (0.8) as g(x) = p 1 {yo13e_6x +. we arrive at the desired interpretation of the r. It is intuitively obvious and not too hard to prove that G is a mixture of two components. yo ^ 1 + oo Q exp {. say. for the storage process {Vt}. say when {Vt} is in state y. It follows in particular that 0(u) = fg(Y)dy. the l.6x and that w(x) < oo for all x > 0. Now obviously. of (1. An attempt of an upcrossing occurs as result of an arrival. (1.8) is the rate of downcrossings (the event of an arrival in [t. say if p(r) goes to 0 at rate 1 /r or faster as r j 0.
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
RT(u) has an exponential distribution with rate (S) and hence E [H(RT(u))I r(u) < oo] L Pe6'r (P/C . 13/E).13 /E) r (./3 log(b + a)] = log ePa/f (a + a ) e which shows that Z is distributed as p/E .pa. . assume that {Wt} is Brownian motion with drift µ and variance v2. As an example.3a/ (5 . and the c.V < x)]0 + f P(V > p/E ) + eby fv (p/E .2) follows by elementary algebra.01'E) + (p/E)al aO l febP/E } IF (0 /0 jF From this (2.3. where V is Gamma(b.b P/E dx /' P/ ' (p/  x)p/e 150/f I' (/3/E) (6P1'E.V.e. i.200 CHAPTER VII.3/E) By the memoryless property of the exponential distribution.3 is also valid if {Rt} is obtained by adding interest to a more general process {Wt} with stationary independent increments. Proof 2 We use Theorem 2. then {Rt} is the diffusion with drift function p+Ex and constant variance a2.5 The analysis leading to Theorem 2. r (j3/E) In particular. From ic(a) = . H(u) = P(Z r < u) = P(V > u + p/E) = (8(p + Eu)/E. 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 .2y +µ ) dy . The process {St} corresponds to {Wt} so that c(a) or2a2/2 . RESERVEDEPENDENT PREMIUMS u from which (2.f.2) follows by elementary algebra.g. with density x(3/e1aQ/e fV (x) _ e 6X ' x > 0. /^ u Example 2 .a) . of Z is IogH[a] = f ytc(y)dy = e fa (0.pa.x) dx e.
write y* for the solution of the Lundberg equation f3(B[ry *] . It must be noted. for a martingale proof.v. write Vi* (u) for the ruin probability etc.. THE LOCAL ADJUSTMENT COEFFICIENT _ Q2a2 pa 4e E 201 I. however. or to nonlinear premium rules p(•). Q2/2E). 3 The local adjustment coefficient. Emanuel et at. Corollary 2. se e.i. [129] and Harrison [185].4 is classical. The solution is in terms of Bessel functions for an Erlang(2) B and in terms of confluent hypergeometric functions for a H2 B (a mixture of two exponentials). it is also used as basis for a diffusion approximation by these authors. Goldie & Griibel [167]. it follows that the ruin probability is Cu) H(u) H(0) 11 Notes and references Theorem 2. see e.p*. not even Erlang(3) or H3. Delbaen & Haezendonck [104].1) . Paulsen & Gjessing [286] found some remarkable explicit formulas for 0(u) beyond the exponential case in Corollary 1. Paulsen & Gjessing [286] and Sundt & Teugels [356]. that the analysis does not seem to carry over to general phasetype distributions. and since RT = 0 by the continuity of Brownian motion. as in the proof of Proposition 2. The formula (2. Paulsen [281].3.g. [129]..g. of the form Ei° p"X" with the X„ i. Z is normal (p/E. 134 (the time scale there is discrete but the argument is easily adapted to the continuous case). Further studies of the model with interest can be found in Boogaert & Crijns [71]. Logarithmic asymptotics For the classical risk model with constant premium rule p(x) . and recall Lundberg 's inequality . [357]. Gerber [157] p.3) was derived by Emanuel et at. Some of these references also go into a stochastic interest rate.d.Y*p* W*(u) < ery*u = 0. [282].8. Gerber [155].2 is a special case of a perpetuity.e. A r.3 is from Harrison [185]. [283].
it holds that f3[s] T oo.1 ). and that p(x) * oo. 1) and for a given E > 0.1.log '(u)/u < ry*(1 .w (u) J dt > c(3)eeu v 1 p(u+ t) .e.*(u) . Let y* < So. Proof of Theorem 3. let p* be a in (3.3) When trying to extend these results to the model of this chapter where p(x) depends on x. choose c(. oo for all E > 0. (x.C*ef*".'y ( x)) = 0 where r. we will use the local adjustment coefficient 'y(x). i. i. a first step is the following: Theorem 3 . e(1o+e)2 (x ) u > 00.i)eex.>o 7(x) > 0. The steepness assumption and p(x) + oo ensure 'y(x) * So..1 Assume that for some 0 < 5o < oo. Then lim sup u>oo u and e E''p(r) + 0. If 60 s f 6o. as will hold under the steepness assumption of Theorem 3. c(. 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 .5) which implies inf. 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) . B(x) > C(2)e(ao+f)x for all x. log ?i(u) < < 00 JO . 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. Letting first E * 0 and next ry * T 5o yields the first statement of the theorem.E).1) . choose uo such that p( x) > p* when x > u0E. as solution of the equation n(x. RESERVEDEPENDENT PREMIUMS and the CramerLundberg approximation V. For the last asssertion . (3.4) we assume existence of y(x) for all x.2) such that p(x) < c(. (3. a) = f3(B[a] .202 CHAPTER VII.. 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). then log u (u) In the proof as well as in the remaining part of the section . x>0 (3. and (for simplicity) that inf p(x) > (3µs . When u > uo. which in turn by Lundberg's inequality can be bounded by ery*(1E)" Hence limsup„.e. x * oo.1.ap(x).
let 0e (u) be evaluated for the process only with 3 replaced by /0/e and U. Therefore the probability that a claim arrives ( before the reserve has reached level u + v is at least c(.6) The second main result to be derived states that the bound in Theorem 3. then Rte) = CRtie for all t so that V). Then . the limit is not u + oo but the slow Markov walk limit in large deviations theory (see e.3.4)eE" Given such an arrival. or (b) Condition 3. Theorem 3 . Then limelog l/ie (u) = I(u). 3. However.2 is also an approximation under appropriate conditions. {Rte)} defined as in (1. 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. 2 Assume that p(x) is a nondecreasing function of x and let I(u) = fo ry(x)dx.7) CIO Remarks: 1. u Obviously. Theorem 3. (3. The rest of this section deals with tail estimates involving the local adjustment coefficient.v. THE LOCAL ADJUSTMENT COEFFICIENT 203 where c. the result is not very informative if bo = oo.' (u) < eI(").13 below holds. UJU > x cannot have a much heavier tail than the claim U itself.3 to be reasonably precise and use e` (u) as approximation to 0 (u). I. and in particular. Bucklew [81]).0 are the same. Condition 3. (u) = O(u/e).g. The form of the result is superficially similar to the CramerLundberg approximation. The first main result in this direction is the following version of Lundberg's inequality: Theorem 3 ..(u) > so. and hence '(u) > c(4)eeuc( 2)e(do+e)u The truth of this for all e > 0 implies lim inf log V. If p(x) = pis constant . . 2.13 is a technical condition on the claim size distribution B.1 only presents a first step.e. 3) = (1 . the asymptotics u * oo and c . one can then assume that e = 1 is small enough for Theorem 3. which essentially says that an overshoot r. by cU2. ruin will occur if the claim is at least u + v.3 Assume that either (a) p(r) is a non decreasing function of r.2).ea°/(ecf1)). noting that in many cases the constant C is close to 1. (3. For e > 0.
8 in terms of I(u) when the claims are exponential: Example 3 .2. 3.1 + b f e.bx} dx fo 00 1 + [exp {/(3w(x) . As typical in large deviations theory.(iw(x) . rather than eI(u)).(3/p(x). However.bx} dx .3.bx} dx = 1 + J0 dodx(x) exp {. say.3. First. we show how to rewrite the explicit solution for ti(u) in Corollary 1.bu}.8. RESERVEDEPENDENT PREMIUMS 4. J0 ^oo g(x ) dx f AX) lexp IOW (X ) bx + b u 1 exp low(x) .bx}]o + b /' oo exp low (x) . the logaritmic form of (3. and r j 1 'Yo v(x)dx = bu  a J0 p(x)ldx = Integrating by parts. One would expect the behaviour in 2) to be important for the quantitative performance of the Lundberg inequality (3. 5.bx} dx oo exp low(x) bx dx 70 Ju r oo = b J exp low (x) .(x) dx.bx} dx 1+0.4 Consider again the exponential case B(x) = eax as in Corollary 1. we consider some simple examples. it is formally needed only for Theorem 3.7) is only captures 'the main term in the exponent' but is not precise to describe the asymptotic form of O(u) in terms of ratio limit theorems (the precise asymptotics could be logI(u)e1(U) or I(u)"e_I(u). we get = 1+ J" AX) exp {(3w (x) . u .exp {/33w(u) .6).204 CHAPTER VII. Then y(x) = b . 3a Examples Before giving the proofs of Theorems 3.
7) follows.2.10) where AE = e log 000 e. applying the inequality 7(x + u) > 7(x) yields immediately the conclusion of Theorem 3. The appropriate definition of the local adjustment coefficient 7(x) is then as the one 2p(x)la2(x) for the locally approximating Brownian motion.1.fory(x+u)dxdy ( 3.fo 7(x)dx/Edy f . oo) with drift µ(x) and variance a2 (x) > 0 at x. For Theorem 3.5 Assume that {Rt} is a diffusion on [0. (X) = µ(x). and (3. BE * 0. we get r 00 e. THE LOCAL ADJUSTMENT COEFFICIENT and hence 205 f°° eI(v )dy .2(X) = ev2(x) so that 7e(x) = 7(x)/e. (u) = I(u) + AE .BE. the integral is bounded by 1 eventually and hence lim sup AE < lim sup a log 1 = 0.. Choosing yo. 0. Similarly. and (3. Be = e log U000 e. (3.2. 3.8) 7(x)dxdy 11 We next give a direct derivations of Theorems 3. In particular. in the definition of AE converges to 0.10 or Karlin & Taylor [222] pp. 70 > 0 such that 7(x) < 7o for y < yo. 191195) that 1P (U) = fu0 eI(v)dy = eI(u) follo e.3 in the particularly simple case of diffusions: Example 3.9 ) 11000 eI(v)dy f000 e.5) is infx>o 7(x) > 0 which implies that f °O . u ..9) yields e log . > lime log e = 0 and AE * 0.e.3.fa 7(x+u)dx/Edy o The analogue of (3. ry(x /b I u o e f0 °° e  e.I ( v )dy fo +u) dxdy .fo 7(x) dx /E dy > av 'yo /Edy = E (1 .ev 0 O /E) J0 70 70 Yo This implies lim inf A.3.0.f y(x)dxd y If 7(x) is increasing .1/8 .I ( u) fool. It is well known that (see Theorem XI. note first that the appropriate slow Markov walk assumption amounts to u. (3. IE(u) = I(u)/e.
5.0. I. ..5.Q/p*.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.4.0/e. E+o e*O By (3. . G. the slow Markov limit a * 0 and the limit u walk approximation deteriorates as x becomes large. As in Example 3. .e. this leads to (3./3 1 AX dx. so our approach is to determine standard functions Gl (u). however . (u) representing the first few terms in the asymptotic expansion of I(u) as u + oo. . Thus 7e(x) _7(x)/e and (3. lim sup Af < lim sup c log(1 .. RESERVEDEPENDENT PREMIUMS The analogue of Example 3. . Ignoring 1/5 in the formula there. Further. + Gq(u) + o(G9(u))• Gi (u) It should be noted . Then the solution of the Lundberg equation is y* = b .. the results are suggestive in their form and much more explicit than anything else in the literature.7o C 15 I I.6/p* so that u 1 I (U) = bu . 0.5) and 7* = 5 . Of course. > .6 Assume that B is exponential with rate S.0) = 0.5 for risk processes with exponential claims is as follows: Example 3 . ) Note that this expression shows up also in the explicit formula for lk(u) in the form given in Example 3. I(u ) = G1(u) + . _ .206 CHAPTER VII.+1 (u) = o( 1). G.6) exactly as in Example 3. Nevertheless . that the interchange of the slow Markov walk oo is not justified and in fact.(u) oo.5.1 3. 7(x) is typically not explicit. we have 5 > 7o and get lim inf AE > lime log e .. G.10) holds if we redefine AE as AE = flog (j °° efo 7(x)dx/edy _ E/5 I and similarly for B. 0 Now (3. the slow Markov walk assumption means 5E = b/c.
1) and 17 = k + 1 if B is the convolution of k uniforms on (0.s)C' f "o o as s T S. u Example 3 .1 =$ f cse8 Sn f e"B(x)dx = e8 Jo s eIB ( 1 .12) with y > 1. e. For example.c2 Su a dx ) Su a<1 Su . the typical case is a = 1 which holds . Hence (3.1) leads to . THE LOCAL ADJUSTMENT COEFFICIENT Example 3 .x)n1.3cse7*I7(77) .y/s)dy sn 1 1 f ' evy'7ldy = cse8r(T7) as s T oc.8). more generally.cs(1 .3.1) leads to (S7T N Ocp a.1/k). x T 1. More precisely.C2p* C2 = (3clr( a))11'. and hence (3. . 1. (3. I(u) Pt. 77 = 1 if B is degenerate at 1.:. B[s] = 1 + s exB(x)dx = 1 +c1SF(a) ('+o(')) (S . y = 2 if B is uniform on (0. 2.7 Assume that B(x) .8 Assume next that B has bounded support.11) with a > 0. say 1 is the upper limit and B(x) . c4 = c2b 1/'/(1 . ry* loge*+ g7loglogp*.c3 logu a= 1 J 0 a + bx 1/ ( c4ul 1/° a > 1 where c3 = c2 /b.. phasetype distributions (Example 1. in the phasetype case .Y . This covers mixtures or convolutions of exponentials or. if the phase generator is irreducible ( Proposition VIII.1/a).clxale5x 207 (3.11) that b[s] * co as s f S and hence 7* T S as p* + oo. It follows from (3.g.ry*°p*. . fu I(u) Su .4) or gamma distributions. u(logu + r7loglogu). Here B[s] is defined for all s and B[s] .
(3.. The assumption implies that ru(t) . (b) 'y(x) <'Yo(x)• Proof That 7(x) is nondecreasing follows easily by inspection of (3.1) .(T1)) > Ee7o(u)(ul+vr»(Ti)).e7o ( u)(ul+u r. of U1 + v .15) Proposition 3.11) and (3.1 Cgs o"O 0 esxex2/2c7 dx = cgsec782/2 f . 1 0 3e. 3b Proof of Theorem 3. of the increment in a small time interval [0.3 (B[s] .208 CHAPTER VII.4).12).c8 log . ec78)2/2c7 dx C7 . I (u) c8u log u 0 where c8 = 2/c7.ru(TI)) . 7 * .9 As a case intermediate between (3.r^.f. This leads to an alternative local adjustment coefficient 7o(u) defined as solution of 1 = Ee''o(u)(vi+u .(t))dt.10 Assume that p(x) is a nondecreasing function of x. assume that B(x) CO x2/2c7.log p*. 1 = E.u is a nondecreasing function of u. By convexity of the m .r„(Ti). h].u .2 We first remark that the definition (3.4) of the local adjustment coefficient is not the only possible one: whereas the motivation for (3. RESERVEDEPENDENT PREMIUMS Example 3 .sp(u).Ote7o( u)(u.13) We get b[s] . Hence for u<V.•.Ul up to the first claim (here ru (•) denotes the solution of i = p (r) starting from ru(0) = u).4) is the formula h logEues ( Rhu) .f. (3. h 10. one could also have considered the increment ru (T1) . this is only possible if 7o(v) 2 7o(u)• . (3. x f oo .14) for the m .css 2%rc7eC782/2.B[7o (u)] . g. Then: (a) y(x) and 7o(x) are also nondecreasing functions of x. .g.
4) considered as function of 7 is convex and 0 for y = 0. Assume (3. the case of 7 then follows immediately by Proposition 3. this is only possible if yo(u) > 7(u). we obtain „I.1) .7o (u)p(u)• Since (3.2 in terms of 7o.16) Proof Define 411(n)(u) = P('r(u) < on) as the ruin probability after at most n claims (on = TI + • • • + Tn).u > tp(u).e70(u)(U1P(u)T1) 209 0 + 7o(u)p(u)' 0 <_ 00['Yo( u)] .17) shown for n and let Fu(x) = P(U1 + u . (3. The case n = 0 is clear since here To = 0 so that ik(°)(u) = 0.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. Hence 1 = EeYo(u)(U1+uru(T1)) < E. Separating after whether ruin occurs at the first claim or not.17) from which the theorem follows by letting n + oo.(n+l) (u) 1 . fa 7o(y)dy < u7o(u) < xyo (u) for x > u. u We prove Theorem 3.es'Yo(u)Fu(dx)} o0 e fo yo( x)dx j. it is easily seen that fu x7o(y)dy < xyo (u).Fu(u ) + J ^(n)(u .3.(n+1) (u) efo Yo(x)dxI^"Q exyo( I u u)Fu(dx )+ J . THE LOCAL ADJUSTMENT COEFFICIENT For (b). Hence „/.11 Assume that p(x) is a nondecreasing function of x. note that the assumption implies that ru(t) . Also.ru(T1) < x). We shall show by induction that (' Y'(n) (u) < e fo 'Yo(x)dx (3.10(b): Theorem 3. Then (u) < efo Yo(x)dx.
3c Proof of Theorem 3.n.. for either of Theorems 3. To this end.n) must first downcross unl.n inf n uk1.E (u/ n) Y'E (un .210 CHAPTER VII.3 is required.nbe C*.11 is sharper than the one given by Theorem 3. given downcrossing occurs. 0.n = sup p(x). yo(u) appears more difficult to evaluate than y(u).11 be reasonably tight something like the slow Markov walk conditions in Theorem 3. 0.n <Z auk}l. and. we have chosen to work with y(u) as the fundamental local adjustment coefficient.x/n. Proof For ruin to occur.2).3/e and U. (3.3).3). we used also Proposition 3.n.n u k}1.n = ku. Lemma 3. RESERVEDEPENDENT PREMIUMS where the last identity immediately follows from (3..n) > k =1 II v ^k n. in . 0 It follows from Proposition 3.. ryk.e (u) = v'. resp. P k. (u). Let Ck. The probability of this is at least n n. (u) < I(u). {RtE)} (starting from u = un. let Op*.E (u/n).n (starting from u/n) without that 2u/n is upcrossed before ruin.n) pn niE (u /n) n n_1 n. the probability that ruin occurs in the CramerLundberg model with p* = pn.2. However. (un2. by €U=. For these reasons. op*.n so that n. 3. x + x/n] by two classical risk processes with a constant p and appeal to the classical results (3.n AX).10(b ) that the bound provided by Theorem 3.E (u/n) Now as e . and here it is easily seen that yo(u) ..15). Further.3 The idea of the proof is to bound { R( f) } above and below in a small interval [x .2..E ( u/n) ^•e. C*e where the first equality follows by an easy scaling argument and the approximation by (3. pk n = uk_l.n. in accordance with the notation i/iE (u). y* evaluated for p* = Pk.12 lim sup4^o f log O. Y*u /E.10(a) for some of the inequalities. define uk..E (u) denote the ruin probability for the classical model with 0 replaced by . the value of {R(E)} at the time of downcrossing is < unl.I.n.: y(u). W O . Also.
211 Clearly.. Indeed . B(x) (3. y > 0 it holds that F(U>x +yIU>x) B(x + y) < F (V > y).n.i.7k. (3.F (2u/n). /' (u/n) 'T nk.13 There exists a r. for all x . k=1 k=1 n u _ nE7 k.! (u/n) n n m 7k. *p•. THE LOCAL ADJUSTMENT COEFFICIENT particular. 0 with n and u fixed. . i. ne7k.n ..E (u/n) OP +^p•. V < oo such that (i) for any u < oo there exist Cu < oo and a (u) > supy <„ 7(x) such that P(V > x) < Cue.a( u)z.nu/en(1 + where o(1) refers to the limit e .E (u/n) Op•. in obvious notation one has tC (x) = y(x)/e. (u) CIO < Letting n 4 oo and using a Riemann sum approximation completes the proof..18) (ii) the family of claim overshoot distributions is stochastically dominated by V. also ryk.2 gives 7PE (u) < eIi"i/f = lim inf Clog 0E (u) > I (u).n + 0(1). so that Theorem 3. 3 now follows easily in case (a).n <X<Uk. In case (b).. uk_1. 40 Combining with the upper bound of Lemma 3. we need the following condition: Condition 3. 11 Theorem 3.n = sup ?'(x).nu /fn( 1 Ck  e.log Ck.e.3.n cE (2u/n) Ck ne7k. since ry' is an increasing function of p'.12 completes the proof.nk=1 limsupelogv). v.19) .nu /En) o(1)). It follows that n log V'C (u) k =1 log Ypk.E (urn) < \ *I.
n V.( . u /en 0(i) _n so that E2 < e2ryl nu/En0(1).. (3.n < ery1. N with EN < 1 = infx>2u/nA(x) = 0(1).E(E) (u.E (0) cf.2y 1 ' . v ) = v . Write EO.E (2u/n . where El is the contribution from the event that the process does not reach level 2u/n before ruin and E2 is the rest. Ei + E2 < e71.V) = e71 nu/Eno(l) (using (3.EV) = e. let v < u and define T(E) (u.^(E) (u. P (T(E) (u.18) for the last equality). RESERVEDEPENDENT PREMIUMS To complete the proof.EV) = El + E2.QEU 1 .E (u/n .^'' = E [ . Then the standard Lundberg inequality yields El < E?. u/n) < oo) . (u/n .eV) • P (T(E) (u.5) and the standard formula for b(0).R<) (u v). T(E) (u.. v ) = inf { t > 0 : R(c ) < v R) = u } .v. (u/n . V < u/En] + P(V > u/En) (u/En . infx>2u /n P(x) .EV) = EiI 1 . u/n) < oo] . .nu/En0(1) . u/n) < oo) EV).nu /EnE [e71. ) (u u /n)) . For E2. we first note that the number of downcrossings of 2u/n starting from RoE) = 2u/n is bounded by a geometric r. u /n) < oo] l = = < E [OE (u/n .. T() (u. u/n)) .1 n. (R. u/n) < oo] E [OE (u/n .212 CHAPTER VII. The probability of ruin in between two downcrossings is bounded by Epp . u/n)) I T(E) (u..of:>2 in n(x). Then Y'E (u) ^(E) (u.
it might be possible to show that the limits e . Bucklew [81]).)Ui } .7(x)) (3.J y(Rs)dR.3. THE LOCAL ADJUSTMENT COEFFICIENT Hence lim inf e log Ali.r.7) then comes out (at least heuristically) by analytical manipulations with the action integral. they also discuss simulation based upon 'local exponential change of measure' for which the likelihood ratio is ( /'t /'t Ns Lt = exp S . 0 and b T 00 are interchangeable in the setting of [89].T) = P „(info<t <T Rt < 0) via related large deviations techniques.20) (with ic(x. 0 ) (= p(x) .21) to pass from u to 0. Similarly. s) as in (3. Whereas the result of [122] is given in terms of an action integral which does not look very explicit. one can in fact arrive at the optimal path by showing that the approximation for 0(u.7) for ruin probabilities in the presence of an upper barrier b appears in Cottrell et al. Typically. T) is maximized over T by taking T as the time for (3. Comparing these references with the present work shows that in the slow Markov walk setup. the risk process itself is close to the solution of the differential equation r(x) _ r (x.3EU) (3.=1 J An approximation similar to (3.) = exp .u/n) < oo) CI  > u n n ryi n' i=1 Another Riemann sum approximation completes the proof. where the key mathematical tool is the deep WentzellFreidlin theory of slow Markov walks (see e . s). the approximation (3. u Notes and references With the exception of Theorem 3. [89]. the results are from Asmussen & Nielsen [39]. u/n) < oo { 40 )I U nryl n+liminfelogP (T(')(u.g. .4) and the prime meaning differentiation w. (u) 40 213 lim inf e log(Ei +E2) + logP (r(`) (u. l o JJJ o .J r(Rs)p(R.)ds + Y(R2.t. whereas the most probable path leading to ruin is the solution of r(x) _ k (x. Djehiche [122] gives an approximation for tp(u. 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 3.13.1.21) (the initial condition is r(0) = u in both cases).
3. For different types of applications of large deviations to ruin probabilities . We should like. the exponential distribution ).214 CHAPTER VII. 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. . RESERVEDEPENDENT PREMIUMS the simplest being to require b[s] to be defined for all s > 0 (thus excluding . e. however.g.. see XI.
on Eo = E U {A} where A is some extra state which is absorbing.1) is 'Here as usual . 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. and not in other cases. A proper knowledge of phasetype distributions seems therefore a must for anyone working in an applied probability area like risk theory. More precisely. we write Pv for the case where Jo has distribution v so that Pv = KER viPi• 215 . P. Typically. 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. that is. a terminating Markov process {Jt} with state space E and intensity matrix T is defined as the restriction to E of a Markov process {Jt}o<t<. This implies in particular that the intensity matrix for { it } can be written in blockpartitioned form as T 0 0 . if v = (vi)iEE is a probability distribution. F (Jt = A eventually) = 1 for all i E E 1 and where all states i E E are transient. refers to the case Jo = i. A distribution B on (0. Note that since (1.Chapter VIII Matrixanalytic methods 1 Definition and basic properties of phasetype distributions Phasetype distributions are the computational vehicle of much of modern applied probability. We often write p for the number of elements of E. if a problem can be solved explicitly when the relevant distributions are exponentials.
e. i. In particular. MATRIXANALYTIC METHODS the intensity matrix of a nonterminating Markov process. that is. the exit rates ti and the transition rates (intensities) tij: tj 3 aj ai i ti tk tjk FkJ ak Figure 1.1 The phase diagram of a phasetype distribution with 3 phases.T)) if B is the Padistribution of the absorption time C = inf{t > 0 : it = A}. 0 2this means that tii < 0. j. B(t) = Fa(^ < t ).e. The initial vector a is written as a row vector.2) The interpretation of the column vector t is as the exit rate vector. Then a = a1 = 1. 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. an exponential distribution with rate parameter .216 CHAPTER VIII. and we have t = Te. k}. T is a subintensity matrix2. and the phasetype distribution is the lifetime of a particle with constant failure rate /3. the ith component ti gives the intensity in state i for leaving E and going to the absorbing state A. A convenient graphical representation is the phase diagram in terms of the entrance probabilities ai. Thus the phasetype distributions with p = 1 is exactly the class of exponential distributions. E = {i. Equivalently.0 = t11. i. C is the lifetime sup It > 0 : Jt E E} of {Jt}. T) (or sometimes just (a. tij > 0 for i 54 j and EjEE tij < 0 . a.3. We now say that B is of phasetype with representation (E. Here are some important special cases: Example 1 .1 Suppose that p = 1 and write . (1. t1 = /3.
.. . 00)) S s o .. 6.. . so that the density is P E ai6ie6. 0 SP 0 and the phase diagram is (p = 2) . .2 corresponding to E = {1.3 The hyperexponential distribution HP with p parallel channels is defined as a mixture of p exponential distributions with rates 51.. 0 •...2 The Erlang distribution EP with p phases is defined Gamma distribution with integer parameter p and density bp XP1 6x (p... 0 ••• 0 0 Sp1 0 0 t= 0 0 00 •.... . a = (1 0 0 . p}.1.1)!e Since this corresponds to a convolution of p exponential densities with the same rate S.. 0 0 0 0 S 6 .. the EP distribution may be represented by the phase diagram (p = 3) Figure 1. 0 S 6 Example 1. 0 0 0 T= t= 0 ••• S S 0 0 0 0 0 0 . .x i=1 Thus E _ Si 0 T 0 S2 0 0 . . PHASETYPE DISTRIBUTIONS 217 Example 1.
the Erlang distribution is a special case of a Coxian distribution. Then for i . Theorem 1 . 36) yields s d.3 .f is B (x) = 1 . a. (c) the m. [APQ ] p. B[s] = f0°O esxB (dx) is a(sI T)lt (d) the nth moment f0°O xnB(dx) is (. the backwards equation for {Jt} (e.218 CHAPTER VIII.4 (COXIAN DISTRIBUTIONS) This class of distributions is popular in much of the applied literature. p:.aeTxe. Then: (a) the c. and is defined as the class of phasetype distributions with a phase diagram of the following form: 1 617 ti t2 2 b2. T). . The basic analytical properties of phase type distributions are given by the following result .g.f.1 tP1 1 Figure 1.g. E t ikp kj = kEE kEE 3For a number of additional important properties of matrixexponentials and discussion of computational aspects . dp. Proof Let P8 = (p ^) be the sstep EA x EA transition matrix for {Jt } and P8 the sstep E x Etransition matrix for {Jt} . the restriction of P8 to E.1)"n! aT"e. ds^ = ds' = ttlaj + tikpkj.d. i.e. j E E. Recall that the matrixexponential eK is defined by the standard series expansion Eo K"/n! 3. 5 Let B be phasetype with representation (E. MATRIXANALYTIC METHODS Figure 1.t2 yt bP.4 For example. see A.3 0 Example 1 . (b) the density is b(x ) = B'(x) = aeTxt.
in which case the time to absorption is 0 with m ..tii and have an additional time to absorption either go to state j which has m . define hi = Eie8S.T) 1t. j#i jEE tijhj + his = ti.6) . (1) n+1n!aT . and (b) then follows from 1: aipF.f.5) Indeed . or w.5) as hi(tii + s) = ti  t ij hj. For (c).tii tii .12) for integrating matrixexponentials yields B[s] = J esxaeTxt dx = a ( f°°e(81+T)dx ) t a(sI .g.s I .s) is the m .g.f.g. Since 1 .1.1 ) n +l n ! a (s I + T ) . hj . After that.f. .tii we go to A.p. we i w.n1t = (1)nn!aTn1Te (1)nn! aTne. 1. (Jx E E) = this proves (a).f.tii is the rate of the exponential holding time of state i and hence (tii)/(tii . and since b[s] = ah. d8 P8 = TP8. the rule (A.e.g.p. of the initial sojourn in state i. ti/ . this means in vector notation that (T + sI)h = t.5) ki = 1 + tii L jj:Ai tii (1.B(x) = 1'a (( > x) = P. Alternatively. Part (d) follows by differentiating the m. h = (T + sI)1t.s j# tii i (1. we arrive once more at the stated expression for B[s]. B(n)[0] = _ Alternatively. d" dsn a (. = aPxe.. i. for n = 1 we may put ki = Ei( and get as in (1. Rewriting ( 1.jEE B'(x) _ cx Pxe = aeTxTe = aeTxt (since T and eTx commute). i. tij / . the solution is P8 = eT8. PHASETYPE DISTRIBUTIONS 219 That is. and since obviously P° = I.n lt .T) 't = (. Then h tit ti + ti3 h j .
are idempotent. see the Appendix. 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. making the problem trivial. One obvious instance is the hyperexponential distribution.6 Though typically the evaluation of matrixexponentials is most conveniently carried out on a computer.h. we get the density as 9 9 6 (1 1) 10 7 1 0 10 2 aeTyt = e x . MATRIXANALYTIC METHODS which is solved as above to get k = aTle. Consider for example 3 9 a= (2 2). 0 Example 1.s.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."n! ( ( l 2 2 ) 17 9 0 \ 1 / 10 10 32 n! 35 6" +n!353 Similarly. T= 2 111 so that 2 2 Then (cf.220 CHAPTER VIII. Example A3. another the case p = 2 where explicit diagonalization formulas are always available. there are some examples where it is appealing to write T on diagonal form.
This is the traditional choice in the literature. < 1. 00 B[Q] = J0 f veTxteQx dx = (v ® I) ( f° eT x edx I (t I) (v (& I) ( (T ®Q)xdx f o" e o )( t ® I) _ (v ® I)(T ® Q)1(t ® I). (1.e 11BIJ = 1laDD < 1. T) is then defined to be oo on a set of probability 1.f.7) Proof According to (A.4. .4b for definitions and basic rules): Proposition 1.1. hail = E=EE a. There are two ways to interpret this: • The phasetype distribution B is defective.T).g. • The phasetype distribution B is zeromodified. or one just lets U be undefined on this additional set. 0 Sometimes it is relevant also to consider phasetype distributions.11aDD.29) and Proposition A4. where the initial vector a is substochastic. and in fact one also most often there allows a to have a component ao at A.hall. B[Q] of B is f3[Q] = J e'1zB(dx) _ (v (9 I)(T ® Q)1(t ® I).e a mixture of a phasetype distribution with representation (a/llall. 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.T) with weight hall and an atom at zero with weight 1 .7 If B is phasetype with representation (v. a random variable U having a defective phasetype distribution with representation (a. then the matrix m. i. i. 5 and serves at this stage to introduce Kronecker notation and calculus (see A.
g. Schmidt & Teugels [307] and Wolff [384]. the text is essentially identical to Section 2 of Asmussen [26]. Then the tail B(x) is asymptotically exponential. but todays interest in the topic was largely initiated by M. All material of the present section is standard. assume that T is irreducible .4c).f. (1. . where C. v.Ce7'. we give a criterion for asymptotical exponentiality of a phasetype distribution B. it is easily seen that the asymptotic form of the tail of a general phasetype distribution has the form B(x) _ Cxkenx. 0 Of course.1 of the Appendix. h be the corresponding left and right eigenvectors normalized by vh = 1 and define C = ah • ve . here k = p1). 77 > 0 and k = 0.g. O'Cinneide [276] gave a necessary and sufficient for a distribution B with a rational m. No satisfying .222 CHAPTER VIII. B(x) . see his book [269] (a historical important intermediate step is Jensen [214]). (or Laplace transform) are often used where one would now work instead with phasetype distributions.8 are far from necessary ( a mixture of phasetype distributions with the respective T(') irreducible has obviously an asymptotically exponential tail. Other expositions of the basic theory of phasetype distributions can be found in [APQ].8 Let B be phasetype with representation (a. In older literature. MATRIXANALYTIC METHODS la Asymptotic exponentiality Writing T on the Jordan canonical form.8). The Erlang distribution gives an example where k > 0 (in fact. distributions with a rational m. let . T). h can be chosen with strictly positive component. i is real and positive. Schmidli. 1.F.8) Proof By PerronFrobenius theory (A. Here is a sufficient condition: Proposition 1. and we have eTx . See in particular the notes to Section 6. cf.. cf. the result follows (with C = (ah)(ve)).hve7x. not only in the tail but in the whole distribution. but the relevant T is not irreducible. Rolski.q be the eigenvalue of largest real part of T. In Proposition A5.. one has k = 0. Lipsky [247]. but in many practical cases. 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.f. Neuts. the Erlang case). Notes and references The idea behind using phasetype distributions goes back to Erlang. let v. x * oo. 2. the conditions of Proposition 1. Example A5. Using B(x) = aeTxe .
Then the renewal density exists and is given by u(x) = ae(T+ta)xt. we denote the density by u(x) and refer to u as the renewal density.+UnEA).T). Jt={Jt?ul}.g.. Lebesgue measure. if U is absolutely continuous on (0. but is in part repeated below. n=O We may think of the U.i. +UnEA} 00 = EEI(U1 +. known. be i. . 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 +.1) Proof Let {Jtk)} be the governing phase process for Uk and define {Jt} by piecing the { J(k) } together.. or the density is available ) is. oo) w...1 of the Appendix.2.1. however.. U2. is Markov and has two types of jumps . the problem has an algorithmically tractable solution if B is phasetype: Theorem 2. as the lifetimes of items (say electrical bulbs) which are replaced upon failure. .r. and U(A) is then the expected number of replacements (renewals) in A. For this reason. A related important unsolved problem deals with minimal representations: given a phasetype distribution . U1<t < U1+U2..1 Consider a renewal process with interarrivals which are phasetype with representation (cr.d. + U0 is 0 ..f. the renewals form a Poisson process and we have u(x) = 0. The explicit calculation of the renewal density (or the renewal measure) is often thought of as infeasible for other distributions. (2. JtJt1) Then { 0<t<U1 . we refer to U as the renewal measure.. .t.... Let U1.. but nevertheless.. 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. RENEWAL THEORY 223 algorithm for finding a phase representation of a distribution B (which is known to be phasetype and for which the m. If B is exponential with rate 0.: U1 + . with common distribution B and define4 U(A) = E# {n = 0.
see Fig.T) where v = aT1 /µB. Then the lifetime is zeromodified phase type with representation (a. u The argument goes through without change if the renewal process is terminating.T).IIBII which is > 0 in the defective case.T + ta). MATRIXANALYTIC METHODS of the last type from i to j occurs at rate tiaj . 2. . this is welldefined. which is phase type with representation (v. Corollary 2.1 Corollary 2. i. u Returning to nonterminating renewal processes .e.3 Consider a renewal process with interarrivals which are phasetype with representation (a. T). the lifetime of the renewal process.224 CHAPTER VIII. the phasetype assumptions also yield the distribution of a further quantity of fundamental importance in later parts of this chapter . The renewal density at x is now just the rate of jumps of the second type. Equivalently.U1 U3 U2 U3 U4 Figure 2. as the time of the last renewal. Hence the intensity matrix is T + ta. B is defective . which is ti in state i. define the excess life e(t) at time t as the time until the next renewal following t.1) remains valid for that case. and the distribution of Jx is ae ( T+t«)x. since Uk = oo with probability 1 .2 Consider a terminating renewal process with interarrivals which are defective phasetype with representation (a. Hence ( 2. + Uit_1 where s.. the density is veTxt = B(x)/µB. i. and hence ( 2. is the first k with Uk = 00. (b) £(t) has a limiting distribution as t * oo. .1. IIafl < 1.1) follows by the law of total probability.T) where vt = ae (T+ta)t . and the jumps of the first type are governed by T. fi(t) U2 U1 . Proof Just note that { it } is a governing phase process for the lifetime.e. This is defined as U1 + .. However. that is. and let µB = aTle be the mean of B. Then: (a) the excess life t(t) at time t is phasetype with representation ( vt.
2) v(T + ta) = 0. u Example 2 .2.q2.2): aT1 e = AB = 1 µB µB a + aT'Tea aT1(T + ta) µB PB a + aea a + a µB µB =0. (ii) First check the asserted identity for the density: since T. Next appeal to the standard fact from renewal theory that the limiting distribution of e(x) has density B(x)/µB. cf. we get B(x) aeTxe aT1eTxTe µB µB PB = veTxt. Here are two different arguments that this yields the asserted expression: (i) Just check that aT1/µB satisfies (2. Al.1. T1 and eTx commute. Hence in (b) it is immediate that v exists and is the stationary limiting distribution of it. (2. hence e(t) is phasetype with representation (vt.T) where vt is the distribution of it which is obviously given by the expression in (a). we first compute the stationary distribution of Q.4 Consider a nonterminating renewal process with two phases.e. The time of the next renewal after t is the time of the next jump of the second type. the unique positive solution of ve = 1.6. = qz ql (x1 xz) = ql + qz ql + q ' and the nonzero eigenvalue A = ql .) ( t2 ) .e. 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). According to Example A3. RENEWAL THEORY 225 Proof Consider again the process { Jt } in the proof of Theorem 2. i. The renewal density is then aeQtt = (al a2) ( 7i 7"2.
6 Let B be hyperexponential. t1B 0 Example 2 . and Example 2. Hence 7r = (1/2 1/2).226 CHAPTER VIII.t2) .tl) 7r2t2 + eat (a17r2 .a27rl) (tl .e2bt) 13 Example 2 . )t (51 .52) 25152 51x2+5251 51a2+5251 Notes and references Renewal theory for phasetype distributions is treated in Neuts [268] and Kao [221]. .52a1. A = 25.a27r1) (t1 . Then _ Q Hence 51 0 0 52 + 51 52 _ 5152 51a2 ) (al a2) 52a1 62a1 Slat + 52a1 51a2 51a2+52a1 A = 51a2 . The present treatment is somewhat more probabilistic. Then Q= 0 55 )+(1o)=( j ad ). and Example 2.(biaz + aza. MATRIXANALYTIC METHODS e.5 Let B be Erlang(2).`t (al a2) + C 11 172 ir12 / \ t 2 ) r1 (7r1 7r2) ( t2 7rltl + J + eAt (al a2) ( 71(t2 .4 yields the renewal density as u(t) = 2 (1 .t2) 1 + eat (a17r2 .4 yields the renewal density as u(t) = 5152 e.
) = F(ST(o) E •. T + to+). Considering the first. cf. Proof The result follows immediately by combining the PollaczeckKhinchine formula by general results on phasetype distributions: for (a). i. Next. T). (b) V. Here we have taken the terminating Markov process underlying B with two states.2. The essence is contained in Fig.1 Assume that the claim size distribution B is phasetype with representation (a. T(0) < oo) the ladder height distribution and M = supt>o St.3. we see that the ladder height Sr+ is just the residual lifetime of the Markov process corresponding to the claim causing upcrossing of level 0. 3. represent the maximum M as the lifetime of a terminating renewal process and use Corollary 2.e. B the claim size distribution. and M is zeromodified phasetype with representation (a+. we shall.i. Since the results is so basic. marked by thin and thick lines on the figure. Corollary 3. use the phasetype representation of Bo. however. THE COMPOUND POISSON MODEL 227 3 The compound Poisson model 3a Phasetype claims Consider the compound Poisson (CramerLundberg) model in the notation of Section 1. and if there is a subsequent ladder step starting in j whic occurs w. T) where a+ is given by a+ = . G+(.(u) = a+e(T+tQ+)u Note in particular that p = IIG+II = a+e. a+j. Thus the total rate is tip + tia+. which occurs at rate ti. with 0 denoting the Poisson intensity. add a more selfcontained explanation of why of the phasetype structure is preserved. . We asssume that B is phasetype with representation (a. the Markov processes representing ladder steps can be pieced together to one {my}. T). For (b).p. {St} the claim surplus process. The stars represent the ladder points ST+(k). 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. Then each claim (jump) corresponds to one (finite) sample path of the Markov process. and rewriting in matrix form yields the phase generator of {my} as T + ta+.f3aT1.3. Within ladder steps. r(u) the time of ruin with initial reserve u. Then: (a) G+ is defective phasetype with representation (a+. Now just observe that the initial vector of {mx} is a+ and that the lifelength is M. itself phasetype with the same phase generator T and the initial vector a+ being the distribution of the upcrossing Markov process at time ST+_.1 on the next page.
T = (3 . see Corollary 2. This is in fact a simple consequence of the form of the excess distribution B0. MATRIXANALYTIC METHODS t . Figure 3.QaT1.... 3e3x + .M {mx} ST+(2)  S . 7e7x 2 2 Thus b is hyperexponential (a mixture of exponential distributions) with a (2 2 ).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 .t t d kkt S.1 . 0 Example 3.3.1 This derivation is a complete proof except for the identification of a+ with .2 Assume that .228 CHAPTER VIII.Q = 3 and b(x) = .
but there the vector a+ is not explicit but needs to be calculated (typically by an iteration).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+. T) for some vector a+ = (a+. and the argument for the renewal case starts in just the same way (cf. 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 . his derivation of +'(u) is different. this was obtained in Section 3. 0(8) (u) (recall that z/i(u) refers to the zerodelayed case and iY(8) (u) to the stationary case).and Markovmodulated models. where a+ is the (defective) .4.1 can be found in Neuts [269] (in the setting of M/G/1 queues. The result carries over to B being matrixexponential.6. 4 The renewal model We consider the renewal model in the notation of Chapter V. In the next sections. see Shin [340]. (a) G+ is of phasetype with representation (a+.2 are taken from Gerber [157]. cf. we encounter similar expressions for the ruin probabilities in the renewal. see Stanford & Stroinski [351] . That is. THE RENEWAL MODEL This is the same matrix as is Example 1.6). For the compound Poisson model.^(u) = a+e( T+ta+)ue = 24eu + 1 e6u 35 35 0 Notes and references Corollary 3.j).4. but that such a simple and general solution exists does not appear to have been well known to the risk theoretic community. 3. It is notable that the phasetype assumption does not seem to simplify the computation of finite horizon ruin probabilities substantially. the duality result given in Corollary 11. see Section 6. Fig. 3.1 In the zerodelayed case. We assume p = PB/µA < 1 and that B is phasetype with representation (a.1): Proposition 4.T). We shall derive phasetype representations of the ruin probabilities V) (u). with A denoting the interarrival distribution and B the service time distribution. if we define {mz} just as for the Poisson case (cf. For an attempt. For further more or less explicit computations of ruin probabilities. the discussion around Fig. The parameters of Example 3. T).
T). which for numerical purposes can be solved by iteration. the Palm distribution of the claim size is just B. B0 is phasetype with representation (aT1/µa.1) Proof We condition upon T1 = y and define {m.1.*'} is Markov with the same transition intensities as {mx}. The key difference from the Poisson case is that it is more difficult to evaluate a+. the form in which we derive a+ for the renewal model is as the unique solution of a fixpoint problem a+ = cp(a+). where a(8) = aT1/PA.Sy} in the same way as {mx} is defined from {St}. with intensity matrix Q given by Q = T + to+.T)• Proposition 4. it follows by integrating y out that the distribution a+ u of mo is given by the final expression in (4. Fig. MATRIXANALYTIC METHODS (b) The maximum claim surplus M is the lifetime of {mx}.2 The distribution G(s) of the first ladder height of the claim surplus process {Ste) } for the stationary case is phase type with representation (a(8). where B0 is the stationary excess life distribution corresponding to B.6. In fact.4 Consider the renewal model with interarrival distribution A and the claim size distribution B being of phasetype with representation (a. Proof Obviously. We have now almost collected all pieces of the main result of this section: Theorem 4 . Then . Nevertheless.5.230 distribution of mo. cf. Since the conditional distribution of my given T1 = y is ae4y. obviously mo = m.T).3. Then {m. CHAPTER VIII.3 a+ satisfies a+ = V(a+). 4.*} from {St+y . But by Corollary 2. Hence by Theorem 11. Also. (4.1). where u w(a +) = aA[T + to+) = a J0 e(T+t+)1A(dy). (c) {mx } is a (terminating) Markov process on E. but with initial distribution a rather than a+. G(') = pBo. the calculation of the first ladder height is simple in the stationary case: Proposition 4.
THE RENEWAL MODEL 231 . and that this is given by Proposition 4.2) where a+ satisfies (4. .1) and a(8) _ aT.. only with initial distribution a(*) for mo. In particular .^(8)(u) = a ( 8)e(T+ta +) xe.1(b).1 by noting that the distribution of mo is a+.•. (4.3) Proof The first expression in (4. a+) > 0 = a+o) implies a+) _ (a+) > W (a+)) = a+) .2.e. by a+ = lim a +n) where a+°) .. (4..3). Furthermore . Hence ^p(.1/pA. i. It remains to prove convergence of the iteration scheme (4. the maximum claim surplus for the stationary case has a similar representation as in Proposition 4..^(u) = a+e ( T+ta+)xe.4. thus .0. The term tf3 in cp(i3) represents feedback with rate vector t and feedback probability vector (3. a+l ) = cp (a+°)) . i y ^ T1= y `•r Figure 4. The second follows in a similar way by noting that only the first ladder step has a different distribution in the stationary case.2 ) follows from Proposition 4.0) is an increasing function of /3. I {mx} .3) (defined on the domain of subprobability vectors . . a+2) = ^p (a+l)) .M.1).1 . a+ can be computed by iteration of (4.
However. 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)..5) yields h = (sI .1 arrivals (n arrivals are excluded because of the initial arrival at time T1 ).f. Then s is an eigenvalue of Q = T + ta+ if and only if 1 =.5 Let s be some complex number with k(s) > 0. Theorem 4. Proof Suppose first Qh = sh. Thus by (4.232 CHAPTER VIII. Then each subexcursion of {St+Tl . so to complete the proof it suffices to show that &+ < a+) for all n.T)1t.ST. let F be the distribution of U1 . we use an argument similar to the proof of Proposition VI. this implies that ahA[s] # 0. 7+ ].2.T)It.T1.4) makes sense and provides an analytic continuation of F[•] as long as s ¢ sp(T). To this end. with B[s].4) whenever EeR(S)U < oo. and let &+". the normalization is equivalent to F(s) = 1. the corresponding right eigenvector may be taken as (sI . Thus .4.g. To prove the converse inequality. limn4oo a ) < a+. MATRIXANALYTIC METHODS and (by induction ) that { a+ n) } is an increasing sequence such that limn. Fn ). F[s] being interpreted in the sense of the analytical continuation of the m. Assume the assertion shown for n . 0 0 We next give an alternative algorithm. For n = 0. (4.} can contain at most n . 0 = a+) < a+ yields a+) _ (a+0)) (a+) = a+ (n and by induction that a(n) < a+ for all n . In that case.1. a+ ) exists .P[s] = A[s]B[s]. Similarly. Then (4. (4. . Then e4'h = e82h and hence sh = Qh = (T + taA[Q])h = Th + A[s]tah. Let Fn = {T1 + • • • + Tn+1 > r+}be the event that {my} has at most n arrivals in [T1. which links together the phasetype setting and the classical complex plane approach to the renewal model (see further the notes). and hence we may assume that h has been normalized such that ahA[s] = 1.) = P(mTl = i. both quantities are just 0 .4).5) Since s $ sp(T). Then F[s] = a(sI . s ¢ sp(T).T)'t • A[s] (4. Obviously. n) &+n) T a+.
This gives d roots 'y.T)lt = sh.. 0).6. letting vi be the left eigenvector of Q corresponding to pi and normalised by vihi = 1 .p1i . Pd in the domain ER(s) > 0 . t(ry) > 0.T)lt + t = s(sI . Notes and references Results like those of the present section have a long history..6) i=1 i=1 Proof Appealing to Theorem 4.T) = 1 ata+ = a+. . yd satisfying R(ryi) > 0. in turn. Let d denote the number of phases. Hence with h = (sI T). hd.. we have IG_ [s] I < 1 . This immediately implies that Q has the form CD1 and the last assertion on the diagonal form . D that with columns p1 hl. .. Then G+ is phase. and the topic is classic both in risk theory and queueing theory (recall that we can identify 0(u) with the tail P(W > u) of the GI/PH /1 waiting time W.6 Suppose u < 0.9) we have G+[s] = 1 which according to Theorem 1.' that the equation F(s) = 1 has d distinct roots p1.. the matrix Q in Theorem 2.. and hence by the WienerHopf factorization identity (A. and the solution is . T) with a+ = a(QT)/at.. pdhd. ..5(c) means that a+(sI T)1t = 1. As in Corollary 4. the classical algorithm starts by looking for roots in the complex plane of the equation f3[y]A[ry] = 1. Q = CD1 where C is the matrix with columns hl. .. hd.1 has the d distinct eigenvalues .. .. Further...lt we get Qh = (T + to+)h = T(sI . ... Corollary 4. . 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 ).5.T)It. (4. Since R(s) > 0 and G _ is concentrated on (oo. Q has diagonal form d d Q = dpivi®hi = dpihivi.type with representation (a+. THE RENEWAL MODEL 233 Suppose next F(s) = 1.4. we get at a(Q . Pd with corresponding eigenvectors hl. and define hi = (piI . W v M(d) in the notation of Chapter V).. 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 .. Given T has been computed.
. The matrix. For further explicit computations of ruin probabilities in the phasetype renewal case . see Neuts [269]. see Dickson & Hipp [118]. Numerical examples appear in Asmussen & Rolski [43].) d (see. 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 solutions are based upon iterations schemes like in Theorem 4. and appears already in some early work by Wallace [377]. a pioneering paper in this direction is Tacklind [373]. and the distribution of an arrival claim is B. The number of elements of El=> is denoted by q. [270] and Latouche & Ramaswami [241]. For surveys .contained derivation). where R is an unknown matrix. the intensity matrix is A and the stationary row vector is ir . In queueing theory. 5 Markovmodulated input We consider a risk process {St } in a Markovian environment in the notation of Chapter VI. Here phase. E(t)). In risk theory. That is .F. the ruin probability can be found in matrixexponential form just as for the renewal model. with representation say (a(' ). e. the fixpoint problems look like R=Ao+RAI+R2A2+ .exponential form of the distribution was found by Sengupta [335] and the phasetype form by the author [18].. the background Markov process with p states is {Jt}.234 then in transform terms CHAPTER VIII..type assumptions are basic. but the models solved are basically Markov chains and processes with countably many states ( for example queue length processes ). We assume that each B.g. is phasetype. starting around in 1975. MATRIXANALYTIC METHODS d F 1 + a J e°" ip(u) du = Ee°w = 11(t. [119]. It turns out that subject to the phase. Asmussen & O'Cinneide [ 41] for a short self. The exposition here is based upon [18]. which contains somewhat stronger results concerning the fixpoint problem and the iteration scheme.4. The distribution of W comes out from the approach but in a rather complicated form . whereas the approach was introduced in queueing theory by Smith [350]. an alternative approach (the matrixgeometric method ) has been developed largely by M. T('). Neuts and his students. involving . similar discussion appears in Kemperman [227] and much of the queueing literature like Cohen [88]. The arrival rate in background state i is a.type assumption .
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. However. say with slope r(i) on intervals where It = i.1. states . We start in Section 5a with an algorithm involving roots in a similar manner as Corollary 4. 5a Calculations via fluid models.4. and the one E(•) for B.6. has states o.1. 5. Section 5b then gives a representation along the lines of Theorem 4. The stationary distribution is obtained by finding the maximum of the Vcomponent of the version of {(It.5. for which the relevant fixpoint problem and iteration scheme has already been studied in VI. MARKOVMODULATED INPUT 235 some parameters like the ones T or a+ for the renewal model which need to be determined by similar algorithms. 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).2. Vt)}t>o such that {It} is a Markov process with a finite state space F and {Vt} has piecewiese linear paths. Diagonalization Consider a process {(It. O.1 In Fig. The key unknown is the matrix K. p = ql = Q2 = 2. (a) 0 0 ♦ o ° tl ♦ • 0 0 o } o o (b) 0 } ♦ • 0 o f o Figure 5. This calculation in a special case gives also the ruin probabilities for the Markovmodulated risk process with phasetype claims. The two environmental states are denoted o. the phase space E(°) for B. 5. The connection between the two models is a fluid representation of the Markovmodulated risk process given in Fig. •.Vt)} obtained by time reversing the I component.
1(b) {(It . Eli) + Proposition 5. 2. •. The fluid model on Fig . of E into components indexed by E. whereas Ee8s' = oo for all t and all s > so where so < oo. '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. Let E denote the matrix .31a(l) (/3i)diag . o. a) = 1. Bi[s] = a(i)(T(i) + sI)it(').A 0 Or 1A/ _ t(i) 0 t(2) 0 0 0 0 0 t(3) 0 T1 0 0 0 . Second. 4}. i E E. V. This implies that in the fluid context.1))diag ) a = sa and the eigenvector b = . a) : i E E. F is the disjoint union of E and the Eli). t. resp. 5. 4. Thus F = {o. 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. a E E(i) } .1(a). F = E U { (i.92a(2) 0 0 T(2) 0 0 0 f33a(3) 0 0 T(3) with the four blocks denoted by Ei„ i.1))diag + sII = 0 (5.Vt)} is then obtained by changing the vertical jumps to segments with slope 1. In the general formulation . a) of {It}. we have more martingales at our disposal. j = 1. A claim in state i can then be represented by an E()valued Markov process as on Fig. 5. < oo for all s.1) if and only if s is an eigenvalue of E. r(i) _ 1. The intensity matrix for { It} is (taking p = 3 for simplicity) I A . If s is such a number.236 CHAPTER VIII. 4.(Ni)diag r(i. corresponding to the partitioning + Epp). First. Recall that in the phasetype case. in the fluid model Eel'.1 A complex number s satisfies 'A+ (f3i(Bi[s] . consider the vector a satisfying (A + (13i(Bi[ s] . MATRIXANALYTIC METHODS 4.
Then (up to a constant) c = a. E(1) + + E(P).Nla(1) 0 0 T 1.sI 0 0 0 T(2) . d correspond to the partitioning of b into components Proof Using the wellknown determinant identity Ell E12 E21 E22 E22 I ' I Ell .sI + E12 (sI . resp .sI ()3i)diag .1).sI)1t)) iag I = 0 which is the same as (5. assume that a is chosen as asserted which means (Ell .E22)1 E21a. MARKOVMODULATED INPUT 237 indexed by E.E22)1 E21a E21a . For the assertions on the eigenvectors.A .5. where c. Noting that E11c + E12d = se by definition. t(1) 0 0 then also 0 t(2) 0 . it follows that Ell E12 ( E 21 E22) (d) = s 1 d I . iEE (a> of 0* 1 AI. with Eii replaced by Eii . Then E21c+E22d = E21a .E22 .T('))1t(i) .sI) (sI . and let d = (sI .sI 0 0 t(3) 0 0 = 0.sI+ ((3ia(i)(T(i) . it follows that if Qla(1) 0 0 .A .E12E22 E21 I .E22)1 E21) a = 0.sI. d = (sI E22)1E21a = E ai(sI . 0 .sI 0 0 0 T(3) . c = a.(sI .32a(2) (/3i)diag .E21a + sd = sd.
v) yields C{V) = e8 . Here E has one state only. j.4 Assume that E has two states and that B1. e89uc(e)) (d(1) .upi(u. w(u. j. j) pi( u . a) = (j.Q.. Example 5 . s2 are the negative eigenvalues of Al +01 A1 E _ A 2 b1 0 52 A2 +32 0 . . I' i( V P2 (w (u) < oo. pi(u.a)d^ ). We can take a = c = 1 and get d = (s + b)16 = 5/(3 = 1/p..v) = Optional stopping at time w (u.238 CHAPTER VIII. w(u)=inf{t >O:Vtu}. Example 5 . v.v.sv)b(v) = 0.. a). B2 are both exponential with rates 51 i b2. j.v}. v > 0. Iw(u.. Thus 0(u) = esu/d = pe7 ° as u should be. MATRIXANALYTIC METHODS Theorem 5.a Solving for the pi(u.pi(u.5. Then .3 Consider the Poisson model with exponential claims with rate 5... q. a) and noting that i1 (u) = >I j..O. the result u follows. define w(u.v) = v) I.. . . sq with $2s.v) = = p i( u . . .v) = (j.. < 0 and let b(v) = I d(„)) be the right eigenvector corresponding to s. a)).. c j.4 that {e"1b(v) is a martingale . Letting v ^ oo and using Rsv < 0 yields e8'u = Epi(u..( u./' u = e' (esiuc ( 1) .v) = j). d("))1 e.v)=inf{t >0:Vtu orVt=. v. a)). u) Iw(u. it follows by Proposition II.j)c v .. Proof Writing Or'Alb( v) = svb( v) as (AI . j. To determine 0 (u). .. For u. j. a) = Pi (Vw(u. a )d(a + e8 °vpi (u . Then we get V)i (u) as sum of two exponential terms where the rates s1.j.2 Assume that E = Or 'Al has q = ql + + qp distinct eigenvalues si. v = 1.. we first look for the negative eigenvalue s of E = I 0 I which is s = ry with yy = b . v.
(y. the Pidistribution of M is phasetype with representation (E(1) + + E(P). (') a T( However .x) 00 f ° (') (j) eT (yy)edx .5 G+(i.h.5.6 For i E E.2.s. oo)) j)ye.k.y = to B k7 j # k In particular.3) . In terms of K. dx)Bj(y . 8^')IT(j)) where e 3^') =. MARKOVMODULATED INPUT 239 5b Computations via K Recall the definition of the matrix K from VI. •) is phasetype with representation (E(i).33(e = 0 a(j))(K ®T ( j))(ej (9 I). is 0 /3 f R(i .Qj eie 0 f e (j) T(') x T(j)y ej a e dx e e 00 00 eKx ® e T(')' dx (ej (& I)e T(')ye eKa®T(')x dx (ej (9 I)eT(') Ye e(i)eT(')ye. 0 Theorem 5 . according to VI. j. (5.xxej • a 00 oo el . U) where t(j) + t(j)O(j j = k uja.( 2. 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 5.3j eye. i.b (u) = Pi(M > u) = 9(i)euue. Proof We must show that G+ (i. j.2) the l. j. 9(').
bn1 bn). However.g.f.. Starting from Jo = i. if b* [0] = b1 +b20+b302 +. Piecing together these phase processes yields a terminating Markov process with state space EiEE E(').p. An alternative characterization is that such a distribution is matrixexponential. This yields the asserted form of uja.. Furthermore. . a) to (k. say. a) is obviously chosen according to e(`). u Notes and references Section 5a is based upon Asmussen [21] and Section 5b upon Asmussen [17]. and lifelength M. 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. MATRIXANALYTIC METHODS Proof We decompose M in the familiar way as sum of ladder steps . 0 1)'. which occurs at rate t^^7. oo) and b* [0] = f °O eBxb(x) dx the Laplace transform.2. and it just remains to check that U has the asserted form.1 Let b(x) be an integrable function on [0. which occurs w. i. and a new ladder step of type k must start in phase y. the current ladder step of type j must terminate. +bn0i1 0n +a10n1 +. (6. intensity matrix U. Associated with each ladder step is a phase process. t) is the representation of the matrixexponential distribution/density): Proposition 6. some square matrix T and some column vector t (the triple (a.) which is rational. +aii10+anI then a matrixexponential representation is given by b(x) = aeTxt where a = (b1 b2 ..2) ..k y. with phase space EU> whenever the corresponding arrival occurs in environmental state j (the ladder step is of type j). t = (0 0 . Bk7 .. we have sofar concentrated on a claim size distribution B of phasetype.e.y) to occur when j # k.. a m.5). equivalently. Numerical illustrations are given in Asmussen & Rolski [43]. For j = k. we have the additional possibility of a phase change from a to ry within the ladder step.. the ratio between two polynomials (for the form of the density. that the density b(x) can be written as aeTxt for some row vector a. see Example 1. which occurs at rate t(i). 6 Matrixexponential distributions When deriving explicit or algorithmically tractable expressions for the ruin probability.e. the initial value of (i. Then b*[0] is rational if and only b(x) is matrixexponential.240 CHAPTER VIII. i.. For a transition from (j. T.
S.47x2 3 1 0 . b(x) > 0 for x > 0.1 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. (6.1 0 . but as follows from Proposition 6. .s.6. t).. For a proof. The converse follows from the last statement of the theorem. personal communication). s) is given by 27r i .4) 0 0 1 c This representation is complex.1. t= 0 . of (6.3) 0 0 0 0 0 . then b*[0] = a(0I T)1t which is rational since each element of (01 .3) was suggested by Colm O'Cinneide. One of his elementary criteria.. 0 1 an an1 an _2 .. 0 0 . (6. Writing b(x) = c(e( 2ni1 ) y/2 . Example 6 . s = c/ 2 .. we can always obtain a real one (a. (6.47r2 3 . T= 0 0 1 . Namely./(0 + bi).e(tai1)x/2 + e'T) it follows that a matrixexponential representation ()3.1) as E 1 c.3) that we can take 0 1 0 0 a= (1 + 47r2 0 0).T)1 is so. u Remark 6.. where c = 1 + 1/47r 2.(6. .an_3 an _ 4 .. Thus.1 0 0 )3 = (111).2). see Asmussen & Bladt [29] (the representation (6. S = f c/2 0 21ri .2). T. a2 a1 Proof If b(x) = aeTxt. since 1 + 4ir2 03 + 302 + (3 + 47x2)0 + 1 + 47r2 it follows by (6.. cannot be phasetype. . u giving b(x) = E 1 ciebiz/bY. . 0 0 0 0 1 0 0 . namely to asssume the roots 6l.h..3 A set of necessary and sufficient conditions for a distribution to be phasetype are given in O'Cinneide [276]. MATRIXEXPONENTIAL DISTRIBUTIONS 241 T = 0 1 0 0 0 ..2 A remarkable feature of Proposition 6. matrixexponentiality implies a rational transform. . 1 . bn of the denominator to be distinct and expand the r. shows that the distribution B with density b(x) = c(1 cos(21r x))ex.
Consider the distribution with density = 15 ((2e2x . recall that t = Te) that if B is phasetype and (a. T the phase generator and t = Te. Then (cf.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)) .1)2 + 6). T.5) Thus. that despite that the proof of (6. then: Proposition 6. leading to matrix calculus in high dimensions when b is small. we shall only consider the compound Poisson model with arrival rate 0 and a matrixexponential claim size distribution B. (6.242 CHAPTER VIII. and present two algorithms for calculating '(u) in that setting. MATRIXANALYTIC METHODS Example 6 . T.6) holds true also in the matrixexponential case.3. q are polynomials without common roots. 7 + 155ex b(x) Then it is known from O'Cinneide [276] that b is phasetype when 6 > 0.1 to get i (u) = f3esus. . For the second algorithm. We recall (see Section 3.6) The remarkable fact is. we use a representation (a. 0. 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. t) a phasetype representation with a the initial vector.1 shows that a matrixexponential representation can always be u obtained in dimension only 3 independently of J. we have represented ti* [0] as ratio between polynomials (note that 0 must necessarily be a root of the numerator and cancels). and can use this to invert by the method of Proposition 6.4 This example shows why it is sometimes useful to work with matrixexponential distributions instead of phasetype distributions: for dimension reasons . Corollary 111.5 (6. (6. we take as starting point a representation of b* [0] as p( O)/q(9) where p. then 5(u) = a+e(T+t+)uTle where a+ = /3aT1. As for the role of matrixexponential distributions in ruin probability calculations. t) of b(x). For the first. and that the minimal number of phases in a phasetype representation increases to 0o as 5 .6) in Section 3 seems to use the probabilistic interpretation of phasetype distribution in an essential way.
6).1t = f3a (0I T)1T1t .7) 9( cf.T)1 + (6I . (6. with A = 91T.1 + b+ = b++ 1 .b* (6.6.'t. 519) (A + UBV ).T). MATRIXEXPONENTIAL DISTRIBUTIONS 243 Proof Write b* = a(9I .to+)1 = (BI .T)1t ( l . Then in Laplace transform formulation . From the general matrix identity ([331] p.T)1 + 1 ib* (91.6b* . this can be verified by analytic continuation from the phasetype domain to the matrixexponential domain . b+ = a +(BI .T . (91.T)1 (91. since (91T)1T . U =.to+)1T . we get b+ = 0aT1(9I T).1t du = .T)1t)1a +(9I .1UB(B + BVA1UB).B=land V=a+. but we shall give an algebraic proof. Presumably.T)1T 2 = and 1 = AB IT2 + 82T .1 = ^(T1 + ( 91T)1). the assertion is equivalent to a+(BI .T)1t.5 ). xb(x) dx = aT2t.T .T)1T1t. Now.1BVA1. we get (91.1 = A1 .T .T)1 so that b* b** b** a+(9I .a+(9I . (6. .to+)1T .1t = b* .A .T)1ta+(OI .t. b+ = a+(9I .1 + 82 (9I .T)1 J0 00 b(x) dx = f aT1t.
which is selfexplanatory given Fig.s. VII.T)1T. In Corollary VII. a key early paper is Cox [90] (from where the distribution in Example 6. see Asmussen & Bladt [29]. For expositions on the general theory of matrixexponential distributions. to piece together the phases at downcrossing times of {Rt} (upcrossing times of {St}) to a Markov process {mx} with state space E.244 CHAPTER VIII. 0 Notes and references As noted in the references to section 4.1. . 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. 3.T)1)t = 8 (1 .7).h. T). 7a Computing O(u) via differential equations The representation we use is essentially the same as the ones used in Sections 3 and 4. A key tool is identifying poles and zeroes of transforms via WienerHopf factorization.8 a(T1 + (01. a. some key early references using distributions with a rational transform for applied probability calculations are Tacklind [373] (ruin probabilities) and Smith [350] (queueing theory). MATRIXANALYTIC METHODS .8. 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. of (6. The proof of Proposition 6. 7 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate 0.1t = /3a (9I .5 is similar to arguments used in [29] for formulas in renewal theory. From this it is straightforward to check that b**/(b+ .1. the ruin probability(u) was found in explicit form for the case of B being exponential. but the argument of [286] does not apply in any reasonable generality). /3aT1(0I .82b*. see the Notes to VII.1) is the same as the r.1.3 is taken). Much of the flavor of this classical approach and many examples are in Cohen [88].b*). cf.3a (1 0 T 2 + 1 T 102 (9I + 02 1 T)1) t P + 7. See Fig.1. (for some remarkable explicit formulas due to Paulsen & Gjessing [286].la. Lipsky [247] and Asmussen & O'Cinneide [41].T)1T2t . 7.
t2) be the matrix with ijth element P (mt2 =j I mtl = i). though still Markov.7. P(tl. Also. we obtain V)(u) = P(m„ E E) = v(u)P(0. Given the v(t) have been computed. Ai(t) = P(mt = i). t + s) . Note that in general >iEE Vi (U) < 1. >iEE Vi (U) is the ruin probability for a risk process with initial reserve 0 and premium function p(u + •). 0 < t < u.e.t)). the A(t) and hence Vi(u) is available by solving differential equations: Proposition 7.1 The difference from the case p(r) = p is that {m2}. Figure 7. In fact. Let P(tl. t2) = exp where Q(t) = ds [P(t. t) is the vector of state probabilities for mt. in contrast to Section 3. Proof The first statement is clear by definition. O<.1 A(0) = v(u) and A'(t) = A(t)(T + tv(u . Since v(u) = (vi(u))iEE is the (defective) initial probability vector for {m8}.u)e = A(u)e (7. RESERVEDEPENDENT PREMIUMS 245 Rt l0 u . the definition of {m8} depends on the initial reserve u = Ro. Define further vi(u) as the probability that the risk process starting from RD = u downcrosses level u for the first time in phase i.tl < t2 < u. i. By general results on timeinhomogeneous Markov processes.1z I.I] I 80 { tq f Q(v) dvl t1 1 . is no longer timehomogeneous.1) where A(t) = v(u)P(0.
t). we get vi(u) = aidt + (1 . whereas in the second case the probability is p(u)dt • tjvi(u). A'(t) = A(t)Q(t) = A(t)(T + tv(u . In the first case. The intensity of a jump from i to j is tij for jumps of the first type and tivj(u . the probability that level u + p(u)dt is downcrossed for the first time in phase j is vj (u + p(u)dt). Proposition 7. from a computational point of view the remaining problem is to evaluate the v(t). Thus.246 CHAPTER VIII.t and being followed by a downcrossing.t) for the second. Hence Q(t) _ T + tv(u .2 For i E E. two things can happen: either the current jump continues from u + p(u)dt to u. {mx} has jumps of two types. Given this occurs. jEE .(tai + vi(u) E vj(u)tjp (u)  Q + vj (u)tjip ( u). 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. dt].t)). the probability of which is 1 . the probability that level u is downcrossed for the first time in phase i is ai. 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.3dt. 0 Thus.(u) p ( u) = . MATRIXANALYTIC METHODS However. Given A'. those corresponding to state changes in the underlying phase process and those corresponding to the present jump of {Rt} being terminated at level u .4) jEE jEE Proof Consider the event A that there are no arrivals in the interval [0. or it stops between level u + p(u)dt and u. the probability of downcrossing level u in phase i is 8ji(1 + p(u)dt • tii) + (1 . Given A.Qdt) vi(u) + vi'(u)p(u)dt + p(u) dt E{tji+tjvi(u)}.Sj i)p(u)dt • tji = Sji + p(u)tji dt. given A. (7. 0 < t < u. vi.
say. say. supRt>v l t<7 I where o. (v) is given by the r. of (7.3 For any fixed u > 0. starting from v"(v) = .) P"(AnBv) = P(AnB.2.0 as v + 00 we have P(A) P"(A) = P(AnBv)+P(AnBv) P"(AnB. we face the difficulty that no boundary conditions is immediately available. To deal with this. Then P(B. vi (U) = lim v= (u). Rt .7. we can first for a given v solve (7. after a certain level v. Now since both P(A n Bv) 3 0 and P"(A n Bv) . RESERVEDEPENDENT PREMIUMS 247 Subtracting v.. P u which implies that v... and similarly P"(Bv) . we have p(r) = p = vi (u) 0aTe. F" etc. refer to the modified process.) is the tail of a (defective) random variable so that P(Bv) + 0 as v 4 oo. (u) on both side and dividing by dt yields the asserted differential u equation. . then P(A n Bv) _ P"(A n BV'). When solving the differential equation in Proposition 7.00 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.5).h. denotes the time of downcrossing level u . From Section 3.)P"(AnB. <oo.^ 0. V .i7rT1/p. Since the processes Rt and Rt coincide under level B.s.4) backwards for {va (t)}v>t>o. Then pv(r) p(r) r < v p r>v ' and (no matter how p is chosen) we have: Lemma 7. Let p" (t).) + 0 as v + oo. Thus. consider a modification of the original process {Rt} by linearizing the process with some rate p.. This yields v. (u) for any values of u and v such that u < v.
7) equals 01 (v . assuming u > v for the moment.e.1. Let ii'( u) = a+'ie(T+ta +^)"e denote the ruin probability for R't where a+ = a+i) = laT1/pi. 7b Twostep premium rules We now assume the premium function to be constant in two levels as in VII. We recall from Propositon VII.. cf.RQ (defined for or < oo only) is defective phasetype with representation (v(u). 3u etc. Recall that q(w) is the probability of upcrossing level v before ruin given the process starts at w < v. The precision depends on the particular quadrature rule being employed.1. MATRIXANALYTIC METHODS Next consider a sequence of solutions obtained from a sequence of initial values {v. To evaluate p1(u). Corollary 3.z51(v)). where.10 that in addition to the O'(•). > 0} and the last term the contribu tion from {R.6) We may think of process Rt as pieced together of two standard risk processes RI and Rte with constant premiums p1. typically the complexity in n is 0(n2) for integral equations but 0(n) for integral equations. numerically implemented in Schock Petersen [288]) and the present one based upon differential equations require both discretization along a discrete grid 0. as well p1(u). for u > v the distribution of v .v v(u)eTat 1 1 . (7.9. i. T).q(v dx +( ) ) = ( ) ( q( )) vueTva (7.1. r<v r > v. p2.1a. v = u.. The algorithm based upon numerical solution of a Volterra integral equation (Remark VII.7) f o (the integral is the contribution from {R. (u)}.. However. which is available since the z/i'(. say. such that Rt coincide with RI under level v and with Rt above level v. the probability of ruin between a and the next upcrossing of v. 1/n.. 2/n. while the fourthorder RungeKutta method implemented in [30] gives 0(n5). Thus we obtain a convergent sequence of solutions that converges to {vi(t)}u>t>o• Notes and references The exposition is based upon Asmussen & Bladt [30] which also contains numerical illustrations. p(r) P. the evaluation of Vi(u) requires q(u) = 1 .x) dx f v(u)eT xt dx .zp1(u)/(1 .248 CHAPTER VIII. let v(u) = a+2ieiT +ta+>)(uv).. 0 < u < v. Then v(u) is the initial distribution of the undershoot when downcrossing level v given that the process starts at u. Therefore u pl(vvueTa t 1. 2u. < 0}).) are so. The trapezoidal rule used in [288] gives a precision of 0(n 3). The f iin in (7.V" M 0 . where v = inf It > 0 : Rt < v}.
x) dx} V 1 1(v) f V v(u) eTxt. B is hyperexponential corresponding to 3 0 3 a(2 2)' T= ( 0 7 t. RESERVEDEPENDENT PREMIUMS 1 . (7.24ev .01 (v . Since µB = 5/21. 01(u) _ 24 u + 35 e6u 1 35 e 4(u) _ 35 .e6u 35 .and A2 = 3 .1 from which we see that pl (u) = 1 + 1 249  1 .to+))11 {e{T®(Ttoy+ ))}„ .(7 The arrival rate is (i = 3. I.7.4) can be written as (Y(u) ®a+)e(T+t°+>)°1 (T ® (T . From Example 3.be the eigenvalues of T + to( 2 ).24e.21 = ? yields 0(u) = 1.jl (t ®e) Thus..x) dx 1 ^(v) ( 1 .v) + (2^ + 3v2 ea'(u " .2.^1(v) 1 .v(u)eTve).2.e.v(u)eTVe . all quantities involved in the computation of b(u) have been found in matrix form. Then one gets X20 20 21 f 1ea1(u v) + 1 3 3 ^ A 2(u e .v) 1eai(u v) + 7 7 1 e\2(u v) 1 3 ^') eA2 (u. so we consider the nontrivial case example p2 = 4 and p1 = 1. Example 7. p2 < 3.2V"2.u .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.4 Let {Rt } be as in Example 3.e6v Let Al = 3 + 2V'2.8) The integral in (7.
pi (u) = p12(u)/p1 l(u) where p1i(u) p12(u) 35e6v .24es" . 192esv + 8 P1 .1.21(35e6v .250 CHAPTER VIII. MATRIXANALYTIC METHODS From (7.24e5v ..24es" . Notes and references [30]./2) ea 1(u .7) we see that we can write pi (u) = v(u)V2 where V2 depends only on v.1 V2 = 4e5"+6 35e6v . and one gets 12e5" .1)' ?.v)esv + 7 4_ 2.1 Thus. ) e sv + ( 2v/2.24e5v . The analysis and the example are from Asmussen & Bladt ./2 ea1(u") .2 35e6v .+ it (3 4'I 1 ea2(uv e1\2(u") 7 + ( 32 +4. 21 3 In particular.b(v) = 192esv +8 35e6v + 168esv + 7* Thus all terms involved in the formulae for the ruin probability have been exu plicitly derived.
oo ) and say then that B is subexponential (B E S) if 251 . B(x) = ex0 with 0<0<1. B(x) = L(x)/x" where a > 0 and L(x) is slowly varying. 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. III. For example. For the definition . A rough distinction between light and heavy tails is that the m. see I. 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.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.B(x).2b. for all t > 0. x 2iror2 (c) the Weibull distribution with decreasing failure rate . For further examples. and instead we shall work within the class S of subexponential distributions . a2)) with density 1 e(logyFh) 2/2az .46 and at numerous later occasions require a light tail.N(µ. x 4 oo. Some main cases where this lighttail criterion are violated are (a) distributions with a regularly varying tail. we require that B is concentrated on (0.f.g.4. (b) the lognormal distribution (the distribution of eu where U . the exponential change of measure techniques discussed in II. L(tx)/L(x) 4 1.
oo). To capture the intuition behind this definition.1(b) is oo.v.1) then means P(X1 +X2 > x) 2P(Xi > x).2 If B E S. B(x) Here B*2 is the convolution square. X2 with distribution B. the r.v. X2) > x} C {X1 + X2 > x}. We later show: Proposition 1. 1). if X1 + X2 is large . HEAVY TAILS B*2\ 2. we have {max(Xi. 1/2 'typical' (with distribution B) and w. Then: (a) P(max(Xi. (b) liminf BB(() ) > 2. Since B is concentrated on (0.B(x)2 . proving (a). 1/2 it has the distribution of X1I X1 > x. In terms of r. then P(X1>xI X1+X2>x)* 2.'s X1. X2 > x) = 2B(x) . one can check that x x where U is uniform on (0. X2) > x}. Thus the liminf in Proposition 1. (1. P(max(Xi. in the subexponential case the only way X1 + X2 can get large is by one of the Xi becoming large.1 Let B be any distribution on (0. Proof By the inclusionexclusion formula. B(x) ax. x 3 00. That is. oo). Thus . . 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. That is. X2) > x) ^' 2B(x). As contrast to Proposition 1.'s. given X1 + X2 > x. P(Xi <yI Xi+X2>x) 1B(y). X2) > x) is P(X1 > x) + P(X2 > x) . that is. X1 is w.2B(x).v. Then X1 +X2 has an Erlang(2) distribution with density yeY so that B*2(x) xex. then (with high probau bility) so are both of X1. In contrast.p. note first the following fact: Proposition 1. X2) > u x)/B(x) = 2. X2 but none of them exceeds x. the distribution of independent r. the behaviour in the lighttailed case is illustrated in the following example: Example 1.252 CHAPTER IX.2.3 Consider the standard exponential distribution.p.F(X1 > x. and thus the lim inf in (b) is at least lim inf P(max(Xi.
The uniformity now follows from what has been shown for y = yo and the obvious inequality y E [0. we get limsupB*2(x)/B(x) < 2. 1 < B(x ) B( x) Y) < B( 0). then the overshoot X . SUBEXPONENTIAL DISTRIBUTIONS Here is the simplest example of subexponentiality: Proposition 1. a contradiction. or they both exceed Sx. Finally lim inf B(x .1(b) we get B*2(x)/B(x) * 2. If lim sup B(x .5)x)' + 0 _ 2 L(x)l xa (16) Letting S 10.4 Any B with a regularly varying tail is subexponential. B( 0 . Hence lim sup a+oo B*2(x) 2B((1 .'s: if X . and combining with Proposition u 1.5 If B E S. we get BZ(x)) > 1 + B(y) + B(B()y) (B(x) .B*n(x . Let 0 < 5 < 1/2. then either one of the Xi exceeds (1 .y)/B(x) > 1 since y > 0.2) B(x) B(x ) B(x) Jo with n = 1 and splitting the integral into two corresponding to the intervals [0. x].B(y) = 2. This follows since the probability of the overshoot to exceed y is B (x + y)/B(x ) which has limit 1. Using the identity B*(n+1)(x) = 1+ + 1)(x) 1+ 2 1 . [In terms of r.] Proof Consider first a fixed y.z B(x) .B E S. we therefore get lim sup B*2(x)/B(x) > 1+B(y)+ 1 . 253 Proof Assume B(x) = L(x)/xa with L slowly varying and a > 0. If X1 + X2 > x.v.6)x)/((1 . yo] as X + 00. We now turn to the mathematical theory of subexponential distributions. then B(B(x)y) * 1 uniformly in y E [0.1. y] and (y.S)x.xIX > x converges in distribution tooo. Proposition 1.S)x + B(Sx)2 < lim sup B(x) xaoo B(x) lim sup 2L((1 x^oo .y)/B(x) > 1.B*(n ) B(dz) (1.yo].B(y)) .
Given e > 0. B(x) \Jo _ B(x .z) B(dz) _y B(x) 111 Lx B ..254 CHAPTER IX.B*2 (x) B(x) (x . 0 Proof of Proposition 1. The case n = 2 is just the definition. Proof For 0 < 5 < e. Proof We use induction. HEAVY TAILS Corollary 1.nI < e for x > y.z ) (x ) = 1 + (^ B(x .5 that B(n) > e6B(n .7 If B E S. so assume the proposition has been shown for n.z) B(x) Here the second integral can be bounded by B*n(y) B(x) .5 and the induction hypothesis. x oo. This implies B(x) > c2e5x for all x. The first integral is y B(x .5 and dominated convergence.2. then for any n B*n(x)/B(x) * n. O The following result is extremely important and is often taken as definition of the class S. then e"R(x) * oo. and this immediately yields the desired conclusions.2).z) B(dz). we have by Proposition 1.z) B(dz) (n + O(e)) ^x JO B(x) (n + 0(0) I B (x) . choose y such that IB*n(x)/B(x) . P(X1 > xIX1 + X2 > x) _ P(Xi > x) _ B(x) 1 P(X1 + X2 > x) B2(x) 2 1 y P(X1<y X1 + X2 > x) B(x .6 If B E 8.B(x . Proposition 1.y) sup v>o B(v) B(x) which converges to 0 by Proposition 1. b[c] = oo for all e > 0.z) B(dz) 2B(x) o rv 2 0 2 using Proposition 1. its intuitive content is the same as discussed in the case n = 2 above. B*(n+1) (x I xy + Jxx y) W. Then by (1.1) for all large n so that B(n) > cle6n for all n.
4) .z) B(dz) x . Then by (1.y))/B(x). 0 Lemma 1.X2 > xv) < A1(xv)A2(x v) . Proof Define 5 > 0 by (1+5)2 = 1+e.X1 > xv. Proof Let X1. 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) .ajB(x)Ai(v) = ajB( x)(1+o„(1)) (j = 3 .'s such that Xi has distribution Ai.y)Ai(dy) = (x)o(1) (1.ala2B(x)2 which can be neglected. oo) such that Ai (x) _ aiB(x) for some B E S and some constants al.9 Let A1. 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.1. a2 with a1 + a2 > 0. Then Al * A2(x) = P(X1 + X2 > x). choose T such that (B(x)B*2(x))/B(x) < 1 + b for x > T and let A = 1/B(T). A2 be distributions on (0. 0 Proposition 1.z) B(x . For any fixed v.3) Using the necessity part in the case Al = A2 = B yields f xv B(x .y)Ai(dy) v) f o . Since P(X1+X2 > x.i).5 easily yields P(X1 + X2 > x.y)B(dy) = B(x)ov (1)• v (1. Combining these estimates and letting a 4. it follows that it is necessary and sufficient for the assertion to be true that JX_VA (x . Then Al * A2 (x) .0 completes the proof. then there exists a constant K = KE such that B*n(x) < K(1 + e)nB(x) for all n and x. an = supx>o B*n(x)/B(x).8 If B E S. X2 be independent r.z) B(dz) < 1 + A + an(1 + d) .B(x .v. Xi <= v Ai (x .2).(al + a2)B(x). Proposition 1. e > 0. an+1 fX B*n( *n(x .z) B(dz ) + sup < 1 + sup f x<T B ( x) x>T 0 B(x .z) B(x) < 1 + A + an sup f x B(x .
In the regularly varying case.9).3) follows if CHAPTER LX. HEAVY TAILS 'VV B(x .5) Here approximately the last term is B(x)o„(1) by ( 1. B1 * B2 (x) . whereas the two first yield B(x)(Ai(v) . with a > 0 and L1. V (1. u Corollary 1. it should hold that B1 * B2 E S and B1 * B2 (x) . A(x) = o(B(x)). L2 are slowly varying.aiB(v)) = B(x)o„(1). u Corollary 1.v)Ai(v) .11 Let B E S and let A be any distribution with a ligther tail.12 Assume that Bi(x) = Li(x)lxa.2aB(x) .Bl (x) + B2 (x) when B1. if q(x) aB(x) for some B E S and some constant a > 0.y)Ai(dy) = B(x)o„(1).Ai(x . Then A * B E S and A * B(x) . u It is tempting to conjecture that S is closed under convolution.2A(x). However.y)B(dy). then A E S. Then B E S provided fo "O exA(x) b(x) dx < oo.10 The class S is closed under tailequivalence. Proof Taking Al = A2 = A. Hence Corollary 1. B1 * B2 E S does not hold in full generality (but once B1 * B2 E S has been shown. L2 slowly varying.B(x) Proof Take Al = A. Then L = L1 + L2 is slowly varying and B1 * B2(x) sim L(x)/x«. B2 E S. a2 = 1.h. the l.v)B(v) + _'U Aq(x .4). .Bl (x) + B2 (x) follows precisely as in the proof of Proposition 1. f " By a change of variables. of (1. That is. That is.13 Let B have density b and failure rate A(x) such that .256 Now (1. We next give a classical sufficient (and close to necessary) condition for subexponentiality due to Pitman [290]. A2 = B so that a1 = 0. i = 1.s.(x) is decreasing for x > x0 with limit 0 at oo. then so is L = L1 + L2. Recall that the failure rate A(x) of a distribution B with density b is A(x) = b(x)/B(x) Proposition 1. a1 = a2 = a yields A*2(x) . it is easy to see that if L1.2.5) becomes x B(x .
To illustrate how Proposition 1. elementary but tedious calculations (which we omit) show that A(x) is ultimately decreasing. Since ) (x . an integrable function by assumption. The middle bound shows that it converges to b(y) for any fixed y since \ (x .1)xcl1 .1 B(x) eA( x)A(xv )A(y)A(y) dy f B(x . Example 1. Thus B*2(x )/ B(x) .A(x .A(y)a(y ) = ev'(y) b(y). L(x) y° (a .12.y ) b(y)dy = B (x) o ox _ J = ox/2 eA( x)A(xy ). subexponentiality has alrady been proved in Corollary 1.1 has limit 1 + 0. Jo For y < x/2. Example 1. replace B by a tail equivalent distribution with a failure rate which is everywhere decreasing).(y) dy. Goldie & Teugels [66]): Proposition 1. proving B E S. Thus..14 Consider the DFR Weibull case B(x) = ex0 with 0 <.(x . Thus by dominated convergence .. Thus A(x) is everywhere decreasing.`(x)b(x) is integrable.y) < A (y) for y < x/2.y) y\(y)• The rightmost bound shows that the integrand in the first integral is bounded by ey"(v). the DFR Weibull distriu bution is subexponential. the first integral has limit 1 . we first quote Karamata's theorem (Bingham.16 For L(x) slowly varying and a > 1. Define A(x) = fo . we can use the same domination for the second integral but now the integrand has limit 0 . SUBEXPONENTIAL DISTRIBUTIONS 257 Proof We may assume that A(x) is everywhere decreasing (otherwise.3 < 1.y) * 0. In the regularly varying case. Further. By (1. the u lognormal distribution is subexponential. Then B(x) = eA(x).y) dy.U) /or) v 2x This yields easily that ex. x .15 In the lognormal distribution. f ' L(y) dy . B*2(x) .A(xy)A ( y). a(x) = ax01.1.y) < yA(x .e009xv)2/2a2/(x 2irv2) logx ( ) 't ((logx . and exa(x)b(x) = (3x01e(10)x9 is integrable.A(y)\(y) dy + fox/ 2 eA(x ). Thus.13 works in this setting. 0 A(x) .2). Then b(x) = Ox0lexp.
L(x)/x" and )t(x) .y(x)B(x). we get (1.1))^ ' (b) Assume that for any yo )t(x + y/A(x)) 1 A(x) uniformly for y E (0.1)])a 1 1 . (c) Under the assumptions of either ( a) or (b). then 7(x) x/(a . HEAVY TAILS Proposition 1. the monotonicity condition in 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.18 (a) If B has a density of the form b(x) = aL(x)/xa with L(x) slowly varying and a > 1.6) EX(x) .1)X(x)/x > y) = P(X > x[1 + y/(a . yo] . However. 'y(x) = EXix>.1/A(x) and P(X ixil'Y (x) > y) * e'. (1 + y/(a . More precisely. Then 7(x) .258 From this we get CHAPTER IX.1) and P(X (.17 If B has a density of the form b(x) = aL(x)/x°+1 with L(x) slowly varying and a > 1.1)]) xa L(x) (x[1 + y/(a . Proof ( a): Using Karamata's theorem. .ea b(x) is integrable.4 is necessary in full generality.1))a . Thus exa(x)b(x) .xjX > x. We conclude with a property of subexponential distributions which is often extremely important: under some mild smoothness assumptions. then B(x) .1)] I X > x) L(x[1 + y/(a . Then: Proposition 1. let X W = X . f O B(y) dy .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 .a/x.13 may present a problem in some cases so that the direct proof in Proposition 1.)/Y(x) > y) (1 + y/(a .E(X .
14. 2 The compound Poisson model Consider the compound Poisson model with arrival intensity /3 and claim size distribution B. THE COMPOUND POISSON MODEL 259 We omit the proof of (c) and that EX (x) . Bo(x) = f0 B(y) dy / µB. r(u) = inf it > 0.t be the claim surplus at time t and M = sups>0 St.. It is trivially verified to hold for the Weibull.f yl 0 0 = exp {y (1 + 0(1))} 0 fY A( x + u /A( x)) a(x) du } The property (1.A(x + y/A(x))} =a(x) a(x + x) dx = ex p ex P .. cf.v. . Notes and references A good general reference for subexponential distribution is Embrechts. 1. Let St = Ei ` Ui . Y2.15. Kliippelberg & Mikosch [134]. and that for each Proof Recall from Section 1 that G*n (u) z > 1 there is a D < oo such that G*n(u) < G(u)Dzn for all u. then Vi(u) P Bo(u). .2 Let Y1. The remaining statement (1. d.1 If Bo E S. 0 G(u) L G(u) .EK G(u). with EzK < oo for some z > 1. be i. P The proof is based upon the following lemma (stated slightly more generally than needed at present). Theorem 2 . nG(u).n0 1•P(K= n)•n = EK.nn..and lognormal distributions . Lemma 2.2.1/. Recall that B0 denotes the stationary excess distribution.(x).A(x) I X > x) = exp {A(x) . St > u}..+YK> u) = ^•P(K = n)G* n(u ) . We get p(yl+. Examples 1. i. u a oo. We assume p = /3µB < 1 and are interested in the ruin probability V)(u) = P(M > u) = P(r(u) < oo).7) is referred to as 1/A(x) being selfneglecting.8) in (b) then follows from P (A(x)X (x) > y) = F(X > x + y/. Then P(Y1 + • • • + YK > u) . with common distribution G E S and let K be an independent integervalued r.
Note that in these examples . The problem is a very slow rate of convergence as u ^ oo. and for the lognormal and Weibull cases it can be verified using Pitman 's criterion (Proposition 1. a].3 If B E S. Bo E S.p) and EzK < oo whenever pz < 1. we have fx B(y)dy = a B0 (x) > lim inf lim inf x+oo B(x) . However. u x+a Notes and references Theorem 2. Borovkov [73] and Pakes [280]..1 is essentially due to von Bahr [56]. P(K = k) = (1. in our three main examples (regular variation .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 ) . The approximation in Theorem 2. (2. then Bo(x)/B(x) + 00. Since EK = p/(1. Proof of Theorem 2.18. The PollaczeckKhinchine formula states that (in the setup of Lemma 2.1.µB(01 . lognormal .2) M = Yl + • • • +YK where the Yt have distribution Bo and K is geometric with parameter p.. Proof Since B(x + y)/B(x) * 1 uniformly in y E [0. Bo E S is immediate in the regularly varying case. .1) In particular . u The condition Bo E S is for all practical purposes equivalent to B E S.?(xµ 8 (x).260 CHAPTER IX.p)p'. Weibull) one has Bo(x ( B(x) . HEAVY TAILS u using dominated convergence with >2 P(K = n) Dz" as majorant. Bo ¢ S.1 is notoriously not very accurate. Bo is more heavytailed than B . mathematically one must note that there exist (quite intricate) examples where B E S. See also Embrechts & Veraverbeeke [136]. r(1/Q) xlQexp B(x) = ex' From this . _ B(x^sx Bo(x) µ8 I aoB(y )dy = (^) . see Abate. The tail of Bo is easily expressed in terms of the tail of B and the function y(x) in Proposition 1. x 4 00. as well as examples where B ¢ S.µ J ) . For some numerical studies. In general: Proposition 2.x^ ) B(x) _ f or ( lox .2. the result follows immediately from Lemma 2.x400 PBB(x) PB Leta+oo.13).
. In [1]. We assume positive safety loading.1 gives 1010. 195 there are numerical examples where tp(u) is of order 105 but Theorem 2. G+ (A) = P(Sq+ E A. + Xn. one may have to go out to values of 1/'(u) which are unrealistically small before the fit is reasonable. T+ < oo) where r+ = T1 + • • • + T.1) this end . M = sup s$ .. THE RENEWAL MODEL 261 Choudhury & Whitt [1].e.. Somewhat related work is in Omey & Willekens [278]. t9(u) = inf {n : Snd> > u} . (3. To Bo(u) u + 00.] The proof is based upon the observation that also in the renewal setting.Ti.y + as usual denotes the first ascending ladder epoch of the continuous time claim surplus process {St}. there is a representation of M similar to the PollaczeckKhinchine formula. Asmussen & Binswanger [27] suggested an approximation which is substantially better than Theorem 2. 1 Assume that (a) the stationary excess distribution Bo of B is subexponential and that (b) B itself satisfies B(x . Then K M=EY. p = iB /µA < 1.g.1 when u is small or moderately large.y)/B (x) > 1 uniformly on compact y internals. [279].9+ < oo) = P(S.} Then ik(u) = F ( M > u) = P(i9 (u) < oo). This shows that even the approximation is asymptotically correct in the tail. Then l/i(u) 1 P P [Note that (b) in particular holds if B E S.1. Thus G+ is the ascending ladder height distribution (which is defective because of PB < PA)..3.... 3 The renewal model We consider the renewal model with claim size distribution B and interarrival distribution A as in Chapter V. The main result is: Theorem 3 . i=1 . Let U= be the ith claim . Define further 0 = IIG+II = P(r9+ < oo). Kalashnikov [219] and Asmussen & Binswanger [27]. Based upon ideas of Hogan [200]. also a second order term is introduced but unfortunately it does not present a great improvement. E. i. {n= 0. T1 the ith interarrival time and Xi = U. . in [219] p.+ E A. let t9+ = i9(0) be the first ascending ladder epoch of {Snd> }. Snd) = Xl +.
Lemma 3 . the contribution from the interval (. x + oo. (b) and does not rely on the structure Xi = Ui .d)) E A) denote the pre19+ occupation measure and let and U_ = Eo G'_" be the renewal measure corresponding to G_.g+ > x. 0] normalized by IPG_ I so that we should have to G+(x) . U_ (dy) is close to Lebesgue measure on (.oo.i.3) and we will prove it in this form (in the next Section.(. P(K = k) = (1 .1) holds for a general random walk satisfying the analogues of (a). Let F denote the distribution of the Xi and F1 the integrated tail. 0] to the integral is O(F(x)) = o(FI(x)). FI (x) _ fz ° F(y) dy. oo) = F(S.1 IPG_ I / F(x .y) dy = 1 Pi (X) oo IPG_ I .y_ E A) the descending ladder height distribution (IIG II = 1 because of PB < P A) and let PG_ be the mean of G_. u a 00. As for the compound Poisson model.B(x).FI(u).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_.N. A.262 CHAPTER IX.2 F(x) . The heuristics is now that because of (b). B(x) _ J O° B(B(x)y) A(dy) f 1 .9)9'' and Y1....1) is equivalent to P(M > u) " . HEAVY TAILS where K is geometric with parameter 9. cf. are independent of K and i.Ti). 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. d+ < oo). Proof Let R+(A) = E E'+ ' I(S.3 G+ (x) . (3.d. Proof By dominated convergence and (b). whereas for large y .2). 0 The lemma implies that (3. G_(A) = P(S.Y2.FI(x) /IPG_I. Lemma 3 . Let further 19_ _ inf {n > 0: S^d^ < 0} be the first descending ladder epoch. x > 0. x * oo. with distribution G+/9 (the distribution of S.PBBo(x). Then 0 0 F( x . we will use the fact that the proof of (3.y+ given r+ < oo). A(dy) = 1. Write G+( x) = G+ ( x. and hence FI(x) .
1)/F(n) < 1 + e for n > N (this is possible by (b) and Lemma 3.F[s] = (1 .1 I .G+[s]) .2) yields 00 F F I (u) P(M > u) _ E(1 . n] < (1 + e)/1µc_ I for n > N.e) z lim inf G+(x)  FI (x) Ip G_ I Letting a 10. Hence using dominated convergence precisely as for the compound Poisson model. n] F1 ( n=N _1 1+e E F(x+n) 0 + limsup xr00 FI(x) FAG.3. and that U_(n . We then get lim sup G+(x) xro0 Fj(x) < lim sup X)00 o F(x . 0] x+00 FI(x) 00 + lim up 1 x) E F(x + n) U_ (n . then by Blackwell 's renewal theorem U_ (n .=1 BIp G_ I (1.I n=N (1 E)2 r00 F(x + y) dy + e) lim sup . choose N such that F(n . F(Y= > x) FI(x)/(OIp _ 1). In the lattice case. Similarly.1. If G_ is nonlattice.1. u Proof of Theorem 3.1.y) U_ (dy) 00 FI (x) < lim sup F(x) U(N.(dy) fN FI ( x) + lim sup ZY00 N F(x . and in the last that FI is asymptotically proportional to Bo E S. the proof is complete.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. By Lemma 3. n] is just the probability of a renewal at n.y) U.3. THE RENEWAL MODEL 263 We now make this precise.0)0k k I(u) A.9) 1 .1.1.O[s])(1 . n] + 1/I µG_ I. Given e. > (1 .UG_ I x.9)IpG_ I Differentiating the WienerHopf factorization identity (A. (3.2). we can assume that the span is 1 and then the same conclusion holds since then U(n .
Proof Let w(u) = inf {n : Sid) E (u . Sty(u) . u).u)) = o(P (M > u)) = o(FI(u)). .a) N P(M > u). Then P(M E (u .SS(u)}n=o.IIG+II)µc_ = (1 .(1 . we have P(M E (u .1 is due to Embrechts & Veraverbeke [136].yiui_1.0)ua_ .. u)). 4 Models with dependent input We now generalize one step further and consider risk processes with dependent interclaim times.4 on the joint distribution of (S..1)6+[0] . Mn < u}.So(u)) are available. S+9(u) .So( u)_1 < a) < P (w(u) < oo)j/i(0) < 0(0) P(M E (u . 10(0) But since P(M > u . HEAVY TAILS µF = (1 . with roots in von Bahr [56] and Pakes [280]. 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.a. Note that substantially sharper statements than Lemma 3. allowing also for possible dependence between the arrival process and the claim sizes. Therefore by Lemma 3.a.4 For any a < oo. Notes and references Theorem 3. see Asmussen & Kliippelberg [36]. must attain a maximum > 0 so that P(M > u.a. FJ(u) UBBO(U) PBo(u) N = (10)Ipc_I JUA .Se(u)_1 < a) = o(Fj(u)).Sty(u)_I < a} we have w(u) < oo. and {Su.(u)+n . P(M > u. on the set {M > u. S+q(u) .264 and letting s = 0 yields CHAPTER IX.AB iP We conclude by a lemma needed in the next section: Lemma 3 .a..2.l. u)) > P(w(u) < oo)(i lp (0))• On the other hand.
. (viewed as random elements of the space of Dfunctions with finite lifelengths) are i. We give here one of them..1 based upon a regenerative assumption.X1 is the generic cycle. examples and counterexamples. G(x) (4.. See Fig. {SX1+t . 2.1 = max k=0... 0o(u) etc.. Theorem 4. For further approaches. We return to this point in Example 4. (corresponding to the filled circles on Fig.4. and apply it to the Markovmodulated model of Chapter VI.Sxk}o<t<xk+1xk is the same for all k = 1.... Define S. Figure 4. The idea is now to observe that in the zerodelayed case.Sxi}0<t<x2Xl . +1. M = sup St.F*(X) = P0(Si > x) . < 0 and EoX < oo where X = X2 .1 except for the first one) is a random walk. 4. t>0 S. Assume that the claim surplus process {St}t>o has a regenerative structure in the sense that there exists a renewal process Xo = 0 < Xl <.X2 < . M.1.1. assume pp. = Sx.n n=0....1) is assumed. such that {SXo+t  SXo}0<t< X 1Xo .. Schmidli & Schmidt [47].i.d. see [47].. 4. E0. M* = max S. Thus the assumption . 4..4 below.. ... and the distribution of {Sxk+t ..1 where the filled circles symbolize a regeneration in the path.1) . The zerodelayed case corresponds to Xo = Xl = 0 and we write then F0. We let F* denote the Podistribution of Si. {Sn}n=o..1 Note that no specific sample path structure of {St} (like in Fig.. MODELS WITH DEPENDENT INPUT 265 Various criteria for this to be true were recently given by Asmussen.
1 Assume that (4. jF11 F* (U). The one we focus on is Fo (Mix) > x) .. Since clearly M(x) > Sl . HEAVY TAILS for some G such that both G E S and Go E S makes (3. the assumption means that Mix) and Sl are not too far away. Fo(Si > X).2) to show F(M* > u) > 1.3) applicable so that F(M* > u) 141 F*(u)..2.2 Theorem 4. See Fig.4) liminf u>oo F(M > u) . (4. Proof Since M > M*. N N Xi=0 N Figure 4..Sxn = sup Sxn+t . Sxn +t .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. it suffices by (4.266 CHAPTER IX.* i o<t<xn+1x. u p 00. (4. (4.3) where Mnx) = sup o<t<xn +1 X.1) and (4.. Then '00 (u) = Fo(M > u) .S. 4.3) hold.
. Theorem 4. 2. Then by Lemma 3.a. S.4)..5) which follows since Po (M > u.4.1 can be rewritten as 00 (U) (4. u))/P(M* = 0) = o(Po(M* > u)).e)Po (MMX> > x).1 = limti00 St/t. To this end. E (u .Sn+1Sn>aV(uSn*)) n=1 00 > (1E)EPo(Mn<u. . u)} < P(M* E (u . MW O(u)+1 < a) IN ( U n=1 A1.( u)1 > a) 00 1: Po(Mn<u.Mn +1 >aV(u n=1 00 S. 0 yields (4.(u) . Letting first u + oo and next e .4.e)Po (M > U). Under suitable conditions .E) Po ( n max St u. M^xu)+l > a) . x > a..Po (M* > u. )) > (1 ..: S. Po(M* > u) . choose a such that Po(Si > x ) > (1 . MODELS WITH DEPENDENT INPUT Define 79* (u) = inf {n = 1 . Given e > 0.Sn 0<t<x„+j ( 1 . assume the path structure Nt St = EUit+Zt i=1 .S.(1 .: Sn > u} ..2..e)Po (M > u. Mn+l > a V (u ..6) 1 p pBo(u) u where B is the Palm distribution of claims and p . We shall use the estimate Po(M > u) Miu^+ 1 < a) = o(Po (M > u)) (4. Let a > 0 be fixed. /3(u) = inf{n=1.+Mn+1>u} 267 (note that {M> u} = {3(u) < oo}).a.
(iii) For some o field Y. oX (see Proposition A1.1 is in force. and also for Mix) since Nx FNX U. since the tail of Zx is lighter than B(x) by (iv). X and N. Then the Palm distribution of claims is B(x) = E N Eo 0 I( U1 < x) .p) Ju P Bo(u) 1p 0 . Proof It is easily seen that the r. and the rest is just rewriting of constants: since p = 1+tlim St = 1+ .2 Assume that {St} is regenerative and satisfies (4.4). Mix) < > UE + i=1 o<t<x Thus Theorem 4. i=1 (4.6) holds with p = .} and satisfying Zt/t N. The same is true for Sl.8) x Write .I u J Po(Sl > x) dx 1 EoNxB(x) dx EoX(1 .3PB. Corollary 4.X both have tails of sup Zt.268 with {Zt} continuous. the proof of Lemma 4.'s order EoNx • B(x).7). Assume further that (i) both B and Bo are subexponential. (ii) EozNX < oo for some z > 1.6 below. are Fmeasurable and NX Po J:U=>x i=1 (iv) Po sup Zt > x / (0:5t<x o(B(x)) Then (4.v. a4' 0. we get 00 (u) 1 IPF.Q = EoNx/EoX. cf. HEAVY TAILS N` U. and ENX Ui . independent of {> CHAPTER IX.
and for some constants ci < oo such that cl + • • • + c. The number N.3 As a first quick application. We now return to the Markovmodulated risk model of Chapter VI with background Markov process {Jt} with p < oo states and stationary distribution 7r.. The arrival rate is /3i and the claim size distribution Bi when Jt = i. Bo E S. Assume that B E S.t} is standard compound Poisson and {Zt} an independent Brownian motion with mean zero and variance constant a2. Example 4 . > 0. we conclude just as in Example 4. (i) holds. In particular.9).0 (thus (iv) is trivial).3 that (4. MODELS WITH DEPENDENT INPUT 269 Example 4 . 1) is Poisson with rate /3 = fo /3(s) ds so that (ii) holds.6) u holds. 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 .. consider the periodic model of VI.. More precisely. Theorem 4.5 Consider the Markovmodulated risk model with claim size distributions satisfying (4.4. i. We consider the case where one or more of the claim size distributions Bi are heavytailed. X2 = 1. we assume that B E S. Again . note that the asymptotics of i/io( u) is the same irrespective of whether the Brownian term Zt u in St is present or not. then (iv) holds since the distribution of supo<t<i Z(t) is the same as that of I Zl 1. . . . in particular lighttailed. X2 = 1..e.6) holds. < 1. Bo E S.6) holds.. of claims arriving in [0. Thus we conclude that (4. Taking again Xo = Xi = 0.(NX). (iii) is obvious. i=1 B = >2 7riaiBi i=1 and we assume p = 014 B = Ep ri/3ipB. The key step of the proof is the following lemma. and taking F = o.t + EN'I Ui where {>N`1 Ui . X3 = 2.4 Assume that St = Zt . Zt . X3 = 2.6 with arrival rate /3(t) at time t (periodic with period 1) and claims with distribution B (independent of the time at which they arrive). . Then (4. 3 The average arrival rate / and the Palm distribution B of the claim sizes are given by P P Q = ir i/i. The regenerative assumption is satisfied if we take Xo = Xi = 0.
c'(x) where c = ciENi . 1. are independent with distribution Fi for Xij. and F a aalgebra such that (N1.5. . oo) such that G E S and some c1. and the rest of the argument is then just as the proof of Corollary 4. .. An easy conditioning argument then yields the result when Jo is u random. . .G( x ) > ciNi . NP ) and X are ..F) = P(Yo > X+x I •^) G (x +x)>2ciNi i=1 . If Jo = i. Then P P(Yx > x) .2. Markovmodulation typically decreases the adjustment coefficient y and thereby changes the order of magnitude of the ruin .X i=1 j=1 where conditionally upon F the Xi.v. . we can define the regenerations points as the times of returns to i. P P P(YX and > x I.. i1 = E\ G(x) In the general case."+Np . and that for some + cp distribution G on [0.. HEAVY TAILS Lemma 4 . Assume EzN1+"'+Np < oo for some z > 1 and all i.. 2 .}P. .. 6 Let (N1. The same dominated convergence argument completes the proof.. For lighttailed distributions. as x a oo.. It follows by a slight extension of results from Section 1 that P P(Yo > x I Y) G( x) ci Ni. i =1 P(Yo > x I ^ ) < CG(x)zN1+ +Np for some C = C(z) < oo. u Proof of Theorem 4. Let {Fi}t=1 P be a family of distributions on [0. i=1 Proof Consider first the case X = 0.Fmeasurable. NP ) be a random vector in {0. Thus dominated convergence yields ( P(Yo>x P(Yo>x .^•) G(x) P ^ E ciNi = C. . oo) and define p Ni Yx = EEX'i .F) < CG(x)zn'1+. i=1 P(Yx > x ^) < P(Y0 > x I.ciG(x). X > 0 a r.270 CHAPTER IX. cp with cl + > 0 it holds that Fi(x) .
this is applied for example to risk processes with Poisson cluster arrivals. Floe Henriksen & Kliippelberg [31] by a lengthy argument which did not provide the constant in front of Bo(u) in final form. Combined with the approximation for O(u). the discussion provides an alternative point of view to some results in Chapter IV. this then easily yields approximations for the finite horizon ruin probabilities (Corollary 5.4.2 and Example 4. 5 was first proved by Asmussen.7).p)Bo(u). cf. and the final reduction by Jelenkovic & Lazar [213]. Within the class of risk processes in a Markovian environment. states that under mild additional conditions.6) to hold in a situation where the interclaim times (T1.T2.. IV. 5 Finitehorizon ruin probabilities We consider the compound Poisson model with p = /3pB < 1 and the stationary excess distribution Bo subexponential. Notes and references Theorem 4.5 that the effect of Markovmodulation is in some sense less dramatical for heavytailed distributions: the order of magnitude of the ruin probabilities remains ft°° B(x) dx. this should be compared with the normal limit for the lighttailed case.7.4. 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. Theorem 4.. We start by reviewing some general facts which are fundamental for the analysis. As usual. Schmidli & Schmidt [47]. That paper also contains further criteria for regenerative input (in particular also a treatment of the delayed case which we have omitted here).. I T(u) < oo). and independent of (T1. i.5. in particular Proposition 2. Essentially. as well as a condition for (4. 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).T2. For further studies of perturbations like in Corollary 4.. The present approach via Theorem 4.4.1 is from Asmussen.3. VI. cf..i. An improvement was given in Asmussen & Hojgaard [33]. see Schlegel [316]. FINITEHORIZON RUIN PROBABILITIES 271 probabilities for large u. The main result of this section.1. Then O(u) . Theorem 2. It follows from Theorem 4.4. cf.. ). Theorem 5.d. i. ) form a general stationary sequence and the U. In contrast.5 shows that basically only the tail dominant claim size distributions (those with c. > 0) matter for determining the order of magnitude of the ruin probabilities in the heavytailed case. m is a (orfinite) . we let PN"N = P(.e. for lighttailed distributions the value of the adjustment coefficient y is given by a delicate interaction between all B. r(u) is the time of ruin and as in IV.pl(1 .
x = 0+ and F = (0. Let G denote the distribution of ENt U.= y. the whole of R and not as usual impose the restrictions x > 0. We let QS be the corresponding distribution and Qx. however . . St is distributed as y .s. a main difficulty is to make sense to such excursions also when Px(w(F°) = 0) = 1. an excursion in F starting from x E F is the (typically finite) piece of sample path' {St}o<t<w(F°) I So = x where w(Fc) = inf It > 0: St 0 F} .t + EI U.2) for all bounded measurable functions h. 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. {Rt}.z. HEAVY TAILS measure on E such that L for all measurable A C E and all t > 0. follows by the substitution y = x . w(Fc) < oo ) 'In general Markov process theory. Sw(F. resp.)k(x .rij = mjsji where r13. say. but the example of relevance for us is the following: Proposition 5. where we can take h.r. r. Then (5. and starting from So = y. Lebesgue measure. y to vary in. . For the present purposes it suffices . t. to consider only the case Px(w(F`) = 0) 0.y = Qx (. The simplest example is a discrete time discrete state space chain.z) dx G(dz) = ffh(y + z) k(y)dy G(dz). to the r. Thus. k as indicator functions. in the terminology of general Markov process theory.t.). Rt is distributed as x + t . Say {St} is reflected Brownian motion on [0.2) with t = 1 means m. oo). a familiar case is time reversion (here m is the stationary distribution).s=j are the transition probabilities for {St}. Proof Starting from Ro = x. j. for states i. The equality of the l.1 A compound Poisson risk process {Rt} and its associated claim surplus process {St} are in classical duality w . u For F C E. m.2) means ffh(a. k on E.s.h.h. y = 0). {St} and {Rt} are in classical duality w.>N` Ui.272 CHAPTER IX.t. (note that we allow x.00). and (5.
FINITEHORIZON RUIN PROBABILITIES 273 y E F (in discrete time. That is. S.. oo) = r(0) x= St y (a) Figure 5. this simply means the distribution of the path of {Rt} starting from y and stopped when 0 is hit.s. Qx y is the distribution of an excursion of {St} conditioned to start in x E F and terminate in y E F.y as a measure on all strings of the form i0i1 .3 The distribution of r(0) given r(0) < oo. Sn = in = y. 5.. w(0. In particular: Corollary 5..). io = x.itt) = P Px(w(Fc) < 00.SS(F.. 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. QR and QRy are defined similarly. Sn+1 E Fc) nx..5. [note that w(z) < oo a. 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).y() = P ({SW(F`)t} 0<t<w(F °) E So = x. Sw(F)1 = y) . and we let Qy y refer to the time reversed excursion . We can then view Qy.2.. in = y. Qx. .= y) Theorem 5 .(0)_ = y < 0 is the same as the distribution of w(y) where w(z) = inf It > 0 : Rt = z}. i1..13AB < 1] Proof of Theorem 5. . the one in (b) is the time reversed path.1 The sample path in (a) is the excursion of {St} conditioned to start in x = 0 and to end in y > 0. Sw(Fo)_ should be interpreted as Sw(F^)_1). z > 0. But in the risk theory example (corresponding to which the sample paths are drawn). in E F.2 Qy. .1 for the case F = (oo.. Thus. when p = . in with i0. x = 0. 0].y(2p21 . /^s x (S1 = Z1.y = Qy Q. The theorem is illustrated in Fig . .
... in = x..... ..rin_1in E Txj jEFC m21 s2120 m2252221 m in Ssn n1 mjSjx Mx m2p mil min1 jEF` 1 Sinin _ 1 ...y(inin _ E SYj jEF` 00 Sxik _1 ..J (i..in1 . S.. in) = Qx.. Silt' E SO k=1 i1. .. = in = y. .(F<)1 = Y) S S and Qx y( ipil .. 2p) when 20.. Si11 S 1 . .. i0) Q x.. Si1y k=1 i1 . note first that Pt' (R l = il. HEAVY TAILS E E Px (Si = 21i .ik_1EF Sxin_1 . 21 . in = x. MY Thus Qx(ioii . in E F.in E F.. R Qy x(2p21 ..... Rn = in = x.. Si1y 00 jEF° E E 5xik_ 1 . Silt' E Sxik_1 .. ... t' y and Qy x are measures on all strings of the form ipi l .... Sn+1 E Fc) n=1 i1.. 2p).. (Fc)1 = y) 00 CHAPTER IX.. ... .. Rn+1 E FC) TioilTili2. .... in with 20. Rn+1 E F`) F (w(Fc) < 00. in) = oo jEF^ Sxin1 ...274 note that Fx(w(Fc) < 00...ik1EF Similarly but easier Sxin_1 ... To show Q y x (i0 i 1 .gilt' k=1 ii .....TI( 2n2n _1 . . S. in)  Pt' (R1 = ii.. 2n) = Qx. i0 = y. Si l io E mjSjx.... 20 = y.ik_1EF ..ii ..i„_iEF Similarly.... Rn = in = x. R .
t. S. 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 .UBBo(u)]. Bo") is also the P(u. the distribution w.t. We are interested in the conditional distribution of T(u) = T(0) given {T(0) < oo.r. y > u. Z follows the excess distribution B(Y) given by B(Y) (x) _ B(y + x)/B(y).2 The distribution of (Y. that is. Z = Zl = ST+( 1)_ the value just before the first ladder epoch (these r.')distribution of Z since P(Z>aIY>u) = 1 °° B(y) B(y + a) dy FLBBo(u) B (y) J°° (z) dy . Y > u). see Fig.(o) > y} = {T(0) < oo. Let Y = Yl = Sr+( 1) be the value of the claim surplus process just after the first ladder epoch . ST(o) > y. the P(u.B(a) +a PBBo(u) . P(") = P(. Y > y} .2.5. 7(0) < oo. To clarify the ideas we first consider the case where ruin occurs already in the first ladder segment .r. the case r (O) < oo.')distribution of Yu is Bo").2. The formulation relevant for the present purposes states that Y has distribution Bo and that conditionally upon Y = y.p. 1 w .v.2. Now the P(u.')density of Y is B(y)/[. That is. Z) is described in Theorem 111. 5. that is.'s are defined w. P(o) ). U T(O) = T (u) Y Figure 5.
Yn_1 'typical'. more precisely. Zn). i. 2... then by the subexponential property Yn must be large. In the proof.1. That is . Then Corollary 5.. Z1). The idea is now to observe that if K(u) = n. Zk be defined similarly as Y = Y1. However....p). we get the same asymptotics as when n = 1. We now turn to the general case and will see that this conclusion also is true in P(")distribution: Theorem 5 . Since w(z)/z a$. and YI. 4 Assume that Bo E S and that (5.e. Zn_1 'typical' which implies that the first n1 ladder segment must be short and the last long. HEAVY TAILS Let {w(z)}Z^. . denote the ladder epochs and let Yk.. Hence Z. are i. cf. Z/'y(u) * W in Pi "' ')distribution . Y1 + • • • + Yn > u} denote the number of ladder steps leading to ruin and P("'n) = P(• I r(u) < oo.. Then 7(u)/y(u) ^ W/(1 .. .. Z = ZI but relative to the kth ladder segment.e. 1/(1 . Then. must be large and Z1. let r+(1) = T(0).. P(Z < a I Y > u) 3 0. ..276 CHAPTER IX. and since its dominates the first n . . conditionally upon r+ (n) < oo. z ^ oo.1) of the last ladder segment can be estimated by the same approach as we used above when n = 1. > u with high probability.p) then yields the final result T(u)/y(u) + W/(1 . this in principle determines the asymptotic behaviour of r(u). Since the conditional distribution of Z is known (viz.3) where the distribution of W is Pareto with mean one in case ( a) and exponential with mean one in case (b). a slight rewriting may be more appealing.. Recall the definition of the auxiliary function y(x) in Section 1.: r+ (n) < oo. (Y. and distributed as (Y. . It is straightforward that under the conditions of Proposition 1. We let K(u) = inf In = 1.o be defined by w(z) = inf It > 0: Rt = z} where {Rt} is is independent of {St}.p) in Pi"'')distribution.. K(u) = n). . in particular of Z. . it therefore follows that T(u)/Z converges in Pi"'')probability to 1/(1 .3 implies that the P("'1)distribution of T(u) = r(0) is that of w(Z). Now Bo E S implies that the Bo ")(a) + 0 for any fixed a. the duration T+ (n) . Z)..d. r(u)/Z 4 1/(1 .p).r+ (n . Bo") ).3) holds.p) in F(u) distribution. i.T+(2). Fig. 5.3.i.18(c) Bo")(yY (u)) + P(W > y) ( 5. the random vectors (YI.
Yn .u) II 0.. Y„1. Proof We shall use the easily proved fact that if A'(u). suitably adapted). +Yn1<u. the condition on A'(u) A A"(u) follows from Bo being subexponential (Proposition 1. then IIP( I A'(u)) Taking A'(u) = {Y..Yn1iYn . .u) E •) .. .5 Ilp(u.2. A"(u) are events such that P(A'(u) AA"(u)) = o(F (A'(u)) (A = symmetrical difference of events). Further.5. > u}..n)..Yl+ +Yf1>u}. P (Yj. . I A'(u)) = P(u.. P(. A"(u) _ {K(u)=n} = {Y1+ P(.n) (y1.Bo (ri1) ®B( ..u) E • I A'(u)) = Bo (n1) ®Bou) . FINITEHORIZON RUIN PROBABILITIES 277 16 Z3 Z1 r+(1) T+(1) T+(1) Figure 5.3 In the following. Lemma 5. I A"(u ))II + 0. II ' II denotes the total variation norm between probability measures and ® product measure. .
2. in our example Y = (Y1... Now use that if the conditional distribution of Z' given Y' is the same as the conditional distribution of Z given Y and JIF(Y E •) .y(u)T) . wk(Zk) has a proper limit distribution as u + oo for k < n.' = y is BM. whereas wn(Zn) has the same limit behaviour as when n = 1 (cf..p) < y). the discussion just before the statement of Theorem 5. .. the density of Yn is B(y)/[IBBO(u)]. 2. Y") u etc.i. ....u has distribution Bout That is...P(Y' E •)II * 0. {wn(z)} be i.7 O (u. Thus F(u'n)(T(u) /7(u) > y) = F(u'n)((wl (Z1) + .4).. Zn). Let {wl(z)}. Similarly (replace u by 0)..278 Lemma 5 . .. and that Yk has marginal distribution B0 for k = 1. By Lemma 5. .t. Z' are arbitrary random vectors.. and clearly Zi. The first step is to observe that K(u) has a proper limit distribution w. Then according to Section 5a..1.+y 1 p"F(Yn > u) P)Pn1 P/(1 . see Fitzsimmons [144]). y > u. Zn are independent.. P(u) since by Theorem 2. k = 1.d.P) Bo(u) for n = 1.1 P PBo(u) • P(W/(1 .. Notes and references Excursion theory for general Markov processes is a fairly abstract and advanced topic.4. be independent random vectors such that the conditional distribution of Zk given Y.1). For Theorem 5. the marginal distribution of Zk is Bo for k < n. Z11). n . + Y" > u) Flul (K (u ) = n) _ Cu) P"F(1'i +. Proof Let (Y11... n.. .P(Z' E •)II > 0 (here Y.6 IIPIu'n ) CHAPTER IX. . .. . Zn) E •) .).P) > y) Corollary 5. . n_1 < u. then 11P(Z E •) . the F'distribution of r(u) is the same as the P'distribution of w1(Zl) + • • • + wn(Zn).. Z. +wn(Z n))l7( u ) > 1y) ^' P(u'n)(wn (Zn)/7(u) > y) 4 NW/(1 . The same calculation as given above when n = 1 shows then that the marginal distribution of Zn is Bou).. It therefore suffices to show that the P(u'")distribution of T(u) has the asserted limit.. HEAVY TAILS ((Z1'. . .r. Y'.1 and Y„ .. Proof of Theorem 5.. (Y.6.. copies of {w(z)}. Y1 +.. in particular his Proposition (2...Bo (n1) ®Bo' 0. .
u (6.2 Define M. that MQ becomes large as consequence of one big jump./3Ea B(u). and premium rate p(x) at level x of the reserve.(u) = P(V > u) = f f (y) dy . Extensions to the Markovmodulated model of Chapter VI are in Asmussen & Hojgaard [33]. the results only cover the regularly varying case. that fo p(x)1 dx < oo. The heuristic motivation is the usual in the heavytailed area. V.y) . Then 0 (u) Qf "O ^) dy. Asmussen & Teugels [53] studied approximations of i (u. cf.(3 u u J B(y) dy . 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). i. max VB>0I Vo=0^ o<s<t J11JJJ Lemma 6 .6. Corollary II. The rigorous proof is. the probability that is exceeds u is then B(u . x > oo. RESERVEDEPENDENT PREMIUMS 279 The results of Section 5b are from Asmussen & Kluppelberg [36] who also treated the renewal model and gave a sharp total variation limit result . claim size distribution B.2. one expects the level y form which the big jump occurs to be 0(1). . 6 Reservedependent premiums We consider the model of Chapter VII with Poisson arrivals at rate /3. We will show that the stationary density f (x) of {Vt} satisfies f (x) /B(x) r(x) We then get V.B(u). Theorem 6 . nontrivial and we refer to Asmussen [22]. More precisely.1.e. 3. and define the cycle as a = inf{t>0: Vt=0. p(Y) and the result follows. however.. Proof of Theorem 6.1 Assume that B is subexponential and that p(x) > 00. = supo<t<0. T) when T + oo with u fixed. Assume for simplicity that {Vt} regenerates in state 0 . Then P(MT > u) .1) The key step in the proof is the following lemma on the cycle maximum of the associated storage process {Vt}.
u Notes and references The results are from Asmussen [22].280 CHAPTER IX. Then D(u) = f(u)p(u) and. Further the conditional distribution of the number of downcrossings of u during a cycle given Mo > u is geometric with parameter q(u) = P(Mo > u I Vo = u).q(u) Now just use that p(x) * oo implies q (x) + 0. by regenerative process theory.q ( u)) 1 . HEAVY TAILS Define D(u) as the steadystate rate of downcrossings of {Vt} of level u and Da (u) as the expected number of downcrossings of level u during a cycle.P(MT > u) $B(u) Ft µ(1 . . It is also shown in that paper that typically. there exist constants c(u) 4 0 such that the limiting distribution of r(u)/c(u) given r(u) < oo is exponential. where also the (easier) case of p(x) having a finite limit is treated . Hence f (u)r(u) = D(u) = Do(u) . D(u) = DQ(u)/µ.
Hence 1.. Ripley [304]. The crude Monte Carlo ( CMC) method then amounts to simulating i. 281 .. ZN.96s z f (1.Chapter X Simulation methodology 1 Generalities This section gives a summary of some basic issues in simulation and Monte Carlo methods . and this is the form in which the result of the simulation experiment is commonly reported. la The crude Monte Carlo method Let Z be some random variable and assume that we want to evaluate z = EZ in a situation where z is not available analytically but Z can be simulated. estimating z by the empirical mean (Z1 + • • + ZN)/N and the variance of Z by the empirical variance N s2 = E(Z{  N 2. where a2 = Var(Z ).. z) 2 = Zit NE ii ii According to standard central limit theory . vrN(z . . Fox & Schrage [77]. Rubinstein [310] or Rubinstein & Melamed [311] for more detail .d. 4Z). We shall be brief concerning general aspects and refer to standard textbooks like Bratley.2) is an asymptotic 95% confidence interval .i. replicates Zl. topics of direct relevance for the study of ruin probabilities are treated in more depth.z) 4 N(0.
b(u. Further. and in most cases this modest increase of N is totally unproblematic. We mention in particular ( regression adjusted) control variates and common random numbers. variance reduction is hardly worthwhile. Conditional Monte Carlo Let Z be a CMC estimator and Y some other r . writing Var(Z) = Var(E [Z I Y]) + E(Var[Z I Y]) . The situation is more intricate for the infinite horizon ruin probability 0(u). conditional Monte Carlo and importance sampling. generated at the same time as Z. 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. Therefore. there are others which are widely used in other areas and potentially useful also for ruin probabilities. We survey two methods which are used below to study ruin probabilities. lb Variance reduction techniques The purpose of the techniques we study is to reduce the variance on a CMC estimator Z of z. an added programming effort. Say that Var(Z') = Var(Z)/2. Typically variance reduction involves both some theoretical idea (in some cases also a mathematical calculation). typically by modifying Z to an alternative estimator Z' with EZ' = EZ = z and (hopefully) Var(Z') < Var(Z). and a longer CPU time to produce one replication. 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. so that Z' is a candidate for a Monte Carlo estimator of z. we then have EZ = EZ = 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'. Z = I inf Rt < 0 (0<t<T = I('r(u) < T).282 CHAPTER X. one can argue that unless Var(Z') is considerable smaller than Var(Z). This is a classical area of the simulation literature. However. v. 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. Letting Z' = E[Z I Y]. and many sophisticated ideas have been developed. Sections 24 deal with alternative representations of Vi(u) allowing to overcome this difficulty.
the argument cheats because we are simulating since z is not avaliable analytically. a crucial observation is that there is an optimal choice of P: define P by dP/dP = Z/EZ = Z/z. and the problem is to make an efficient choice.3) Thus. using the CMC method one generates (Z1. L1). one would try to choose P to make large values of Z more likely.zrs.E [Z Z]2 = z2 .zrs) = 2 1 N 2 2 2 i=1 i=1 N > Lt Zi . i.v. but tentatively.3). . GENERALITIES 283 and ignoring the last term shows that Var(Z') < Var(Z) so that conditional Monte Carlo always leads to variance reduction. . Then z Var(LZ) = E(LZ)2 .1. Thus we cannot compute L = Z/z (further. Variance reduction may or may not be obtained: it depends on the choice of the alternative measure P. In order to achieve (1. Importance sampling The idea is to compute z = EZ by simulating from a probability measure P different from the given probability measure F and having the property that there exists a r. it may often be impossible to describe P in such a way that it is straightforward to simulate from P). (1. Thus. even if the optimal change of measure is not practical.[E(LZ)] = E Z2 Zz . LN) from P and uses the estimator N zrs = N > L:Zj i=1 and the confidence interval zrs f 1. L = z/Z (the event {Z = 0} is not a concern because P(Z = 0) = 0).96 sis v^ N 2 1 where srs = N j(LiZi . it gives a guidance: choose P such that dP/dP is as proportional to Z as possible. it appears that we have produced an estimator with variance zero. the obvious possibility is to take F and P mutually equivalent and L = dP/dP as the likelihood ratio.z2 = 0.e. L such that z = EZ = E[LZ].. To this end. (ZN. This may also be difficult to assess .. Nevertheless. However. .
B = iP(AB) = P(BIA). u + oo. large sample sizes are required.96 2 z2 z increases like z1 as z .284 CHAPTER X.5 or even much smaller . as is the case of typical interest. Again.e. Z z V5 In other words . assume that the A(u) are rare in the sense that z(u) * 0. SIMULATION METHODOLOGY 1c Rare events simulation The problem is to estimate z = P(A) when z is small . In ruin probability theory. i. This leads to the equation 1. but if the point estimate z is of the order 105.e. Another way to illustrate the problem is in terms of the sample size N needed to acquire a given relative precision . The CMC method leads to a variance of oZ = z(1 . Thus. Z = I(A) and A is a rare event. we may try to make P look as much like P(•IA) as possible.z) which tends to zero as z ^ 0. the optimal P is the conditional distribution given A. in terms of the halfwidth of the confidence interval. 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). 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. We then . However. An example where this works out nicely is given in Section 3... the issue is not so much that the precision is good as that relative precision is bad: oZ z(1 .z) 1001. I. N .0. Two established efficiency criteria in rare events simulation are bounded relative error and logarithmic efficiency. say of the order 103 or less. To introduce these. However. The optimal change of measure ( as discussed above) is given by P(B) = E [ Z] i.1.1. it does not help telling whether z is of the magnitude 104. let z(u) = P(A(u)). say 10%.e. a confidence interval of width 10 4 may look small. 10 . A = {T(u) < T} or A = {r(u) < oo} and the rare events assumption amount to u being large. and further it is usually not practicable to simulate from P(•IA). if z is small. assume that the rare event A = A(u) depends on a parameter u (say A = {r(u) < oo}).96oz /(zV) = 0.96 2Z ( 1 . z I.z) 1 > 00.100 . For each u.
where M = X1 + • • • + XK. X2..log z(u) of (1. The term logarithmic comes from the equivalent form . .X1 + + XK. 3. let Z +.1. P(K = k) = (1 .i. Generate K as geometric. F(K = k) = (1 . and in practice.d. see Asmussen & Rubinstein [45] and Heidelberger [190]. 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. Thus. Let M .. 2 Simulation via the PollaczeckKhinchine formula For the compound Poisson model. it is appealing to combine with some variance reduction method .e.log Var(Z(u)) lim inf > 2 u+oo . it is not efficient for large u . O (u) = z = EZ.0.. 2. However. but as a CMC method . SIMULATION VIA THE POLLACZECKKHINCHINE FORMULA 285 say that {Z(u)} has bounded relative error if Var(Z(u))/z(u)2 remains bounded as u 3 oo.1) may be written as V) (u) = P(M > u). Therefore . where X1. Notes and references For surveys on rare events simulation.p)pk.4). this means that the sample size N = NE(u) required to obtain a given fixed relative precision (say a =10%) remains bounded. so that NE (u) may go to infinity. Generate X1. If M > u. with common density bo(x) = B(x)/µB and K is geometric with parameter p.(2. We shall here present an algorithm developed by Asmussen & Binswanger [ 271. let Z +. the PollaczeckKhinchine formula III. Otherwise. The algorithm gives a solution to the infinite horizon problem . . the mathematical definition puts certain restrictions on this growth rate. i.2. Var(Z(u)) hm sup U+00 z (u) 2E < oo (1. are i. logarithmic efficiency is almost as good as bounded relative error.p)pk.4) for any e > 0. where Z = I(M > u) may be generated as follows: 1. which gives a logarithmically efficient estimator . XK from the density bo(x). According to the above discussion.. . This allows Var(Z(u)) to decrease slightly slower than z(u)2.
.. Z(1)(u) is defined as 0).. XK. SIMULATION METHODOLOGY when the claim size distribution B (and hence Bo) has a regularly varying tail. 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.XK_1 and let Z( 1)(u) = Bo (Y) (if K = 0. we generate only X1...L(x)/x`' with a > 0 and L(x) slowly varying. asymptotically it presents no improvement : the variance is of the same order of magnitude F(x). 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..X(K1)) .2.X(2).X(2)..S( K_1)) V X(K1)) / Bo(X(K 1)) where S(K_l) = X(1) + X(2) + • • • + X(K_1).. For the simulation. compute Y = u . and considering only the remaining ones. and that Bo(y) = 1..XK_1).... conditional probability.+XK > uIXl. we thus generate K and X1i ..X(n_1)) Bo(X(„_l) V X) Bo(X(n1)) . XK1. form the order statistics X(1) < X(2) < .....p)Bo(x)..b(u) = P (Xl +•••+XK>u) = EF[Xl + . .1) V)(u) . < X(K) throw away the largest one X(K).p/(l . Thus. Xl > x. just note that EZ(1)(u ) 2 > E[Bo (x . note first that To check the formula for the P(X(n) > x I X(1). As a conditional Monte Carlo estimator .X1 . .. Z(1) (u) has a smaller variance than Zl (x). So. assume in the following that Bo(x) ..286 CHAPTER X. Then (cf. . The idea of [27] is to avoid this problem by discarding the largest X. X1 + + XK_ 1 > x when X1 > x. However. ......SK1)2... Theorem IX.. y < 0).XK1] = EBo(uX1 . and let Z(2)(u) = _ P (SK B0((u > u I X(l).Xl . To see this. A first obvious idea is to use conditional Monte Carlo: write i. K > 2] = P2p(Xl > x) = P2Bo(x) (here we used that by positivity of the X.
B. 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. 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)) . Then the algorithm given by { Z (2) (u) } is logarithmically efficient. 3 Importance sampling via Lundberg conjugation We consider again the compound Poisson model and assume the conditions of Ce7". for the purpose of recording Z(u) = erysr(u). X(2). 111... and that paper contains one more logarithmically efficient algorithm for the compound Poisson model using the Pollaczeck. Binswanger and HOjgaard of [28] give a general survey of rare events simulation for heavy tailed distributions . in the renewal or Markov. the algorithm for generating Z = Z(u) is: 1. and simulate from FL._1) > P(X(n) > _ X X(1).3. For practical purposes. . Asmussen .Khinchine formula and importance sampling . X(2). BL by I3L = /3B[y].y. . However .. IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 287 We then get P(S" > x I X( 1). use the the CramerLundberg approximation so that z(u) = '(u) = e7"ELe7E(") where ^(u) = ST(") . BL instead of 0. X (. .1) . Thus. Compute y > 0 as solution of the Lundberg equation 0 = K(y) = )3(B[y] . and define )3L. .u is the representation 0(u) = e7sr(u) overshoot (cf.5). using 13L.. BL(dx) = e7sB(dx)/B[y].S(n_1) I X(1).S (n1)) V X (. the continuoustime process {St} is simulated by considering it at the discrete epochs {Qk} corresponding to claim arrivals. The algorithm is sofar the only efficient one which has been developed for the heavytailed case. X . and we refer to [27]. X(n1)) Bo((x . X(2). l)) BO(X(n1)) Theorem 2 . that is. 1 Assume that Bo (x) = L(x)/x° with L(x) slowly varying. X(n1)) P(X(TZ) + S(. Notes and references The proof of Theorem 2. ..modulated model P(r+ < oo) and G+ are not explicit ). .1 is elementary but lengty..
i. B) is not logarithmically efficient when (/3. return to 3.(u)) are asymptotically coincide on {r(u)} < oo. and we have: Theorem 3. We may expect a small variance since we have used our knowledge of the form of 0(u) to isolate what is really unknown. cf. There are various intuitive reasons that this should be a good algorithm.7 tell that P(. to deal with the infinite horizon problem . r(u) < oo) and FL (both measures restricted to. b) # (/3L. let S. Ti. 4.1) (simulated with parameters ^3..l3 and U from B. Proof Just note that EZ(u)2 < e .3. BL could improve the variance of the estimator . Generate T as being exponential with parameter .2ryu _ z (u)2/C2. and avoid simulating the known part e7". In detail . Otherwise. M(u) = inf {n : S„ > u}. b different from . so that changing the measure to FL is close to the optimal scheme for importance sampling . Xi = U. X2. A > AL as in Chapter V. Let FL (dx) = 'For the renewal model. and assume that µF < 0 and that F[y] = 1. with distribution F. the discussion at the end of Section 1b.Q.1 The estimator Z(u) = e'rs* "u) (simulated from FL) has bounded relative error. and the change of measure F r FL corresponds to B > BL. BL). Let S . . be i. Let X1. The answer is no.2 The estimator (3. In fact: Theorem 3..d. the results of IV. Let Sf0 CHAPTER X. It resolves the infinite horizon problem since FL(.. If S > u. 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. = X1 + .F. u It is tempting to ask whether choosing importance sampling parameters .288 2. The proof is given below as a corollary to Theorem 3. We formulate this in a slightly more general random walk setting '. . + X. let Z F e_'s. More precisely. The algorithm generalizes easily to the renewal model .r(u) < oo) = 1.S+U .... one must restrict attention to the case 4µB > 1.. P'[y] < oo for some ry > 0.QL.T. namely ELery£("). SIMULATION METHODOLOGY 3.
= c'. Since ELM(u)/u + 1/ELXi. let F be an importance sampling distribution equivalent to F and M(u) dF Z(u) _ I (Xi) .3 The estimator (3. F(XM(u)). EFZ(u)2 = EeW2(FIF) = Ep [W2(FIFL)W2(FLIF)] = EL [W2 ( FIFL)w(FLIF)] = ELexp {Kl+.2ryELXi. 1.. where e' = EL Iog dFL (Xi) > 0 by the information inequality.... write W(F IF) _ F(XI).2) dF Theorem 3. EFZ(u)2 EFZ(u)2 lim sup z(u)2eeU = lim cop C2e2...yu = G. are i.2ryELXi)} .2 > 0. is not logarithmically efficient. IMPORTANCE SAMPLING VIA LUNDBERG CONJUGATION 289 e7yF(dx). it thus follows that for 0 < e < e'/ELXi. + KM(u))} = exp {ELM(u)(E . For the second. where Kl og (X) (j) 2 ) = log dFL (Xi) .. (3.d... By the chain rule for RadonNikodym derivatives. .P = FL. Since K1.i. K2.+KM(u)}. The importance sampling estimator is then Z( u) = e'rSM( ).2'X1 .3.yu+elu u +oo etry' 1 > lim up C2e2. When F # FL. More generally. Here ELK. . Proof The first statement is proved exactly as Theorem 3 . Jensen's inequality and Wald's identity yield EpZ(u)2 > exp {EL(K1 + .2) (simulated with distribution F of the X3 has bounded relative error when .
3. /3". yu) is close to zk(u). optimality is discussed in a heavy traffic limit y 10 rather than when u + oo.T".T' has a left exponential tail with rate /3' and U" . then the estimator Z(u) = e7Sr(°)I(r(u) < yu) (simulated with parameters /3L. see e. In fact: Proposition 4. The results of IV. In [13].T" has a left exponential tail with rate /3'. . This immediately yields / = 3". The queueing literature on related algorithms is extensive .1 If y > 1/ic'('y). all that needs to be shown is that if U' . U' .4 indicate that we can expect a major difference according to whether y < 1/r. B" and generic claim sizes U'.2. The extension to the Markovian environment model is straightforward and was suggested in Asmussen [ 16]. Then according to Theorem 3. T".1 is from Lehtonen & Nyrhinen [244].290 which completes the proof. Next. then /3' B' = B". we conclude by differentiation that Bo(x)=B' (x)forallx > 0. U".'(y) or y > 1/r.T". /3'eQ'YR'( x + y) dy = .g. claim size distributions B'. As in IV. 4 Importance sampling for the finite horizon case The problem is to produce efficient simulation estimators for '0 (u. SIMULATION METHODOLOGY u Proof of Theorem 3.e. we write T = yu. The easy case is y > 1/k'(y) where O(u. u Notes and references The importance sampling method was suggested by Siegmund [343] for discrete time random walks and further studied by Asmussen [ 13] in the setting of compound Poisson risk models .T' D U" . the references in Asmussen & Rubinstein [45] and Heidelberger [190]. T) with T < oo.'(y). so that one would expect the change of measure F 4 FL to produce close to optimal results.B'=B".T" > x) J /3"e0 yB (x + y) dy = .4. U' . The optimality result Theorem 3.T' = U" . Further discussion is in Lehtonen & Nyrhinen [245].3'eO'x f f P (U" . First by the memoryless distribution of the exponential distribution . BL) has bounded relative error. Consider compound Poisson risk process with intensities /3'.3"eQ x 0 J eQ zB (z) dz x (x > 0) and /3' = /3". with the present (shorter and more elementary) proof taken from Asmussen & Rubinstein [45]. CHAPTER X. from 3' P(U'T'>x) ^ = ^ eQ'zB (z) dz. generic interarrival times T' .i.
Let Qy2 = . one would expect that the change of measure F Pay is in some sense optimal. lim inf uoo 27yu .3) (simulated with parameters /gay.2 The estimator (4.4.(u) * 1 (Theorem IV.4. 7y (4. T(u) < yu] . (4.log Var(Z(u)) _ . yu) is of order of magnitude a71.2) Since the definition of ay is equivalent to Eay r(u) .1.1).yk(ay) determines the order of magnitude of z'(u. yu) = eayu Eay Leay^(u)+r(u)K(ay).1) which is all that is needed here can be showed much easier .8 has a stronger conclusion than (4. Proof Since ryy > y. (4. 3 Theorem IV. and in fact.1). yu)/z. T( u) < yu] e2ryyuEay le 2ay^(u). The corresponding estimator is Z(u) = eavS' ( u)+T(u)K (ay)I(T( u) < yu).yu. and that ryy > ry.4. IMPORTANCE SAMPLING FOR THE FINITE HORIZON CASE 291 Proof The assumption y > 1/n'(y) ensures that 1fi(u.yy> 2 . We recall that ay is defined as the solution of a'(a) = 1/y. Bounding u ELZ(u)2 above by a7u.log Var(Z(u)) l im of .'(7). that ryy = ay . We next consider the case y < 1/r. yu) in the sense that . we have ic(ay ) > 0 and get Eay Z(u)2 = Eay [e  2aySr( u)+2r(u )r. Bay) is logarithmically efficient.1) so that z(u) = zP(u.to g x ( u ) u u so that (1. Remark 4 . the result follows as in the proof of Theorem 3.8).3) and we have: Theorem 4.4. Further . T(u) < yu] e Hence by (4.O(u. .5) follows.(ay).log 4')u) 4 u (Theorem IV.
1) (see Proposition IV.2).b(u. yu . (5.a yu +l/ur' (av)Ei`av reav^(u)+(T(++)(U) yu . SIMULATION METHODOLOGY Vara„ (r(u))/u so that (T(u) .yu)/(uyu1/2) . 0 Notes and references The results of the present section are new.1) may be useful. > u) = E f I(VV > u) dt 0 (5.T) = P O<t<T inf Rt < 0 = P(VT > u). and (5. we believe that there are examples also in risk theory where (5. One main example is {Vt} being regenerative (see A.. Hence lira inf log ryyu + vyu 1/2 tc(ay) . the algorithm in Section 3 produces simulation estimates for the tail P(W > u) of the GI/G/1 waiting time W). N(0. Z (u)2 above.3.1) is used to study Voo by simulating {Rt} (for example. 0 rather than when u 3 oo. there exists a dual process { V t} such that i. yu . related discussion is given in a heavy traffic limit q J. Then z(u) = Eay Z(u) > Eay avS'(u)+T( u)k(av 1 ).u1/2 < r(u) < yu Le ] l = e.1) where the identity for Vi(u) requires that Vt has a limit in distribution V. In most of the simulation literature (say in queueing applications).yu1/2 <1 T(u) < yu l r > e7vu +avul/ 2r.4.a vt(u).4.o. zi(u) = INV.1): then by Proposition A1.Qyu1/2 < T(u) C yu e.o .292 CHAPTER X.4).7y x(u) > hm inf u+Oo U . However. '%(u) = P I info Rt < 0) = P(VV > u). 5 Regenerative simulation Our starting point is the duality representations in 11.(av)Eav l e. In Asmussen [13]. the object of interest is {Vt} rather than {Rt}.ryyu +oy u1/2K'(av)Eo l v 1/2) where the last step follows by Stam's lemma (Proposition IV.2) .uaoo U That lim sup < follows similarly but easier as when estimating En.3: for many risk processes {Rt }.
record Zi'i = (Z1'). which we survey below .d. z2)) > N(O. Z1 = (Zl1i +. Vh = (8h/8z1 8h/ 8z2). Simulate a zerodelayed version of {V t } until a large number N of cycles have been completed. EZ1'i = z1 = Ew.+Z(N) z 1. let E denote the 2 x 2 covariance matrix of Z('). + Z2N)) . the regenerative estimator z%(u) is consistent. For details .. Z2 .h (zl. For the ith cycle.. REGENERATIVE SIMULATION 293 where w is the generic cycle for {Vt}. Thus the method provides one answer on to how to avoid to simulate { Rt} for an infinitely long time period. Z2'> the time during the cycle where { Vt} exceeds u and zj = EZJ'). provides estimates for F ( V. j = 1. Therefore .E).. Z2 a4* z2. (u) ?2 = E fo I(Vt > u) dt = 0( u ) zl Ew as N > oo. Thus. letting J0 'o I (Vt > u) dt . Then Z(1).5... To derive confidence intervals . )).. z2) z2/z1 yields Vh = (z2/z2 1/zl).3) . a standard transformation technique (sometimes called the delta method) yields 1 V 2 (h (Zi. Taking h(zl. i (^(u) . Then (Z1z1i Z2z2 ) 4 N2(0. . oh) for h : R2 ^ R and Ch = VhEVh. is the cycle length. 02) (5.. .. 2.. Z(N) are i . > u) (and more general expectations Eg(V. Zl the LLN yields Z1 a$' Z(1) +. The method of regenerative simulation. EZ2'i = z2 = E Thus. Z2 = N (X21' + . + Z1N>) . consider first the case of independent cycles .t(u)) 4 N(0. i. and Z2'>) where Zi'i = w...
The regenerative method is not likely to be efficient for large u but rather a brute force one. with distribution depending on a parameter (. Here are the ideas of the two main appfoaches in today 's simulation literature: The score function ( SF) method . 9.Z) ^Z(=) . see e.g. Let X have a density f (x.C)dx = f w(x) d( f ( x. Notes and references The literature on regenerative simulation is extensive. () dx so that differentiation yields zS d( fco(x)f(x. In 111. v. There is potential also for combining with some variance reduction method. () dx f Ax) (dl d()f (x' () f ( z. in some situations it may be the only one resolving the infinite horizon problem .9.v. Rubinstein [310] and Rubinstein & Melamed [311]. We here consider simulation algorithms which have the potential of applying to substantially more complex situations.2.g S12 (5.294 where 01 2 CHAPTER X. Z of the form Z = ^p(X) where X is a r . asymptotic estimates were derived using the renewal equation for z /i(u). However . the expectation z = EZ of a single r. say risk processes with a complicated structure of the point process of claim arrivals and heavy tailed claims .z^ i=1 so a2 can be estimated by 2 2 = 72 S11+ 12 S22 .96s/v"N. () dx = E[SZ] f(X. to evaluate the sensitivity z/i( (u ) = (d/d() 0(u) where ( is some parameter governing the risk process . Then z(() = f cp(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 .0 .5) Z1 Z1 Z1 and the 95% confidence interval is z1 (u) ± 1. SIMULATION METHODOLOGY 2 Eli = Z2 z1 + 2 E22 . 6 Sensitivity analysis We return to the problem of 111 . Before going into the complications of ruin probabilities . () depending on C. consider an extremely simple example .
()) is 0 for C < Co and 1 for C > Co so that the sample path derivative cp'(h(U. SZ is an unbiased Monte Carlo estimator of z(. C)). = E [`d (h(U. for some Co = (o(U). () _ (eSx. () = . ()). Example 6 . For example . zc = E [d co(h(U. A related difficulty occurs in situations involving the Poisson number Nt of claims: also here the sample path derivative w.1).t. For IPA there are. () is an unbiased Monte Carlo estimator of zS. the Poisson rate /3 in the compound Poisson model. IPA will estimate zS by 0 which is obviously not correct. /3 is 0. To see this. however . ()) d( hc(U. ()) h((U. giving h( (U. ( where h( (u. For the SF method.r.() d( is the score function familiar from statistics . () = log U/(2. r(u) = Tl + • • • +TM(u)). () is increasing in C. one can take h (U. this is usually unproblematic and involves some application of dominated convergence .6. this phenomenon is particularly unpleasant since indicators occur widely in the CMC estimators . if f (x. just take cp as an indicator function .t. /3o is M(u) Oe (3T: < oo) . 11 /3oeOoT. So assume that a r. The following example demonstrates how the SF method handles this situation. C).v. The likelihood ratio up to r(u) for two Poisson processes with rates /3. with density f (x. Then z(() = Ecp(h(U. nonpathological examples where sample path derivatives fail to produce estimators with the correct expectation.r. The derivations of these two estimators is heuristic in that both use an interchange of expectation and differentiation that needs to be justified.log U/(. I(r(u) . Infinitesimal perturbation analysis (IPA) uses sample path derivatives. C) f(X. () Thus. () where U is uniform(0. In the setting of ruin probabilities . cp(h(U. Thus. SENSITIVITY ANALYSIS where 295 S = (d/d()f (X.1 Consider the sensitivity tka(u) w. () = (8/8()h (u. () = d log f (X. Thus . say W(x) = I(x > xo) and assume that h(U. p. one. ()) is 0 w . () can be generated as h(U. Let M(u) be the number of claims up to the time r(u) of ruin (thus. cp' (h(U. Then .
) we have VarL(ZQ(u)) ZO(u)2 O(u2)e2 u2e2ryu yu .9 .3L. We recall (Proposition 111.T(u)) I(T(u) < co) ] . Thus. However. since ELZp(u)2 < (M(U) _T(u) \ 1 2 a2ryu = O(u2)e27u.t. j3 and letting flo = 0. change the measure to FL as when simulating tp(u).3 (u) is of the order of magnitude ue7u. a relevant reference is VazquezAbad [374]. Example 6.r.296 CHAPTER X. In the setting of ruin probabilities. the estimation of z(ip(u) is subject to the same problem concerning relative precision as in rare events simulation . We then arrive at the estimator ZZ(u) = (M(u) . There have been much work on resolving the difficulties associated with IPA pointed out above.r. the risk process should be simulated with parameters .T(u)) e7uerVu) for ?P.0(1) so that in fact the estimator Zf(u) has bounded relative error. in part for different measures of risk than ruin probabilities. 0 Notes and references A survey of IPA and references is given by Glasserman [161] (see also Suri [358] for a tutorial). . for different models and for the sensitivities w.t. we get 1 M(u) 00(u) = E (_Ti)I(T(U)<) E [(M(u) .1 is from Asmussen & Rubinstein [46] who also work out a number of similar sensitivity estimators. BL). To resolve the infinite horizon problem . SIMULATION METHODOLOGY Taking expectation. different parameters. 4) that V5. differentiating w.3 (u) (to generate Zp (u). whereas for the SF method we refer to Rubinstein & Shapiro [312].
1}valued . }).d.. Y'a(U) = P(T (u) = r+(a)) = 1 . .(u) is defined as the probability of being ruined (starting from u) before the reserve reaches level a > u..... Consider first a Bernoulli random walk.g. '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 ).. a) = r(u)). Oa(U ) can also be a useful vehicle for computing t/i(u) by letting a * oo.. as e. either this makes no difference (P(R.i. in the Bernoulli random walk example below. with P(Xk = 1) = 9. X2.Chapter XI Miscellaneous topics 1 The ruin problem for Bernoulli random walk and Brownian motion. wherel T(u) = inf {t > 0 : Rt < 0} . where X1.P(•r(u. R„ = u+X. That is..+• • •+X.1. defined as Ro = u (with u E {0. T(u. 297 . The twobarrier ruin problem The twobarrier ruin probability 0. Besides its intrinsic interest . in most cases . are i... a) = r(u) A T+(a). and {1. T+(a) = inf It > 0 : Rt > al.(u) = 0 ) = 0) or it is trivial to translate from one setup to the other.
and insertion shows that ( 1..1) is solution. i..o)'t/1a(a .2) Oa(a . u + 1. We choose a = ry where ry is the Lundberg exponent.a(u)).. and in view of the discrete nature of a Bernoulli random walk we write z = e7. one elementary but difficult to generalize to other models.. z and the solution is z = (1 .. MISCELLANEOUS TOPICS Proposition 1. (1.y] = 1. zu = EzRO = EzRT(u. Proof 1. = (1 .1) o If 0 = 1/ 2.a) = 0) + zap ( R.0)/0.(4. Conditioning upon X1 yields immediately the recursion 'a(1) = 19+00a(2). = z°Va(u) + za(1  ... In a general random walk setting .e. C1_0\a.1 For a Bernoulli random walk with 0 0 1/2.1.o)T/la (1) + 8z/'u(3).(1B)u oJ 0. u Proof 2. Wald's exponential martingale is defined as in 11. By optional stopping.298 CHAPTER XI..(u) I\ e = 1 oa ' ()i a = u. the solution of F[.. where a is any number such that Ee°X = F[a] <oo.2).o» = z°P (RT ( u.r(u. then 'Oa(u) _ au a We give two proofs .4) by ea(u+Xl+.+Xn) F[ a]n n=0. The martingale is then {zuzXl+•••+X„ } = {zR° }. 7/la(a .1) = (19)4/'0(a3)+9ba(a1). and the other more advanced but applicable also in some other settings.a) Y. tba(2) _ (1 . The Lundberg equation becomes 1=F[ry]=(19)+9z.
1).zu)/(za .e7u)/(e7° ..1 If p = 0. RANDOM WALK.1) for p # 0. 1h (u) = a el u \1 If 9 < 1/ 2. TWO BARRIERS 299 and solving for 4/la(u) yields t/ia(u) = (za . thenz1 (u)=1.} is then itself a martingale and we get in a similar manner u = ER° = ER ra( u) = 0 • Y'a (u) + all  au Y'a( u)).1 yields 't/la(u) = (a . pa( u) _ u Corollary 1. .4 For a Brownian motion with drift u > 0. u Proposition 1.0a(u)). } yields e7u = Ee7R° = e°Wa(u) + e7a(1 . If p<0. then Vi(u) = 1. Applying optional stopping to the exponential martingale {e7R.u) = et(a2 /2 +aµ) the Lundberg equation is rye/2'yp = 0 with solution y = 2p. Proof Let a+ oo in (1. If p = 0.3 Let {Rt} be Brownian motion starting from u and with drift p and unit variance . Corollary 1.2) is trivial (z = 1). BROWNIAN MOTION. Then for p 0 0.2 For a Bernoulli random walk with 9 > 1/2. {Rt} is itself a martingale and just the same calculation as in the u proof of Proposition 1.u)/u. (1.• a2µa e2µu . {R. and solving for 9/la(u) yields Z/)a(u) = (e 76 .ba(u) = e2µa . If 9 = 1/2.5) . (1.1). However. then Proof Since 'Oa (U)  au a Eea(R°. i1(u) = e211 .1.
a) = a on {r (u.300 Proof Let a * oo in (1.e7a Again . letting a * oo yields the standard expression pe7u for .+^a(u))^(a) If 7k(a) < 1.616). and hence e7u = Ee7Ro E [e7R(.a) = a ) + e ' ° ( 1 .a) = r+ (a)} and similarly for the boundary 0.6 If the paths of {Rt} are upwards skipfree and 7//(a) < 1. CHAPTER XI.7) . VIII. 5). 1 . For most standard risk processes . Ic 5ry 'pa(u) Using y = 6 . 0. valid if p < 1 (otherwise .7)..7/la(u)). and thus one encounters the problem of controlling the undershoot under level 0. the paths are upwards skipfree but not downwards. Here is one more case where this is feasible: Example 1.0 (u) (where u p =. we obtain 'Oa a7u .5 Consider the compound Poisson model with exponential claims (with rate.5a). However. (1.4). however. a) I R(u a ) < 0] P (R(u . It may then be easier to first compute the onebarrier ruin probability O(u): Proposition 1. (u) _ O(u) . 7/'(u) = 1).vi(a) Proof By the upwards skipfree property.3. say.0(a) 0 < u < a. this immediately yields (1. 7O(u) = 7/la(u) + (1 . Here the undershoot under 0 is exponential with rate 5. .a) = a) = 5 y = P (R (u. MISCELLANEOUS TOPICS u The reason that the calculations work out so smoothly for Bernoulli random walks and Brownian motion is the skipfree nature of the paths.e7a (u) = 6 /0 .a ) < 0) + e 7aF ( R (u. implying R(u.a) < 0) + e7°P (R(u. passing to even more general cases the method quickly becomes unfeasible (see.
1a for computing ruin probabilities for a twostep premium function. (1. For µ # 0.µ so that {St} is Brownian motion with drift µ .8) Proof In terms of the claim surplus process { St} = {u .r(u).ST<u) = P(MT>u. of r(u) are ( U2 Pµ (T(u ) E dT) = 2^T 3/2 exp µu .T) P(MT > u) where MT = maxo<t<T St..11) VIT ) Proof For p = 0. and (1 . We now return to Bernoulli random walk and Brownian motion to consider finite horizon ruin probabilities.ST>U). MT > U) = P(ST > u) + P(ST > u) (1.2 . we have ili(u. P(MT > u) = P(ST > u) + P(ST < u. Hence P(MT>u. ( 1.10) Pµ (T(u) < T) !. (1..d. + µ2T) } .. 10) follows then by straightforward differentiation. (i).11 ) is the same as (1. Here {St } is Brownian motion with drift 0 (starting from 0). BROWNIAN MOTION. TWO BARRIERS 301 Note thas this argument has already been used in VII. RANDOM WALK.µ T I + e2µ"4) ( .f. For the symmetric (drift 0) case these are easily computable by means of the reflection principle: Proposition 1. the density dPµ / dP0 of St is eµsttµ2/2. Corollary 1.4) I = .1.8 ). 0(u.9) = 2P(ST > u).)_ _( u)µ2 /2.7 For Brownian motion with drift 0. 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 . T(u) E dT. Then the density and c. T ) = P(T(u) < T ) = 241.. = 1 .. and hence Pµ('r(u) E dT) = Eo [e µsr(. = eµuTµ2/2Po (T( u) E dT) 2 eµuTµ2/2 u T3/2 ex p u 27r p 12 T .Rt}. MT > u) = P (ST > u) + P (ST > u.8 Let {Rt} be Brownian motion with drift .µ%T (1.
9 For Bernoulli random walk with 9 = 1/2. Proof The argument leading to ( 1. close to x {Rt} behaves as Brownian motion with drift µ = u(x) and variance a2 = a2(x).T (1.T)dx. Vi(u. is (1.T+2.T) = P(ST = u) + 2P (ST > u). oo) with drift µ(x) and variance a2 (x) at x.11) then follows by checking that the derivative of the r. 0 0 (1.g.9).10) and that the value at 0 is 0.10).12) P(ST = v) = 0 otherwise. We finally consider a general diffusion {Rt} on [0.10 Consider a diffusion process {Rt} on [0.8 also applies to the case 9 54 1/2. u Small modifications also apply to Bernoulli random walks: Proposition 1. We assume that u(x) and a2 (x) are continuous with a2 (x) > 0 for x > 0. MISCELLANEOUS TOPICS which is the same as (1. is finite for all x > 0. (1. that 0(u).T2. Thus.. is zero for all u > 0 but that nevertheless Rt ^4 0 (the problem leads into the complicated area of boundary classification of diffusions.. such that the drift µ(x) and the variance a2(x) are continuous functions of x and that a2(x) > 0 ..g.9) goes through unchanged. The expression for F ( ST = v) is just a standard formula for the u binomial distribution. 226). e.13) with 0 as lower limit of integration. Let s(y) = ef0 ry(. the behaviour at the boundary 0 is more complicated and it may happen. S(oo) = f c s(y)dy. Breiman [78] or Karlin & Taylor [222] p. If this assumption fails. Here {2T( (v}TT)/2) v=T.3 we can define the local adjustment coefficient y(x) as the one 2µ(x)/a2(x) for the locally approximating Brownian motion. as defined in (1.302 CHAPTER XI. whenever u. The same argument as used for Corollary 1. and in a similar spirit as in VII.h. and (1. S(x) = f x s(y)dy.s.13) The following results gives a complete solution of the ruin problem for the diffusion subject to the assumption that S(x). but we omit the details.12) is the same as ( 1. see e. Theorem 1. oo). as defined above as the probability of actually hitting 0.. T are integervalued and nonnegative.
b(u) be the probability that {Rt} hits b before a starting from u.10. S(oo) < oo separately u completes the proof. In view of (1. where Lq(u) = 0'22u) q "(u) + p(u)q(u) is the differential operator associated with the diffusion.b (Rdt) = Oa.16) S(a) . Wa. 15) i.S(b) Proof Recall that under mild conditions on q. see Asmussen & Perry [42].S(u) (1. Letting b J.0(u) = 1 for all u > 0.b = 0 implies that VQ b/s is constant. Using s'/ s = 2p/a2. If b < u < a. Further references on twobarrier ruin problems include Dickson & Gray [116]. Letting a T oo and considering the cases S(oo) = oo. 191195 for material related to Theorem 1. 0 Proof of Theorem 1. 1'.b(u) + L. If (1.ba.11 Let 0 < b < u < a and let t&0.b(a) = 0 then yield the result. we can ignore the possibility of ruin or hitting the upper barrier a before dt. The obvious boundary conditions '0a. BROWNIAN MOTION. Then YIa. b = 0.b(u)dt.b = a+/3S.b(u) = S(a) .1. [117]. and we get Wo. . RANDOM WALK.b(b) = 1.b('u) = Eu &0. see in particular pp. then. A classical reference for further aspects of Bernoulli random walks is Feller [142].16) yields 4b (u) = 1 . i.13) is finite for all x > 0.17) Hence L.(u) < 1 for all u > 0 and ^ S^ Conversely.e. Notes and references All material of the present section is standard. TWO BARRIERS 303 for x > 0.14) fails. A good introduction to diffusions is in Karlin & Taylor [222]. O.16). E„ q(Rdt) = q(u)+Lq(u)dt. so that Y)n. Assume further that S (x) as defined in (1.S(u)/S(a). (1 .e LVa. elementary calculus shows that we can rewrite L as Lq(u) d 1a2 (u)s(u)d [ s (u) ? ] .10.14) S(oo) < 00. 0 in (1.b(Rdt). Lemma 1. For generalizations of Proposition 1.6 to Markovmodulated models .b('u) = Eu . then 0 < 2l. if (1.ba. the function S(x) is . (1.
5) A martingale proof of (2. They all use the fact that ( tx(a) l ( eaRt = eau + aSttx(a) < e7yu.2) C_e7u < t(u) < C+e _7u.)AT . Markovmodulated Brownian models .1. yu) where W (ay) = y. (2. which is motivated from the study of modern ATM (asynchronous transfer mode ) technology in telecommunications.(a) (2.1) (2.17). Lo is a martingale (cf. MISCELLANEOUS TOPICS referred to as the natural scale in the general theory of diffusions (in case of integrability problems at 0.4.3.3) < e 7yu.(.aR. (2. and here are alternative martingale proofs of the rest . y > . one works instead with a lower limit 5 > 0 of integration in (1. information on ruin probabilities can be obtained . See Asmussen [20] and Rogers [305] for some recent treatments and references to the vast literature.ytc (ay).t&(u.9 ) and optional stopping applied to the stopping time r(u) A T.2. Another basic quantity is the speed measure M . The emphasis is often on stationary distributions . defined by the density 1/va(u)s(u) showing up in (1.4.1 ) was given already in II. equivalently.4) I. variance 0.(T(u)AT) r.304 CHAPTER XI. but by duality. (2.. much of the literature dels with the pure drift case. IV. with the drift and the variance depending on an underlying Markov process . Lo I.5): _ z/'(u) < e 7u. where C_ = B(x) _ B(x) sup 2no fy° e7(Y )B(dy)' f2e7(Y2)B(dy)' C+ i/i(u. 7y = ay . 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. 1 y < k (y). is currently an extremely active area of research.5. 111 . Remark 11. yielding eau = Ee. (2.(7) .6) .o•K(a) = Ee .aRo . correponding to piecewise linear paths or . yu) '+/1(u) .13)).6.
T)  V.6) below by 1 E Le7Rr(. it follows easily from (2. For (2. u Proof of (2.4). (B(y) . Let H(dt.T(u)K(ay) I yu < r(u) < T] F(yu < r(u) < T) > e. dr) e 7( yr)B(dy) B(r) f oo o 0 r > H(dt. Rt has distribution B(r + dy)/B(r). (2. dr) denote the conditional distribution of (T(u).1 .1.3).)r(u)r.yu))• b(u.7R. dr JO Zoo ) f e7'B(r + dy) B(r) Jo ^00 ^00 H(dt. dr) 1 = 1 I0 /o C+ C+ From this the upper inequality follows.2): As noted in Proposition II. yu))• Letting T + oo yield e_ayu > eyur4ay)(0(u)  Notes and references See II.1. RT(u)_) given r(u) < oo. eyuk (ay) = e7yu e > eyu"(ay ) ij(u. Hence E [e7Rr (u) Jr(u) < ool ^00 H( dt.E [e. Proof of ( 2.d.(u.(ay)I T(u) < yu] P(r(u) < yu) (using RT(u) < 0).4): We take a = ay in (2.yu) Y Similarly for (2. A claim leading to ruin at time t has c. Equivalently. we have ic(ay ) < 0 and use the lower bound E [e7Rr („).6). y > r. we have tc(ay) > 0 and we can bound (2.B(r))/B(r).( u ) I T(U) < 00] . and the proof of the lower inequality is similar. . so that i/1(uL yu) < eayu .f.6) with = 'y that eyu ..yuk (ay)(u&(u.3). 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). when Rt_ = r.2.
1) but only the dominant exponential term . if x > EX1. e.1 We will go into some more detail concerning (3. However .2).g.3na with a < 1.. Example 3. its generality. gn with fn + 0 .306 CHAPTER XI. such that the cumulant generating function r. Accordingly.?n typically only give the dominant term in an asymptotic expression . the parameter will be u rather than n).. logarithmic asymptotics is usually much easier to derive than sharp asymptotics but also less informative . cle . nroo n n /// Note in particular that (3.nn or C2e. The classical result in the area is Cramer's theorem. however . og For sequences fn. v2 later.(B) = log EeOX 1 is defined for sufficiently many 0. in being capable of treating many models beyond simple random walks which are not easily treated by other models .means (as at other places in the book) that the ratio is one in the limit (here n * oo).1) amounts to the weaker statement lim 1 log P I Sn > x I = 17.1) is an example of sharp asymptotics : . and gave sharp asymptotics for probabilities of the form P (S.1) does not capture the \ in (3. Cramer considered a random walk Sn = X1 + . ri. and that a considerable body of theory has been developed. (3. The limit result (3. (3. we will write fn 1. For example.2) can be rewritten as F (Sn/n > x) 1g a'fin. . The last decades have seen a boom in the area and a considerable body of applications in queueing theory. not quite so much in insurance risk. large deviations results been. logarithmic asymptotics . MISCELLANEOUS TOPICS 3 Large deviations The area of large deviations is a set of asymptotic results on rare event probabilities and a set of methods to derive such results.. The advantage of the large deviations approach is. large deviations results have usually a weaker form. Thus. then P C S. + X. . which in the setting of (3.gn if nioo lim 109 fn = 1 log gn (later in this section. Thus .^ e nn 1 > x n 0o 2xn (3. gn 4 0./n E I) for intervals I C R..1).the correct sharp asymptotics might as well have +.1) where we return to the values of 0.
rc(0) where 0 = 0(x) is the solution of x = rc'(0). and hence for large n P(Sn/n > x) > E [e.3) is put equal to x. we get P(Sn/n > x) E [e9nx +nK(9)9" '.the mean rc'(0) of the distribution of X1 exponentially tilted with 0.4) n Next.sseo f which in conjunction with (3.q = rc* (x). the LegendreFenchel transform or just the Legendre transform or the large deviations rate function).(e)i XI E dx]. In fact. More precisely.96o /] > 0. we have P(nx < Sn < nx + 1. P with mean nx and variance no. since Sn is asymptotically normal w.4) immediately yields (3. replacing Sn in the exponent and ignoring the indicator yields the Chernoff bound P Sn > x 1 < e°n (3. exponential change of measure is a key tool in large deviations methods. of P(X1 E dx) = E[e9X1K. 2 where o2 = o2(x) = rc"(0).960/) * 0.tin f o') o e9o^y 1 1 ey2/2 dy 21r = etin 1 Bo 27rn . Most often.r. Define . i. (3. which is a saddlepoint equation .425.r.(0)) e 307 (other names are the entropy. if we replace Sn by nx + o / V where V is N(0.9S„+n' ( 9). nx < Sn < nx + 1.2). the sup in the definition of rc* can be evaluated by differentiation: rc*(x) = Ox .1). Since P nn > x) = E {e_8 ' ( 9).3. LARGE DEVIATIONS Define rc* as the convex conjugate of rc. S rtn > x 1.4 enn +1.t.e. V > 0 e. rc*(x) = sup(Ox .
which is a version of Cramer's theorem where independence is weakened to the existence of c(O) = limn.s. r(u) = inf {n : Sn > u} and o(u) = P('r(u) < oo). that is.'(u) )Ng a"u.. X2. n Icn(0) exists and is finite for ry .. We shall need: Lemma 3 .1). . . is differentiable at ry with 0 < K'(y) < 00./^ >7 < zn n for n n0. Xn) and sn = x1 + • • • + xn (note that the r. Then i/.e < 8 < y + e. For the proof... .dxn) where Fn is the distribution of (X1i . 1) and no such that Sn . and the WentzellFreidlin theory of slow Markov walks. Pn Sn1 .h.3. The substitution by V needs. integrates to 1 by the definition of Icn).3 For each i > 0.. be a sequence of r.. (ii) lim supn.2 (GLYNN & WHITT [163]) Let X1.e < 8 < y + e.. and write Sn = X1 + • • • + Xn.. we introduce a change of measure for X1... MISCELLANEOUS TOPICS which is the same as (3. Mogulskii's theorem which gives path asymptotics.. Xn given by Fn(dxl. Sanov's theorem which give rare events asymptotics for empirical distributions.o log Ee9Sn /n. Assume that there exists 'y. e > 0 such that (i) Kn (0) = log Ee°Sn is welldefined and finite for 'y .p > 7 < zn. In the application of large deviations to ruin probabilities. .. which is of similar spirit as the dicussion in VII. 1]. Further main results in large deviations theory are the GartnerEllis theorem. Ee9X n < oo for e < 0 < e.v. to be made rigorous..dxn) = 05nKn(7)Fn(dx1.. (iv) tc(ry) = 0 and r.308 CHAPTER XI. (iii) #c (8) = limn. there exists z E (0. 260 for details. asymptotics for probabilities of the form P ({S[nti/n}o<t<l E r) for a suitable set r of functions on [0. We further write µ = tc'(ry). we shall concentrate on a result which give asymptotics under conditions similar to the GartnerEllis theorem: Theorem 3 . however. see Jensen u [215] or [APQ] p.. .. commonly denoted as is the saddlepoint approximation.'s..
> . in particular the r. u Proof of Theorem 3. The corresponding claim for Pn(Sn/n < µ .s.h.m(7).91) + o(O ) as 0 J.n > u ) = [ Em [em Em 1e. S. log zl'(u)/u > 'y. Let r7 > 0 be given and let m = m(77) = [u(1 + 77)/µ] + 1. Clearly. for Sn. 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. This establishes the first claim of the lemma .. can be chosen strictly negative by taking 9 small enough.YS.s. The rest of the argument is as before..n e(µ +o)w"(7) [Eep(B +7)Sn]1 /p [EegoX. > 1 +17] m(7).2. is of order .+r7) < zn for n > no.s. ( U) P(S. can be chosen strictly negative by taking p close enough to 1 and 0 close enough to 0.Bµ .ne(µ limsup 1 log Pn (Sn/n > µ + 17) < ic(9 + ry) .. it is easy to see that the r.077 n^oo n and by Taylor expansion and (iv ).. mµ Sm > u] km e7Sm+n. LARGE DEVIATIONS Proof Let 0 < 9 < e where a is as in Theorem 3. P n(Sn/n > {c+77) < e no(µ 309 +n)Enees n +n)elcn(B +7). h.µ?7 .W.r (7) n = e. the r . Then V. We first show that lim inf„_. h. Since I EeqOX „ ] 1/q is bounded for large n by (ii). 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.+r.ne(p+ 17).3.n m µ 1 + rl ..]1/q = e. This proves the existence of z < 1 and no such that Pn (Sn/n > µ. 0.2.y) . S.Kn(7)e'n (p(O +7))/p I Ee geXn]1/q where we used Holder's inequality with 1/p+ 1/q = 1 and p chosen so close to 1 and 0 so close to 0 that j p(0 +.77) follows by symmetry (note that the argument did not use µ > 0).71 < e and jq9j < e.
(iv) and Lemma 3. For lim supu. I2 = F(T(u) = n). 14 = = E Lu(16)/aJ+1 Lu(1+6)/µJ+l = n) and n(S) is chosen such that icn('y )/n < 6 A (.n Yµ 1 + m + r ('Y) } U n \ 77 m µ µ7 1 < 1+ 77 ) Here E..+wn(7). MISCELLANEOUS TOPICS (7).. this is possible by (iii). n=1 .7) so that n(b) I1 < e'Yu E en. Pn \ > la+ 8 I < zn (3. 3.log z) /2 and Sn Fn\ n >lb+S) <Zn. 0 yields liminfu __..YS +^c CHAPTER XI. Obviously..3. n=1 n=n(b)+1 00 Lu(1 +6) /µJ 13 F( T (u) = P(T(u) n).(•) goes to 1 by Lemma 3.6) for some z < 1 and all n > n(E).(Y).. I > IL exp `S. Sn > u] < eYu+Kn(7)pn(Sn > u) (3. P(T(u) = n) < P(Sn > u) = En [e7S. we get lum inf z/i(u) 1 +12r7 >_ ry + 77 Letting r7 J.. logO(u)/u > ry. and since Ic.n(ry)/u 4 0andm/u* (1 + r7)/µ.310 ]Em I e. we write P(T(u) = n) = Il + I2 + I3 + I4 'i/I(u) _ E00 n=1 where n(b) Lu(10/µJ Ii = 1: F(T(u) = n)..I < µl1 1+77 I M 1_ 1+277 S.0 log i'(u )/u < 'y.
' 1 + b) n e7u x 1 /2 1 n x n / 2x (3. eryu en logz/2p n nt n. 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. > u) Lu(1+6) /µJ +l 00 )^n 'YSn+kn (7) .10) 00 I4 < E F(Sn_1 < u. 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 . Sn1 C U.3. Sn > U] [ e(u(1+6)/µJ+l < eYu (u(1+6)/µJ+1 7u r 0 0 e L^ en('Y ) fPn (I Sn 1 . u . S. LARGE DEVIATIONS Lu(16)/µJ 311 I2 < e"u n=n(6)+1 e'n(Y)P(Sn > u) < Lu(16)/µJ ^. C 26u `p / +1 I e6u(1+6)/µ (3.
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.b)/i(7).('+'Y). MISCELLANEOUS TOPICS The following corollary shows that given that ruin occurs. it suffices to show that for j = 1.(u) = I1+I2+I3+I4'^ ery( u). it holds for each b > 0 that 0(u) 1' g F(T(u) E (u(1 ..z 1/z For I1.u(1+b)/rc'(7)). 2. For I.2. Then for n large. u .9) can then be sharpened to x LQuJ /2 I2 < e7u 1 .4/3rc'(y) > 0.11 ) can be sharpened to x 4 [u(1+6)/µJ /2 1 . (7 + a) < 2arc'(7).4.3ui where . cf. we have rcn (a + 7) < 2n^c(7 + a) < 4narc' (7)..8) by P(S.7' a"ju.312 CHAPTER XI. I2. we replace the bound P(Sn > u ) < 1 used in (3. > u) < e"' E eIsn = ectueKn (a+'Y)Kn(7) where 0 < a < e and a is so small that r. For 14. 4 there is an aj > 0 and a cj < oo such that Ij < c3e. For 12. Corollary 3.4 Under the assumptions of Theorem 3. u(1 + b)/i(7)) Proof Since V.Q is so small that w = 1 . Letting c11 = maxn<n.xl/2 to give the desired conclusion. we need to redefine n(b) as L. IV. the typical time is u/rc'(7) just as for the compound Poisson model. this is straightforward since the last inequality in (3. 13 = P(T (u) E (u(1 b)l^ (7). say n n1. e'. ryue«iu . the last steps of (3.
13) One would expect this to hold in considerable generality.e.g. (iv) becomes existence of a limit tc(9) of tct(9) _ log Ee8S° It and a y > 0 with a(y) = 0. r..5 Assume the Xn form a stationary Gaussian sequence with mean p < 0. The problem is whether this is also the correct logarithmic asymptotics for the (larger) ruin probability O(u) of the whole process. Assuming that the further regularity conditions can be verified. and we conclude that Theorem 3 .3.. Theorem 3.2 shows that the discrete time structure is used in an essential way.. the key condition similar to (iii).v. t] is Rt = E V (Un) n: o„ <t . If {St}t> 0 is the claims surplus process. 11 Inspection of the proof of Theorem 3.f.1. whether P ( sup St > u ltg a ^" 0<t<oo // (3.2 then immediately yields the estimate log F( sup Skh > u) a7u (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.. we shall give two continuous time examples and tacitly assume that this can be done.3(s) at time s. 09(9). An event occuring at time s is rewarded by a r. and in fact. i. but nevertheless..(O) = 9µ+02 for all 9 E R.LARGE DEVIATIONS 313 Example 3 . Hence z z\ 2 z nrn(9) _ n Cn0p+BZn/ * .12) k=0. 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.. 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.14) is needed in both examples . Let {Nt}t>0 be a possibly inhomogeneous Poisson process with arrival rate .. Obviously many of the most interesting examples have a continuous time scale.'(y) > 0. Thus the total reward in the interval [0.1.. 2 is in force with y = 2p/wz. V(s) with m. criteria are given in Duffield & O'Connell [124]. Xk+l) k=1 00 naoo n provided the sum converges absolutely.
14). e.s).1) ds .1) ds rt (3. one would take p(t) = (1 + rt)At/ t. It is interesting and intuitively reasonable to note that the adjustment coefficient ry for the shot . then the payments from the company in [on. Most obviously. At = . if the nth claim arrives at time a. An apparent solution to this problem is to calculate the premium rate p = p(t) at time t based upon claims statistics .15) . the Cramer.314 where the an CHAPTER XI. = U„ ( t .2 are trivial to verify.1) ds .'`1 U. Then logEeOR° = J0 /3(s)(^8(9) .14) (to see this . 0 Example 3 .6 We assume that claims arrive according to a homogeneous Poisson process with intensity 0 .. derive . the best estimator of /3µB based upon Ft.Lundberg model has the larger ruin probability.It.d. Kt (0) t (Ee9U"it8i J0 . we have rct (9)/t 4 ic (9). Since the remaining conditions of Theorem 3. We let ic (9) = 3(EeWU° . this is not realistic . a differential equation in t). nondecreasing and with finite limits Un as s T oo ( thus. Of course. the CramerLundberg model implicitly assumes that the Poisson intensity /3 and the claim size distribution B (or at least its mean µB) are known. (9) < oo for 9 < 'y + C.noise model is the same as the one for the Cramer Lundberg model where a claim is immediately settled by the amount Un. O'n +S] is a r . n: o. We further assume that the processes {U1(s)}8>0 are i. . the above discussion of discrete skeletons). where Ft = a(A8 : 0 < s < t). e > 0 such that ic('y) = 0 and that r.g. Thus by (3.. Un(s). is At . If the nth claim arrives at time Qn = s. (3. we have S. assuming a continuous premium inflow at unit rate. 0 and since EeOUn(8) + Ee°U^ as s * oo. it contributes to St by the amount Un(t .9t = /3 J t (Ee8U° i8l . but that a claim is not settled immediately. More precisely. 7 Given the safety loading 77. i. Thus.9t. Un represents the total payment for the nth claim).v. we conclude that Cu) log e7 u (cf. MISCELLANEOUS TOPICS are the event times. Thus.t. Example 3.Q„) ..1) . leading to St = At(1+77) Joo t S8 ds..0 and assume there are y. Of course . <t which is a shotnoise process.
uniform (0./3.18) Thus (iii) of Theorem 3.2 hold. i.14) that rt _ 13 Jo _ (a [1_( i+77)log]) ds_flt = t (a) (3.20) (3. rewrite first rc as te(a) _ /3E 1 1 +(1+77)aUJ eau 1 . we conclude that t.3. It then follows from (3. standard exponential .e.i. one has y > y' (3.(1 +i) f > i= 1 s ds = E Ui 1 .d. (3. and since the remaining conditions are trivial to verify. Thus. equivalently.(1 + 17)0µB = 0. Ui Nt / t 01i 315 St = Ui .16) i=1 o i=1 Let ict (a) = log Eeast .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 . typically the adaptive premium rule leads to a ruin probability which is asymptotically smaller than for the CramerLundberg model .log Oi are i. the Vi = . 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. To see this .d.17) K(a) f o 1 O (a[I + (1 + 77) log u]) du )3.b(u) IN a'Yu (cf. 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 . we have Nt t N.(1 + r7) log t (3. LARGE DEVIATIONS With the Qi the arrival times.19) with equality if and only if U is degenerate. the solution of /3(Eelu .1) or .1) . (3.i. Indeed.
[257] and Nyrhinen [275]. say one year. In addition to Glynn & Whitt [163]. 4 The distribution of the aggregate claims We study the distribution of the aggregate claims A = ^N' U. and k(x) < 0.d. This is a topic of practical importance in the insurance business for assessing the probability of a great loss in a period of length t. the proof of (3.19). with common distribution B and independent of Nt. this in turn yields y > y*. .(1 + ri)y*x is convex with k(oo) = 00. the study is motivated from the formulas in IV.7. though we do not always spell this out.i. using that Ek(U) = 0 because of (3. MartinL6f [256]. a* (s) are convex with tc'(0) < 0 . are i. Further. k'(0) < 0. assuming that the U. MISCELLANEOUS TOPICS Next. see Nyrhinen [275] and Asmussen [25]. 11 Notes and references Some standard textbooks on large deviations are Bucklew [81]. rc*' (0 ) < 0. at time t. The main example is Nt being Poisson with rate fit.2 expressing the finite horizon ruin probabilities in terms of the distribution of A. = P(N = n) = e(3an However. [245]. we then take t = 1 so that p. This implies n(y*) < 0.1 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. the function k(x) = e7*x . For Example 3. we are interested in estimating P(A > x) for large x.2. much of the analysis carries over to more general cases. In particular.20) is due to Tatyana Turova. see also Nyrhinen [275] for Theorem 3.. Lehtonen & Nyrhinen [244].1 . For notational simplicity. Dembo & Zeitouni [105] and Shwartz & Weiss [339]. and since tc(s). k(0) = 0. 0 < x < x0. Therefore e7'U _ k(U) E [1+(1+77)y*U] . x > x0. Further applications of large deviations idea in risk theory occur in Djehiche [122]. y = y* can only occur if U .xo. so there exists a unique zero xo = xo(r7) > 0 such that k(x) > 0.316 CHAPTER XI.
1 Assume that lim8T8.4.e. only with 0 replaced by a9 and B by B9. 818' where s' = sup{s : B[s] < oo}.9(Ax).3B"[9]. we define the saddlepoint 9 = 9(x) by EBA = x.1) where )30 = .1).1. A E dx] . e9x+K(°) P(A > x) B 2ir /3 B" [9] Proof Since EBA = x. i. A > x)] = eex+K( e)E9 [e . This shows that the Pedistribution of A has a similar compound Poisson form as the Fdistribution.[s])3/2 = 0."(0) = . Then as x * oo.3B[9] and Be is the distribution given by eox B9(dx) = B [9] B(dx). THE DISTRIBUTION OF THE AGGREGATE CLAIMS 317 4a The saddlepoint approximation We impose the Poisson assumption (4. The analysis largely follows Example 3. Proposition 4. Hence P(A > x) = E e [e9A+ ic(9). B"[s] = oo. For a given x. In particular. 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 .3e(bo[a] . The exponential family generated by A is given by Pe(A E dx) = E [eeA K(9).2) implies that the limiting Pedistribution of (A . Then Ee"A = e'(") where x(a) _ 0(B[a] .ic(9) = . B"' [s] lim (B". K'(0) _ ic'(9) = x. Vare(A) = s. (4.1). no(a) = logE9e'A = rc(a + 9) .x)//3B"[9] is standard normal.
3) and related results u for the case of main interest . 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 . it is quite questionable to use (4. bounded with b(x) . or.4) .1 yields: Proposition 4.1). Thus .v. large x. The present proof is somewhat heuristical in the CLT steps.x') where x' = sup {x : b(x) > 0}.e.2) is often referred to as the Esscher approximation.EN B(x). under the Poisson assumption (4.1 goes all the way back to Esscher [141].(3µB)/(0µB^))1/2 has a limiting standard normal distribution as Q ^ oo. For details. more generally. A covers the exponential distribution and phasetype distributions. For example.2 If B is subexponential and EzN < oo for some z > 1. In particular. In fact. For a rigorous proof. 1 . 2).3) The result to be surveyed below improve upon this and related approximations by taking into account second order terms from the Edgeworth expansion. The (first order) Edgeworth expansion states that if the characteristic function g(u) = Ee"`}' of a r. Var(A) _ ^3p. then P(A > x) . Furthermore 00 b(x)Sdx < oo for some ( E (1. 4b The NP approximation In many cases . b is gammalike. it holds that EA = .l3pB.318 CHAPTER XI.(D X . B covers distributions with finite support or with a density not too far from ax° with a > 1.2i and that (A . b(x) = q(x)eh(z). i. just the same dominated convergence argument as in the proof of Theorem 2. b is logconcave. Remark 4 . Notes and references Proposition 4. where q(x) is bounded away from 0 and oo and h (x) is convex on an interval of the form [xo. [138]. Y satisfies 9(u) ti eu2/2(1 + ibu3) (4. For example. some regularity of the density b(x) of B is required. leading to P(A > x) :.Q{AB (4. the distribution of A is approximately normal . Jensen [215] and references therein. see Embrechts et al. and (4. MISCELLANEOUS TOPICS It should be noted that the heavytailed asymptotics is much more straightforward.ycix °ie6x B. either of the following is sufficient: A.
the NP (normal power) approximation deals with the quantile al_E. zl_e be the 1 .5) may be negative and is not necessarily an increasing function of y for jyj large. K3 = E(Y .1).i 6 r 1 3 so that we should take b = ic3/6 in (4. Let Y = (A .c2i.5). A particular case is a. and from this (4... (4. u5. which is often denoted VaR (the Value at Risk).2 2 . and so as a first approximation we obtain a1_E = EA + yle Var(A) . are small. resp.: EA + zl_E Var(A) .6) .. where Kl . as u2 u3 u4 9(u) = Ee'uY = exp {iuci .f. then P(Y < y) 4(y) ..5) is obtained by noting that by Fourier inversion.. so that 1(u) 3 exp { . the standard normal distribution.4.. the density of Y is 1 °° _ eiuy f(u) du 2x _..3& (y). defined as the the solution of P(A < yle) = 1 . .5 (y3 . s. Thus if EY = 0.EY)3..6(1 .l = EY. one expects the u3 term to dominate the terms of order u4. Remark 4.equantile in the distribution of Y. ylE should be close to zl_E (cf.EA)/ Var(A) and let yl_E. . Rather than with the tail probabilities F(A > x).5) follows by integration. however.h.99. K2 = Var (Y).s. Heuristically. K4 .2X2 . . f °o 9(y) = 1 e'uye u2/2(1 + iSu3) du 27r _ cc(y) .2 ^ \1 . one needs to show that 163. (4. Var(Y) = 1 as above . in particular. of (4.y2)^P(y)• 319 Note as a further warning that the r.e. THE DISTRIBUTION OF THE AGGREGATE CLAIMS where b is a small parameter. In concrete examples . are the cumulants . the CLT for Y = Y6 is usually derived via expanding the ch. If this holds .i 3 K3 } Pt^ exp .3!). If the distribution of Y is close to N(0..2K3 + 4i 64 + .
Using Y = (A ..E . We can rewrite (4. b such that EN 1 U%.E)A1 l E) 1 E 4)(yl E) ^' . 21 . For example. that [101] distinguishes between the NP and Edgeworth approximations.zlE)W(zlE) 1 .1)! n ^eQ .zlE )w(zl _E) = which combined with S = EY3/6 leads to q^ 1 Y1 .E = z1E + S(zi_E .zl E)^o(zl E) .6pBki) d/2.zl E)V(zl_E) . In particular .. as required .zi. this yields the NP approximation 6(Z1 _E .yi.1)EY3. n = 1.1)..(y) terms dominate the S(1 .3ni /3 . [101].S(1 . . however.1)^ 2) µ'E Notes and references We have followed largely Sundt [354]. K5 .S(1 .7) as 1 (3) a1E = Qµa +z1 . MISCELLANEOUS TOPICS A correction term may be computed from (4. k3 is small for large /3 but dominates 1c4.y2)cp( y) term..EA)3 a1_E = EA + z1_E(Var (A))1/2 + 1 Var(A) Under the Poisson assumption (4.E )Azl E) 4(z1E) + ( ylE .6 (1 ... b = /3 for the Poisson distribution with rate /3 since Pn = Pn1 n! n (n . the kth cumulant of A is /3PBk' and so s.. this holds with a = 0.E + (yl. Note. Another main reference is Daykin et at. let pn Pn = (a+ = P(N = n).1) E (A . and assume that there exist n ) Pn_i .EA ) / Var(A).E(/3PB^1 )1^2 + s(z1E .k = /3µB^1 / (.5(1 .320 CHAPTER XI.5) by noting that the 4. 4c Panjer 's recursion Consider A = constants a. This leads to t( yl E) .
THE DISTRIBUTION OF THE AGGREGATE CLAIMS 321 Proposition 4.4. then j (a + b!) 1ag k_1 3 gkfj.. Then fo = >20 9onpn and fi = 1 E In particular.4 Assume that B is concentrated on {0.4. j = 0. which would consist in noting that (in the case go = 0) fj = pn9jn n=1 (4. n. the value of (4. .13) Namely. fj = P(A = j). and calculating the gj*n recursively by 9*1 = 9j. .. (4... (4. . (4. The expression for fo is obvious.k . (4... Since the sum over i is na + b.12).14) is therefore a + b/n. By symmetry.9).5 The crux of Proposition 4. 2.12) where g*n is the nth convolution power of g. . ..4 is that the algorithm is much faster than the naive method.11) Remark 4.14) is independent of i = 1.13) but only O(j2) for Proposition 4. the complexity (number of arithmetic operations required) is O(j3) for (4. fj = E (a+ b k =1 )9kfi_k . n = k=n1 9k(n1 )9j k • (4.} and write gj = 2 .... if go = 0. j = 1. 2.4. . 1. E[a +bU=I >Ui =j l i=1 J (4. j1 g. u Proof of Proposition 4. .1.. . 2. j = 1. .12) we get for j > 0 that fj n a b + n p nlgj *n 00 U I n 1 *n = E a+bUi=j pn19j n=1 j i=1 CC) n Ui EE n=1 Ia +b Ul i=1 =j pn_1 . Hence by (4..10) f o = po.
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 N 1 = Eexp ry E[Ui ii . that 01[ry] < 0[y] where 0[y] = Ee'r(U^') .P I = EC [7]N.d.h * (X) . Eexp 7 [E [Ui .. this implies 7* > 7.6).4) and u g(x) = e7x in Ohlin's lemma.h(U)]. (6. Proof As in the proof of Proposition 6.d. REINSURANCE Proof Define N 329 h* (x) = E > h(Ui) X = x . N E X .5 Because of the independence assumptions .h(Ui)P JJJ l:='l {ry ] or.h(u) and any P satisfying (6. Applying the inequality Ecp(Y ) > EW(E (YIX )) (with W convex ) to W(y ) = eryy.h(U) (as in the proof of Proposition 6.6) u where C[ry] = Ee'r(u4(u)). The arrangement used in practice is.4).i. however.b)+ with b determined by E(U .4). u But since ry > 0.6.3). ry* > 0 because of (6. expectations like those in (6.5) holds trivially.b)+ = Eh(U) (and the same P) satisfies 71 > ry. then (6. Local reinsurance with h(u) = (u . (6.P]}.4.h(Ui)] .P } < 1 = Eexp E[Ui. and so on.b)+ is referred to as excessofloss reinsurance and plays a particular role: Proposition 6. .P.h(Ui)] .h(Ui)] .5) reduce quite a lot. Then for any local retention function u .h( UU) = EN • E[U . we get EX = EN • EU. Remark 6. This follows by taking Xl = U A b.4). i. it suffices to show that Eexp {ry ii 'UiAb. Assuming for simplicity that the Ui are i. ' ii (6. as often local as global.P > EexP{7[X .6 Assume the Ui are i. y = Ei [Ui . the excess ofloss rule hl (u) = (u . X2 = U . appealing to (6.
[76].many texts on insurance mathematics. See further Hesselager [194] and Dickson & Waters [120]. . e. Bowers et at. see also Sundt [354]. Heilman [191] and Sundt [354].330 CHAPTER XI. MISCELLANEOUS TOPICS Notes and references The theory exposed is standard and can be found in. The present proof is from van Dawen [99]. The original reference for Ohlin's lemma is Ohlin [277].g.
.. + U2 where U1 is a finite measure and U2(dt) = u(t)dt where 331 .T„_1).U(t) is the expected number of renewals in (t. t +a]). t 00 (A. denoted by F in the following and referred to as the interarrival distribution. 2h. The number max k : Tk_j < t of renewals in [0. That is. = T„ .. Y2. i.. The mathematical representation is either the ordered set 0 < To < T1 < ..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... all have the same distribution. If Yo = 0.t. of epochs or the set Y1.. when t is large. stating that U(t+a)U (t) ^ a. of interarrival times and the time Yo = To of the first arrival (that is.. Then Blackwell 's renewal theorem holds. note in particular that U({0}) = 1. t] is denoted by Nt.1) (here U(t) = U([0. Y. The associated renewal measure U is defined by U = u F*" where F*" is the nth convolution power of F. U(A) is the expected number of renewals in A C R in a zerodelayed renewal process. .e. Y2. The renewal theorem asserts that U(dt) is close to dt/µ. then Stone 's decomposition holds : U = U. .r. If F satisfies the stronger condition of being spreadout (F*' is nonsingular w . Lebesgue measure for some n > 1).. the renewal process is called zerodelayed. Technically. Y1. are independent and Y1. Lebesgue measure dt normalized by the mean to of F. . . not concentrated on {h. The point process is called a renewal process if Yo. t]) so that U(t + a) . some condition is needed: that F is nonlattice. the distribution of Yo is called the delay distribution..} for any h > 0.
Then Z(u) 4 z(oo). resp.2) Z(u) = J0 u z(x)U(dx).e.3) Further. z(u) a known function. stating that U(t)/t > 1/p. oo).i". Under weak regularity conditions (see [APQJ Ch. Equivalently.a. the asymptotic behavior of Z(u) is given by the key renewal theorem: Proposition A1.EN(t) .1 if F is nonlattice and z (u) is directly Riemann integrable (d. u u PF 4 00. that z(u) has a limit z(oo) (say) as u 4 oo.5) 2This condition can be weakened considerably . (A. U Z(u . in convolution notation Z = z + F * Z. then it suffices for (A. the statements being EN(t + a) .x)F(dx).2). (A. Both result are valid for delayed renewal processes. and F(dx) a known probability measure . z(x) = 0. see [APQ] Ch. then Z(u) i f0 z(x)dx . IV).out.4) that z is Lebesgue integrable with limZ.4) If F is spread.332 APPENDIX u(t) has limit 1/µ as t 4 oo.2) has the unique solution Z = U * z. (A.i. µF (A. In 111. (A. Note in particular that F is spreadout if F has a density f. wee shall need the following less standard parallel to the key renewal theorem: Proposition A1.9.R. and that F has a bounded density2. IV).2 Assume that Z solves the renewal equation (A.. A weaker (and much easier to prove) statement than Blackwell's renewal theorem is the elementary renewal theorem. i. 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 . but suffices for the present purposes .
asymptotic properties can easily be obtained from the key renewal equation by an exponential transformation also when F(dx) does not integrate to one. {Tn} if for any k. The property of independent cycles is equivalent to the postTk process {XTk+t}t>0 being independent of To. Y1 . Z(u) U = 1 u 1 u f z(u . 0 PF µF 11 In risk theory. we let µ denote its mean. . • . To this end.r. of Yo.3) satisfied by the ruin probability for the compound Poisson model.} be a renewal process. The kth cycle is defined as {XTk+t}o<t<Yk .(3.5a. Here the relevant F does not have mass one (F is defective). results from the case fo F(dx) = 1 can then be used to study Z and thereby Z. or many queueing processes..d. . F(dx) = e7xF(dx).. However. . Assuming that y can be chosen such that f °° Ox F(dx) = 1. however. T1.2) by e7x to obtain Z = z +P * Z where Z(x) = e'Y'Z(x). Note. multiply (A. this expression is to be interpreted as a random element of the space of all Evalued sequences with finite lifelengths. 1c Regenerative processes Let {T.. refer to the zerodelayed case. Hence by dominated convergence. The distribution F of Y1. Tk (or.. and its distribution does not depend on k. Yk ). A stochastic process {Xt}t>0 with a general state space E is called regenerative w.t))u(ut) dt 0 0 J f z(oo) • 1 dt = z(OO). cycles.APPENDIX 333 Proof The condition on F implies that U(dx) has a bounded density u(x) with limit 1/µF as x * oo. equivalently. i. This program has been carried out in III. this covers discrete Markov chains where we can take the Tn as the instants with Xt = i for some arbitrary but fixed state i. However. the postTk process {XT..i. where the Tn are the instants where a customer enters an empty system (then cycles = busy cycles). A regenerative process converges in distribution under very mild conditions: . Tk and {Xt }o<t<Tk • For example. a basic reason that renewal theory is relevant is the renewal equation II. The simplest case is when {Xt} has i...k+t }t>o is independent of To.. that the existence of y may fail for heavytailed F. is called the cycle length distribution and as before. Y2.e. that F is a probability measure. z(x) = e7xz(x). T1. the present more general definition is needed to deal with say Harris recurrent Markov chains. . . Eo etc.t.x)u(x) dx = z(u( 1 .. We let FO. .
where the distribution of X. P(C ( t) < a) 4 0 for any a < oo) and ij (t) * oo. and we have: holds more generally that (rl(t). r. are i.. e(t )) . under the condition of Blackwell's renewal theorem. fi (t) = inf {Tk .ZTOI < 00. in total variation. We denote the limiting r. Then {e(t)}. Otherwise ...e..tEU1/µ)/f has a limiting normal distribution with mean 0 and variance Var(Ui) + (!)2Var (Yi)_ 2EU1 Cov(U1.e. {Tn}... C). 2.ZT Then: (a) If E sup I ZTo+t . Then {Zt}t^...r. cycles (we allow a different distribution of the first cycle). {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. then Xt . i. {Tn} if the processes {ZT +t . (b) If in addition Var(Ul ) < oo. µ 0 If F is spreadout.v.3. Y1) le Residual and past lifetime Consider a renewal process and define e ( t) as the residual lifetime of the renewal interval straddling t.. {Tn}.r..3 Consider a regenerative process such that the cycle length distribution is nonlattice with p < oo.0 is called cumulative w. 0<t<Yi then Zt /t a$• EU1/µ. If p = oo. but in fact.+ X.ZT }0<t<Y„+.oo (i.d. [0. Then it (ii. (A.d. Then Xt Di X. resp . is given by Eg(Xoo) = 1 E0 f Ylg (Xt)dt. assume that p < 00 and define Un = ZT}1 . just the same proof as there carries over to show: Proposition A1.t. then e (t) .t : t < Tk}.i.334 APPENDIX Proposition A1.Tk : t < Tk} as the age. This is the case considered in [APQ] V. An example is Zt = fo f (X8) ds where {Xt} is regenerative w.r. oo).4 Let {Zt}t^. and q(t) = sup It .i.t. oo)..'s by e. for n = 1. then (Zt ..t.6) id Cumulative processes Let {Tn} be a renewal process with i.0 be cumulative w.
APPENDIX 335 Theorem A1.. U(x + 1) . Since the maximum Mn of n i.U(x) < U( 1)). Hence t t lt ) = f U(dy)z(t . Then fi(t)/t a4' 0 and. V is uniform on (0. ^ > y) = 1 f +Y (z)dz. use t E^(t)/t = E[Yo . .5 Under the condition of Blackwell's renewal theorem. 0 If Markov renewal theory By a Markov renewal process we understand a point process where the interarrival times Yo . and the equivalence of (a) with (b)(d) is an easy exercise.d. (c) the marginal distribution of q is FO.4. the first statement follows. Y1 > t] 4 0. Since z ( k) < E[Yi . and the conditional distribution of ri given l.6 Consider a renewal process with µ < oo. For the second. assume first the renewal process is zerodelayed. EC(t)/t + 0. Proof The number Nt of renewal before t satisfies Nt/t a4' p.d.y)P(Yo E dy) .^(t))} as a regenerative process.v. (1 V)W) where V.. Y1i Y2. but governed by a Markov chain {Jn} (we .(t).t. (b) the joint distribution of (ri. Then Eo^(t) satisfies a renewal equation with z(t) _ E[Y1 . and the conditional distribution of given 17 = y is the overshoot distribution R0(Y) given by FO(Y) (z) = Fo (y+z)/Fo(y). W are independent. Hence for t large enough.. if in addition EYo < oo. ^) is given by the following four equivalent statements: (a) P (77 > x. Yl > t]. 1) and W has distribution Fw given by dFw/dF(x) = x/pF.i.t. we can bound e(t) by M(t) = max {Yk : k < 2t/p}.y) = f U(t . Yo > 0] + f Eo^ (t .i.dy )z(y) < c ^ l z(k) Eoe(t 0 0 k=o where c = sup. In IV. = z is Foz) The proof of (a) is straightforward by viewing {(r. the joint distribution of (rl. In the general case. r. l:) is the same as the distribution of (VW.'s with finite mean satisfies Mn/n a$• 0 (BorelCantelli).U(x) (c < oo because it is easily seen that U(x + 1) . (d) the marginal distribution of ^ is FO. we used: Proposition A1. are not i. the sum is o(t) so that Eo£(t)/t + 0 .
r+ < oo).. T_=inf{n>0: Sn<0}. A Markov renewal process {Tn} contains an imbedded renewal process.jEE is a family of distributions on (0.) and (Fij )i. A stochastic process {Xt}t>o is called semiregenerative w.. distribution ofjXt}t>o itself where Pi refers to the case Jo = i. the conditional distribution of {XT„+t}t>o given Yo. .. . . be i. Further: Proposition A1. .. Assume that uj = EjYo < oo for all j and that {J„} is irreducible with stationary distribution (v3)jEE.t.and regenerative processes. These facts allow many definitions and results to be reduced to ordinary renewal.336 APPENDIX assume here that /the state space E is// finite) in the sense that P(Y. .. namely {Twk } where {Wk } is the sequence of instants w where Jo. G+(x) = P(S.r. J1 i . . .+ < x. the semiregenerative process is called nonlattice if {T... G_(x) = P(ST_ < x. Jo. IT.g.. and define r+=inf{n>0: Sn>0}. Y1. X2. Let X1.i .T_ < oo). the Markov renewal process if for any n. . We call r+ (T_) the strict ascending (weak descending) ladder epoch and G+ (G_) the corresponding ladder height distributions. Jn = i is the same as the P. . oo)..} is nonlattice (it is easily seen that this definition does not depend on i). Alsmeyer [5] and Thorisson [372]...r. < yIJ) = Fij( y) on {Jn= i.}. with common distribution F.. = io for some arbitrary but fixed reference state io E E. Yn. Jn +1=j} where J = a(JO.7 Consider a nonlattice semiregenerative process...t. . Then Xt 4 Xo. is given by Eg(X00) = 1 YO vjEj f g(Xt) dt µ jEE o where p = ujEEViAj. For example. oo). where the distribution of X. Jn_1. A2 WienerHopf factorization Let F be a distribution which is not concentrated on (oo. The semiregenerative process is then regenerative w. Notes and references Renewal theory and regenerative processes are treated.d. Sn = X1 + • • • + Xn the associated random walk. e.. 0] or (0 . in [APQ].
7) follows since G+(A) = 0 when A C (oo.=n w=m i Figure A.T_=n} = {S. we consider the last such time (to make w unique) so that {w=m.=EGn. (d) R+ = U_. oo). A C (oo. G+. Proof Considering the restrictions of measures to (oc. oo). A C (oo. >0. we may rewrite (a) as G_ (A) = G+(A) = F(A) + (G+ * G_)(A).x)R+(dx).7) (A. (A. oo) (A. F(A) + (G+ * G_)(A). m<j<n}. .and r_ preoccupation measures T+1 r_1 R+(A) = E E I(Sn E A).APPENDIX 337 Probabilistic WienerHopf theory deals with the relation between F. n=0 n=0 00 00 and the T+.8) (e. 0]. F(A) is the contribution from the event {T_ = 1} = {X1 < 0}.S.S.7).1 .r. (c) G+(A) = f °. u . define w as the time where the preT_ path S1.G+ * G_: (b) G_ (A) = f °° F(A ..g. . 0] and (0. n=0 The basic identities are the following: Theorem A2. More rigorously. F(A . S. n 0 R_(A) = E I(Sn E A). 0<j<m. U. 0). A C (0.. G_. the renewal measures U+=>G+.. A C (0.1 (a) F = G+ + G_ . On {T_ > 2}.>0. 0]). In (A.x)R_ (dx). (e) R_ = U+. Sr_ _1 is at its minimum .
0 < k < n.8) is similar. (b) follows from 00 G+ (A) _ E F(Sn E A.x)R+(dx)..= n.1._ = n . r+ = n) n=1 n=1 0  C0 E fF(Sk< 0. Aso.7) follows. and the proof of (A. m it follows (see Fig.Sn_1Edx. ST+Edu). A.3 8 APPENDIX Reversing the time points 0.. SnEAIS..1) that P(Sj Sn.F(r_n_mSrEA_u). .du) (G+ * G)(A)• C llecting terms.XnEAx) 00 f 0 f 0 00 00 1: F(A . . m < j <n. .1).+ E du) E P(S..u) f0m m=1 n=m+1 00 J0 OO P(S. E du) = P(T_=nm. Sr_ E Adu) (s ee again Fig . S.0<k<ri . Sn1 E dx) n=1  F(A .m.3. A._ E A) n1 f P(r_=nw=m Sm EduSrEA) m=1 n1 F(r+=mSr+Edu). It follows that for n > 2 F (7. m=1 f S mming over n = 2. SmEdu) = P(T+=m..+ E du)P(S.>0. ST_ E A . (A..x)P(Sk < 0. and reversing the order of summation yields P(T_ > 2. 0<j<m._ E A . clearly (Sj Sm>0.. ST_ E A) P(T+ = m. S..
oo). Then for A C (oo. .f.1). and using timereversion as in (d) to obtain the explicit form of R+ (Lebesgue measure).0<k<n. P(SnEA . and similarly H_ (s) = 1 . consider a fixed n and let Xk = Xn_k+l.O<k<n.9) whenever F[s]. there is no direct analogue of Theorem A2. being concentrated at 0. this holds always on the line its = 0. u Notes and references In its above discrete time version. it serves as model and motivation for a number of results and arguments in continuous time. In continuous time.0<k<n.O<k<n. cf.G_[s]) (A. However.F[s] = (1 . For (d).6. we can rewrite (a) as 1 . the analogue of a random walk is a process with stationary independent increments (a Levy process. Nevertheless.1. The present proof of Theorem A2. Again. 0]. The classical analytical form of the WienerHopf problem is to write 1 .SnEA) = P(Sn<Sk.P as a product H+H_ of functions with such properties. Since G+ is concentrated on (0. which is basic for the PollaczeckKhinchine formula. H+ (s) = 1G+[s] is defined and bounded in the halfplane Is : ERs < 0} and nonzero in Is: Rs < 01 (because IIG+lI _< 1). u Remark A2.1(a) is from Kennedy [228]. and G+.0+[s])(1 .T+> n) = P(Sk < O. is based upon representing G+ as in (b). WienerHopf theory is only used at a few places in this book. 11.. Sk = X1 + • • • + Xk = Sn . and sometimes in a larger strip. see for example Bingham [65]. a number of related identities can be derived.SnEA) is the probability that n is a weak descending ladder point with Sn E A. then T+ = inf It > 0 : St = 0} is 0 a.g.1.g.s. the derivation of the form of G+ for the compound Poisson model (Theorem 11.2 In terms of m. see e. if {St} is Brownian motion.. such developments motivate the approach in Chapter VI on the Markovian environment model. the survey [15] by the author and the extensive list of references there. For example. there are direct analogues of Theorem A2. 6+ [s].Sn_k. In discrete time.g.4). and the proof of (e) is similar.APPENDIX 339 and the proof of (c) is similar. Another main extension of the theory deals with Markov dependence. Summing over n yields R+ (A) = U_ (A).SnEA) = P(SnSn_ k. E. G_ [s] are defined at the same time.G_ [s] is defined and bounded in the halfplane is : ERs > 01 and nonzero in Is : ERs > 0}.SnEA) = P(Sn<Sk.'s. G_ are trivial. In this generality of.
one needs to compute matrix inverses Q1 and matrix exponentials eQt ( r just eQ ).1 (SCALING AND SQUARING) The difficulty in directly applying t e series expansion eQ = Eo Q"/n! arises when the elements of Q are large.5 that when handling phase type distributi ons.13) henever A is a diagonal matrix with all diagonal elements nonzero.340 APPENDIX 3 Matrixexponentials T e exponential eA of a p x p matrix A is defined by the usual series expansion 00 An eA n=0 n! he series is always convergent because A' = O(nk Ialn) for some integer k < p. Here are.11) A f eAtdt = eA. 0 . To circumvent this. It is seen from Theorem VIII. Here it is standard to compute matrixinverses by GaussJordan el imination with full pivoting .12) eA'AO = Ale AA (A. three of the c rrently most widely used ones: xample A3.10) d dteAt = AeAt = eAtA (A. 1. ere A is the eigenvalue of largest absolute value. write eQ = (eK)m where = Q/m for some suitable integer m (this is the scaling step). whereas there is no similar single established a proach in the case of matrix exponentials. if m is s fficiently large. Thus. however . Some fundamental properties are the following: sp(eA) = {e' : A E sp(A)} (A. Eo Kn/n! converges rapidly and can be evaluated without p oblems. and eQ can then be computed as the mth power (by squaring if = 2). _I 0 (A. JAI = max {Jjt : µ E sp(A)} and sp(A) is the set of all eigenvalues of A (the spectrum). 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).
. The probabilistic reason that (A. and we may consider a new Markov process {Xt} which has jumps governed by P and occuring at epochs of {Nt} only (note that since pii is typically nonzero . some jumps are dummy in the sense that no state transition occurs ).. Zo = a (Z = QZ. Ap.2 (UNIFORMIZATION) Formally.e. Let vi. One then can reduce to p linear differential equations by noting that k = ZQ. condition upon the number n of Poisson events in [Olt])  Example A3. However . Zo = h). the procedure consists in choosing some suitable i > 0. assume that Q is the intensity matrix for {Xt} and choose q with rt > max J%J = max qii• 1.3 (DIFFERENTIAL EQUATIONS) Letting Kt = eQt.3 i (A. what is needed is quite often only Zt = TreQt (or eQth) with it (h) a given row (column) vector. Here is a further method which appears quite appealing at a first sight: Example A3 . construction of {Xt} by realizing the jump times as a thinning of a Poisson process {Nt } with constant intensity 77.4 (DIAGONALIZATION) Assume that Q has diagonal form.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}. The approach is in particular convenient if one wants eQt for many different u values of t.14) holds is therefore that the tstep transition matrix for {fft} is eQt = E ent (. 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.. i.. .]t)n (A. p different eigenvalues Aj i .15) Then it is easily checked that P is a transition matrix .. To this end. we have k = QK (or KQ) which is a system of p2 linear differential equations which can be solved numerically by standard algorithms (say the RungeKutta method) subject to the boundary condition Ko = I.APPENDIX 341 Example A3. letting P = I + Q/i and truncating the series in the identity = e17t 00 Pn(.7t) n=0 n! u °O n Pn (to see this.e. In practice.. vp be the corresponding left . i.
and vihi ¢ 0. hi have been computed. The phenomenon occurs not least when the dimension p is large. we can take H as the matrix with columns hl. under the conditions of the PerronFrobenius theorem). hp the corresponding right (column) eigenvectors. In view of this phenomenon alone care should be taken when using diagonalization as a general tool for computing matrixexponentials. D = ) 2 2 .. i= 1 i=1 P P (A. Nevertheless. say Al. and hence A2 is so because of A2 = tr(Q). this last step is equivalent to finding a matrix H such that H1QH is a diagonal matrix. i=1 i=1 Thus.g H1. v5Q = Aivi.. Qhi = vihi.5 If Q= ( 411 ( q21 q12 q22 is 2 x 2. say A = (Ai)diag. i # j. of largest real part is often real (say. Example A3. and writing eQt as eQt = He°tH1 = H (e\it)di.. (A. hp. we have an explicit formula for eQt once the A j. vi..17) eQt = E e\`thivi = E ea:thi ® vi.18) contains terms which almost cancel and the loss of digits may be disasterous.18) Namely. There are.. however.. not all ai are real. (A. Complex calculus : Typically. and we need to have access to software permitting calculations with complex numbers or to perform the cumbersome translation into real and imaginary parts. two serious drawbacks of this approach: u Numerical instability : If the A5 are too close. Everything is nice and explicit here: 411+q2+D' )12_g11+q2^^ where (411422z + 4412421... the eigenvalue. Then vihj = 0.342 APPENDIX (row) eigenvectors and hl. some cases remain where diagonalization may still be appealing. Then P P Q = > Aihivi = E Aihi (9 vi. and we may adapt some normalization convention ensuring vihi = 1.16) (A.
However. eqt = eNlt ( ir1ki i2k1 \ ir1 k2 72 k2 + e azt 7r2k2 i2k1 7ri k2 7r1 k1 (A.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 .e. where (A. Then 7r = (ir1 7r2 ) = a (q21 Al . l ab (g12g21 + (A1  411) 2) = 1.6 A particular important case arises when Q = q1 qi ) q2 q2 J is an intensity matrix.q.20) ir = q2 ql qi +q 2 9l +q2 (A. it is easier to note that 7rh2 = 0 and v2k = 1 implies v2 = (k2 . 1) . v2 and h2 can be computed in just the same way. replacing ai by A2.19) Example A3 . i. k  C k2 ) =b ( A1 q 1 Q11 / where a .7 Let 3 9 2 14 7 11 2 2 . b are any constants ensuring//Irk = 1.Q2i and after some trivial calculus one gets eQt = 7r 1 112 + eat 7r1 7r2 / (7fl 7r2) = ( 7r2 1r2 7r1 IF.k1).21) Here the first term is the stationary limit and the second term thus describes the rate of convergence to stationarity. Of course. Then Al = 0 and the corresponding left and right eigenvectors are the stationary probability distribution 7r and e. h2 = Thus. u Example A3.
11/2 . APPENDIX x1 3/2 . for example AA+A = A. A+AA+ = A+.22) Note that in this generality it is not assumed that A is necessarily square. They are most often constructed by imposing some additional properties .satisfying AAA = A. (A+A)' = A+A.6. A2 = 3/2 . 2 2 1=ab(142+(1+2)2 ) = tab..344 Then D= 2+ 11)' 7 T4 2 =52. and a generalized inverse may not unique.5 . but only that dimensions match .23) ..11/2 + 5 1. (AA+)' = AA+. Generalized inverses play an important role in statistics. e_6u A4 Some linear algebra 4a Generalized inverses A generalized inverse of a matrix A is defined as any matrix 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_.
ew. and can define /ail 0 0 0 0 0 0 A+ = C A' 0 0 0 C' .1Q = Q(Q . and define D = (A .e. _ A. and exists and is unique (see for example Rao [300]).eir )1. Am+1 = . one is also faced with singular matrices .e ® 7r)1.APPENDIX 345 A matrix A+ satisfying (A.25) .g.D + O(ebt).g. are ordered such that Al > 0.23) is called the MoorePenrose inverse of A... = 0 where m < p is the rank of A.I) (A.P + e7r)1 (here ( I . if A is a possibly singular covariance matrix (nonnegative definite).1 goes under the name fundamental matrix of the Markov chain). Then for some b > 0.eir)1 = I ..= (I . (A. most often either an intensity matrix Q or a matrix of the form IP where P is a transition matrix. 0 01 In applied probability. . (I .. These matrices are not generalized inverses but act roughly as inverses except that 7r and e play a particular role . E. Rather than with generalized inverses .. one then works with Q = (Q .. . Assume that a unique stationary distribution w exists . Am > 0. ( Q . Here is a typical result on the role of such matrices in applied probability: Proposition A4.P).1 Let A be an irreducible intensity matrix with stationary row vector it.24) = te7r .P + e7r ).eir ). then there exists an orthogonal matrix C such that A = CDC' where 0 0 D = AP Here we can assume that the A . lt o eAx dx = te7r + D(eAt .
the r. the formulas involving O(e6t) follow by PerronFrobenius theory.91a(2) .eir)eAt = eAt = A'(t)..I)}.I) . Then A(O) _ B(O) = 0. h as 1 x m and k x 1 matrices. Interpreting 7r. I. resp.s. . Note that h ® it has rank 1.24). of (A. it follows that h ® it is the k x m matrix with ijth element hi7rj .I)} dx. Equivalently. the rows are proportional to it.J {xe^r + D(e . (A.h. (A. in block notation i2h A®B= ( a11B a21 B a12B a22 B Example A4.346 t APPENDIX 2 xe Ax dx = eir + t(D + e7r) + D(eAt .e.3 Let 2 A= 4 3 Vf' N7 5 )' B= ( 8 ). ()®(6 f 6/ 7f 8^ 7 8 )=! ^)( 6 7 8 )=(6^ 7^ 8^) \ u Example A4.s. B'(t) = e7r + DAeAt = eir + (I . and in fact any rank 1 matrix can be written on this form.27) Proof Let A(t).26) follows by integration by parts: t f t /' xeAx dx = [x {xe7r + D(eAx .h.2 Let it be a row vector with m components and h a column vector with k components.2e7r .I) (A.26) 2 = 2 e7r + tD . 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. 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 the columns to h. B(t) denote the l. see below.D + D2 + O(ebt).DZ(ent . respectively. . o Finally. For example.
Using (A.3V8.(A.50 6 7 6 4f 4.5v'8 5vf9 11 A fundamental formula is (A1B1C1) ®(A2B2C2) = (A1 (9 A2)(B1 (9 B2)(C1®C2). A2 = v2 are row vectors and C1 = h1. (AED B)1 = (A®I+I(9 B)l is the sum of all products of t factors.4vf.3v'6.4 eA® B = eA ®eB. C2 = h2 are column vectors. each of which is A ® I or I ® B.31). such a factor is Ak (&B 1k according to (A.k)! ( n0 n=0 t=0 k=0 J _ ® Ak ®Blk r ^. then the Kronecker sum is defined by A(1) ®A(2) = A(1) ®Ik2 + k ®A(2).APPENDIX 347 Then A®B = 2 f 20. (A.28) In particular.29) If A and B are both square (k1 = ml and k2 = m2).31) Indeed.3f 4v/. if Al = vi. 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). and v1B1h1 • v2B2h2 = v1B1h1 ® v2B2h2 = ( v1(&v2 )( B1(&B2 )( h1(&h2 ) .30) eA+B = eAeB function generalizes to Kronecker notation (note that in contrast typically only holds when A and B commute): Proposition A4.5v/. it follows that e® ® e B An _ 0o oo oo Bn 7 I F n! = ` k! (I . if A ® I occurs k times.3vV/72f 20. and the number of such factors is precisely given by the relevant binomial coefficient. (A B)' = eA®B e! L 1=0 0 . (A. then v1B1h1 and v2B2h2 are real numbers.A9.
resp . in the definition (A. Q(2). (A. A special case of Proposition A4. Ps 1) = exp {sQ ( 1) } > p(2 ) = exp {sQ(2) } can therefore be rewritten as Taking s = 1 for simplicity .3 < 0 Lemma A4 .32). and Q = Q(1) ® Q (2) = Q(1) ® I + I ® Q(2) (A. Let further it.32) is the intensity matrix of the bivariate continuous Markov process {Yt(1).33) . {Yt(1). first term on the r . v whenever a is an eigenvalue of A and 0 is an eigenvalue be any row vectors and h. the same time. Yt(2 ) }. and the form of the bivariate intensity matrix reflects the fact that Yt(2) } cannot change state in both components at due to independence . where transition matrix of the bivariate Markov chain {X n1). Let P8f P(Sl). h. 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 +. Yt(2) where independent Markov processes with intensity matri{y(2) } are {Y(1) }. { On the other hand. Thus . X ) }. P8 = exp {sQ} = exp {s (Q(1) ®Q(2)) } .5 Many of the concepts and results in Kronecker calculus have p(2) is the intuitive illustrations in probabilistic terms. Then 2 0 ire At h • ve Bt kdt = (^®v)(A®B)1(e A®Ba . p = P(1) ® {X }. From what has been said about matrices of {Yt( 1).s. P(2).I)(h ® k). the {Yt(2) } transitions in the {Yt(1) } component and the second transitions in the component . P(t) Yt(2) }. k any column vectors.348 APPENDIX Remark A4. n2 n1 ) {X(2) } are independent Markov chains with transition matrices P(1). we have P8 = Pal) ® p(2).4 can easily be obtained by probabilistic be the sstep transition reasoning along the same lines . { 1't(1) }. represents ces Q( 1). independent Markov chains.6 Suppose that A and of B.
= j and atk_li.. h can be chosen with 3By this. h such that vh = 1.. and the corresponding left and right eigenvectors v.. Then: (a) The spectral radius Ao = max{JAI : A E sp(A)} is itself a strictly positive and simple eigenvalue of A. . h = e and v = 7r (the stationary row vector). (b) if in addition A is aperiodic.8 Let B be an irreducible3 p x pmatrix with nonnegative offdiagonal elements. so that by asssumption A ® B is u invertible. . i. Then the eigenvalue Ao with largest real part is simple and real.1 and references there (to which we add Berman & Plemmons [63]): Theorem A4. and the corresponding left and right eigenvectors v. ao).34) Note that for a transition matrix.. That is. in such that io = i. then An = Aohv+O(µ") = Aoh®v+O(µ") for some u. we have AO = 1. E (0.3 whenever a is an eigenvalue of A and 3 is an eigenvalue of B. we mean that the pattern of nonzero offdiagonal elements is the same as for an irreducible intensity matrix. > 0 for k = 1. and if we normalize v. ..7 Let A be a p x pmatrix with nonnegative elements. . . Similarly. .12).. Here is the PerronFrobenius theorem.The PerronFrobenius theorem has an analogue for matrices B with properties similar to intensity matrices: Corollary A4. 4c The PerronFrobenius theorem Let A be a p x pmatrix with nonnegative elements. A is called aperiodic if the pattern of zero and nonzero elements is the same as for an aperiodic transition matrix. . h can be chosen with strictly positive elements.g. . . then IN < Ao for all A E sp(A). [APQ] X. We call A irreducible if the pattern of zero and nonzero elements is the same as for an irreducible transition matrix. f o r each i. Now note that the eigenvalues of A ® B are of the form a +. n. (A.APPENDIX 349 Proof According to (A. p there should exist io.29). the integrand can be written as ( 7r (9 v)( eAt ® eBt )(h ®k ) = ( 7r ®v)(eA (DBt)(h (& k). and appeal to (A. j = 1. il. which can be found in a great number of books. see e..
. h = e and v = 7r (the stationary row vector).8. it was shown that under mild conditions the tail of a phasetype distribution B is asymptotical exponential.1. T(°)) is asymptotically exponential with parameter t* _ r EiEE aiti as a 4 oo. h such that vh = 1. To this end. not only in the tail but in the whole distribution. Then for any (3. A5 Complements on phasetype distributions 5a Asymptotic exponentiality In Proposition VIII. Proposition A5. 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). Note that for an intensity matrix. Example A3. Corollary A4.(ti)ding. if we normalize v. Ao).35) for some p E (oo. I.8 is most often not stated explicitly in textbooks.(ti)diag where Q = T + (ti)diag is a proper intensity matrix (Qe = 0).(3. 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.2). relate the eigenvalues of B to those of B via (A. the condition is that t is small compared to Q. 10) and use the formula me at e Bt = e 00 Antn = e . then eBt = ea0thv + O(eµt) = eA0th ® v + O(et t) (A.1 Let Q be a proper irreducible intensity matrix with stationary distribution a. note that we can write the phase generator T as Q . let t = (ti)iEE # 0 have nonnegative entries and define T(°) = aQ . For example. let {Yti°i } be a Markov process with initial distribution a and intensity . the analogy of this procedure with unformization. Bi° (x) + at*x Proof Let { 4 } be the phase process associated with B(a) and (°) its lifelength. Furthermore. The next result gives a condition for asymptotical exponentiality.e. but is an easy consequence of the PerronFrobenius theorem.350 APPENDIX strictly positive elements. the phasetype distribution B(a) with representation (. we have A0 = 0.
Then a(a'V)/a (aV) a' 1. v/ t. Hence O ((a) aa. prove a somewhat more general result which was used in the proof of Proposition VI.9.Yj(av) = j f .bij) Hence the intensity matrix of { Ix} is (qij/ti)i.g. in fact . and that Yt(a) = Yat for all t.1. from which the phase process is terminated . we get dx F (Idx = j) = (1 + qij t )Sij + qij dt. 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}. and this easily yields a(x)/x a' 1/t*. it states that the state. dx/ti] or not.x (1 . fo tY dv/t a$' t*.)_ = Y(a) = 1'aS(a) = Ya(av)^ it follows that Pi ((.(a) > x . Conditioning upon whether { Yt} changes state in [0. We shall .a' + oo (e. from which it is easily checked that the limiting stationary distribution is (aiti/t*)iEE• Now let a' 4 oo with a in such a way that a' < a. a'/a + 1. a . and write Yt = Yt(1). a' = a . Proof Assume first ti > 0 for all i and let I. {t Y( a) } v>0 . We can assume that Jta) = Yt(°). t < (a).APPENDIX 351 ((1) etc. = YQ(x). By the law of large numbers for Markov processes .2 Pi (c(a) > x.jEE. Since JJ(. Let further V be exponential with intensity V and independent of everything else. Then {Ix} is a Markov process with to = Yo. In addition to the asymptotic exponentiality. J(()) _ = i) + at•x t tt' . has a limit distribution: Proposition A5. We can think of ( ( a) as the first event in an inhomogeneous Poisson process ( Cox process ) with intensity process matrix aQ . J^O)_ = j) Pi (v(aaV) > x.YQ(av) = j) Pi ( ci(a'V) > x.aE where 0 < e < 1).
. However. a). so is the geometric distribution. a) if B is the lifelength of a terminating Markov chain (in discrete time) on E which has transition matrix P = (p. See also Korolyuk. .. say bk = 0. (b) the generating function b[z] _ E' ..5 Let B be discrete phasetype with representation (P. Indeed.. Keilson [223]. let E and Pkj j=k1.. a = b = (bk)k=1. Et II I a(a^V) > x) at' . k>1..352 rr Ia(a'V) Ei I ( > x) P APPENDIX L at (Yo (aV) .+ at*x • a't' L ` at t* t* J Reducing the state space of {Ix } to {i E E : t.g. > 0}.1 and A5.Pe.2 do not appear to be in the literature.. an easy modification of the argument yields finally the result for the case where t.. 2. P. zkbk is za(I . . = 0 for one or more i. and thus the parameter p of the geometric distribution u can be identified with the exit probability vector p. so we shall be brief.} is said to be discrete phasetype with representation (E. A distribution B on {1. K}.j) and initial distribution a..3 As the exponential distribution is the simplest continuous phasetype distribution. 5b Discrete phasetype distributions The theory of discrete phasetype distributions is a close parallel of the continuous case. u Notes and references Propositions A5. Penev & Turbin [238]. Example A5. with point probabilities bk = (1 . (c) the nth moment k 1 k"bkis 1)"n!aP"p.p)k1 p.x k > K. the simplest discrete phasetype distribution: here E has only one element. ' pk 0 k>1 11 Theorem A5.zP)'p. k = 1. .. 1 k=1 1 0 otherwise. Example A5. these results are in the spirit of rare events theory for regenerative processes (e. 2. Then P is substochastic and the vector of exit probabilities is p = e . Then: (a) The point probabilities are bk = aPklp. Gnedenko & Kovalenko [164] and Glasserman & Kou [162]).4 Any discrete distribution B with finite support.. is discrete phasetype.
a' . Jt t > U1 + U2.a(2). resp. and hence the negative binomial distribution is discrete phaseu type. 11 Example A5. initial distribution a and phase generator T. A. T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2).6 is the Erlang distribution Er which is the convolution of r exponential distributions. T= ( 0 T(2) ) (A. and a=1). A reduced phase diagram (omitting transitions within the two blocks) is am E(1) t(1) a(2) (2) t(2) Figure A.36) in blockpartitioned notation (where we could also write a as (a (1) 0))...6. B2 be phasetype with representations (E(1).T(1)). as is seen by minor modifications of Example A5. Then {Jt} has lifetime U1 + U2 .6 (CONVOLUTIONS) Let B1.T(2)). a. { Jt 2) } with lifetimes U1 .a(1).7 (THE NEGATIVE BINOMIAL DISTRIBUTION) The most trivial special case of Example A5. _ i E E(1) T(1) t(1)a(2) i E E(2) . The discrete counterpart is the negative binomial distribution with point probabilities bk k1) (1 k = r. U2.APPENDIX 353 5c Closure properties Example A5.{ 0. Then the convolution B = B1 * B2 is phasetype with representation (E.r + 1. and piece the processes together by it = 41) 0<t<U1 U1 < t < U1 + U2 2U. (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) }. resp. r .
.a(2).37) (1) (1 .. Example A5.0)ai2).0)a(2) E(2) Figure A. resp.3 In exactly the same way. (E(2).10 (GEOMETRIC COMPOUNDS) Let B be phasetype with representation (E. a reduced phase diagram is f a E t Figure A..4 . a. Then it is trivial to see that B(") is u phasetype with representation (a("). T) where E = E(1) + E(2) is the disjoint union of E(1) and E(2).0)a(2))). and o'i Oa. we need to restart the phase process for B w. one obvious interpretation of the claim u size distribution B to be a mixture is several types of claims. then C is the distribution of Ul + • • • + UN. if U1. P(N = n) = (1 . Let B(") be the corresponding phasetype distribution. Then the mixture B = 9B1 + (1 .p)pn1B*n. In risk theory. Equivalently.O)B2 (0 < 0 < 1) is phasetype with representation (E. this means that a = (Oa(1) (1 .p)pn1. U2. Thus. a mixture of more than two phasetype distributions is seen to be phasetype.354 APPENDIX Example A5.i. B2 be phasetype with representations (E(1). a. are i.T(2)).a(1). p at each termination.d.p. To obtain a phase process for C. i E E(1) T 0 I (A.E) where a(°) = fAa(a)v(da). i E E(2) 0 T(2) =IT (in blockpartitioned notation. with common distribution and N is independent of the Uk and geometrically distributed with parameter p.T. Example A5.9 (INFINITE MIXTURES WITH T FIXED) Assume that a = a(°) depends on a parameter a E A whereas E and T are the same for all a. and consider B(") = fA B(a) v(da) where v is a probability measure on A.'). T) and C = EO°_1(1 .8 (FINITE MIXTURES) Let B1.T(1)). A reduced phase diagram is 0a(1) E(1) A .
let {Jtl)}.x)+.T) where F[T] = J0 "o eTx F(dx) u is the matrix m.2. Then the minimum U1 A U2 and the maximum U1 V U2 are again phasetype..TWWW).. Example A5 . say with distribution F. if B is defective and N + 1 is the first n with U„ = oo.11 (OVERSHOOTS) The overshoot of U over x is defined as the distribution of (U . say v. B2 of phasetype with representations (E('). j E F}. then U1 +• is phasetype with representation (E.. a. X independent of U. with common distribution B and N is independent of the Uk with P(N = n) = f. v. T(1) ® T(2)). we then let the governing phase process be {Jt} _ {(411 Jt2))} 2) interpreting exit of either of {4 M }. If U1 has a different initial vector.. Proposition VIII. { 4 } as exit of {Jt}.T + pta).aF[T]. are i.f. Example A5 . f2.g. but the same T. T) and C = F.°_1 f„ B*?l. Equivalently. Example A5.1. be the point probabilities of a discrete phasetype distribution with representation (E. of F. 12 (PHASETYPE COMPOUNDS ) Let fl. It is zeromodified phasetype with representation (E. (E(2). let the phase space be E x F = {i j : i E E. Indeed. Minor modifications of the argument show that 1. 13 (MINIMA AND MAXIMA ) Let U1.7. then C is the distribution of U1 + • • • + UN. For U1 A U2.2.d. then U1 + • • + UN is zeromodified phasetype with representation (a.X)+ is zeromodified phasetype with representation (E.. .°. T(2) ). +UN 2. if U1. a. To obtain a phase representation for C . then Jy has distribution aeTx. let the initial vector be a ® v and u let the phase generator be I ® T + P ® (ta). resp.. v. { Jt2) } be independent with lifetimes U1. U2 be random variables with distributions B1. To see this.v. i.9) that (U . T). U2. P). if {Jt} is a phase process for U. Note that this was exactly the structure of the lifetime of a terminating renewal u process. let B be a continuous phasetype distribution with representation (F.. If we replace x by a r.APPENDIX 355 and C is phasetype with representation (E.aeTx.T) if U is phasetype with representation (E. it follows by mixing (Example A5. U2. Thus the representation is (E(1) x E(2). a.a(1). a(2). . resp. a(1) ® a(2 ). E). Corollary VIII. T + ta.. cf. T + pta). cf.
2) } to go on (on E(2)) when { i 1) } exits.356 APPENDIX For U1 V U2. we need to allow { Jt. Now we can find first a sequence {Dm} of distributions with finite support such that D. oo). By the diagonal argument (subsequent thinnings).} of phasetype distributions such that Bn 3 B as n + oo. Then from above.14 To a given distribution B on (0.8... i= 1 C. relies more on matrix algebra than the probabilistic interpretation exploited here).. r # oo.(Sn) with Sn = n/b. and let Bn be the Erlang distribution E. Let the support of Dn be {xl(n). there is a sequence {B. with weight pi(n) for xi(n). and vice versa. see Neuts [269] (where the proof.(n) = D. however. we can assume that ID.(bk) + B(bk) for all k as n * oo.(bk)'. q(n) q(n) pi(n)a .. That is. cf. Example A5. and the closedness of the class of phasetype distributions under the formation of finite mixtures. oo) can be approximated 'arbitrarily close' by a phasetype distribution B: Theorem A5. The mean of B„ is n/Sn = b and the variance is n/Sn = b2/n.. The general case now follows easily from this.B(bk) I < 1/n for n > k.(bk) + B(bk) for all k. any distribution B on (0. say degenerate at b. Here are the details at two somewhat different levels of abstraction: (diagonal argument . 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). the fact that any distribution B can be approximated arbitrarily close by a distribution with finite support.. Then we must find phasetype distributions Bn with B. Hence it is immediate that Bn 4 B. the initial vector is (a(1) (& a (2) 0 0). 5d Phasetype approximation A fundamental property of phasetype distributions is denseness . elementary) Let {bk} be any dense sequence of continuity points for B(x).n = I:pi(n)Er v ( __ ) n) ) a= 1 ..xq(n)(n)}.. Proof Assume first that B is a onepoint distribution.
i = 1. compute W(B) and use this quantity as an approximation to cp(B0). for some a < oo. however. Let E be the class of functions f : [0. Hence G C PET and L = PIT. say on the claim size distribution B in risk theory. In particular.n (bk) . It should be noted. the class CO of all discrete distributions. oo) * [0.(x)Bf. one would use the B given by some statistical fitting procedure (see below). E E.. we can then approximate Bo by a phasetype B. f2.. in at least two ways: insensitivity Suppose we are able to verify a specific result when B is of phasetype say that two functionals Cpl (B) and W2 (B) coincide. there is a sequence {Bn} of phase type distributions such that Bn Di B as n 4 oo and f ' f..i. For a general Bo. u 2 (abstract topological ) The essence of the argument above is that the closure (w. oo) such that f (x) = O(e«x). u Theorem A5. But To is the class G of all distributions on [0.e. k < n. then it is immediate that WI(B) = p2(B) for all distributions B on [0. If Cpl (B) and ^02(B) are weakly continuous.d. k < n. that this procedure should be used with care if ^p(B) is the ruin probability O(u) and u is large. the topology for weak convergence) PET of the class PET of phasetype distributions contains all onepoint distributions.n( b k ) . oo) and any fl.t. 2.. Then ICr( n ). if information on Bo is given in terms of observations (i.n. oo). Corollary A5. oo) approximation Assume that we can compute a functional W(B) when B is phasetype. i. and that cp is known to be continuous.15 To a given distribution B on (0 ..r. Since PET is closed under the continuous operation of formation of finite mixtures.. .14 is fundamental and can motivate phasetype assumptions.D(bk)I < n.( dx) * f r f{(x)B(dx).. PIT contains all finite mixtures of onepoint distributions. replications). x 4 oo.B(bk )I < .APPENDIX 357 Hence we can choose r(n) in such a way that ICr( n). and we can take Bn = Cr(n)..
there is a sequence {Bn} of phase type distributions such that Bn Di B as n + oo and all moments converge. n B=az.2 ..f (z) = f = 1 1 1 1n/ o .f ' f (x)B(dx).. TO (A. and hence it is sufficient to show that we can obtain limsup n4oo fi(x)Bn(dx) < Jo 0 f fi( x)B(dx ). .... we may assume that in the proof of Theorem A5.(dx) > J fi(x)B(dx). if f (x ) = e°x. 2. for each i.14 Dn has been chosen such that 00 1 °° f fi(x)D n(dx ) < 1++ '  o \ n o f fi(x)B(dx). By (A. i = 1.358 Proof By Fatou' s lemma.oo J fi(x)B.39) Indeed. and the case of a general f then follows from the definition of the class E and a uniform integrability argument. i=1. . .16 To a given distribution B on (0 . . n.. \\ 0 Corollary A5.  APPENDIX B implies that 00 o o 00 n. liminf B. i = 1.. f00 fi(x)Cr. Bn=En z f f (x)Bn(dx) fof (x)B(dx) = ° (A.. ..n(dx) < 1+. i = 1.. and hence we may choose r(n) such that L 9l) f (x)Cr(n).38 ). oo). Now returning to the proof of (A.38) We first show that for each f E E. f° xtBn(dx ) * f °° x`B( dx). then cc f (x)Bn ( dx) = (?!c ) e'= ...n(dx) + f 0 fi(x)Dn(dx).39).f (x)B(dx). n..
oo) with B[y +e] < oo for some e > y = 7(B. .17 To a given /3 > 0 and a given distribution B on (0. (N or a given distribution Bo. This is motivated in part from the fact that a number of nonphasetype distributions like the lognormal. . and therefore the following result is highly relevant as support for phasetype assumptions in risk theory: Corollary A5. The present section is a survey of some of the available approaches and software for inplementing this. 5e Phasetype fitting As has been mentioned a number of times already. there is substantial advantage in assuming the claim sizes to be phasetype when one wants to compute ruin probabilities. If ei > 0.. from a more conceptual .3). .16. For practical purposes. .e. lim inf > is proved similarly./3). Proof Let fi(x) = el'r+E. one can obtain 7(Bn. the problem thus arises of how to fit a phasetype distribution B to a given set of data (1. then Bn['Y + ei] * B[y + ei] > 1 + 7 Q implies that 'yn < ry + ei for all sufficiently large n . I. (N. However.l3µb < 1.14 is classical.APPENDIX 359 In compound Poisson risk processes with arrival intensity /3 and claim size distribution B satisfying ./3) is defined as the unique solution > 0 of B[y] = l+y/j3. . O We state without proof the following result: Corollary A5. the adjustment coefficient 'y = 7(B. .18 In the setting of Corollary A5. but are certainly not unexpected. /3) = ry for all n. The adjustment coefficient is a fundamental quantity. e ) and ei J. We shall formulate the problem in the slightly broader setting of fitting a phasetype distribution B to a given set of data (1i . . the loggamma or the Weibull have been argued to provide adequate descriptions of claim size distributions. lim sup ryn < 7. Notes and references Theorem A5. 0 as i * oo.> y for some sequence {ei} with ei E (0.. the remaining results may be slightly stronger than those given in the literature. there is a sequence {B.. and in part from the fact that many of the algorithms that we describe below have been formulated within the setup of fitting distributions.} of phasetype distributions such that Bfz + B as n * oo and Yn 4 ry where ryn = y(Bn.
[70]) restrict attention to acyclic phase type distributions ..f.. one could argue that the results of the preceding section concerning phasetype approximation contains a solution to our problem : given Bo (or Be). giving mass 1 /N to each S=. a program package written in C for the SUN workstation or the PC is available as shareware.. It seems therefore a key issue to develop methods allowing for a more general phase diagram. The earliest such reference is Bux & Herzog [85] who assumed that the Erlang distributions have the same rate parameter. and as fitted distribution we may take B. defined by the absence of loops in the phase diagram .} of phasetype distribution such that Bo.g. .360 APPENDIX point of view the two sets of problems are hardly different : an equivalent representation of a set of data (1 . A number of approaches restrict the phase type distribution to a suitable class of mixtures of Erlang distributions .g.d. d. we do not not want to perform matrix calculus in hundreds or thousands dimensions). [317] ) has considered an extension of this setup.g . Schmickler (the MEDA package. (N is the empirical distribution Be. risk theory. and in practice this sets a limitation to the usefulness (the curse of dimensionality . we have constructed a sequence { B. g. and used a nonlinear programming approach . reliability or queueing theory. The observation is that the statistical problem would be straightforward if the whole ( EAvalued) phase process { Jtk)} o<t<( k associated with each observa . cf..f. [202]. for some suitable large n. at a a number of selected points . The likelihood function is maximized by a local linearization method allowing to use linear programming techniques. 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. B„ The problem is that the constructions of {B„} are not economical : the number of phases grows rapidly. . Asmussen & Nerman [38] implemented maximum likelihood in the full class of phasetype distributions via the EM algorithm . 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.'s). and this is what matters when using phasetype distributions as computational vehicle in say renewal theory. A method developed by Bobbio and coworkers (see e.g. Johnson & Taaffe considered a mixture of two Erlangs (with different rates ) and matched (when possible ) the first three moments . and we next describe two such approaches which also have the feature of being based upon the traditional statistical tool of like maximum likelihood. In a series of papers (e. Of course. The characteristics of all of these methods is that even the number of parameters may be low (e. the number of phases required for a good fit will typically be much larger. e . three for a mixture of two Erlangs ). the L1 distance between the c . [216] ).
.(k] (Ti is the total time spent in state i and Nii is the total number of jumps from i to j)..T(n) k=1 I (Jti) dt o \f a(n)eT(n )(kt(n) N f:i a(n)eT(n)xei .APPENDIX 361 tion Sk was available.. . . = j) f k=1 k =1 tE[0.T(n) (Ti ^^ 1. Nii = = .g.x)t(n) 1 and this and similar expressions are then computed by numerical solution of a set of differential equations.. E.T (n)(TiI(1. (N) tJk Ea ( n).(N) = E Ea(n).. . it is easy to see that N (k Ea(n).. eieT(n)((k. The general idea of the EM algorithm ([106]) is to replace such unobserved quantities by the conditional expectation given the observations.g. e. (n+1) _ Ea (n). N Ti = I(J= i) dt. it seems open whether the restriction to the acyclic case is a severe loss of generality. (N ) (^ 54 k )+ and similarly for the cn+1) The crux is the computation of the conditional expectations. In fact. EN where ai = N 1 I ((k) = i) tii=i iEE. jEEA. Thus. the methods of [70] and [38] appear to produce almost identical results. . one is lead to an iterative scheme. then the estimators would be of simple occurenceexposure type. since this is parameterdependent. .T(n) (Nik IC1.... In practice.
This page is intentionally left blank .
3157 . Alsmeyer (1991) Erneuerungstheorie. [13] S. New York. [15] S. Sparre Andersen (1957) On the collective theory of risk in the case of contagion between the claims . Abate & W. pp. ibid. New York. Appl. Choudhury & W. 1984 . Whitt (1998) Explicit M/G/1 waitingtime distributions for a class of longtail service time distributions . Aktuar Tidskr. Arfwedson ( 1955) Research in collective risk theory. Act. 20. [10] K. AT&T. New York. Queueing Systems 10.). Skand. 37. [6] V. Teubner. Transactions XVth International Congress of Actuaries. Dover. G. [14] S.L. 311338. Chichester New York. 253267. 191223. 38. Borokov. Whitt ( 1992) The Fourierseries method for inverting transforms of probability distributions . Asmussen (1984) Approximations for the probability of ruin within finite time . Arndt (1984) On the distribution of the supremum of a random walk on a Markov chain . Abramowitz & I. 587. [11) S. 57. [2] J. [12] S. 219229. Stoch. Skand. 363 . Aktuar Tidskr. [4] M. The case of equal risk sums. 213229. Probab. Scand. Optimizations Software.Bibliography [1] J. Queueing Systems 5 . Preprint. 1985. Asmussen ( 1982 ) Conditioned limit theorems relating a random walk to its associate. Proc. J. 14.A. II. Asmussen (1989a) Aspects of matrix WienerHopf factorisation in applied probability. Abate & W. Stuttgart.G. Anantharam (1988 ) How large delays build up in a GI/GI11 queue. Adv. In: Limit Theorems and Related Problems (A. Asmussen (1985) Conjugate processes and the simulation of ruin problems. B. [7] E. [8] G. Whitt (1994) Waitingtime tail probabilities in queues with longtail service time distributions . with applications to risk reserve processes and the GI /G/1 queue. 345368. [5] G. [3] J. Stegun ( 1972 ) Handbook of Mathematical Functions (10th ed. John Wiley & Sons.). Appl. 143170. Queueing Systems 16. Arfwedson ( 1954) Research in collective risk theory. ed. Asmussen (1987) Applied Probability and Queues. 101116. Abate. The Mathematical Scientist 14. [9] G. 53100.
69100. Bladt (1996) Renewal theory and queueing algorithms for matrixexponential distributions. Asmussen & B. SIAM Review 40.). 1996. [21] S. Ann. 193226.364 BIBLIOGRAPHY [16] S. Asmussen & B. 1936. Probab. Asmussen (1995) Stationary distributions via first passage times. 303322. Asmussen (1992b) Light traffic equivalence in single server queues. 315.K. 189201. [25] S. [19] S. 4966. Dshalalow ed. [30] S. [32] S. Binswanger (1997) Simulation of ruin probabilities for subexponential claims. Probabilistic Analysis of Rare Events (V. [20] S. [34] S. Astin Bulletin 27. Asmussen (1992c) Stationary distributions for fluid flow models and Markovmodulated reflected Brownian motion. L. 106119. Advances in Queueing: Models. Probab. Asmussen. J. Act. 2149. 27. 297318. Ann. New York. MatrixAnalytic Methods in Stochastic Models (A. 772789. 25. Florida. Asmussen. eds. Kalashnikov & A. J. Asmussen (1998c) A probabilistic look at the WienerHopf equation. Rolski & V. 137168. Methods V Problems (J. Asmussen (1998a) Subexponential asymptotics for stochastic processes: extremal behaviour. Probab. [27] S. stationary distributions and first passage probabilities. [33] S. 1999 . K. J. 313341. Scand.). Scand. Asmussen (1999) On the ruin problem for some adapted premium rules. Scand. [31] S. Asmussen (1991) Ladder heights and the Markovmodulated M/G/1 queue. Asmussen (1992a) Phasetype representations in random walk and queueing problems. 96107. Alfa & S. Hojgaard (1999) Approximations for finite horizon ruin prob abilities in the renewal model. 2943. Boca Raton. Marcel Dekker. [23] S. [24] S. Frey. Appl. Scand. 313326. [18] S. T. Riga Aviation University. Floe Henriksen & C. Asmussen (1989b) Risk theory in a Markovian environment. Statist. CRC Press. [26] S.M. Appl. Th. [22] S. 37. [17] S. A. Stoch. [28] S.S. Asmussen.). 1989 . Asmussen (2000) Matrixanalytic models and their analysis. 354374. 2. Proc. Asmussen & K. Bladt (1996) Phasetype distributions and risk processes with premiums dependent on the current reserve. Hojgaard (2000) Rare events simulation for heavytailed distributions. 555574. 79102. Asmussen & M. . Asmussen & M. Appl. Hojgaard (1996) Finite horizon ruin probabilities for Markovmodulated risk processes with heavy tails. Appl. Stoch. Extremes 1 . Andronov. Proc. Act. Asmussen (1998b) Extreme value theory for queues via cycle maxima. Chakravarty. Kliippelberg (1994) Large claims approximations for risk processes in a Markovian environment. Stochastic Models 11. [29] S. Bernoulli 6. 20. Act. 54. Ann. J. eds. 8. Binswanger & B. Schmidt (1995) Does Markovmodulation increase the risk? ASTIN Bull. Random Processes 2.
Appl.Y. Asmussen . CRC Press. . Asmussen & H. Asmussen & V.). Asmussen & R. 58. Adv. Asmussen & T. [48] S. [45] S. 115. Kliippelberg (1996) Large deviations results for subexponential tails. Res. 299324. Probab. Read eds. 64. Rolski (1991) Computational methods in risk theory: a matrixalgorithmic approach.BIBLIOGRAPHY 365 [35] S. 10. 11251141. J. Proc. O'Cinneide (2000/2001) Matrixexponential distributions [Distributions with a rational Laplace transform] Encyclopedia of Statistical Sciences. O'Cinneide (2000/01) On the tail of the waiting time in a Markovmodulated M/G/1 queue. Prob 32 . Taksar (1997) Controlled diffusion models for optimal dividend payout. Koole (1993). Taksar (2000) Optimal risk control and dividend distribution policies . B. Math. Inf. Probab. Methods 8 Problems (J. Dshalalow ed. [43] S. Management Science 45. Sc. J. H. Res. Rubinstein (1999) Sensitivity analysis of insurance risk models. first passage problems and extreme value theory for queues. [44] S. Opns. Appl. Rubinstein (1995) Steadystate rare events simulation in queueing models and its complexity properties. [52] S. Stochastic Models 8. Asmussen & G. 421458. [37] S. Advances in Queueing: Models. Asmussen & M. Stoch. Fitting phasetype distributions via the EM algorithm. Appl. Oper. [41] S. Schmidt (1995) Ladder height distributions with marks. 429466. 259274. Probab. [36] S. Johnson.M. Asmussen & C. [40] S. Schmidli & V. J. Th. Nielsen (1995) Ruin probabilities via local adjustment coefficients . Hojgaard & M. 31. 419441. Proc. Nerman & M. Scand. Asmussen. Rolski (1994) Risk theory in a periodic environment: Lundberg's inequality and the CramerLundberg approximation. Appl. [47] S. Statistica Neerlandica 47. Adv.A. Schock Petersen (1989) Ruin probabilities expressed in terms of storage processes. Sigman (1996) Monotone stochastic recursions and their duals.Y. [46] S. 20.A. [38] S.). 913916. 736755. Appl. Florida. Olsson (1996). Asmussen & C. Wiley. 30. 105119. 365372. 120. 410433. [49] S. with applications to insurance risk Stoch. 422447. [39] S. Example of excessoffloss reinsurance for an insurance corporation. Asmussen. Boca Raton. Asmussen & D. 103125. [51] S. Eng. Insurance: Mathematics and Economics 10. Statist. [42] S. Perry (1992) On cycle maxima. Asmussen & S. Finance and Stochastics 4. Asmussen & R. Probab. Asmussen & K. Schmidt (1993) The ascending ladder height distribution for a class of dependent random walks. Appl. Insurance: Mathematics and Economics 20. Supplementary Volume (Kotz. Marked point processes as limits of Markovian arrival streams. [50] S. Asmussen & V. Schmidt (1999) Tail approximations for nonstandard risk and queueing processes with subexponential tails. (to appear). Asmussen & C. O. 23. Asmussen & T. 19.
Bloomfield & D. [70] A. [64] P. 77111. Borovkov (1976) Asymptotic Methods in Queueing Theory.H. Insurance: Mathematics and Economics 9. Berman & R. 1975.366 BIBLIOGRAPHY [53] S. Crijns (1987) Upper bounds on ruin probabilities in case of negative loadings and positive interest rates. Grandell (1985) An insensitivity property of the ruin probability. [63] A. [59] O. Scand. 33.J. Act. J. Teugels (1987) Regular Variation. Beekman (1985) A series for infinite time ruin probabilities.L. 1985 . Act. Boogaert & V. [56] B. J. 1974. [62] J. Athreya & P. Plemmons (1994) Nonnegative Matrices in the Mathematical Sciences. J. [72] P.T. Billingsley (1968) Convergence of Probability Measures. Bjork & J.B . Probab.L. [58] O. Wiley. Act. von Bahr (1975) Asymptotic ruin probabilities when exponential moments do not exist. Grandell (1988) Exponential inequalities for ruin probabilities in the Cox case. BarndorffNielsen (1978) Information and Exponential Families in Statistical Theory. 7. [57] C. 148156. Clarendon Press. [60] J. J. Ney ( 1972) Branching Processes . 661667. Scand. [69] P. 11811190. Appl. [71] P. BarndorffNielsen & H. Bjork & J. C.A. . Cambridge. Stochastic Models 10. Act. Trans. Actuaries 21. J. Bingham. Baker (1977) The Numerical Solution of Integral Equations. J. [55] B. 9. Adv. von Bahr (1974) Ruin probabilities expressed in terms of ladder height distributions. Oxford.R. SpringerVerlag. 9599. [61] J. 610. 221232. [66] N. Scand. J. Schmidli (1995) Saddlepoint approximations for the probability of ruin in finite time.A. Berlin. Halsted Press. Appl. Bingham (1975) Fluctuation theory in continuous time. 1995. Insurance: Mathematics and Economics 6. Act. de Waegenaere (1990) Simulation of ruin probabilities. [67] T. Boogaert & A. 1988 . [65] N. New York. Teugels ( 1996) Convergence rates for M /G/1 queues and ruin problems with heavy tails . Springer . Beekman (1969) A ruin function approximation.H. Scand. 705766. Cambridge University Press. [73] A. New York. Appl. Telek (1994) A bencmark for PH estimation algorithms: results for acyclic PH. Goldie & J. Beekman (1974) Two Stochastic Processes. Asmussen & J. Soc. 169186. SIAM. New York. 129134.H. Cox (1972) A low traffic approximation for queues. 4148. 832840. Probab. Wiley. 275279. Bobbio & M. [68] T. Insurance: Mathematics and Economics 4. Scand. [54] K. 190204.M. Probab.
Hickman. Boxma & J. [86] K. Oper. 51. Cohen (1998) The M/G/1 queue with heavytailed service time distribution. Chung (1974) A Course in Probability Theory (2nd ed. Appl. Fort & G. [80] F. Bowers . Illinois.Y . [85] W. SpringerVerlag. Burman & D. 62. Cambr. Aut. Bell.A.BIBLIOGRAPHY 367 [74] O. Boxma & J.M. 392433.J. Skandia Jubilee Volume. 749763. Simulation and Estimation. AddisonWesley. Cohen (1999) Heavytraffic analysis for the GI/G/1 queue with heavytailed service time distributions . Geb. New York. IEEE J. Computer Performance (K. [91] H. Sel. Smith ( 1983) Asymptotic analysis of a queueing system with bursty traffic. 2338. New York San Francisco London.R. [83] D. [90] D. S. Control AC28.L. Chandy & M. [75] O. Cottrell. Syst. Insurance : Mathematics and Economics 9. Fox & L. . Tweedie (1982) Storage processes with general release rule and additive inputs . [93] K. Jr. Biihlmann ( 1970) Mathematical Methods in Risk Theory. [92] H. H.I. Philos. Probab. Berlin. Resnick & R. Broeck .L.W. Area Commun. [89] M. 24. 105119. 14. M.R. Stockholm. 313319.Y. Academic Press. B. Amsterdam. Veraverbeke (1990).) NorthHolland.). Reading. 34. [81] J. Wiley. [79] P. [76] N. [77] P. Cramer (1930) On the Mathematical Theory of Risk. Wahrscheinlichkeitsth.L. J. Smith ( 1986) An asymptotic analysis of a queueing system with Markovmodulated arrivals . 3539. Burman & D . Bucklew ( 1990) Large Deviation Techniques in Decision. IEEE Trans. [88] J. Brockwell . Adv. Bux & U. Jones & C.L.U. 127130. The Society of Actuaries . [84] D. verve. Springer. [87] E.R. [78] L. Cramer (1955) Collective risk theory.C. Bratley. Goovaerts & F. J. 16. Soc. Stockholm. NorthHolland. Insurance : Mathematics and Economics 5. Reiser eds. 14331453. Gerber. D. Croux & N. Malgouyres (1983) Large deviations and rare events in the study of stochastic algorithms . Proc..W. Res. Z.A. II. Herzog (1977) The phase concept : approximations of measured data and performance analysis. J. Nesbitt (1986) Actuarial Mathematics . 907920. Queueing Systems 33. Amsterdam. Breiman ( 1968) Probability. De Vylder (1986) Ordering of risks and ruin probabilities. Nonparametric estimators for the probability of ruin. Cohen (1982) The Single Server Queue (2nd ed. Itasca.W. Cox (1955) Use of complex probabilities in the theory of stochastic processes.J. New York. Tech.C. 177204. 93121. The Jubilee volume of Forsakringsbolaget Skandia. [82] H. Qinlar (1972) Markov additive processes.). Schrage (1987) A Guide to Simulation.
[103] F. . Teugels ( 1990) Empirical Laplace transform and approximation of compound distributions . 201206. Csorgo & J. N. Roy. [98] A. De Vylder & M. Math. Laird & D. 624628. 121131. Zeitouni (1992) Large Deviations Techniques and Applications. Haezendonck (1987) Classical risk theory in an economic environment. Insurance: Mathematics and Economics 3. 583602. Springer. Soc. Daley & T. [107] L. [111] F. Dassios & P. De Vylder & M. J. Delbaen & J. [95] M. Pesonen (1994) Practical Risk Theory for Actuaries. 114119. T. Res. 3. De Vylder (1996) Advanced Risk Theory. Daykin. Suppl. [108] F. Haezendonck (1985) Inversed martingales in risk theory. Rolski (1991) Light traffic approximations in queues. Insurance: Mathematics and Economics 6. Scand. 9 . Boston. 22. Davidson (1946) On the ruin problem of collective risk theory under the assumption of a variable safety loading (in Swedish). [99] R. Rolski (1984) A light traffic approximation for a single server queue. [100] A. 85116. 16. Scand. De Vylder (1978) A practical solution to the problem of ultimate ruin probability. Forsiikringsmatematiska Studier Tillagnade Filip Lundberg. Steinebach (1990) On some alternative estimators of the adjustment coefficient in risk theory. [105] A. De Vylder (1977) A new proof of a known result in risk theory. J. Embrechts (1989). Math. Comp. [110] F. Rubin (1977) Maximum likelihood from incomplete data via the EM algorithm. [106] A. Devroye (1986) Nonuniform Random Variate Generation. 227229. English version published in Skand. Dembo & O. Chapman & Hall. Stockholm. Probab. J. J. Daley & T. [102] P. 181217.P. van Dawen (1986) Ein einfacher Beweis fur Ohlins Lemma. Goovaerts (1984) Bounds for classical ruin probabilities. Oper. 3750. Blotter der deutschen Gesellschaft fur Versicherungsmathematik XVII. Jones and Bartlett. Pentikainen & E. Aktuar. [101] C. Appl. 88101. Goovaerts (1988) Recursive calculation of finitetime ruin probabilities. 7083 (1969). Insurance: Mathematics and Economics 4.368 BIBLIOGRAPHY [94] M.D. [112] F. Stochastic Models 5. [97] D. 5771. 435436. Insurance: Mathematics and Economics 10. [96] D. Act. Steinebach ( 1991 ) On the estimation of the adjustment coefficient in risk theory via intermediate order statistics . Insurance: Mathematics and Economics 7. Res. Dempster. Deheuvels & J. 135159. 17.B. Tidskr. Statist. Oper.M. Martingales and insurance risk. Appl. 27. [104] F. Math. Csorgo & J. Delbaen & J. [109] F. New York.. Actuarial J. 1990 ..L. Editions de 1'Universite de Bruxelles.
C. Math. Hipp (1998) Ruin probabilities for Erlang (2) risk processes.C. Djehiche (1993) A large deviation estimate for ruin probabilities. Ruin estimates for large claims . Insurance : Mathematics and Economics 22. van Doorn & G. Shiu (1991) Risk theory with the Gamma process.G. J. Dickson & J.C.BIBLIOGRAPHY 369 [113] D. 105115. 131138. J. Act. [121] D.S.M. [120] D. Regterschot (1988) Conditional PASTA. Duffield & N. H. 6180. Schmidli (1993) Finitetime Lundberg inequalities in the Cox case . Dickson (1995) A review of Panjer's recursion formula and its applications. Camb. J. Dickson & C. Oper. 1984 . [118] D. [123] E. Act.M. O'Connell (1995) Large deviations and overflow probabilities for the general singleserver queue.J. with applications. Res. Dickson (1992) On the distribution of the surplus prior to ruin . Dickson & H. Act. Act.J.C.M. 177192.C. Embrechts ( 1988 ). Scand. [115] D.C. J. Springer . [130] P. 147167. Gerber & E. 4259. 1993 . 251262.M. Gerber (1988) The surpluses immediately before and at ruin. Proc. 1975 . Gray (1984) Exact solutions for ruin probability in the presence of an upper absorbing barrier. Dickson & H.R.R. Taylor ( 1975) A diffusion approximation for the ruin probability with compounding assets. Astin Bulletin 21. Dufresne & H. 229232. [119] D. 1741.B.C.R. Letters 7. Gray (1984) Approximations to the ruin probability in the presence of an upper absorbing barrier. and the amount of claim causing ruin. Insurance: Mathematics and Economics 25. [128] E. Hipp (1999) Ruin problems for phasetype(2) risk processes. Embrechts . Dufresne & H.C. Insurance : Mathematics and Economics 11. J. Act. Waters (1996 ) Reinsurance and ruin . .R.M. Insurance: Mathematics and Economics 7. Scand. J.U. 1984 . Grandell & H. British Actuarial J. [117] D. Insurance : Mathematics and Economics 7. Dickson & C. Act. Gerber (1991) Risk theory for the compound Poisson process that is perturbed by a diffusion. [127] F. Harrison & A. [126] F. Dufresne. [131] P.C.M. [114] D. 118. Scand. 191207. Soc. J. J. 3745. 107124. [125] F. [129] D . Scand. Act.U. [124] N. Insurance: Mathematics and Economics 19.W. Dynkin (1965) Markov Processes I. Dickson & J. [122] B. 4962. Scand.M. Scand. 269274. [116] D. 174186. Scand.M. 2000 . 5159. 1994 . 1.U.M.M. Philos. Insurance : Mathematics and Economics 10. Waters (1999) Ruin probabilities wih compounding. Dickson (1994) An upper bound for the probability of ultimate ruin.K. Berlin Gottingen Heidelberg.C. Emanuel . 1993 . 193199. J. 363374.
Trans. Feller (1971) An Introduction to Probability Theory and its Applications II (2nd ed. Probab. 29. On the excursions of Markov processes in classical duality. [139] A. [145] P.). 38. Wiley. Proschan & D. Probab.W. Schmidt (1992) An insensitivity property of ladder height distributions. Appl. F. Erlang 2. Villasenor (1988) Ruin estimates for large claims. J. Pitts (1993) Some applications of the fast Fourier transform in insurance mathematics . [142] W. Schmidli (1994) Ruin estimation for a general insurance risk model. 566580. [146] E. Adv. Insurance: Mathematics and Economics 10. Sci. [134] P. [141] F. 695712. Astin Bulletin 16. J. [135] P. Tidsskr. Feller (1966) An Introduction to Probability Theory and its Applications I (3nd ed. Mikosch (1991). Danish Academy Tech. Appl. Schmidt (1982) Queues and Point Processes.L. Embrechts & H. Reprinted 1948 as 'The theory of probabilities and telephone conversations' in The Life and work of A. 1998 . Insurance : Mathematics and Economics 1. A bootstrap procedure for estimating the adjustment coefficient. Fields 75. Th. Ann. Embrechts & T. Adv. Embrechts & N. [136] P. with applications. Insurance: Mathematics and Economics 7. Scand. [140] J. Act. [143] W. Furrer (1998) Risk processes perturbed by astable Levy motion. Fuh (1997) Corrected ruin probabilities for ruin probabilities in Markov random walks. 26.K. Appl. 29. C. J. Teugels (1985). 616624. D. Wiley.370 BIBLIOGRAPHY [132] P.W. Erlang (1909) Sandsynlighedsregning og telefonsamtaler . Springer. Statist. Esary. Appl. Embrechts. Probab. first passage times and moments of ladder variables in Markov random walks. Embrechts & J. Embrechts. Jensen. J. 131137. Maejima & J. Akt. [133] P. [138] P. Griibel & S. U. Skand. Walkup (1967) Association of random variables. Adv. Probab. Statistica Neerlandica 47. Appl. Lai (1998) Wald's equations.K.J. 404422. 8190. 35. Fitzsimmons (1987). Wiley. 17.D. Math. 181190. 5572. 14661474. 5974. Kliippelberg & T.L. M.D. Veraverbeke (1982) Estimates for the probability of ruin with special emphasis on the possibility of large claims . Frees (1986) Nonparametric estimation of the probability of ruin. 623637. Approximations for compound Poisson and Polya processes. [137] P. Fuh & T. 175195.D. Nyt Tidsskrift for Matematik B20. .M. Rel. Probab. New York. [147] M. Frenz & V. [149] C. Konig. New York. Franken. Heidelberg.). Mikosch (1997) Modelling Extremal Events for Finance and Insurance. 3339. 159178. [148] C.A. Embrechts. [144] P. Esscher (1932) On the probability function in the collective theory of risk. Arndt & V. Probab. 269274. [150] H. 5975. R.
Michna & A. Grandell (1977) A class of approximations of ruin probabilities. 424445.U. Schweiz. Scand. 205210. 463480.N. [163] P. Goovaerts. Glasserman & S. Vers. 6370.S. Gerber (1988) Mathematical fun with ruin theory. Gani. Kaas. [156] H.U. eds. Mitt. Probab. [162] P. Scand. Insurance: Mathematics and Economics 20. Schweiz. [161] P. Schmidli (1994) Exponential inequalities for ruin probabilities of risk processes perturbed by diffusion. Act. Aktuarietidskrift 1970 . [168] J. Glasserman (1991) Gradient Estimation via Perturbation Analysis. [157] H. Goldie & R. Adv. [165] M.71. Amsterdam. Math. [154] H. Life Insurance Mathematics. Act. 105115. Ver.U. [158] H. Skand.U. [153] H.U. [155] H. Gerber (1979) An Introduction to Mathematical Risk Theory. Weron (1997) Stable Levy motion approximation in collective risk theory. Gerber (1973) Martingales in risk theory. S. Huebner Foundation Monographs. Amsterdam.M. Gerber (1986). Furrer & H. A..BIBLIOGRAPHY 371 [151] H. Birkhauser . Griibel (1996) Perpetuities with thin tails. Gerber (1981) On the probability of ruin in the presence of a linear dividend barrier. 31A. van Heerwarden & T. Mitt. [164] B. Insurance : Mathematics and Economics 7. Furrer. J. J. Gnedenko & I. Grandell (1978) A remark on 'A class of approximations of ruin probabilities'. R.G. University of Pennsylvania. 28. Dordrecht. Insurance: Mathematics and Economics 15. Scand. Appl. Act. Haezendonck (1984) Insurance Premiums.J. . [167] C. In Studies in Applied Probability (J. [160] H. ETH Zurich. Unpublished.). Whitt (1994) Logarithmic asymptotics for steadystate tail probabilities in a singleserver queue. Probab. 5. Ann. Gerber (1971) Der Einfluss von Zins auf die Ruinwahrscheinlichkeit. Furrer (1996) A note on the convergence of the infinitetime ruin probabilities when the weak approximation is astable Levy motion. Bauwelinckx (1990) Effective Actuarial Methods. Glynn & W. 97114.). 131156. [152] H.V. 7778. 1978 . 205216. [166] M.W. Kou (1995) Limits of first passage times to rare sets in regenerative processes. Appl. Vers. contained in the authors PhD thesis.U. Goovaerts.E. F. 1981 . Galambos & J. Suppl. 2336. NorthHolland. J. J. Springer. Probab. Basel. NorthHolland. 1977. de Vylder & J. 3752. Kluwer. 1523. Gerber (1970) An extension of the renewal equation and its application in the collective theory of risk.U. Kovalenko (1989) Introduction to Queueing Theory (2nd ed. Math. [159] H. 73.J.. Boston. Ver. Appl. Z. [169] J. London.
SpringerVerlag. Stoch. . 6779. KTH.and Sozialforschung 6. Probab. AkademieVerlag. [180] R. Grnbel (1991) G/G/1 via FFT. Appl. Archiv v. [189] A. [177] B. [174] J.A. Proc . Rink. [182] A. 355365. 8. Grandell (1999) Simple approximations of ruin functions. Grandell (1997) Mixed Poisson Processes. Statistical Algorithm 265. Appl. Gyllenberg & D. Winter School on Stochastic Analysis and Appl . 3041. Probab. [186] J. 33. M. Proc. [175] J. Grigelionis ( 1993) Twosided Lundberg inequalities in a Markovian environment. Ibragimov. Liet. Grigelionis (1996) Lundbergtype stochastic processes. Resnick (1976 ) The stationary distribution and first exit probabilities of a storage process with general release rule . 197214. M.S. Grigelionis ( 1992) On Lundberg inequalities in a Markovian environment. Res. [181] D. 5 . 14. Res. Segerdahl (1971) A comparison of some approximations of ruin probabilities. M. Th. Gut (1988 ) Stopped Random Walks . Appl. van Heerwarden (1991) Ordering of Risks: Theory and Actuarial Applications.O.. Hermesmeier (1999) Computation of compound distributions I: aliasing errors and exponential tilting. Harrison & S. New York. [187] J. Berlin. Gusak & V.V. Lemoine (1977) Limit theorems for periodic queues. Appl. [171] J. 5766. Silvestrov (2000) Cram€rLundberg approximations for nonlinearly perturbed risk processes . Opns . [188] J. Matem.243255. 3 .I. 566576. Applied Statistics 40. Opns .). [172] J. Grandell (1990) Aspects of Risk Theory. [185] J. Math. 332. Insurance: Mathematics and Economics 26. Math. Harrison & A. 54. Grandell & C. 14. E. Probability Theory and Mathematical Statistics (I. [173] J. Skand. Grandell ( 1992) Finite time ruin probabilities and martingales .372 BIBLIOGRAPHY [170] J. J. Astin Bulletin 29. Chapman & Hall. [184] H . Informatica 2. Resnick (1977) The recurrence classification of risk and storage processes . 143158. Amsterdam. [183] M.J. Aktuar Tidskr. [176] B. [179] R. [178] B. Hadwiger (1940) Uber die Wahrscheinlichkeit des Ruins bei einer grossen Zahl von Geschiiften . mathematische Wirtschaft . 1. 167176.I. SpringerVerlag. Tinbergen Institute Research Series 20. M. ed. 7590. Grandell (1979) Empirical bounds for ruin probabilities. Preprint. Griibel & R. 347358. Harrison & S.131135. Proc. Stoch. Harrison (1977) Ruin problems with compounding assets . Korolyuk (1969) On the joint distribution of a process with stationary increments and its maximum . Theory and Applications. 400409.
305313. Studies in Statistical Quality Control and Reliability 1992 :4. Karlsruhe. 23.L. The ruin problem for finite Markov chains. Huskova. 1998 . Mathematical Statistics. Heilmann (1987) Grundbegriffe der Risikotheorie. Hoglund (1991). . [199] R. 123151. J. J. S. Herkenrath (1986) On the estimation of the adjustment coefficient in risk theory by means of stochastic approximation procedures. ACM TOMACS 6. 19. Hipp (1989a) Efficient estimators for ruin probabilities. [204] T. Hesselager (1990) Some results on optimal reinsurance in terms of the adjustment coefficient. Foruth Prague Symp. Act.L. 3. [192] U. Astin Bulletin 19. Iversen & L. Iglehart (1969) Diffusion approximations in collective risk theory.). Hogan (1986) Comment on corrected diffusion approximations in certain random walk problems. 5770. verve. 285293. 166180. Wiley. Klugman (1984) Loss Distributions. Proc. 285292. Verlag Versicherungswirtschaft e.BIBLIOGRAPHY 373 [190] P. 25. 259268. Nerman (1992) EMPHT . New York. Karlsruhe. Wahrscheinlichkeitsth. Hoglund (1974) Central limit theorems and statistical inference for Markov chains..V. Michel (1990) Risikotheorie: Stochastische Modelle and Statistische Methoden. Hipp & R. Ann. Taksar (2000) Optimal proportional reinsurance policies for diffusion models. Insurance: Mathematics and Economics 5. [205] T. Statist. 11631174. Probab. Prob. Staalhagen (1999) Waiting time distributions in M/D/1 queueing systems. 12981310. Nachrichten 30. [194] 0. Probab.R. Hill (1975) A simple general approach to inference about the tail of a distribution. [191] W. Z. Hoglund (1990) An asymptotic expression for the probability of ruin within finite time. [202] 0. 8095.M. Ann. [203] T.V. [201] L.B. Asmussen & 0. 4385. on Asymptotic Statistics (P. Act. Hojgaard & M. 377381. Hogg & S. Math. 8996. [197] C. [200] M. Hermann (1965) Bin Approximationssatz fiir Verteilungen stationarer zufalliger Punktfolgen. 1990 . [198] C. Willekens (1986) Estimates for the probability of ruin starting with a large initial reserve. Horvath & E. J.a program for fitting phasetype distributions. J. Versicherungswirtschaft. [206] B. [208] V. No. eds. Insurance: Mathematics and Economics 5. Electronic Letters 35. Hipp (1989b) Estimators and bootstrap confidence intervals for ruin probabilities. 6. Scand. Appl. Scand. Heidelberger (1995) Fast simulation of rare events in queueing and reliability models. Ann. [207] D. [195] B. [196] C. Chalmers University of Technology and the University of Goteborg. Mandl & M. Geb. 378389. Haggstrom.A. [193] U. 18. 29. Probab. Appl.
[228] J. University of Chicago Press. [210] D. [226] J.M.B. Kalashnikov (1996) Twosided bounds of ruin probabilities. Lazar (1998) Subexponential asymptotics of a Markovmodulated random walk with a queueing application. J. Munksgaard. Jagerman (1991) Analytical and numerical solution of Volterra integral equations with applications to queues. Johnson & M. Appl. Manuscript. 187193. Addenda to for processes defined on a finite Markov chain. New York.J. 31. Clarendon Press. 6.G. Kalashnikov (1999) Bounds for ruin probabilities in the presence of large claims and their comparison. Kaas & M. 37. ibid.M.C. Soc. Kennedy (1994) Understanding the WienerHopf factorization for the simple random walk. [227] J. J. Janssen (1980) Some transient results on the M/SM/1 special semiMarkov model in risk and queueing theories. ibid. Statist. Keilson & D.M.P. [218] V. [221] E. Probab. 173190. Proc. Oxford. [214] A. J. Janssen & J. Reinhard (1985) Probabilities de ruine pour une classe de modeles de risque semiMarkoviens. Jelenkovic & A. Taylor (1981) A Second Course in Stochastic Processes.J. [217] R.G. Wishart (1964). Soc. 11.G. 118. Ann. Astin Bull. 3.R. 711743. [223] J. Act. Astin Bull. Chicago. Insurance: Mathematics and Economics 5. . Soc. Keilson (1966) A limit theorem for passage times in ergodic regenerative processes. Jagerman (1985) Certain Volterra integral equations arising in queueing. Wishart (1964) A central limit theorem for processes defined on a finite Markov chain. Wishart (1964) Boundary problems for additive processes defined on a finite Markov chain. Kao (1988) Computing the phasetype renewal and related functions. Math. N. [220] V. Kluwer. Amer. Appl. 561563. Karlin & H. [225] J. [213] P. Scand. 116129. Cambridge Philos. 547567. Cambridge Philos. [211] J. Goovaerts (1986) General bound on ruin probabilities.M. Cambridge Philos.L. Act. 63. J. 60. Probab. Academic Press. Jensen (1995) Saddle Point Approximations. Proc. 123133. 866870. [219] V. Copenhagen.M. [215] J.H. NEC Research Institute. 283306. 35. Stochastic Models 5. Stochastic Models 1. Kalashnikov (1997) Geometric Sums: Bounds for Rare Event with Applications. Keilson & D. Taaffe (1989/90) Matching moments to phase distributions. Jensen (1953) A Distribution Model Applicable to Economics.374 BIBLIOGRAPHY [209] D. 8793. N. 15. [224] J. Technometrics 30. [222] S.A. Princeton. 325247. [212] J. 6. Kemperman (1961) The Passage Problem for a Markov Chain. 61. 165167. 1996. 259281. Proc. 239256. 4151. [216] M. Keilson & D.
Appl. Klnppelberg (1989) Estimation of ruin probabilities by means of hazard rates. [246] D. Kunita (1976) Absolute continuity of Markov processes. 283284. 359361. Penev & A. [239] H. Philos. [232] C. Turbin (1973) An asymptotic expansion for the absorption time of a Markov chain distribution. Lecture Notes in Mathematics 511. Bernoulli 1 . 125147. 889900. Lucantoni (1991) New results on the single server queue with a batch Markovian arrival process. [231] J. [247] L.BIBLIOGRAPHY 375 [229] J. Seminaire de Probabilties X . Lindley (1952) The theory of queues with a single server. Klnppelberg & T. Math. J. 18. J. Macmillan. Probab. J. Kliippelberg (1988) Subexponential distributions and integrated tails. Lemoine (1989) Waiting time and workload in queues with periodic Poisson input.F. SIAM. Latouche & V.C. Korolyuk. Sorensen (1997) Exponential Families of Stochastic Processes. Mikosch (1995) Explosive Poisson shot noise with applications to risk retention. Scand. 132141. Proc. (235] C. 48. Cambr. 390397. I. Stadtmiiller (1998) Ruin probabilities in the presence of heavytails and interest rates . Kliippelberg & U. Klnppelberg (1993) Asymptotic ordering of risks and ruin probabilities. 858874. Lemoine (1981) On queues with periodic Poisson input. [236] C.J. Lehtonen & H. Quart. SpringerVerlag. J.P. [234] C. Cybernetica 4.C.S. Prob. SpringerVerlag. [245] T. 4958. Statist. Proc. Act. 383392. Scand. [230] J. Actuarial J. Philos. Ramaswami (1999) Introduction to MatrixAnalytic Methods in Stochastic Modelling. 6075.F. 259264.F. Insurance: Mathematics and Economics 8. Oxford 12. Appl. Lipsky (1992) Queueing Theory . J. Kingman (1961) A convexity property of positive matrices .a Linear Algebraic Approach. Act. Appl. 1995 . . Appl. [240] U. Nyrhinen (1992b) On asymptotically efficient simulation of ruin probabilities in a Markovian environment. Probab. Klnppelberg & T. Quart. 279285. 60 . 277289. [241] G. Kuchler & M. [248] D. 133135 (in Russian). Nyrhinen (1992a) Simulating levelcrossing probabilities by importance sampling.J. [237] C. 154168. 4477.C. B24. [242] A. Scand. [244] T. [238] V. [243] A. J. 25. 1998 . Soc. 26. 24. Camb. Kingman (1962) On queues in heavy traffic.F. Roy. J. Lehtonen & H. Soc. Insurance : Mathematics and Economics 12. New York. [233] C. Adv. 146. Probab. Mikosch (1995) Modelling delay in claim settlement.. Stochastic Models 7. Kingman (1964) A martingale inequality in the theory of queues. Soc.
K. 676705. a useful concept in risk theory. 370377. J. 834841. 11 . [259] Z . Math. [266] P.D. 29 . Hoist eds. Soc. Act. Proc. Cambridge Philos. 268285.D. Scand. Cambridge Philos. 2939. Proc. Moustakides ( 1999) Extension of Wald 's first lemma to Markov processes. Scand. J. J. Adv. [253] V. Appl. . [257] A. 147149. MeierHellstern & M . [261] H. [256] A. 1261 (0?)1270. MartinL6f (1983) Entropy estimates for ruin probabilities . Appl. Probab. Malinovskii ( 1994) Corrected normal approximation for the probability of ruin within finite time.V. Appl. Eigenvalue properties and limit theorems . with an application . 1996 . Lundberg (1903) I Approximerad Framstdllning av Sannolikhetsfunktionen. [250] F. Gut & J. Scand. [254] V . Europ. A fitting algorithm for Markovmodulated Poisson processes having two arrival rates . Malinovskii ( 1996) Approximation and upper bounds on probabilities of large deviations of ruin within finite time. 15. [258] K . Nummelin (1987) Markov additive processes I. F. [251] F. 4859. Mammitzsch ( 1986 ) A note on the adjustment coefficient in ruin theory. Insurance : Mathematics and Economics 5. 2. Probab. Miyazawa & V. 561592. 124147.K. Englunds Boktryckeri AB. Probab. Appl. Statist. 32. MartinL3f (1986) Entropy.F. Almqvist & Wiksell. Appl. Probability and Mathematical Statistics (A. J. 286298. K. Miller (1962 ) A matrix factorization problem in the theory of random variables defined on a finite Markov chain . Michna ( 1998) Selfsimilar processes in collective risk theory. 1986 . J. 1994. 58. Miller ( 1962 ) Absorption probabilities for sums of random variables defined on a finite Markov chain . Ann. Miller ( 1961 ) A convexity property in the theory of random variables defined on a finite Markov chain. [260] H. Soc. Ann. Anal. Meier ( 1984). Schmidt ( 1993) On ladder height distributions of general risk processes . Act. [263] M . J.V. [262] H . Makowski ( 1994) On an elementary characterization of the increasing convex order . 22. [255] V. Appl. . [265] S. Stockholm. [252] A.376 BIBLIOGRAPHY [249] D.S. J. Ney & E. Ann. 31.D. 763776.M. 223235. 429448. Lundberg (1926) Forsdkringsteknisk Riskutjdmning. Uppsala. Res. Act. Opns.). Lucantoni . Neuts ( 1990) A single server queue with server vacations and a class of nonrenewal arrival processes . II Aterforsdkring av Kollektivrisker . Stoch. Th. 378406. Nagaev (1957) Some limit theorems for stationary Markov chains. [264] G. 36. 161174. Probab. Probab.S. Probab . 58 . 3. Math.
Statistics 21. Stoch. 249266. [272] J . 316. 327361. Astin Bulletin 5 . Stoch. Neuts ( 1981 ) MatrixGeometric Solutions in Stochastic Models. 733746. 75. Nyrhinen ( 1998 ) Rough descriptions of ruin for a general class of surplus processes. Appl. E75B . Stoch. Abh. [284] J. 135148. Proc.K. Proc.F. Neveu ( 1961 ) Une generalisation des processus a accroisances independantes. Commun. Gjessing (1997b) Present value distributions with applications to ruin theory and stochastic equations. [270] M . Appl. Willekens (1987) Second order behaviour of distributions subordinate to a distribution with finite mean . 339353. Math. Willekens (1986) Second order behaviour of the tail of a subordinated probability distributions. 764779. [278] E. Stochastic Models 3 . J. 30.F. Probab. 555564. 12551265. Insurance: Mathematics and Economics 22. Neuts ( 1978) Renewal processes of phase type. Appl. O'Cinneide (1990) Characterization of phasetype distributions. London. 12.F. Stoch. [282] J . Omey & E. Models 6.F. Johns Hopkins University Press. Omey & E. Neuts (1977) A versatile Markovian point process. [285] J. Appl. Neuts (1992) Models based on the Markovian arrival process. Naval Research Logistics Quarterly 25. Probab. Insurance: Mathematics and Economics 20. 16. Appl. 311342. . [269] M . [273] R. 36. [281] J. Paulsen & H. Paulsen & H. New York. Ohlin (1969) On a class of measures for dispersion with application to optimal insurance. IEICE Trans. 273299.F.a survey. Gjessing (1997a) Optimal choice of dividend barriers for a risk process with stochastic return on investments. Stoch. Pakes (1975) On the tail of waiting time distributions. J. Probab. Paulsen (1993) Risk theory in a stochastic economic environment. [275] H. 123144. [271] M.A. 157. 21.BIBLIOGRAPHY 377 [267] M . Paulsen ( 1998) Sharp conditions for certain ruin in a risk process with stochastic return on investments. Norberg ( 1990) Risk theory and its statistics environment .G. Probab. Adv. [280] A. Hamburg. J. Appl. [277] J. Baltimore .K. [283] J. Nyrhinen (1999) Large deviations for the time of ruin. Paulsen (1998) Ruin theory with compounding assets . Marcel Dekker. 46. [274] H . Appl. Neuts ( 1989) Structured Stochastic Matrices of the M/G/1 Type and their Applications. [279] E. 10081026. [268] M . Proc. Appl. [276] C. 215223. Sem. Proc. 445454. 71.
11. 568576. Embrechts (1996) Confidence bounds for the adjustment coefficient. Math. 29. 965985. Wiley. [287] F. [303] S. [292] S.M. Ripley (1987) Stochastic Simulation . [300] C. Aktuar. Soc. Appl. Rao (1965 ) Linear Statistical Inference and Its Applications. Adv. Pellerey (1995) On the preservation of some orderings of risks under convolution. Stat. 32. Math. Springer.. Oper. Probab. [304] B. [288] S. 139143. Appl. [293] N. and Dams. Pitman (1980) Subexponential distribution functions. Adv. Insurance Risk. Regterschot & J. de Smit (1986) The queue M/G/1 with Markovmodulated arrivals and services. Inst. [289] C. Prabhu & Zhu (1989) Markovmodulated queueing systems. Probab . Probab. Wiley. Ann. 691707. Pitts. 19. [306] T. Ann.L. Prabhu (1965) Queues and Inventories. New York. Opns. 29A. 235243. Heidelberg. Resnick & G. Ann. [298] V. Res. Gjessing (1997c) Ruin theory with stochastic return on investments. [297] R.M. Math. J. Ramaswami (1980) The N/G/1 queue and its detailed analysis. Queues. 46. 537555. Pitts (1994) Nonparametric estimation of compound distributions with applications in insurance. [301] G. On the ruin problem of collective risk theory. 2330. Math. 4568. [302] J. Astin Bulletin XIV. Act. Statist. [305] C. Adv. Ramsay (1984) The asymptotic ruin problem when the healthy and sick periods form an alternating renewal process.R. Statist.K.A.M. Ann.G. 337347. Res. 215246. Math. Schock Petersen (1989) Calculation of ruin probabilities when the premium depends on the current reserve. Griibel & P. [295] N. [296] N. 30. [294] N. Berlin. Rogers (1994) Fluid models in queueing theory and WienerHopf factorisation of Markov chains. Insurance: Mathematics and Economics 16. J. 147159. Insurance: Mathematics and Economics 3. [299] C. Austr.U.U. Tidskr. 465483. .U. R. 757764. Prabhu (1980) Stochastic Storage Processes. Paulsen & H.U.K. Rolski (1987) Approximation of periodic queues. 390413.378 BIBLIOGRAPHY [286] J. Appl.H. [291] S.J. Prabhu (1961). Queueing Systems 5. Scand.J. 12. 61. Samorodnitsky (1997) Performance degradation in a single server exponential queueing model with long range dependence. 28. New York. 1989 . Philipson (1968) A review of the collective theory of risk. Appl. Probab. 117133.G. Pyke (1959) The supremum and infimum of the Poisson process. 2343. [290] E. Appl. J. 820827. Wiley. Skand. 222261. Probab. Reinhard (1984) On a class of semiMarkov risk models obtained as classical risk models in a markovian environment. 4.
V. [324] H . [318] H. Stochastic Models 10. Appi. [326] H. Schmidli ( 2001 ) Optimal proportional reinsurance policies in a dynamic setting . [313] M. Wiley. Wiley. 227244. Singapore. 417421. [321] H. Stochastic Models 6. 4857. Lecture Notes of the 8th Summer School on Probability and Mathematical Statistics ( Varna). Rubinstein & A. Stoch. [309] H. 5.J. Schlegel ( 1998) Ruin probabilities in perturbed risk models.M. 1997. 145165. [315] T. Wiley. [311] R. Schmidli ( 1999a) On the distribution of the surplus prior and at ruin. Seal ( 1969) The Stochastic Theory of a Risk Business. Data Anal. Astin Bulletin 29. Math. Schmidli ( 1997a) Estimation of the Lundberg coefficient for a Markov modulated risk model . Stochastic Models 10 . Science Culture Technology Publishing . 93104. Ryden (1994) Parameter estimation for Markov modulated Poisson processes. 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. Rolski. [310] R.BIBLIOGRAPHY 379 [307] T. [314] T. [308] S. [325] H . [320] H. 262286. 155188. Rudemo (1973) Point processes generated by transitions of a Markov chain. Th. Operationsforsch. 131156. 4068. Rubinstein (1981) Simulation and the Monte Carlo Method. 135149. [312] R. 2001 . Schmidli ( 1995 ) CramerLundberg approximations for ruin probabilities of risk processes perturbed by a diffusion . [319] H. Probab.Y. 5. 795829. 121133. Scand. Insurance: Mathematics and Economics 22. Schmidli ( 1997b) An extension to the renewal theorem and an application in risk theory. Ann. Probab . Ryden (1996) An EM algorithm for estimation in Markovmodulated Poisson processes. [322] H. J. 431447 [316] S. Ross (1974) Bounds on the delay distribution in GI/G/1 queues. Act.L. Melamed (1998) Modern Simulation and Modelling. [323] H. Appl. 687699. 21. Shapiro (1993) Discrete Event Systems: Sensitivity Analysis and Stochastic Optimization via the Score Function Method. [317] L. Schmidli ( 1994) Diffusion approximations for a risk process with the possibility of borrowing and interest. J. H.Y. Rubinstein & B. Scand. Proc. Schmidli ( 1996) Martingales and insurance risk. 7. 11. Teugels (1999) Stochastic Processes for Insurance and Finance. Insurance: Mathematics and Economics 16. Comp. Statist. Wiley. Appl. Probab. Adv. Schmidt & J. 365388.Y. Schmidli ( 1999b) Perturbed risk processes: a review . Schmickler ( 1992 ) MEDA : mixed Erlang distributions as phasetype representations of empirical distribution functions . Schmidli. Statist. Wiley. Act. 5. . J.
Segerdahl ( 1942) Uber einige Risikotheoretische Fagestellungen . 1974. 61. [337] M. 701719. [343] D. Probab . Appl. [331] G.W. Act. [347] K.the Harald Cramer volume (ed. Risk teory and the single server queue . [335] B . 4. t). Academic Press. In Probability and statistics . Aktuar Tidsskr. Stochastic Models 6. [333] C: O. Siegmund (1976) Importance sampling in the Monte Carlo study of sequential tests. Mitt. Seal (1974) The numerical calculation of U(w. Versich. 72. Sengupta (1990) The semiMarkov queue: theory and applications. 4. Verein Schweiz Versich. [339] A.W.L. Skand. Th. [348] K . Adv. Math. 1955.S. Wiley. Shantikumar (1993) Stochastic Orders and Their Applications. the probability of nonruin in an interval (0.F. 4383. 914924. Act. Shin (1987) Convolution of uniform distributions and ruin probability. Sigman (1992) Light traffic for workload in queues. 11. . Seber (1984) Multivariate Observations. Mitteil. 171178. Ann. pp. Sengupta ( 1989) Markov processes whose steadystate distribution is matrixgeometric with an application to the GI/PH/1 queue. [332] C : O.S. [341] E. Siegmund (1979) Corrected diffusion approximations in certain random walk problems. Grenander). Probab. Weiss (1995) Large Deviations for Performance Analysis. Ann. 1987. 383413. 191197. . Math. Shwartz & A. Probab. 243249. Shtatland ( 1966) On the distribution of the maximum of a process with independent increments. Scand. Siegmund (1975) The time until ruin in collective risk theory. Verein Schweiz. J. 121139. [340] E. 75. Siegmund ( 1985) Sequential Analysis . Versich. [330] H. 21. Sigman ( 1994 ) Stationary Marked Point Processes: An Intuitive Approach. Seal ( 1972). 159180. 11. Aktuar Tidsskr. [336] B. Almqvist & Wiksell. 483487. Seal (1972) Numerical calculcation of the probability of ruin in the Poisson/Exponential case. Verein Schweiz. Insurance: Mathematics and Economics 8.L. Chapman and Hall. J. Probab.G. 2236.380 BIBLIOGRAPHY [327] H. [346) D . L. 673684. [342] D. [334] C: O. 72. Statist. Segerdahl (1955) When does ruin occur in the collective theory of risk? Skand. Appl. Seal (1978) Survival Probabilities. Stockholm. Scand. 429442.S. Mitt. Segerdahl (1959) A survey of results in the collective theory of risk. Shin (1989) Ruin probability by operational calculus. t). New York. [344] D. Adv. Shaked & J. [338] E. [345] D. [329] H. 279299.L. Math. 157166. Chapman & Hall. 77100. Siegmund (1976) The equivalence of absorbing and reflecting barrier problems for stochastically monotone Markov processes .A. Queueing Systems 11. [328] H. SpringerVerlag. Appl. Wiley.
Stanford & K. Teugels ( 1995 ) Ruin estimates under interest force . Teugels (1982) Estimation of ruin probabilities . Stoyan (1983) Comparison Methods for Queues and Other Stochastic Models (D.C. [353] E. Proceeedings of the IEEE 77. Cambridge Philos. Act. [350] W.L. 65102. Proc. Springer. [354] B. . 8199. 118. Scand. Taylor (1976 ) Use of differential and integral inequalities to bound ruin and queueing probabilities . [360] G. [362] G.C. Taylor (1979) Probability of ruin with variable premium rate. Sundt & J. 722. Karlsruhe. Astin Bulletin 12. 449461.L. Insurance: Mathematics and Economics 16. Scand. Taylor (1979) Probability of ruin under inflationary conditions or under experience ratings .). J. 1982 . [364] G. NorthHolland.L. [351] D . 1976. Scand. 163175. 8594. [361] G. 149162. Thorin (1986) Ruin probabilities when the claim amounts are gamma distributed. Sundt ( 1993) An Introduction to NonLife Insurance Mathematics (3rd ed. Suri ( 1989) Perturbation analysis : the state of the art and research issues explained via the GI/G/1 queue. Scand..). Hoesman ( 1989) Moderate.. New York.L. Daley ed. Appl. Sundt & W. Scand. 114137. Straub (1988) NonLife Insurance Mathematics. J. Takhcs (1967) Combinatorial Methods in the Theory of Stochastic Processes.S. 235254.C. 197208. Verlag Versicherungswirtschaft e. Thorin (1977). Taylor (1978) Representation and explicit calculation of finitetime ruin probabilities. Astin Bulletin 24. Ruin probabilities prepared for numerical calculations. [365] J.C. John Wiley & Sons. Unpublished manuscript. Taylor (1986) Claims reserving in NonLife Insurance. Act. 1974. 1978. 2739. [358] R. Adv.A. Smith ( 1953) Distribution of queueing times.J. Act. J. Slud & C. [363] G. 65102. J.C. New York.BIBLIOGRAPHY 381 [349] E. Act. Astin Bulletin 16 . [357] B. 5776. Actuarial J. Insurance: Mathematics and Economics 1. 1980 . [368] O. Insurance: Mathematics and Economics 19. Stroinski ( 1994) Recursive method for computing finitetime ruin probabilities for phasedistributed claim sizes. [355] B . Jewell ( 1981 ) Further results on recursive evaluation of compound distributions . [369] O. Probab. Soc 49. Teugels (1997) The adjustment coefficient in ruin estimates under interest force . Thorin (1974) On the asymptotic behaviour of the ruin probability when the epochs of claims form a renewal process. Sundt & J.J. 21.and largedeviations probabilities in actuarial risk theory. Scand. Wiley.V. [356] B . [366] O. Thorin (1982) Probabilities of ruin. J. 725741. Act. [367] O. [352] D. [359] L.
Willmot (1994) Refinements and distributional generalizations of Lundberg's inequality. [380] W. Wald (1947) Sequential Analysis. 4963. Thorisson (2000) Coupling. Act. Unpublished Ph. Appl. Sherman & D. Goovaerts. Wallace (1969) The solution of quasi birth and death processes arising from multiple access computer systems.E. 6785. 1942. Haezendonck eds.382 BIBLIOGRAPHY [370] 0. [373] S. Insurance: Mathematics and Economics 15. 743756. Thorin & N. [371] 0. Tidskr.). 148164. Tacklind (1942) Sur le risque dans les jeux inequitables. Astin Bulletin VII. . Insurance : Mathematics and Economics 13. Reidel. Wikstad (1977) Numerical evaluation of ruin probabilities for a finite period. Probab. J. Willinger. 1983 . Veraverbeke (1993) Asymptotic estimates for the probability of ruin in a Poisson model with diffusion. [375] N. Wolff (1990) Stochastic Modeling and the Theory of Queues. R. M. 31. [377] V. 142. J. Wilson (1995) Selfsimilarity in highspeed traffic: analysis and modeling of ethernet traffic measurements. Stationarity and Regeneration. [381] W. Waters (1983) Probability of ruin for a risk process with claims cost inflation. Goovaerts (1983) The influence of reinsurance limits on infinite time ruin probabilities. Lin (1994) Lundberg bounds on the tails of compound distributions. SpringerVerlag. M. [374] F. PrenticeHall. F. De Vylder. Res. thesis. University of Michigan.E. [383] G.W. D. 936952. Thorin & N. In: Premium Calculation in Insurance (F. [378] H.R. [382] G. Wiley. De Vylder & M. Scand. 231246. Dordrecht Boston Lancaster. Willmot & X. Oper. Statistical Science 10. Astin Bulletin IX. Wikstad (1977) Calculation of ruin probabilities when the claim distribution is lognormal. Skand. Aktuar. VazquezAbad (1999) RPA pathwise derivatives of ruin probabilities. [372] H. 5762. Van Wouve. J. Whitt (1989) An interpolation approximation for the mean workload in a GI/G/1 queue. Taqqu. [376] A. 37. Submitted. 137153. [384] R. [379] M.
341.308 CramerLundberg model: see compound Poisson model cumulative process 334 dams: see storage process differential equation 16.135.269. 245248. 7475. 111117. 117127 corrected 121127 duality 1314. 361 diffusion 3.150. 301 Kronecker product. 170173.121129.281.249250 integral equation 16 Lindley 143 renewal 64.86. 360 excursion 155156.Index adjustment coefficient 17. 39.217. 5796. 205.100. 1415.4447.293294.178184. 3436. 217. 110113. 3032. 7179. 7079. 17.307312 compound Poisson model 4. 97. 201214. 86. 37.287292. 15.346349 383 . 17.299.203. 162164. 239. 207 heavytailed distribution 6.314316. 316323 Bessel function 102. 227229. 302303 diffusion approximation 17. 14.318320 change of measure 2630.328330. 122. 226. 2526.301 central limit theorem 60 . 119. 308.272. 196201 inverse Gaussian distribution 76. 3334. 278 gamma distribution 67. 318319 Erlang distribution 7. 323 Coxian distribution 147. 9496. 201 Brownian motion 3 . 91. 1112.203. 79. 189. 9293. 138139. 4851. 218 Cox process 4.359 aggregate claims 103106. 135. 8283 hyperexponential distribution 7.285292. 332333 Volterra 192194. 141144. 248 WienerHopf 144 interest rate 190. 5. 12 CramerLundberg approximation 1617. 271274.6779. 80 81.242. 89. 3839.9899. 283. 9396. 40. 7879.228229.226.259261. 117128.200201.182. 1819. 2425.and sum 221.137141. 97129.185187.251280 heavy traffic 76.292293 Edgeworth expansion 113.249.160167. 180182.
16. 3947. 7179. 245 M/G/1 13.336339 . 261264. 251. 35.148.128129. 149.350361 Poisson process Markovmodulated 12 periodic 12. 52 53.174. 96.160161.384 ladder heights 4756.178182. 267269 Panjer's recursion 320323 Pareto distribution 910.259261. 41. 100.302.297299. 203 Markov additive process 12. 108 life insurance 5.348 terminating 215216.161164. 6162.215250.349 350 perturbation 172173. 14. 7576. 185187 GI/G/1 141144 M/D/1 6667 equation 16. 295. 4446. 137139. 176185. 132133. 86 periodicity 12. 65. 25. 260 Lundberg conjugation 6979 . 25.234240. 35.201.139141. 3947. 5758. 227228. 157. 179 NP approximation 318320 Palm distribution 5253. nonlinear 155. 80. 71. 133.287291 INDEX matrix equation .288290.298299. 16. 9899. 175 light traffic 8183 Lindley integral equation 143 process 3334. 304305 random walk 3336. 134.161. 234 matrixexponential distribution 240244 matrixexponentials 14. 141144.339 large deviations 129.275278.285287 queue 14 . 42. 142 likelihood ratio : see change of measure lognormal distribution 9. 144. 203204. 3639.152160.238. 138. 108109.336339 Laplace transform 15. 32. 59.146148.134135. 213214. 145187.161.123.269271. 6970. 39. 230. 15. 162.304 process 2830. 134135. 171. 112113. 44.261264.234. 271274. 257.180. 44. 306316 Levy process 3.240244. 133. 99. 229 M/M/1 101 Markovmodulated 185187 periodic 187 martingale 2426.218221. see also sensitivity analysis phasetype distribution 8.340350 multiplicative functional 2830. 113114. 2730. 38. 178 modulation 12. 38. 176185 nonhomogeneous 60 PollaczeckKhinchine formula 6167.287.315 inequality 1718.108. 37. 154. 269 PerronFrobenius theory 4142.227230. 106108.
332333 model 12. 233234. 281296 stable process 15. 292294. see also matrixexponential distribution regenerative process 264 268. 191192. 172173. 186187 virtual: see workload rational Laplace transform 8. 177 timereversion 14. 168172 storage process 13. 31. 11. 141144. 4950. 335336 sensitivity analysis 8693. 174. 189214. 3032.314. 317318 semiMarkov 147. 261264 reservedependent premiums 14. 229234.154157. 260 reinsurance 8. 280. 294296 shotnoise process 314 simulation 19. 5455. 233. 256258. 152. 244. 120 statistics x. 222. 12. 89. 251. 338 utility 324. 257. 147.262263. 60. 9693. 8386. 331336 equation 64. 240. 253.186. 186187 renewal process 131. 7475. 307308. 123. 251280 time change 4. 107. 37. 279280 subexponential distribution 11. 238 saddlepoint method 115117. 162.244250. 251.279280 Rouche roots 158.273274.INDEX 385 waiting time 141. 213. 223226. 131144.359361 stochastic control x stochastic ordering 18. 260 WienerHopf theory 144. 327 . 160. 333334 regular variation 10. 326330 Weibull distribution 9. 146. 87.336339 workload 13. 1819.
"This book is a must for anybody working in applied probability.. exact solutions.. Some i (l I JL I J r of the topics are Lundberg's inequality. phasetype distributions as a computational vehicle and the connection to other applied probability areas like queueing theory. y finite horizon ruin probabilities. worldscientific. 2 A I 11 JjVb l' i  i Yj . for heavytailed claim size distributions). I 1! Ruin Probabilities .g.Advanced Series on Statistical Science & Applied Probability .Vol. extensions of the classical compound Poisson model to allow f o r reservedependent premiums. the ^W A l \ i l ' ''' CramerLundberg approximation. Special features of the book are the emphasis on change of measure techniques." Short Book Reviews ISBN 9810222939 mi u inn i nun I I I I I I i in u www. P'i yfliother approximations (e.com 2779 he 9 "789810ll22293211 . It is a comprehensive treatment of the known results on ruin probabilities.T [Ail i The book is a comprehensive treatment of  I i I \ classical and modern ruin probability theory. Markovmodulation or periodicity.